Using samples and inspiration to link cultures and eras

James Blake is one of the best live artists I’ve ever seen. Midway through watching him chop up his own voice for “Fall Back”, my friend decided it was the perfect moment to diagnose me with a new personality trait.

How do you keep track of all the samples these guys use?

Extremely weird timing, as the only sample I can identify in “Fall Back” is James Blake himself, saying “Fall back”. He’s not wrong though, I have an obsession with samples, melodies, and traceable musical elements between songs. I think it’s my version of looking for the Akira Bike Slide.

I must have pointed out one or two of these during the gig. As we’ll find out, a reference doesn’t always have to be as obvious as a full blown sample.

A tour guide once described the city of Chicago to me as “constantly in conversation with itself”. The architecture of one building was a response to another, which itself might have been a response to any number of other small influences. Material prices, the flow of the river, the arrival of the World Fair, or even the use of a building for a film are compacted into the landscape that these architects make their decisions in. Chicago’s also an important part of another landscape in constant conversation with itself: the world of hip-hop.

Hip-hop is the standout example of how tracking musical elements and influences can enrich the music itself. A sample is worth so much more than its sound (though it doesn’t always have to be), and can say a lot about the person who picked it. One of my favourite things to do is compare how the same samples are used in different tracks, like “Runnin’” by the Pharcyde and “DIKEMBE!” by JPEGMAFIA. They both sample “Saudade Vem Correndo” by Stan Getz and Luiz Bonfá.

“Runnin’” was produced by J Dilla and released in ‘95. It inteferes minimally with the sample, with only added drums and bassline to add a bit of movement. I’m not convinced this is all the way normal for him; A lot of other tracks that Dilla’s made really chop up the sample (thinking about “Little Brother” by Blackstar, or “Players” by Slum Village). This leaves me wondering what made Dilla decide to “underprocess” this cut. It could be something to do with who he’s producing for, as his remix of “She Said” on the same album (Labcabincalifornia) also has that same light touch feel to it. Perhaps he was trying to fit into their broader sound, or maybe he was really taken with Saudade.

JPEGMAFIA was both the producer and rapper on “DIKEMBE!”, which is how he likes to do things. He takes Saudade and dices it to fit into a thumping industrial beat, crushed against a rattling bass and knocking drums (these are all good things I love this song). This is in stark contrast to “Runnin’”, where the beat is clearly built around the sample. After his verse, JPEGMAFIA switches up and almost abandons Saudade for a set of electronic sounds and keys. This use of the cut is much more instrumental, wielded like a sound effect bank, and it overall signals that this beat is digital. It’s synthetic, it’s from the world of robots and computers, and I truly adore it.

This demonstrates one of the things I love about sample hunting. Keeping track of these links between songs quickly creates a dense, featured network that can tell stories. A track doesn’t get sampled once, in one way, and then left alone. Dorothy Ashby’s “Cause I Need it” has been sampled 6 times (7 if we include remixes) by 5 different artists! JPEGMAFIA, who used to go by Devon Hendryx, did an incredible cover of Call Me Maybe by Carly Rae Jepsen! A Tribe Called Quest used the same sound effect/sample thing in “Bonita Applebum” and “Enough!”, bookending their career in a really sweet way.

Chicago’s own contributions to hip-hop are key parts of the medium’s metacommentary. Kanye West is one of the defining artists of his time, who changed the kinds of people who got to be stars within hip-hop (F.D Signifier has a wonderful video that covers how he enabled Drake’s rise here). He’s also said things that are anti-semitic, racist, and misogynistic, which make it hard to cite him as an influence without embarassment. Good art, it seems, doesn’t just come from good people. I haven’t yet figured out how to deal with the fact that I may subconsciously carry some of his style in my own art, just from having listened to him. I’m interested to hear from people who have that worked out.

This goes far beyond hip-hop, and I have a particular love for catching similar melodies in different genres. Let’s take this to the extreme and jump straight to my favourite example.

This is “Jab Bhi Koi Haseena Dekhu”, made for the movie Hera Pheri. This is one of the first Bollywood movies I’d ever watched, and I really was not familiar with anything about Bollywood music.

This is “Return of the Mack” by Mark Morrison. I grew up hearing this song everywhere, and when I heard the melody to this warble out of Akshay Kumar’s dubbed lips I shrieked. To find a cultural reference so close to my own experience at the beginning of a movie in a language I couldn’t speak, from a place I’d never been, was just incredible. Top ten moments in history.

The first draft of this post had a slot for another song that I was convinced had interpolated “Return of the Mack”. I had a really good joke about the Return of the Return of the Mack. I can’t for the life of me find this track, so we’ll all have to find a way to live our lives in absence of this really good joke.

There is a notable phenomenom in Bollywood cinema of interpolating and referencing songs from other cultures, but I just wasn’t emotionally prepared for this. Did you know that George Benson has been interpolated for Bollywood? Eireann?? (Side note but the Eireann interpolation is in a movie called “Kambakkht Ishq”, or Corrupted Love. You will never guess who’s in this movie. Never.)

To be clear, this is brilliant. I love this beyond measure. India had a really solid disco phase in the 80’s, and clearly it’s love affair with Black art (amongst others) and music continues today with interpolations and a burgeoning rap scene. This was the gateway into so much Indian culture for me, seeing threads of familiarity that I could grasp onto and explore from. I’ve used a love of samples to become a huge fan of Brazilian music and deepen my knowledge of funk and soul as well, picking up on threads such as the use of Edu Lobo’s “Zum Zum” in “Tribe” by Bas and J. Cole and the myriad cuts of “Inside My Love” by Minnie Ripperton (my favourite has got to be “Rope // Rosegold” by Isaiah Rashad).

By the way, it’s not fair for me to pretend as if the West hasn’t also used and been inspired by Desi music. Britney Spears’ “Toxic” samples “Tere Meere Beech Mein” by Lata Mangeshkar and S. P. Balasubrahmanyam from the movie Ek Duuje Ke Liye. I grew up hearing Panjabi MC’s “Mundian To Bach Ke” everywhere (as did every other UK millenial), one of the first songs I liked in a language that wasn’t English. I’m sure everyone has heard “Get Ur Freak On” by Missy Elliot, which uses samples of drums and a string-sounding instrument that originate in traditional music from the Indian subcontinent (a random website tells me it’s probably a dhol and a tumbi. I have it on good authority that drum is a tabla, and not a dhol. Believe what you like).

Samples, interpolations, and references connect the islands of individual songs and genres together. Copyright makes things hard, but I wish we had a better approach to sharing our inspiration and building upon the work of others. I’d like to finally find my Return of the Return of the Mack, if nothing else.

I’m writing this in 2024, and others might be wondering how I feel about the rise of sample use in pop music. You’ve got Nicki Minaj using “Can’t Touch This” in “Super Freaky Girl” (and also having used “Baby Got Back” in “Anaconda”), Elton John licensing out a bunch of his discography for “Cold Heart” in collaboration with Dua Lipa, and Bebe Rexha and David Guetta somehow making Eiffel 65’s “Blue (Da Ba Dee)” actually depressing in “I’m Good (Blue).

It is tempting to lean into anti-corporate screeching (especially when I think about the David Guetta one). There is a definite monetary incentive for the corporate music mogul’s to get on board with this trend; Old music is trying to kill new music and I think it’s because the margins on old tracks are ridiculous. New musicians need studio time, aggressive marketing, and the slow buildup of a strong career on the back of small gigs at small venues. Old songs are just there for the taking, with established fanbases and an existing legacy, and the largest labels have entire oceans of fantastic music in their walled gardens.

I think part of the reason I’m turned away from “I’m Good (Blue)” and “Super Freaky Girl” is the unshakeable feeling that these songs are cynical, pandering to people who like to pick up on musical references (and that’s insulting to me, being one of those people). I also don’t have the same learning experience with these songs, and that’s partly because of those doubts about sincerity. Do David Guetta and Bebe Rexha have a relationship with Eiffel 65’s work? Did they select it because they love the original, because it means something special to them, or even just because the sound was a perfect fit for what they wanted to do? It’s not their fault that I doubt the authenticity of such a move. The understanding that they, or the label execs they work with, have a clear business reason for interpolating just cheapens the act. I have looked behind the curtain of the music business and, shock horror, the Wizard of Aux was a middle-aged white guy who needs to make a stock ticker go up.

I suppose that samples are still helping me learn here. I’m not entitled to always enjoy what they teach me.

James Blake uses no samples in “Fall Back”, and in a lot of his other work too. That doesn’t mean he simply pulled his sound out of the void. Music is never conceived immaculately, and the booming electronic beats he sings over could be inspired by any number of tracks that he loves and listens to. I make my own music, and I count James Blake proudly amongst my inspiration alongside a pantheon of talented artists (many of whom are the same artists I’ve talked about above).

Samples and explicit references do something that interviews and scraps of homage can’t recreate. They centre influence at the discussion of music. Its impossible to talk about artists like Pete Rock, A Tribe Called Quest, and Dilla without mentioning how Minnie Ripperton, Ronnie Laws, and Cal Tjader appear in their work. I like that it brings the complex interrelation down to my level, the level of people who don’t study musical history formally.

If you’re interested in becoming more aware of samples, I have a few tips to help you get started.

1. Check some of your favourite tracks on places like whosampled.com. You might be surprised by what you find!

2. Listen to your favourite artists favourite artists. Get the Jones Girls, Stan Getz, Eddie Kendricks, all kinds of old music pumping out your speakers. Go as far and wide as you can.

3. Listen with other people! The amount of times my mum has told me what sample is used in a track, I think she must have some kind of secret DJ career.

Even if you’re a fan of a genre where sampling isn’t massive, I’m sure you’ll learn something. There’s this thing in jazz called “The Lick” (or the licc if you’re Gen Z) that keeps being referenced in solos and tracks from across generations. It really could be anything.

That’s all for this episode of blog. Next time, we’ll talk about the maths behind competition reality TV. I challenge you to make a coherent connection because I can’t for the life of me.

Can a computer find love without understanding it?

DISCLAIMER - THIS BLOG POST RECEIVED AN UPDATE ON THE 21st NOVEMBER 2024 WITH CORRECTED RESULTS.

Have you ever been transfixed? Hypnotised to the point of paralysis? Maybe you’ve experienced this with beauty, unable to do anything but stare at transcendental visual paradise as it swells to fill every corner of your vision. Art, people, nature, the things that stun us into silence are myriad. Maybe it was true horror that froze you. The glacial flow of swollen seconds as you watch your worst nightmare coalesce into reality. We become transfixed by ideas, more so than stimulus. Is this the worst/best/most we have experienced? Could we ever be more sensationally saturated?

The tide of stimulation always recedes, but sometimes the ideas themselves remain. In the grim darkness of the Covid Lockdown’s, I fished a dire show called “Are You The One?” from the putrid depths of Netflix’s decaying catalogue. I watched it through a fog of intense boredom and became transfixed, and the ideas that were left clinging to my brain have kept me close to the brink of mania (at least when it comes to competitive dating reality TV shows) for over four years now.

Take a leap of faith with me, as I try to explain how the producers of “Are You The One?” accidentally invented one of the most interesting maths questions on screen.

Sidebar: the fact that this is a genre is weird. It’s hilarious to me (and maybe me alone) as a fan of Snow Crash, Synners, and most anything else that can trace its lineage back to Mirrorshades. Sci-fi, and cyberpunk is no exception to this, has its fair share of uber-misogynist hypergender freaks who find it easier to imagine a society where consciousness is effortlessly transferred between bodies than to imagine a woman with agency. There are plenty of examples of demeaning, regressive opinions about gender that somehow don’t come close to the comic exaggeration of gender presentation and norms that occur in Competitive Dating Reality TV (CDRTV). I don’t watch Love Island on principle, despite it being one of the UK’s most popular recent cultural exports. But the few bits I have seen present a formula for gender expression that feels lab-designed to embarrass straight cis men and women into hyper-heteronormative conformity (not to mention just completely erase the existence of anyone else). By intention or accident, artists couldn’t come up with something as insane as this genre. There’s some poetry there.

I can’t believe the genre known for settings rife with sexism (whether intentional or not) was topped by a few Essex boys.

There’s a lot of just dating reality TV shows that don’t clear this bar of bizarre. Yes, it was weird when the Truman Show came out to imagine televising every aspect of your life. It’s still “oooooh, panopticon bad” when you really take the time to think about what the Kardashians and the Celebrity Dating Agency and the Real Estate Agents (sidebar to the sidebar but real estate agents uniquely excelling at reality content was similarly unpredicted) are doing to themselves and the world by squeezing their manicured, semi-authentic existence down the fibres. But I remain convinced that the underlying principle of “I’m putting my life on film” doesn’t come close to matching the terror of CDRTV. Forming (or faking) intense physical, emotional, and sometimes sexual relationships where you are judged on how much everyone else likes your relationship must involve walking the thinnest mental tightropes a person can willingly subject themselves to. You’re already in the destabilising environment of “unscripted” reality, where authority figures (producers) are actively working to manipulate your emotions to get quotable moments, but you have no way of knowing whether your romantic partner(s) are also playing the game. Sure, there might be moments where two contestants willing make the devil’s bargain of “I will love you as much as this camera needs to get my cash”, but to spend weeks receiving signals (however convincing) that the other person is actually falling in love with you would still confuse most people. There’s no reason to believe that all contestants are making these kinds of agreements either, as reality cast members occupy a weird double role as half-actors and half-documentary subjects. This is like trying to build a mind palace in quicksand.

Watching contestants on CDRTV is the reality equivalent of watching amateur extreme sports on LiveLeak, and this article shouldn’t be taken as a call to watch more of these (I don’t watch any CDRTV anymore). It should be no surprise to anyone to see people getting hurt doing this.

It’s 2014. You are an nonthreateningly attractive and caricature-capable young person living in one of the major cities of the U.S., who perhaps has aspirations to become an influencer or a surprisingly successful D-list actor. You’ve charmed a casting director who thinks you’d be a perfect fit for one of those competitive dating reality shows, that unexpected incarnation of Pat Cadigan’s cyberpunk entertainment hallucination. This is what success feels like.

On your first day of shooting, the producers (allegedly after ramping up your star quality with a lot of alcohol) set you loose in a villa full of other hopefuls. They watch you from a dark room filled with screens as you banter and bond (if you’re female, they’ll allegedly watch you do a lot more, and if you’re black and female they allegedly might not like that they “have” to watch you so much).

Doesn’t this make you feel nice?

You’re drinking and dancing and a host with an ultraviolet-reflecting smile comes out to tell you that one of these people has been assessed to be the one for you (perfect match has already been trademarked). The aim of the game is for all of you to find out who. Twenty people. Ten “correct” couples. Ten weeks to do it. One trillion dollars on the line (inflation has been killer recently).

In the boozy Hollywood haze of it all, you probably don’t realise it. This is one of the greatest maths questions a TV show has ever asked.

You could reasonably accuse me of hyperbole here, but also let me live. I was transfixed after all, and there was scant else on TV that I hadn’t already had a go at in lockdown. I’ve left out one key detail that really makes this problem drive me wild. Each week(? I have no idea how long they spend making this), the group has to make a guess for all ten couples.

They are only told the number of correct guesses.

They aren’t told which guesses are correct. This turns an otherwise pedestrian game into something fascinating enough for a whole episode of blog! Each time you guess, you have a substantial chance of reversing your progress, and you could be within inches of the answer and never know. You could get seven correct guesses one week and still have no idea how to win, because you don’t know which seven are the ones to keep.

If you’re not convinced yet, lets take a look at the raw probabilities. For each individual out of the twenty, there is a one in ten chance that the couple they’re in is the “correct” couple for them (this would be one in nineteen purely theoretically, but the show in question had only selected straight people and there were ten of each gender. Sad.) At each pairing ceremony (christ alive this is the most het thing you can imagine), each pair is picked sequentially. The probability that a correct full ten is randomly picked is the sequence of 1/10 times 1/9 times 1/8… all the way down to 1/1 for the last pair to go.

I’ll save you the switch to your calculator, that is 1 in 3,628,800. In scientific notation, that is 2.76e-7. So across the ten weeks, your chance of just lucking into the right combination is microscopic if we don’t consider things like compatibility. It doesn’t matter if the group has ten weeks or ten years, random chance will leave them in the dust.

There is a twist in the tail here I forgot to mention. This show has a mechanic called the truth booth, which allows the group to discover whether a couple is a perfect match or not. The first episode has some absolute top tier comedy where a couple, convinced that they’re the ones for each other, walk in and find out they’re not the producers idea of a perfect match. This couple then proceed to keep on canoodling for multiple episodes despite the entire rest of the cast having the results streamed to a TV in the villa. Love finds a way.

The scenario where you find out a couple isn’t a match is less desirable than finding out a matched pair, but still improves the overall odds slightly. For one couple, the chance of selecting a correct match jumps up to one in nine from the baseline (unless you pick the wrong person on purpose??? Which I think happens????). It doesn’t really propagate any information out to the rest of the group however, so I believe the rest of the probabilities remain the same. This is where the sequential aspect of the pairing ceremony (ugh) interferes with my nice easy statistics: depending on when the non-matched pair are selecting, their chances of selecting a correct match go up (If everyone else has also selected correctly). If the two bad matches have got the final four, with everyone else having picked correctly (without realising) then they’re guaranteed to make a correct choice. If they go first, then they’re only down to 1 in 9. This means the probability of random success varies between 1 in 362,880 and 1 in 3,265,920.

If a matched pair is found, the probabilities change for the whole group. Everyone has received information about who not to pick, and the amount of correct guesses drop to nine. We end up with one over nine factorial as the probability of random success, or 1 in 362,880. Still not exactly awe inspiring. It makes it clear (to me at least) that the truth booth is best used to find matched pairs rather than to eliminate bad matches. Which sounds an awful lot like the beginnings of a strategy.

I see the irony in a person suggesting a strategy for competitive dating after discussing the inherent ickiness of competitive dating mere paragraphs ago. Fear not, for

1. This strategy has nothing to do with love or emotional manipulation

2. I am okay with being a hypocrite about this.

With that moral quandry resolved, let’s start thinking about how we make this less insane to win. Strategy one is easy: The contestants can just leave it to random chance. I’ve already discussed the odds associated with this, so all strategy two has to do is be better than 1 in 362,880, right? Well, we have to consider the effect of the truth booth, but it’s unlikely to make the odds substantially better.

Perhaps this is the algorithm design equivalent of jumping over a bar on the floor.

Let’s build the second strategy without thinking about any human aspects to this problem that would make it easier (personalities, shared backgrounds, values, dreams etc.). It’ll be as if we are pairing identical pet rocks. God knows why no one has hired me to write a TV show yet. In our new game show, “Will You Rock With Me?”, the rocks make up for their dry and somewhat emotionless performance with a basic grasp of competitive decision making.

Hire me. Hire me to write for you NOW-

On the first week, there is no information with which to make any kind of decision. Each rock knows nothing about which pairs are cosmically destined and which should be consigned to the bucket of bad exes. Therefore, any decision they make is equivalent to a random set of guesses. The rocks know this and don’t bother thinking, and our first week of strategy two looks very similar to strategy one. But here’s where things get interesting.

On week two, the rocks know three things.

1. Which rocks were in what pairs

2. The number of correct pairs

3. How many weeks of the game are left

The probability of any one of the pairs being correct is equal to the number of correct pairs over the total number of pairs. The rocks, thinking like true masters of competition, take this grain of information and write it down (somehow). They also know that, if they were able to improve by one pair every week, how long it would take to win the game. Unless they rolled really badly on the first week and only got one correct pair, they have the time to relax and be calm. All they have to do is be one better next week, which means they only have to change two pairs (swap one partner in pair one for a partner in pair two). If they get lucky, they’ll jump up the correct number of guesses by two and have two confirmed correct pairs. If they jump down by two, then they don’t really lose out because they still have confirmed two pairs. Of course, they may only end up making or breaking one correct pair (and not knowing which one they broke). Worse, the number of correct guesses could stay exactly the same, and they could have no information.

Week three rolls around, and the rocks now have the same set of data points for two weeks. How many correct pairs were there, and who was paired? If they made one change, and the number of correct pairs changed, they have a great comparative dataset. For the pairs that were the same as in week one’s selection as in week two, there are now two estimates of their probability of success. This can be accounted for by taking the mean average of the two probabilities, and this principle can be extended across many weeks.

However, coming towards the end of the game the rocks have a problem. Sure, they’ve inched closer to guessing all of the pairs, but by week eight they don’t have enough time to keep swapping one for one. Oh no! This means panic stations for the rocks, who pull up the notes they’ve been keeping of all of the results so far. The probability for each pair that appears (unless they’ve been confirmed) is the average of the probability they were correct for each week. The pairs with the highest probability are selected, and the rocks hope this is enough to push them over the edge to win the one trillion dollars.

Lets sum this up without the tortured rock metaphor:

1. Guess randomly on week one

2. For each week where you have enough time to win by mixing two pairs, mix two pairs

3. Calculate the probability of each pair being correct by taking the mean average of each weekly probability they were correct

4. Once you run out of time, select the pairs with the highest probability.

As a final point, the rocks could submit couples with a higher likelihood to the truth booth. But in the show version I saw, the contestants didn’t really have that power. They do still have to reevaluate each attempt as they find out good and bad matches from the truth.

We can test this against strategy one by making a computer model. I managed to calculate analytical odds for one week of strategy one using a range-based approach (the truth-booth makes things a bit stochastic), but that doesn’t leave with an estimation of where the accurate probability of success sits for a whole season. So, I have turned to ye olde faithful modelling method: Monte Carlo simulation.

The programme I have written (rewritten because I wrote a version back in 2021 that wasn’t quite right) will run a whole load of seasons of the show using either strategy one (random chance) or strategy two (the Rock algorithm). A singular season has a set number of weeks, where two things happen:

1. The truth booth identifies if a couple is matched or not.

2. A guess is put forward for the ten couples.

In a season with strategy one, the only information kept is a list of matched and non-matched couples from the truth booth. You could compare two weeks if you happened to roll almost exactly the same set of couples, but that’s not in the spirit of strategy one. In a season with strategy two, a record of all the submitted couples is kept with their number of occurrences and the average probability of correctness. Each individual has an ID for tracking, which also enables the one-pair swapping in strategy two to yield proper information. If a one-pair swap raises or lowers the correct count by two, then the two identified correct pairs are recorded.

On any week, if the ten correct pairs are guessed the season finishes and a win is recorded. If the season reaches the end with no recorded win, it outputs a loss. The season is then reset and run again. And again. And again, until the number of requested simulations are met. The results object then simply divides the number of wins by the number of simulated seasons, and an estimate of the chance of success is reached.

I wrote this in C++ so my computer didn’t explode running so many seasons, but the above instructions should give you enough to try your own version of this if you don’t want to run my code (which is linked here). You could do the fediverse classic and rewrite it in Rust! (if you do this, or make any other modifications, please show it to me I would love to learn from you).

Rocks are not particularly smart. I feel the need to remind you of this, in case you’re expecting this algorithm to suddenly make this game winnable. Deep in the code I wrote in 2021 is a comment stating “no point trying to solve the Travelling Salesman Problem on steroids”, which is a statement I stand by. At least with the Travelling Salesman Problem, you aren’t (theoretically) constrained in the number of solutions you can try. I’m not a mathematician so I can’t tell you if this problem is NP-hard, but if you can then hit me up on shonk app to let me know. I get the distinct feeling that, even with unlimited weeks, coming up with a good algorithm to solve this is easier said than done.

I feel compelled to remind you of this because my attempt yielded a mean per-game solution probability of 7.019% (to 3sf, calculated over 10 simulations of 1,000,000 seasons each with 10 couples and 10 weeks).

7.019%

Of course, this must be compared against the benchmark of no strategy at all. For the same set-up (10 sims, 1,000,000 seasons each), my toy programme calculated a mean per-game solution probability of 0.00919%.

0.00919% is a lot smaller than 7.019%

This means my strategy does improve the chances of success meaningfully; you’ve got just over 50% probability of seeing a success within 10 games and a 90% probability of seeing a success within 32 games by being a rock. By doing nothing at all you would only have a 50% chance of seeing one success after 7,500 games, and the 90% chance only comes after 25,000 games.

That makes me feel a bit better about my go at solving this. But alas, it’s not quite good enough.

In the real show, they solved it in seven weeks. It turns out people aren’t rocks after all.

UPDATE: These results are a bit inaccurate, and you can read the correction here. The actual success chance is 8.25%.

I really like getting to use statistics like this in the wild. Clearly I am the target audience for the maths teacher who puts on “21” in class.

This was really fun to do, much more so than actually watching the show (which got boring really quickly). Rocks turned out to be much more interesting to me than humans, who knew! I could see this being the basis for a kind of video game, which would be much nicer than seeing real people put themselves at risk doing this. I’ll keep you posted when I find time to get started on that.

It would also be interesting to find other shows, books, movies, or anything that presented questions like this, so get in contact on my shonk app page if you know of any. Maybe you’ve been transfixed too.

Next episode of blog might be about slow food and “pretty” food. It might be about something else though idk

A non-condescending guide for non-computer people

If anyone has ever told you to “just use Linux”, this guide is for you. This guide is for you if computers aren’t your thing and you like it that way, but there’s something about Windows that just rubs you the wrong way. It could be the rampant ads, or the CrowdStrike fiasco, or the Recall password-stealing machine that is now automatically scraping every Windows 11 machine (even if you don’t have the Copilot chip).

Your computer is your digital home. Unlike in the real world, you actually get to decide how that home gets built, and you can change your mind any time you like. Adding rooms, changing paint, rearranging and decorating as you please is a powerful ability, but I imagine you’re less concerned about optimising the plumbing and testing the electricity. In a house, there are methods of expression (which are fun to play with) and then there are components you just want to work. An operating system is, in this wonky analogy, the boring bits of the house done for you. A plain, functional shell with competently built essentials and maybe some basic built-in furniture, but you still bring everything that makes it home. But what do you do if that shell just doesn’t do what you need it to anymore? You walk in one day, and your boring old fridge has been replaced with a SmartFridge 3000 that gets unlocked with an app that sometimes breaks, has a huge screen to advertise Trader Boe’s Blue Sauce to you, and charges you for every item of food you remove.

As a user, regardless of how “good” you are at computers, you have the power to say no to decisions you don’t like. There are any number of reasons to move on from Windows, but as an ecosystem Windows is built to make that hard for you. So is Mac OSX, and the whole Apple ecosystem. This is the style du jour for a lot of companies today, who realise that getting you trapped in one way of doing things is a great way to stiff you on price later on. You don’t need to take it lying down. What you need is a friendly, practical guide on how to get out of the rut they’ve put you in. This is that guide.

If you’re a Mac user who wants a version of this, don’t despair. I’ll be working on one of those soon I hope!

The process

This guide will assume that you’re starting off as a Windows user on a home computer. It will take you through the following steps:

1. Preparing the important parts of your computer for being moved. (This will benefit you by creating a copy of your important files, passwords, and other things in case your computer ever breaks)
   • (Optional) Making room for the new operating system to live alongside Windows.
2. Creating an installer for the new operating system.
3. Booting up the installer and having an initial play with Linux.
4. Installing Linux onto your device (either instead of or alongside Windows).
5. Installing software and adding files.
   • (Optional after getting used to Linux) Shrinking Windows
   • (Optional) Deleting Windows

You will have every opportunity to abort the process if you don’t feel comfortable, and you never have to completely remove Windows from your machine if you don’t want to. This guide is here to empower you, and while I do have some opinions about what will make your Linux experience good I am not here to force you to do anything to your machine that you cannot reverse.

By the end of step 5, you will have a working Linux device with everything you need to use it as your daily driver. You will optionally have a spare Windows partition to retreat back to whenever you need something you haven’t figured out on Linux yet. You will know where to get new software on Linux and where the system settings are. This is all possible with:

• No knowledge of code or programming
• No knowledge of “the command line” or “the terminal”
• No confusing choices between the different types of Linux (I have selected a beginner friendly, stable, and safe type for you to get started with)
• No knowledge of computer hardware

Ultimately, I want to give you the smooth experience of using a computer that Windows achieves, without having to be locked into Microsoft’s plans for monetising you and your data. I think it would be great if, once you were successfully using Linux, you looked into some of what I mentioned in that list, but I don’t think it should be required for you to object to business practices you don’t agree with.

Step 1: Packing up the old house (Preparing your important things)

Before we start eyeing up where we want to go, lets take a good look at where we are now. I’ve talked about how you as the user make your computer into your digital home. You do this with your data, which can be split (roughly) into these categories:

1. Your files
2. Your passwords
3. Your internet browsing data
4. Your apps/programmes/software
5. Your customisations

These are the things you’ll want to take with you to the new operating system. These categories are ordered by importance and portability. For each category I’ll go over how to “pack up” your important things so that they’re ready for a safe transfer. The nice thing about computers is that you can do this without losing access to them: if you decide to stay on Windows, I still recommend having these things “packed up” in case you find yourself having to move unexpectedly.

For the more technically literate of you reading this, the phrase “pack up” is deliberately similar to backing up, because that’s basically what we’re doing.

Files

These are probably the most important thing on your computer. If they’re not, they’re definitely the hardest to replace. Files living on your computer might not live anywhere else, so if you lose them they’re gone. Recently Microsoft has been working on getting people into using OneDrive, which is a cloud storage service, but if you’re moving on from Windows you might not want to use their cloud service either (also it sucks and I hate using it at work we lose files all the time). Other cloud storage services are available and you might already be using them, but they’re limited in the space they can provide to you and might be hard to access if you lose any passwords. This is why I do not recommend using only cloud storage services like OneDrive to keep your files safe when you’re moving operating systems.

The best way to “pack up” your files is to use an external drive. Using an external drive will allow you to keep your files safe away from the computer you’re working on, so you can try the future steps without being concerned about losing everything. You won’t need a password to access them, you can load your files from another computer safely, and if you are in an area with low internet speeds it is much faster than using a cloud service. You can buy external drives on Amazon, and I would recommend buying from a reputable brand like Toshiba or Crucial a drive that has a lot of storage and a USB 3 or above connection (so USB 3.1 or 3.2 would also work). Check the ports on your device before you buy. If your device only has a USB-C connection port, as some newer laptops do, you must buy a different drive that has the correct connection or buy a separate adapter from the drive connection to the laptop connection (this will be a USB-A to USB-C adapter for the drives I’ve recommended). If your device only has USB 2.0 or lower connections you do not need to worry, these drives will still work (they’ll just be slower).

External drives can be expensive. If you’d like to save money, you can buy a hard drive enclosure and a hard drive intended for use inside a computer.

The easiest way to back up your files on Windows is to use the File History feature. Windows does also have a “Back up” feature, but this tried to sync all of my files to OneDrive rather than use an external drive (and as I’ve previously mentioned, this doesn’t help us out loads). If you search “File History” in the start menu, you should see a dialog that states whether File History is on or off. If it’s on, make sure that you run it again. If it wasn’t previously on, select your external drive and hit the button to turn it on. It will let you know that it’s running for the first time, which will take a while.

You can test out if this has worked by opening the File Explorer while the drive is plugged in. You should see it in the sidebar on the left, and it should contain a folder called “FileHistory”. This will contain a snapshot of all of the files targeted by FileHistory, which should be all of your important files.

Passwords

Your passwords are probably just as important as your files, maybe even more so. I’ve put them second because, in the absolute worst case, you should be able to recover a lot of them if you maintain access to whatever email addresses you use to sign up to services. The following recommendations will take you through a process of finding, collecting, and saving your various passwords to a secure location that you can access from any operating system and any browser.

Modern browsers come packaged with a basic password manager that remembers where you type in a password and offers to save it for later. This is true for Chrome, Firefox, Safari, and Edge. It is separate from the system that websites use to log you in immediately if you have visited recently, so you do not need to save your passwords to the browser’s built-in password manager to skip login for websites such as Netflix, Facebook, or most modern websites that have a login page. For the curious, websites do that using “Cookies”, which are small files that websites leave as a reminder that you recently provided the correct password (they do many other things as well).

The trouble with a browser-based password manager is that it only works in one specific browser. If you also don’t want to “sync” your browser across devices using an account, then those passwords only live in one device on that one browser. This is an issue if you ever lose access to your device, or your account for the browser, or if you ever want to switch anything major about your computer (like a browser, or even an operating system!). You could store your passwords in multiple browser-based password managers, but this means that you are at risk in multiple locations. So you should export them and save them to your external drive, before deleting them from the browser manager.

For Chrome:

• Click “Profile” in the top right and then “Passwords”
  • If you can’t find the Passwords icon, at the top right, select More and then Passwords and autofill and then Google Password Manager.
• To export, click “Settings” and under export passwords click “Download File”
• To delete, close all the menus that were opened for export and click “More” in the top right.
• Click “Delete browsing data” and select passwords (do this after you’re done with any browsing and signed out of any accounts).

For Firefox it’s easier:

• Click the top right hamburger menu (three horizontal lines) and click “Passwords”
• Click on the ellipsis in the top right and click “Export passwords”

For Edge:

• Go to settings
• Go to profiles > passwords
• Click the ellipsis next to passwords
• Select “export passwords”

Save the files you need from this in the external drive (separately from the FileHistory).

I am going to say something that will make nerds around the world scream: if you are keeping your passwords for a personal computer in a physical book, you don’t need to change anything. Just don’t bring the book somewhere silly and leave it out where people can read it, and make sure your passwords aren’t saved on your computer anywhere else (like that browser-based manager). You’re done for this section.

If you are storing all of your passwords in meat memory (i.e. your own head), then you don’t need to do anything either. I would suggest reading this article on why using a password manager is a better idea, but for the purposes of switching from Windows to Linux you don’t have to do more.

If you are already using a separate password manager service like 1Password, Bitwarden, or Nord (other brands are available) that is cloud-based, you also don’t need to make any changes. If you use a password manager that is local to your PC like KeePass, you should be able to export the local database (it might be as simple as finding where a file lives). Consult the documentation for your password manager on how to do this, and save the exported file(s) to the external drive.

So to recap: Check for passwords in your browser and any password managers you have.

Internet browsing data

Congratulations: if you have completed the instructions in the above two sections, you’ve done the really important stuff. Your reward is a convenient little disaster recovery kit that, if you ever need to get set up on a new computer, will let you keep all your important stuff. The next three sections cover optional activities and a few things to think about before you start moving into your new digital home.

Consider, for example, your browser history. Some of you may be shuddering at the thought. It contains a record of websites that you might find very important, or you might not be bothered about bringing it along. If you’re not actively using it, I really wouldn’t bother trying to bring it along. If you are actively using it, consider using bookmarks in whichever browser you like instead.

If you’re a Chrome user, this page has instructions on how to export your bookmarks. In case that link breaks, here’s a recreation:

1. Open Chrome
2. Select the three vertical dots on the top right > Bookmarks and lists > Bookmark Manager
3. At the top select “More” with the three dots again > Export Bookmarks

If you’re on Edge then bookmarks are now Favourites and the world is now on fire I guess (seriously why does Microsoft have to make everything a bit different). This page has some answers. It’s very similar to the Chrome process.

1. Open Edge
2. Put the following the address bar: “edge://favourites/”
3. Select the three dots in the top right > Export Favourites

If you’re a Firefox user then these instructions will get you started.

1. Click the hamburger menu in the top right > Bookmarks > Manage bookmarks
2. Go to Import and Backup > Export bookmarks to HTML

You might also hear about browser “Sync” capabilities where you sign into an account (I’ve linked the Chrome page but this exists on Firefox and Edge as well). You can do these if you like, but you don’t have to, they’re on all major browsers.

Applications, programmes, and software

This is where you as a user will start to run into some hard differences between Windows and Linux. There are some fundamental under-the-hood differences between how Windows and Linux run that mean that you cannot 1-for-1 copy and paste the software that you use in Windows into your new Linux operating system. Trust me when I say that this will not stop you from getting everything you want in Linux except for very specific circumstances.

I recommend you make a list of the important software you use. This could be office software like Microsoft Word/Excel/Powerpoint, video games, or creative tools for a specific hobby you do. For example, I use FL Studio to make music (though writing this guide made me figure out how to do that in Linux!). Maybe you use Photoshop to edit photographs, or Blender to make 3D models. You just need a starting checklist of all of the software you want to move to your new operating system.

Once you have your checklist, you can go down each item and make sure you’ve found a Linux-compatible replacement. I am going to provide some common replacements:

• Microsoft office → LibreOffice
• Google Chrome → Firefox or Google Chrome (yeah you can use the same browser! Maybe consider Ungoogled Chromium instead though?)
• Steam → Steam (Wow, also the same as the Windows one! you can look at Lutris)
• Outlook → Thunderbird
• Photoshop → Krita or GNU Image Manipulation Program
• Any Video Editing Software → Shotcut or DaVinci Resolve (You might already be using these)
• Any 3D Modelling software (non-CAD) → Blender (same as the Windows one)
• Any CAD software → FreeCAD

I’m spoiling a future section here, but you’ll notice that all of these links are for a place called “Flathub”. Flathub is the software catalogue/default app store for our new Linux operating system, and all of the apps are free!

Now, it’s entirely possible you’ll run into software with no close replacement in Linux (it could be argued that Krita and GNU Image Manipulation Program are both different enough to Photoshop that they don’t count as replacements). I have this issue with FL Studio, and it’s compounded by the fact that I paid money to have it (so obviously I want to get my money’s worth!). But don’t despair. Following the optional steps in this guide will leave you with a Windows installation that you can still use all your old software on. This will give you time to get comfortable with Linux before learning about how you can run (some) Windows software in it.

Customisation

At this point in packing up, you’re at the stage where what’s left behind is very specific to the operating system you’re using now. Did you move the taskbar? Customise your wallpaper? Change any of the sounds when you click something, or maybe the mouse pointer you click with? These things are nice ways to customise your system, but very difficult to pack up and redeploy when they’re changes to the specific way Windows “feels”.

The great news is that Linux is just as customisable as Windows, if not more! You should make sure any wallpapers you want to keep are saved in that external drive we used to back up all of your files, but in terms of other customisations I recommend seeing what your new home lets you do first.

Optional Step 1A: Clearing land for the new house (Making room for Linux)

If we want to put Linux on our computer, it needs a place to go. If you want to move straight to Linux and leave Windows completely behind, ignore this step. If you still need Windows for something then make sure you do this. Seriously consider if you need to keep Windows here, this is a non-trivial amount of work and risk.

A partition is a division of space on your computer’s storage drive. We will be making a new partition on the drive so that we can put Linux in a separate place from Windows, which allows you to choose between Windows and Linux every time you switch your computer on. The software that we use to do this is Disk Management.It is a tool built into Windows, provided by Microsoft to manage your storage drive (so we aren’t doing anything that breaks Windows). We need to do one thing: Shrink the Windows partition to leave some unallocated room.

To shrink the Windows partition, you must have some free space on your drive. I recommend having at least 50-100GB free, because this will translate directly into how much space your Linux partition will have. If you leave no room for it, then your experience using Linux will be very frustrating. Use a utility like Disk Cleanup to remove files that Windows has identified as good candidates for deletion (temporary files from downloads and things like that), and then consider uninstalling large software packages if that doesn’t get you far enough (big video games are an easy target).

To shrink your Windows partition once it has spare room, open Disk Management as an administrator. Right click on the volume labelled as “C:” which should also be a “Primary Partition” and select “Shrink Volume”. The on-screen instructions will take you through the process of reducing the size of the partition, and at the end you should see a new block on the partition diagram entitled “Unallocated space”.

Congratulations! You have learnt how to manage your storage drive, which is very cool. This is a skill that will benefit you going into the future as you explore Linux, and you’ll be able to make more and more space for the Linux partition as you use it more.

Step 2: Ordering the kit for the new house (Creating an installer)

Where do you get Linux from? This is a non-trivial question - there are lots of different types of Linux you could download that you then have to do different things with. I have a specific recommendation for you: Ultramarine Linux.

Ultramarine Linux is what I use, and it is what got me comfortable using Linux as my main operating system. It is very simple to install, easy to use, and comes with a small amount of great software to get you started. It allows you to do everything you need to do without the command line except for one thing, but we don’t need to do that one thing until after you’ve installed Linux and given it a real good try. It is also not the version of Linux you will normally see recommended for beginners.

If you’ve looked at switching to Linux before, you will have seen many guides recommend a thing called Ubuntu. This is a different type of Linux, and it’s actually the version of Linux that I first tried. The thing is, Ubuntu does a lot of things very differently to both Windows and many other types of Linux.

• It uses a more Mac OSX feeling user interface, with limited options for customising
  • You can pick different flavours if you want a different desktop feel on that distro though
• It uses a different software catalogue/app store called snapd
• It has a long term support release model, meaning it can lag behind the cutting edge of what Linux has to offer by quite a bit (In contrast, Ultramarine uses a regular release model that has a faster tempo).

These differences don’t make it “bad”. But for moving on from Windows, Ubuntu offers a different enough experience to feel uncomfortable. It also is rigid in its experience, so if a user wants to make changes they may find it difficult to do so. Ultramarine Linux is able to provide a better experience for people moving from Windows, which is exactly what you’re looking for! Ultramarine also benefits from two sets of documentation: it is very close to the Fedora Linux type that it is based off of, and if you struggle to find anything in the Ultramarine documentation you are very likely to find it in the Fedora docs.

(if you’re a technical person who disagrees with my choice here, just copy this guide and rewrite it with a distro of your choice. I’m not precious about this, it’s computer wars rubbish that means very little in the grand scheme of things)

The first thing you’re going to do is pick up a copy of Ultramarine KDE Plasma Edition from the download page. I actually use the Flagship edition, but the Ultramarine devs recommended me to point people to KDE Plasma owing to its strong accessibility features and high customisability without using command line tools. You’re also going to download a copy of BalenaEtcher, a free program that will write the Ultramarine KDE Plasma .iso file to a USB stick. You will also need a spare USB stick to write the .iso file to that should be at least 16GB in size.

Once both of your files have downloaded, you should install BalenaEtcher and start it. It will ask for the file you’re burning (that’s the .iso file for Ultramarine Linux) and the drive you’re burning it to (that’s the USB stick). After a bit of time for the .iso file to be written to the stick, BalenaEtcher will let you know that it’s finished.

Congratulations! You now have a USB stick that can install Ultramarine KDE Plasma Edition to any computer with a USB port! Now we move onto the fun part: trying out Linux for the first time.

Step 3: Seeing the show home (Booting the USB and trying Linux)

In order to boot your new Linux USB, you’ll have to access the BIOS of the computer. The BIOS does a lot of things to do with running your computer, but we do not need to change any settings except for one. We need to instruct the computer to activate the operating system in your USB, not the Windows operating system in your storage drive. I am linking to an HP guide on how to do this here, but here’s the tried and tested method for how to do this. (I’d recommend loading this webpage on a phone so you can have the instructions open while you’re doing this)

1. Shut down your computer.
2. Plug in your Linux USB.
3. Turn your computer on.
4. IMMEDIATELY begin repeatedly pressing one of the following keys on your device:
   • The “del” or “delete” key
   • The “F2”, “F10”, “F1”, or “F12” key
   • The “Esc” or “escape” key
   • If you can’t find the correct key, check what your computer says on the splash screen as it’s starting up. Otherwise you can consult this guide
5. Navigate to the menu called “Boot” or “Boot Options”
6. Find the option called “Boot Order” or similar
7. Set the first option to be “USB Storage”, “Removable Devices”, or the name of the Linux USB by moving it to the top of the list
8. Save your changes and exit (This is usually done by pressing F10 but can also be done by navigating to the Save menu or Exit menu)

Your mouse might not work in the BIOS menu. Use your arrow keys on your keyboard, and the enter key to select things.

Once you’ve done this, your computer will reboot and activate the Linux USB. Pat yourself on the back: you have successfully booted Linux!

Linux has not installed itself to your storage drive yet (that partition we prepared earlier is still empty). You are now in what’s called a “Live Environment”, which is like a demo copy of Ultramarine Linux. This is a perfect chance for you to play around with how the new operating system feels with no consequences! Before you begin exploring, navigate to the bottom right of the window and click on the icon that looks like a “No Wifi” symbol. Connect to your home network in the way you usually do (if you use a wired connection it will already be connected!)

Let’s start with an easy one: You’re going to browse the web. Double click on the Firefox shortcut (Or access it from the start menu by hitting the “Windows” key on your machine) and type “frog” into the search bar. You should get a set of search results for frog! The search engine will probably be Google, but don’t fret if it’s not. Now try going directly to a website: “rustredriver.com”. This should bring you to the same site that contains the guide you’re reading now!

There should also be some programs and settings for you to play around with. Just take your time and see how you like it, try customising a few things (Change a wallpaper, maybe!), and really get comfortable before the next step. If you’re not happy with Ultramarine Linux, now is the time to retreat back to Windows and look up a guide on how to install another type of Linux. My secondary recommendations are Linux Mint or ZorinOS.

Step 4: Buying the new house (Installing Linux to your device)

If you’re ready to commit, I recommend reading the instructions on installing Ultramarine on their site first. I will provide detailed instructions here, but in case I confuse you their instructions are available to fall back on.

There are 5 things you need to do before you can hit the “begin installation” button. We’ll go through them together.

Keyboard

You should select a keyboard layout that matches the keyboard your device has. If you don’t know what layout your keyboard has, you should select the layout that corresponds to the country where your device was made or bought from (so if you bought a device in the U.S. select “English (U.S.)”). Don’t sweat this too much, you can change it after installation if you select the wrong one.

Date and time

This is another low stress one. Just select where you are in the world on the interactive map. You can also change this after installation. If you’re connected to the internet this might sort itself out automatically!

Storage

This is one that you can’t change easily, so you need to get it right. Select the main storage drive of your device, not the connected external drive. Do not check the box that says “I would like additional space available” if you did step 1A. This will delete your Windows partition. But if you did not do Step 1A and you don’t want to keep Windows, check this box. As long as you select the correct drive and use the automatic option for storage configuration, the installer will correctly install Ultramarine in the way you want it.

If you check the “Encrypt my data” box, you’ll need to come up with a very strong passphrase (using the guidelines in the section below entitled “User Accounts”) that you’ll need every time you boot up your computer. This type of drive encryption is effective if you regularly turn your machine off, but if you don’t then you might not see much value.

Networking

This is actually just where you set the name of your computer. You can have fun with it. It’s also changeable post-installation.

User accounts

This is another tricky one to change after installation. Provide a name (doesn’t have to be your real name) and a username, and make sure the two checkboxes remain ticked. Administrative privileges are important to make sure you can keep full control of your computer, and that’s why your account needs to be password protected. When you put in a password, make sure it’s strong and memorable. Password guidance changes frequently and I understand you might read conflicting accounts about what a good password is, so here are my recommendations:

• It should be long. A passphrase is preferred over a single password because it’s easier to remember a long phrase than a singular long word/character soup.
• It should be memorable. Pick something that has meaning to you that isn’t immediately obvious to others. I’m sure you’ve seen movies where a characters birthday or child’s name is the password, and that makes it easy for someone else to guess.
• It should be unique. Don’t use a passphrase that you’ve used anywhere else.

As I said earlier in Step 1, you are allowed to write down your passphrase as long as you keep it somewhere very safe. I’m talking about somewhere that can be locked inside your home. If you’re using a password manager to store all of your other passwords, that should make it easier to remember this one. Do not store this passphrase locally on your computer anywhere.

Now you’ve done the five things necessary to begin the installation. Just to recap:

1. Keyboard layout - Fixable post-install
2. Date and time - Fixable post-install
3. Storage - Must be correct before you install
4. Networking - Fixable post-install
5. User accounts - Must be correct before you install

Take your time here, make sure your device is plugged into power if it’s a laptop, and when you’re happy press the “Begin installation” button. Then go and do something that isn’t computers for a bit. Make a cup of tea. Figure out what’s good to watch right now. Have a snack. Take the dog on a walk. After your well-earned break, you should come back to a window prompting you to reboot your device. You should do that through the start menu, accessed with the “Windows” key or the bottom left badge like you did earlier when you were exploring.

Step 5: Moving in (Installing software and setting up your new Linux system)

When your device wakes up, it will come up with a screen like this.

(the text will look different, it will have an entry for “Ultramarine Linux” instead of “Linux Mint”)

If you want to enter Linux, you don’t have to do anything. If you want to go back to Windows ever and you did Step 1A, you do it from this screen. Simply hit the down arrow key until you’ve highlight the entry that says “Windows Boot Manager” or similar and press enter. (You can change the default order later if you like, but isn’t the point to get used to Linux? :D)

After this screen, you’ll see Ultramarine Linux booting up before coming to a login screen.

Type in that strong passphrase you put into the installer before, and voila! You’re logged into to your own local Linux installation! At this stage give yourself an enormous high five because you have done all of the hardest stuff. Let’s recap:

1. You got your digital life together and created a copy of it all on an external drive. (Bonus: now you have a recovery drive!)
   • You learned how to manage storage space on a Windows machine (And probably did some spring cleaning)
2. You made a Linux USB
3. You learned how to access your device’s BIOS, and set it up to boot from USB’s if it needs to
4. You tested out Ultramarine Linux…
5. And you decided you liked it enough to install it onto your device!

Now comes the easy bit: moving into your new digital home! Just like in real life, this is something best done in stages and with the goal of figuring out what you actually need to bring along. Let’s start with that list of software you drew up back in step 1. If you hit the “Windows” key and start typing “Software” you’ll find a programme called “Software”. Open that and you’ll find the app store/software catalogue that I linked to earlier. If you click the magnifying glass in the top left OR if you just start typing, you can search for the software you have on your list. LibreOffice might be a good one to start with, or maybe Thunderbird if you want to start with an email client.

It’s worth noting that, in Linux, there are multiple different ways to distribute software. The “Software” programme is a great way to find new software, but just like with all software you should be careful about what you put on your computer. The best way to check if you’re getting genuine software is to find the website for it and look at their recommended installation method, but the “Software” programme also has verification markers for genuine software as well. LibreOffice and Thunderbird both have these.

You might here references to “FlatPak’s” and “FlatHub”. FlatHub is the name of the software catalogue that the “Software” programme is connected to, and FlatPak’s are the applications that you download from FlatHub.

One thing to bear in mind is that some software on Windows might be taking money from you right now. Set yourself a reminder to check if you’re using the free software on Linux or the paid-for software on Windows in a month. If there’s any paid-for software that you aren’t really using that takes a subscription fee, you should consider cancelling that subscription.

You can also try moving some of your files into Linux. Simply open the file explorer by opening the start menu and typing “Files”. There will be a programme called “Files” that acts as your file explorer on Linux, and it works very similarly to the Windows one. You have a search, a sidebar on the left, and a main window that shows the contents of the folder you’re in. In the “Devices” part of the sidebar, you should see that external drive (I hope you left it plugged in! But no worries if not, just plug it in). Just open it up like in Windows, and have a look at your files. Copy a few over if you like. Drag and drop, ctrl-c and ctrl-v, and right clicking all work here.

If you want a full guide on how to customise KDE, check out this IT’S FOSS guide.

This is you done, basically. I’d honestly set a reminder in a week or even a month to come back to this guide. Let yourself really get used to Linux, playing around with it, using all the software. If you need more space, or if you’re finding yourself coming back to Windows less and less, then it’s time for step 5A. If you didn’t do step 1A and you’re frustrated you can’t run Windows programmes, look into WINE to try and get your favourite Windows software working in Linux (Or Lutris if you’re looking at video games specifically). If you really really hate Ultramarine Linux, have a look at all of the other types of Linux you could try like Linux Mint or ZorinOS (or others, there are so many out there!). If you really hate all of those, then I am willing to concede that Windows may be the better fit for you. Instructions to reinstall it can be found on Microsoft’s site.

If you are enjoying Linux, you should make sure you look after it! Get yourself sorted with a backup solution using Pika Backup. Look into one of those password managers I was talking about! Read step 5C for some notes on things you should do to upkeep your system (This is where that one thing we have to do on the command line comes in).

Optional Step 5A: Downsizing (Shrinking Windows)

So, it’s time to admit it. You’re a Linux user now. You might still occasionally pop onto Windows, but you can confidently get almost everything you need done with the tools you have in Linux. Why not resize your partitions to match how you actually use your computer?

We’re going to do the first part of this in Windows because you already know how to use the disk management tools there, and there is one important order of business we have to take care of first. You need to clear out your files and software in Windows as much as you can. This might feel a bit scary, but remember! You have a copy of everything on your external drive, so you can delete without worrying about losing anything. Let’s start with all of the software that you have replacements for on Linux. Microsoft office - gone. Outlook - gone. All the video games that can run on Linux (which is 90% of them these days) - gone. Let’s move onto files as well. Everything in your user area, such as Documents and Downloads, can be gotten rid of. Desktop - cleared. As long as you do a quick mental check of “Did I definitely do my backup after I made this?”, you can delete to your hearts content. If you’re ever unsure, just open up your backup drive and check if there’s a copy in there.

Once this is done, you can open up Disk Management again. Just like last time, we’re going to shrink the Windows partition so it’s just a bit bigger than the content we have now. You should be aiming for as small as possible, so if that Windows partition is still taking up more than 20% of your drive you want to figuring out why it’s so large (if there is a legitimate reason, like a specific software package you use, then that’s fine).

Now, reboot your computer back into Linux. Open the software catalogue, and search for a tool called “GParted Partition Editor”. Install it, and run it (you’ll need to type in your password for it to run). You should see a window that looks like this.

There should be a large amount of unallocated space to the right side of the NTFS partition called “Basic Data Partition”. You should also see an EXT4 partition to the right of that, followed by a BTRFS partition. The EXT4 partition is the “boot” partition for the Ultramarine Linux operating system (like the starter motor of a car), and the BTRFS partition is the actual “data” partition i.e. where all your stuff lives.

GParted works by queueing up actions first, before executing them all at once: you have time to figure out if you’ve correctly queued up what you want to do. We’re going to queue up two actions:

1. We’re going to move the EXT4 partition to the other side of the unallocated space (You can think of this as moving the unallocated space to be next to the BTRFS partition.)
2. We’re going to extend the BTRFS partition to fill the unallocated space.

To move the EXT4 partition, click on it to select it. You should see a button in the top toolbar that, when you hover over it, says “Resize/Move the selected partition”. You can also right click on the EXT4 partition to get the same button. In the menu that appears, you should see three numbers: “Free space preceding”, “New size”, “Free space following”.

The free space preceding should be equal to the unallocated space you left from shrinking the Windows partition. Copy that value and paste it into the free space following box. Make sure the free space preceding value is zero after you do this (you’re essentially swapping the two values). Then press the “Resize” button in the bottom right.

Do not press any buttons that say “apply all operations” or “undo last operation” after doing this. You must queue up the operation to resize the BTRFS operation first.

To resize the BTRFS partition, click on it to select it. Find the same button as you pressed for moving the EXT4 partition, and you’ll see the same menu as before with different values in each of the fields. You should see a large non-zero number in “Free space preceding”, and you should set that number to zero. This should increase the size of “New size” by the exact number that was in “Free space preceding”, and “Free space following” should remain at the value it was at before (which should be zero). Once you’ve done this press the “Resize” button in the bottom right.

This has queued up two actions that will increase the size of your Linux partition. Double check that they’re in the right order, and that the partition diagram reflects how you want the drive to be set up (no changes have occurred yet). If the diagram has any new partitions or unallocated space remaining, cancel the operations with the red curving arrow in the toolbar that says “Undo last operation” (you will have to click it until the list at the bottom is gone). If you have no unallocated space and a nice large BTRFS partition, you can press the green tick to apply operations.

Optional Step 5B: Demolition Day (Deleting Windows)

Please read step 5A first to become acquainted with the GParted tool. We will be using it again here.

If you’ve made it this far, then congrats! You’ve become a serious hardcore computer user without even realising. Drive management, switching operating systems, and dealing with the BIOS are all things that you might never have had to do in your life. Props to you for not sticking to the easy path. It also probably means that you’re not using that Windows partition any more. That’s great! Let’s get it out of your way so you can fully enjoy your whole computer. Since you’re familiar with GParted this shouldn’t take long, but it is a few steps longer than Step 5A. Same as before, queue up all of your operations first before applying any. Make sure you’re happy with the resulting partition arrangement before you commit.

We are going to delete two partitions:

1. The “Microsoft Reserved Partition” that should have an “Unknown” filesystem.
2. The “Basic Data Partition” that should have an NTFS filesystem.

The Fat32 “EFI System Partition” must not be touched, and the rest of the partitions are for your Linux installation.

To delete a partition, select it and navigate to the rubbish bin/trashcan icon in the top toolbar (Or right click and select delete). If this option is not available to you it means that partition is mounted, and you can access the files inside it. Partitions must be unmounted to be deleted, which you can do from the right click menu (but please make sure you don’t unmount any of your Linux partitions or the EFI system partition!). This should leave you with a large amount of unallocated space once you have deleted the two partitions for Windows.

To add that unallocated space to Linux we follow the same process as in Step 5A. This will work best if you reboot into your live USB, as otherwise the Linux partitions will be mounted (and you can’t unmount them while you’re using them). We first move the EXT4 partition to the left side of the unallocated space, and then we resize the BTRFS partition to swallow up that space. Jump back up to 5A and follow the steps to queue that action up now, but don’t apply any operations yet.

Once you have completed this, check the partition diagram. You should

1. See two less partitions than you did before (Microsoft Reserved Partition and Basic Data Partition should be gone)
2. Have a much larger BTRFS partition
3. Have no unallocated space
4. Still have a fat32 EFI System Partition

If any of these things aren’t true then undo all of your operations and start again. If it looks correct, then you can apply all operations and wave goodbye to Windows!

Optional Step 5C: Maintenance

Before you start, make sure you have a backup solution sorted for your Linux installation like Pika Backup. Make sure you’ve backed up your system at least once.

We’re going to dip our toes into the slightly scary part of Linux now by using the command line. The command line is a way of using text to interact directly with a computer, as opposed to using a graphical user interface. You’ll also see references to “terminals” and “shells” so let’s do some disambiguation.

• The command line is the abstract concept of using text to interact with a computer. Command line interfaces (CLI’s) are specific applications of this concept to software or your computer directly. You can have a command line interface that moves and changes files, or a CLI for a specific programme.
• The terminal is a programme that gives you access to a command line for your computer. When you open or run a terminal, a window appears that you can type text into. This is where you type commands to be executed by the computer. Terminal’s are also a very old school term for a physical computer. More specifically, a terminal was a keyboard/display combo (sometimes with a mouse) that connected to a mainframe computer (it’s kind of like connecting to a server from a laptop or desktop).
• The shell actually processes the commands typed into the terminal. There are many different specific kinds of shell, but they all have the same job: to manage processes and execute commands that the computer receives (which can be from the user itself or other software).

So any time you’re running software and executing processes using text, that’s using a command line interface. You can use a command line on your computer by opening up the terminal programme, which will pass whatever you type to the shell for it to read and process.

Let’s give this a try. Hit the start menu key (what used to be your Windows key, sometimes it’s called the “Super” key) and start typing “Terminal”. Open the programme by that name and you’ll get a small window with your username and your hostname followed by a blinking cursor. This is where you type commands. We’ll start with a really simple command:

echo “Hello, world\!”

Copy and paste this command into your terminal by using the right click menu. Ctrl-C and Ctrl-V do not work in the terminal the same way that they regularly do (Ctrl-C is the command to interrupt and stop a command!). You’ll also notice that the exclamation mark has a backslash in front of it - this is because certain characters are special and get read differently by the shell. A full list can be found here. If you want to use a special character in a regular way, like as part of a name or other text, you have to use the \ backslash to “escape” the character and let the shell know not to attach the special meaning (this includes the backslash itself by the way!).

Your terminal should reprint the phrase “Hello, world!”. That’s what the echo command does: it “echoes” text and prints out what you tell it to (the text in the quotes is given to the echo command as an argument, which is what we call the parameters we provide to a command). There are loads of commands and they all have their own instructions associated with them, but we’re learning this for maintenance purposes so I’m going to teach you one specific one.

sudo dnf update is the command to update your installation of Linux to the next minor version. Breaking this command down, sudo is the command to execute the next command as a superuser or administrator. The terminal will ask you for your password to do this, which you should provide. The dnf first argument is actually a second command (sudo executes the first argument as a superuser) that calls out to the DNF software, which is a package manager. It looks after a bunch of software on your computer, similar to FlatHub but more for system components than apps. The update second argument is passed to dnf to tell it to run the update process: it goes to check all of the lists of software it has (so make sure you’re connected to the Internet) against the latest versions of all of that software and comes back with what it needs to update. It then downloads and installs the new versions. After you do this, you’ll have to reboot your system (make sure it’s finished before you do this).

In Windows, there is a background process that runs to watch out for updates. When it spots any, it will let you know that it needs to download and install them. In Linux we do this ourselves (you can set it up automatically but starting off this way is simple). This is nice because your computer will never decide to update itself at a bad time, but you do have to remember to do this regularly and make time for it. You should back up before each update for safety.

Correcting a mistake in my code

You should read the first post in this category to get up to speed.

I was doomed the moment I said “rocks are not particularly smart”. They have turned out to be smarter than I, because I messed up my simulation of their algorithm. This is a short post clarifying how I had to change the code and what difference that makes to the results. The original post will contain a link to this update as well, in the spirit of scientific transparency. I know this research is the most important thing in the world to a great deal of people.

In the original code, there was a line that was supposed to check if a new pair was already registered as a failure or a success. I use a text editor that has autocomplete, and I’m also apparently an idiot, because the line said this

if(!contains(truth_booth_successes, ele) && !contains(truth_booth_successes, ele))

you know, you can always make a boolean check more secure by just doing it again that should do loads for your code quality

Anyway, fixing that meant some of the other bits of my code that weren’t being used actually got used. This was the change I needed to find out that those bits were broken. Woo. You can litigate the full change on codeberg, but the only other thing I added was a small rule that if 8 or more correct couples get identified with more than 2 weeks left a win is automatically possible.

We come now to the bit that matters; did the results change? Just as a reminder I measured 7.019% as the chance of success per game with the incorrect method. The new results? You’re not ready for the new results. You were never ready for them.

The new, true, chance of winning is

8.25%

Yeah look this didn’t change much I just had to make the correction. Still would love to work on algorithms for reality programmes though!

The UK government suggests a plan to help AI firms access copyrighted material

On the 17th of December in the year we call 2024, right around the final weeks of work before Christmas and New Years holidays, the UK Government released a consultation on altering copyright law. As a person who’s been on the internet for one or two decades, this should be exciting to me. Copyright law has been brandished with all the grace and precision of a firehose in the digital world. It has menaced

• Fan games
• Video and film makers
• Authors of satire and fanfiction
• Authors of anything else that happens to fall afoul of rights holders sensibilities
• Musicians, especially remix and sample artists
• Painters and visual artists
  • Animators, forever the mistreated darlings of the internet, get double whammy for also creating video content
• Journalists poking corporations who don’t like it
  • SLAPP lawsuits are hardly digital-native but are alive and well here
• People who want to stream a song they like in a country that their streaming service doesn’t have the rights to broadcast it in
  • Ad naseum for TV shows, movies, books, yada yada

… to name a few. Any efforts to engage with the practical realities of dealing with copyright law have years of legitimate grievance and illustrative experience to explore the issue through, and it should be clear that the system is imperfect at realising its ultimate goal: to protect the rights of anyone who wants to create things.

If you’ve spent any time on YouTube in the last decade, you have heard about Fair Use. You have heard so much about it that you’ve become an armchair copyright lawyer.

It is bizarre and quite a shame that the UK’s first bold step to improve this takes it straight off the edge of a cliff.

The precipitating event for this consultation is not the wholesale abuse of copyright law to monetise old corporate owned art at the expense of new individual art. The UK government seems to struggle to admit this, but this is not an earnest, heartfelt attempt to nurture a creative economy that is one of our islands only remaining positive contributions to the world. I’m not saying that the government doesn’t care about art; £124.8 billion in gross value added to the economy means that no government can afford the luxury of being unfussed. I’ll be bold and lay the blame at the hands of two bullet points.

1. We’re fucking broke.
2. There’s this new industry that seems magnetised to investment cash: AI.

A consultation like this would never have originated if the AI sector did not possess absolute black-hole levels of gravitational pull on the wallets of the investor class. That cash doesn’t (exclusively) disappear up the rear ends of founders. It pays for servers to host all of the very important slide decks, rent for open-planned offices in impressive postcodes, and people (!!!) to try and make sense of underspecified ideas. The UK government, starved of tax take and charged with empowering us to surmount the treacherous climb to a fat Greggs breakfast, has noticed that some of these people are employed in the UK. They pay taxes in the UK. It would be pretty great if a lot more of them lived and worked in the UK.

A good place to make AI products

However, it’s not as simple as asking them if they enjoy the idea of navigating our insane class system for the rest of their lives. One of several obstacles to making the UK uniquely attractive to AI developers is our insistence on enforcing copyright law, which prevents AI developers from consuming vast quantities of protected material without consideration for the rights of the producers of that material. At the moment, AI developers are legally obliged to pay the owners of the rights to books, newspapers, music, films, and a large part of the “high-quality” portion of the databases they use to train their models (in the UK). It’s not easy to catch them, mind: their training data is very rarely public.

This is the why, the meaning of the meeting between the Saville Row Suits and the SilValley Seed Fund Sweethearts. The central big idea of the consultation is a change to copyright law that would exempt users of copyrighted material from having to license that material for “text and data mining”. You gotta steal all of it, or none of it at all (and it actually might not matter if you’re not using it for AI specifically). Rights holders would be empowered to reserve the right to this exemption on the basis of explicit expression, so if you’ve made something and you haven’t outwardly communicated that you do not allow the use of that something for “text and data mining” then it’s fair game. This would be an “opt-out” deal, a tacit agreement-in-principle that you’re okay if your art ends up getting trained with.

The consultation mentions that transparency would be the price of compromise for these AI developers. How that transparency would be realised or enforced is left up the imagination, but they do say the word a lot. The consultation is also clear that AI outputs that are copyright infringing should, indeed, continue to be recognised that way. You can’t generate my hit album Drifter with a music generator and escape my wrath. My two fans will be sure to let me know so I can get your ass, don’t even THINK about it. The proposed exemption will be for developers to train their models with copyrighted material only, not to indemnify them from reproducing it. To be clear, I don’t believe anyone was seriously considering any other scenario.

I think we have an understanding of why the UK might entertain this. The stated aim is to encourage AI developers to “train their models in the UK”. The implicit aim is to generate tax revenue and employment by having UK residents doing the training on hardware in the UK. That’s the metric of success here before we start to zoom out. Will this have a net positive economic and social benefit?

I would wager that if this goes ahead, the vast majority of UK rights holders would just reserve their rights. Unless the process was made so incredibly frictional as to put off everyone involved, it is absolutely in the interest of anyone who has a copyright to do so. The gains for us creatives (and even the business moguls who frequently take ownership of our rights) are difficult to understate. Allowing data mining of our work enables machines to either reproduce or integrate that work into their product, with little hope of stimulating positive engagement with the original down the road. In a world where this happens, we’re back to square one. AI developers will still be unhappy with the large costs associated with licensing, and all the UK government will have accomplished is a fantastic exercise in bridge burning with every artistic and creative community on the island.

Of course, there is no guarantee that this mass reservation will occur. It could be drowned by a paperwork process of suitably crushing intensity, for example. When considering the effect a “free to mine” body of UK copyrighted material would have on the AI developers, we are forced to remember that the UK is supposed to respect the copyright of other jurisdictions as well. Maybe AI developers would welcome this small addition to their burgeoning piles of training data, but it does them little good when this exception doesn’t exist in most of the rest of the world. Training data drought is the defining issue of the AI industry today, with new models struggling to rise above their older counterparts without access to exponentially increasing human output. Generated material doesn’t cut it here without introducing model collapse, and I’m sorry to say that on a quantity level the UK’s move to redistribute its creative output would do little to make a dent in this issue.

I note that the consultation does not include a Theory of Change, even without the capital letters, which is an important part of the UK Civil Services approach to assessing potential policy interventions. I seem to only have the ability to theorise negative changes from this intervention, so I would greatly appreciate the guidance of the policy professionals who came up with this in proving me wrong.

Lets talk a bit more about this “train in the UK” goal. Developing these sorts of products is a complex affair that involves intersecting expertise, but it is not the exclusive purview of high paid, power-wielding “skilled workers” that it would need to be to really motivate this kind of politcal self-harm. The UK is stuck chasing the crumbs with this approach. Work around “training” is laborious, harmful, and low paid. It is ClickWork, micro-tasks around labelling and processing huge lakes of data that ends up on Amazon Mechanical Turk and other labour crowdsourcing platforms. What it is not is the sexy, big-salary work of developing model architectures and doing exciting R&D, or even the safe-bet sales jobs. It doesn’t produce those nice upper-middle class workers who can do things like buy houses and invest savings.

In short, it isn’t work that the UK government should want. It also isn’t work the AI developers want to move to the UK, because it’s already happening in places with lower labour costs. This mad dash for nothing is made hilarious when you think about the AI jobs that are already in the UK; high paid research, development, and deployment roles that centre the UK as a hub for technological progress and innovation while that word is still fashionable. It is true that there are much more of those jobs in places like the U.S., but it is also true that those places can pay talent much more and provide access to much greater investment. Funnily enough though, most of the locations where AI jobs concentrate have a similar approach to copyright as we have right now.

Compare this to the creative sector employees that this move would destabilise. There is a wide range of tax brackets and job types that live and work all across the UK, generating enormous economic and social value for their communities and the whole country. It feels uncharitable to limit the scope of assessment to just “how much economic value do you add” when a huge part of our islands continued global relevance is down to the art and entertainment we produce. British actors are such a common sight in Hollywood it’s become a meme. British authors have written world changing books (and only some of them have spiralled into a disastrous crusade against the existence of trans people). British musicians probably aren’t as good for stan twitter as Korean artists but they still have worldwide cultural relevance. There are children in the U.S. who sound like Peppa Pig. These are not preserved artifacts of our past either, but vibrant contemporary additions we’re making to the world every day. If we automated all of that to nurture a few more jobs in a sector struggling to find its raison d’etre, it would be at the cost of much of our remaining appeal to the rest of the world. You know, in addition to upending untold thousands of jobs across the country.

This is the face of thousands of distressed American parents

This consultation mentions the phrase “transparency” ad nauseum. It is, apparently, a necessary component to make this whole proposal work. How do you know if an AI developer has only used material that hasn’t had its rights reserved? It has to tell you all the material it has used or resign itself to drowning in requests for that information. This isn’t something that AI developers actually do right now, at least not the businesses creating products, and the consultation acknowledges this. It references “practical” issues around doing so, implying that the only reason we don’t have a full searchable database of OpenAI or Microsoft’s training data is because they haven’t figured out how to do that cheaply and easily.

This is, to use the technical term, complete bollocks.

An AI developer is in a tough spot business wise, because it has to create something meaningfully better than all of the other AI developers out there. It may be possible to do this with a novel technique in machine learning, but finding these new ways involves spending lots of money on research that isn’t guaranteed to pan out. It’s risky, it takes time, and it yields incremental improvement in most instances. A much more attractive method of exerting competitive control would be to refuse to share their training data. This is particularly attractive if you’re a larger company that already has a great training set; what you want to do is to deny future entrants and current competition access to as many resources as you can, forcing to burn cash and time recreating what you already have.

Training data transparency disrupts one of the most important tools you have as an AI developer to compete by forcing you to inform the world of how you built that key training dataset. New entrants can come in and copy your homework, leapfrogging you by not doing the hard graft of thinking about where to get all of that data and how it should be used. Maybe you sign deals for exclusive access to data sources to really make it difficult for anyone else to make the product you do, but then you have to negotiate with the data providers with everyone else. Not to mention, if you’ve been sloppy and “let” some unapproved data into your training set, you’ll have to deal with those annoying rights holders who want to know why you took their shit without asking. You might also get nosy parkers analysing your training set for awkward things like race and gender bias, which would be another thing to sort before the investor reports go out.

It is fair to say that you don't necessarily need the training data to identify biased models. The original research here was done by Julia Angwin et al, 2016.

Lack of transparency is a decision to support “good business” for AI developers. It strengthens their market position and ultimately reduces their bottom line. This doesn’t change with the proposed copyright law exemption, and the UK government should not be surprised to see a disappointing effort to avoid any kind of transparency at all. There is no explicit provision for how these AI developers would transition to a “transparent” approach to training data in the consultation, or how they would be regulated if they couldn’t meet standards, and I’m not surprised because there is no “transparent” approach that allows AI companies to hold on to that competitive lever. The UK might be willing to cut its nose off to landscape its own face, but AI developers are unlikely to go down that same path.

There is a really telling sentence in the consultation that the UK government hasn’t thought through this proposal.

Models trained in other jurisdictions which do not meet UK standards may be difficult to restrict from the UK market.

This tells me that the UK government is scared. They are worried that AI developers will take their ball and go home if they aren’t given what they want, seemingly without any thought or hardship. The response from the UK government has been to originate a policy proposal that in the short term will damage the creative industries and, in the long term, not do enough damage to them to empower the AI developers to want to stay. I should point out that AI companies are already in the UK under the current copyright regime, and that threats to move production to a more permissive regime should be met with “ok lol where”. The U.S. copyright holders will still need AI developers to pay for licenses to their stuff. So will the EU ones. So will anywhere that doesn’t exempt data miners from their responsibility to compensate the creators of that data.

The consultation presents this as a “win-win” approach. In a sense, I agree with them. AI developers win in the short term as they can exploit UK copyright material while the rest of us are navigating whatever system they put in place. In the long term, they win again when they can just move themselves back out of the country when they decide they no longer like the rules and too many people are reserving their rights.

Creatives and artists will continue to suffer under a dysfunctional copyright system. The UK’s only major move here has been to change it so they suffer more. In part 2 of this indulgent screed we’ll explore some other issues around suffering and AI. I would like to end this part with a suggestion. Protect the sector that has proven it will generate value through some of the worst funding periods it has ever seen. Protect the sector that employs UK residents everywhere in unique, vibrant, precious jobs that can’t easily be picked up and plonked in a fresh oasis campus backed by oceans of U.S. funding. Keep making kids in the U.S. speak like the funny animated pig.

Please god, just give a shit about the arts.

The copyright system will not save us

I find myself surprised at the full-throated evangelism for copyright holders to come and rescue the fate of the artistic world from the slop sheds of generative machines. Perhaps I shouldn’t. Though the issues in the intro to part 1 are real and important, the advent of machine learning consumer services has dropped a few very frightening issues into our laps that go beyond the scope of respecting the rights of creators and artists. In a strange parallel, the consultation on AI and copyright law goes far beyond the scope of the suggested intervention in its own text, wading into the mud of deepfakes and discerning AI content from human creation.

This is an odd alignment between some of those who hate the government’s proposed copyright fix and the government itself. Copyright has become a beacon of hope for those worried about the damage we’re starting to see come from generative machines, the weapon we need to rein in dangerous software. In some sense, I believe that people have witnessed the devastation left in the wake of copyright abuses on the internet specifically and wish to see that awful destructive power turned squarely towards plagiarism parrots, mockingbird music generators, and skin stealing software. It’s a “get the nearest biggest gun because that thing looks terrifying” response. From the government perspective, copyright is a battle-tested legal framework. If it can stop the abuses of generative machines that otherwise warrant new case law and possibly even the passing of legislation, that sees off a problem before it has the time to snowball into something they have to include in a manifesto.

I get in, on some level. But I can’t shake the feeling we’re stuck trading one melon for another. Copyright is not designed to protect us from the harms of generative machines and, as evidenced by the consultation, isn’t up to the job in practice either. We need specific systems designed for that, a set of laws, technologies, and social understanding that are focused on stopping the harms of generative machines.

This is not the destructive power you are looking for.

Let’s talk about your face. You own it, right? You carry it around on your body and show it to people, you wash it and keep it clean (i hope), you use it to express emotions and dazzle your enemies with radiant beauty.

Yeah, I’d say you own your face. So would the copyright framework of a lot of countries, as far as I know. Well, kind of.

Photographs of you are an original artistic work, and would be the property of either you or whoever took the photo of you. But your actual, physical face? Nah. You don’t have any copyright protection for that, because it’s a physical thing. This leaves us with an interesting question; if a generative machine spits out a picture of your face, to an accurate enough level that someone would recognise you, do you own the picture? If it screams and wails in your voice, begging for deliverance from assured death and mutilation at the hands of druglord kidnappers, with enough accuracy to terrify and traumatise your closest loved ones into immediately wiring every pound they can spare to try and save “you”, will you be able to sue the cyberskinwalkers for infringing on your copyright?

I’m getting ahead of myself, and the modern copyright framework we live in. These are questions that we just haven’t answered to my knowledge. I don’t expect we’ll have good answers for a long time.

The consultation states that “It is possible that an approach which enables individuals to reserve their rights for copyright works at the input stage may also help them control the use of their voice or appearance at the output stage.” This is, as I talk about above, a position that relies on the proved lethality of the copyright system to bring decades of case law down on the heads of the infringer. Such a great and terrible force of nature would surely cow those deviants who were planning on making the machines spit out pictures of real faces and voices, right?

I would like to point out that, if you were planning on committing worse crimes anyway, the spectre of being pulled up on copyright infringement isn’t going to make a meaningful difference. Some of the major harms of this technology were already blatantly criminal, such as non-consensual sexual images and phishing scams. Yet we still see people harmed in these ways, and we see the people causing this harm facing little consequence for doing much more often than a well functioning system should allow. Copyright is not the nuclear-grade deterrent it needs to be to change this. If you weren’t planning on being caught doing any of this, then it changes nothing.

Harms arising from generative machines are a deeply serious issue. On an individual level and a societal one, these technologies are causing real pain to people from all walks of life. Sometimes these are as sharp and undeniable as a picture of you that was never taken in a position you would never want to be seen in, or your voice saying things you would never want to say. Sometimes they’re part of the unseen ocean of bureaucracy and opaque machinery of life, a notice of a rejected job application or a denied credit line that happened to be issued by an algorithm that noticed you had a Black-looking name. I’m not the expert here and I don’t need to badly rephrase the brilliant research in this area: just read anything published by Timnit Gebru, Emily Bender, or the wider Distributed AI Research Institute sphere as a start. That’s a recommendation for the reader and a command for the UK government, who would benefit from giving these issues their proper time in the sun. Folding them into this consultation is misguided, and would feel like a cheap attempt at poisoning the well against criticism of their proposed intervention if I was more conspiratorial.

Above all else, the frustrating part of this consultation for me has been a failure to grapple with why these systems need the data they need. The reason why they’re built to consume, and the purpose of their existence. It’s really hard not to feel confused at what the proposal writers think AI even does; part of the consultation talks about limiting the scope of the proposed exemption to certain models based on their purpose or their size, which sounds to me like creating room for discussion about “enabling models that are built for the public good”. It feels like a line is deliberately not being drawn between the consumer-facing models that create text/sound/images and the analytical, relationship-describing models that might, I don’t know, detect cancer cells in an image.

To be clear, I know this line is a blurry one. All models are relationship-describing, after all. The “generative” ones describe the relationship between an input and the likely distributions of text/pixels/sound that input was assigned to.

What kinds of model do you think rely most on copyrighted material? Does a cancer cell detector benefit from training on Kandinsky and Rothko? Would a health monitoring model for axles on a train benefit from training on The Pharcyde and Eddie Grant? There are specific applications and products that are really hungry for this data, and I think it’s mostly the ones that want to produce “competition”. It’s the LLM’s that are vying to wall off internet users from the internet and keep them in dark grey chat windows, or the image generators who want to churn out endless stock imagery for corporate presentations and press releases.

“Competition” is perhaps charitable for me to say. A line in the consultation reads

For example, where a song is AI generated and labelled as such, music listeners would benefit from this information to make informed choices on the music they listen to.

Pray, what choice would you imagine a music listener to make with this information? How would their listening experience be altered? A brief tangent: I got got by those “(a) sip.” style videos on YouTube. If you’re not familiar, they’re hour-long playlist-style videos full of middling lounge jazz music that have no credited performers. This is because they’re fully generated, but I wasn’t really paying attention when I was busy working and needed some inoffensive noise to keep the brain on track. However, on learning that they were generated, I (rather dramatically) felt ill. I felt lied to. I felt cheated, because I would never willingly choose to listen to model muck when there are so many real people making music for my enjoyment. Surely the only choice available to “music listeners” is to decide how much of their listening diet is generated vs human created?

This is why this post is very light on images by the way. I don’t think you should have to have generated images shoved in your face to make my point, they’re very off-putting.

Where these inferior products succeed is in the business-to-business (b2b if you’re super lame) context. AI developers could offer a cheap endless stream of just-above-crap to play over another business speakers, or in the back of their videos and podcasts, or to have as B-roll, or to fill out reports and articles and all kinds of business detritus that barely gets engaged with. This is the only context where I can see AI labelling as any kind of positive, but only if it becomes associated with a lower price point, a thing that tech companies are famously quite good at abandoning when they get enough users. At whatever price point gets set, that’s money that could otherwise be paying for real human artists and creators.

The purpose of these systems is to replace human creation, at least partially. That is the mission that all this copyrighted material is being marshalled for. It is below the bare minimum of moral consideration to ask that the developers of these systems be forced to meet the price that human creators set for their own sidelining. The consultation uses the word moral one of four times to say this.

Copyright underpins the success of our creative industries because it guarantees their economic and moral rights.

There is very little that feels moral about any of this.

Writing code is just as much “writing” as “coding”

I do a few different kinds of writing. I blog, I write lyrics, I’m writing a novel and some shorter form prose. None of these are the main kind of writing I have ever done. My primary word-weaving activity is writing software, and I am beginning to realise that treating it separately from other narrative activities is harming my ability to do it well.

No, I’m not just talking about design

This is distinct from observations about the wider process of creating software, the point about “coding ≠ programming”. That’s true and could do with repeating, but it leaves room to conceptualise the actual process of creating code as a mechanical affair. You do your thinking in the design stage, and then you just press the buttons on your keyboard that makes it do what you’ve designed. Politely, I think that’s rubbish.

For almost all modern programming languages, any language where source code contains words, you are engaging with what you type semantically. You are making choices about vocabulary and meaning in addition to registers and bytes. You are writing in the same way that I’m writing this blog, and you’re probably doing it with very little of the mindset that you need to really excel at it.

STEMLord programmer bad?

Don’t take offence at the last bit. I’m not insulting you personally. I’m calling the popular approaches to writing software uncaring and ill-equipped to deal with this. There are decisions common to lots of different styles of software creation that fly in the face of what we know to be important parts of the writing process.

• Programmers will rarely be given space to recreate an existing functionality or feature without addressing a new requirement. That could be fixing a bug, adding new things, or making a change to the way software behaves.
• Common practice for adding new code is to create a merge/pull request that relates to a single “branch” of changes. Reviewers will ask for new changes to the branch if dissatisfied with the changes/if something doesn’t work, but I have almost never seen a request rejected and the developer asked to start from scratch. If this ever does happen it is a sign of extremely bad code.
• Old code is very frequently repurposed as the basis for new code. “Iteration” for programmers can often mean changing the shape and behaviour of what you already have/what you’ve inherited from something else. In some frameworks, like Agile, this is by design. You make something work for one set of requirements, and then you change it so it works for the next set as well.
• Breaking down complex software into manageable chunks is crucial to get the whole working, but it does mean a complete-context view of what a programmer works on is optional. A point of pride for some programmers is creating completely context-insensitive code that “would work anywhere”, even if the likelihood of it being used like this is slim-to-none.

Let’s compare this to the processes most writers use.

• Writers create drafts. Sometimes it’s a few, sometimes it’s enough to drown in. A writer is expected to create the “same” narrative several times over until it is at the standard an editor or client will accept.
• Drafting does not prevent old ideas from being reused in new narratives, but it does encourage you to discard what no longer works. “Iteration” for writers can involve starting from scratch multiple times, and you’re unlikely to waste effort recreating something bad or even only okay.
• Writers need to maintain a complete-context view of what they’re working on. Even in a distributed team where you might be in charge of a single characters dialogue, your work must meet a minimum standard of cohesion with the broader whole to be accepted.

Creative destruction

The popular mindset for creating software leave almost no room for experimentation, flying in the face of all the evidence that humans almost never write something perfectly the first time. If you can make something good enough to work and be accepted once, it is quickly calcified by the next feature that builds on top of it. You don’t start again unless there is no other option. There is a reluctance to discard completed work and re-evaluate the importance of things written long ago without the external pressure of bugs and breakage.

This manifests in so many small ways throughout a programmers day. Project schedules with no time built in for rewriting. Creating issues that prompt you to create merge requests paired with branches, so you immediately start work on what you will one day try to add to the codebase. Linear burndown charts and time trackers. It is easy to be shaped by the environment we work in into prioritising the next feature and avoiding the introspection that good writing requires.

This is only partially addressed by refactoring, which does not happen in practice the way it should happen in theory. If you have “working software”, you don’t have to go back and improve it, even if you really really should. Refactoring is mistimed by design because it’s supposed to occur on software that already meets requirements in some way (and may already be deployed!). You are encouraged to make changes in a small, non-breaking way, and I’m talking about giving yourself the room to completely gut a project and start over. You need that space to really think about what you need to keep.

This blog post was inspired by the latest thing I’m working on at my job, and my own bad habits when it comes to programming. We require a whole host of new features that, if I was committed to preserving the code as it currently is, would create a bad programme. It would make life “easier” for some of the others developing on it, as they wouldn’t have to redo any of their own work, but I know that the added difficulty in navigating the resulting mess would do all kinds of horrific things to our budget and deadline. So I’m ditching the old code.

I also have a bad habit of git commit regret, where I notice a missing change or a mistake three femtoseconds after I finalise changes. My commit histories are messy, full of reversions and indecision, and altogether indicative of an evolving understanding of how the narrative needs to be written. So I’m going to make a draft branch. It will eventually result in code that works, but it will never see the light of day. Once I’ve marshalled my thoughts and understand exactly how the object needs to work, I’ll do it again from a fresh start.

Measure twice, cut once

“Writing is thinking. To write well is to think clearly. That’s why it’s so hard.” - David McCullough

You’re thinking when you write code. Without the ability to rethink and redo, you are stuck with the first thoughts you had when you started. They’re bulging underneath the surface of all that you do, forcing you to navigate around unseen mountains of old junk memory.

Maybe this says something about how much some programmers value their own intellect, that they believe that every idea they birth is worth holding onto. Maybe the popularity of LLM’S with programmers indicates our collective disrespect for source code as narrative. We are being out-thought by machines that can’t think at all.

Writers draft, edit, cut, and rephrase. If we embraced these behaviours, and built tools that encouraged them, we would create more precise narratives. We would think clearly.

How my life and the world was shaped by a single set of chords

Disclaimer: this article is about a song called “Gypsy Woman (She’s Homeless)” by Crystal Waters. The word “Gypsy” is a racist slur and I don’t support it’s use in regular communication. This article has much to say about a song with that word in the title, but none of what I say will defend or excuse the use of a racist slur. So you don’t have to read the word over and over again, I have abbreviated it to GW in all but one instance.

Reflex

You’re at a concert. A large one, seven people deep either side of you and enough distance from the stage that the performers look like matchstick people. Wisps of light suggest faces in your peripheral but your eyes are barely getting through to your brain. Your ears are in control, and your mind is dominated by a short saturating silence. Then you hear this sound.

A roar erupts in every direction around you as the crowd collective nervous system is electrified. You know these chords. Every one around you knows these chords. You’ve heard them your whole life in every place imaginable, on morning radio shows on the way to school, blasting from tinny shop speakers, at a million other gigs and festivals and club nights and house parties and in the scores of TV shows that can afford licensed music and Doppler shifted from car windows as they fly past. More than any national anthem or other stilted musical culture-building exercise, you feel a connection with these chords in your bones.

I think this is how people feel about the chords from Gypsy Woman, a 1991 dance track by Crystal Waters. I’ve been in that concert crowd. I’ve heard the roar. This track still lives in the hive mind of the U.K. over thirty years later (the world as we’ll find out later), and I thought it would be cool to experience its journey through music for myself.

So I listened to every single version of it I could find.

Awakening

There are 116 songs that sample GW registered on friend of the blog https://whosampled.com (plus 42 covers and 13 official remixes). I made a graph of when they were released.

I find a few things interesting about this.

1. The drop-off from 1991 is precipitous. We go from 14 samples in the year it came out to 3 in 1992, and we don’t see a full recovery until 2013. That’s over two decades of GW barely seeing homage in other music, after a huge glut of references in the year it came out. Did it oversaturate the ears of the listening public?
2. Then we get to 2013, with an enormous 10 recorded samples. This is also the year that remixes shot up to their all-time high of 6. What happened exactly 21 years after GW came out?
3. Up to the present day we have sustained interest in using this as a sample. If we take some basic 3 and 5 year moving averages we can see what I would described as a continuing increase in the usage of the sample, and it looks a bit too noisy to call whether we’ll see another GW Winter soon. As we’ve seen, when people lose interest they lose it fast.

I had a few things I wanted to know beyond this though. How many times were the chords sampled, and in what years? What about the vocals? Did I like every song that used these crack-laced keys? In order to answer these, I listened to every single song registered on WhoSampled as sampling GW.

All of them. And then some.

Mania

“Und wenn du lange in einen Abgrund blickst, hörst du noch einmal der Akkorde von GW.”

• me, right now. No I don’t know who this Neat Chi is.

I deliberately didn’t record the length of these tracks because I have no business knowing how many seconds of my life I’ve wasted on this. Great scholars don’t blink at the sacrifices our craft demands of us, the lengths to which we need to go to really advance human knowledge. I chose to view this as my monthly “Possessed by the spirit of stupid” moment and dive right in.

• For most songs, I listened to the whole thing. The exceptions were particularly long dance tracks and medleys.
  • For the long dance tracks, I struggle to believe the intended experience was for me to sit at my desk/schmoove in my living room on my own for the whole 6-8 minutes. These are normally cut into a set for less time, and once I felt as though I had the vibe down I made sure there were no midpoint switchups and moved on to the next one.
  • For the medleys, I was essentially just listening to the original again. Didn’t think they was giving me much useful information. I confirmed what they were doing with GW and moved on.
• I had pre-created the list of all the songs to listen to before I really got started, and I listened in chronological order.
• For each track, I took some notes and tried to figure out what it wanted from me. A lot of these were dance tracks, and so if you want to imagine me doing a little shape cutting as I was having my meltdown go right ahead.

That’s about as complex as the process gets. If you want the whole list of results go here. Instead of subjecting you to All Of That, I also noted which ones I wanted to highlight. Apparently there are only… 21 of those.

Strap in.

Parce Qu’On Est Jeunes - Benny B

So things were different in 1991, I’m told. This sort of thing was not the complete headfuck it is for me today. I highlight this as the first non-English language offering on the list and boy. They really said what they said. I don’t think I could understand it even if I was a French speaker, and the delivery makes me feel out of breath, but you’re hear to hear about GW chords and not my breakdown of flow. The chords here are missing their spice, and ⅔ of their original variety, and you can just see how it turns them into this background boring musical feature. I doubt anyone who listened to this was inspired to go “ooooh those chords are really nice, I wonder where they’re from?”. This is an example of early stage treatment of GW: nobody knows what they have yet, and nobody knows what to do with it.

Megamix - Crystal Waters (1991)

For the majority of 1991, the person who was trying to get GW into their music was… Crystal Waters. This is one of a number of “Megamix”/medley style tracks this will end up on, and this one is exclusively her own music. Interestingly we see GW sampled for use in remixes of her other hit from around this time, Makin’ Happy.

Boliyan (Dance Hall Mix) - Anaamika (1992)

Starting off with O Fortuna is a wonderfully dramatic way to begin your EuroDance/Dancehall twin vibe warper. The first half is an “eh” dance remix but the hard pivot into full on Jamaican Dancehall deals me unrecoverable psychic damage every time. Is this an elevation of GW? Does it rely on it? The answer to both of these is a hard “Christ on a turntable”.

Oh La La La - 2 Eivissa (1997)

We jump to 1997. History is almost over, and in Europe people are celebrating with the cheesiest, campiest dance music you could possibly imagine. In the depths of GW Winter, with the last sampling before this happening in 1994, 2 Eivissa have the bravery to be the ones to bring these chords out of the darkness and into the strobelights. This track shows the power those chords have: for the part of the song where the chords aren’t playing I could honestly take it or leave it. But as soon as those chords come on I get that feeling, that reflex, that soul-shaking joy at that particular wiggling of the air in my head. I would go feral if this came on in a club.

Favourite Drug - Styles P (2006)

Wow, we’re already six years into the new millennium? Did we solve Y2K? Anyway, this is one of two songs in 2006 that sampled GW. Both of them are hip hop, and I’m highlighting this one because fuck T.I. and this one is better anyway. Styles P is one of the first artists to really move these chords into a genre outside of dance and house music, and crucially to a setting where you’re not exclusively expected to listen in a club. The chords here are clearly the backbone but they fit snugly into a well produced, smooth track with great vocal performance from Styles and Rashad. Clearly GW has places it can go, and this song is a harbinger for much genre experimentation to come.

What That Neck Do - DJ Rashad (2007)

This is one of a few songs that has had a second life as a meme. In this case, Gorbachev didn’t want this work. I think it’s pretty cool to see how GW ends up with an internet presence through samples and memes like this, and this footwork take on the chords is another new space to push them into. If I’m honest the better footwork song for this is Gypsy Theme by DJ Tameil, which is more danceable, but this is definitely a better meme song than A Pimp Named Slickback by Lakim (which ended up on TikTok as a slowed and reverb kind of deal).

Another Dimension - Timmy Vegas feat Bad Lay-Dee (2008)

I remember hearing this on the radio when I was a kid, so I had to let it slide into the highlighted list. I think this has the same experience that Oh La La La does with the chords: as soon as they come on I love the track, so it’s no wonder they had them for the chorus.

Walking - Mary Mary (2010)

I’M WALKING

Funny chorus aside, this is actually quite a fun track. Similar to What That Neck Do, this has definitely seen some life as a meme. This is also not the only Christian song on this list, which is unexpected (yeah I didn’t realise it was about Jesus either I thought it was just about walking).

Down for the Count - Girl Talk (2010)

This is one of a few mashup tracks that feature GW, and it’s the best one I know of to be honest. It has a clear mission: contrasting instrumentals with rap vocals. What’s interesting is that we really start seeing GW originating as a “classic of the 90’s” with items like this.

Honey Baby (Demo) - Kali Uchis (2013)

This is NOT the “SPOILED!” version that released later on TO FEEL ALIVE, but the one that was cut from Por Vida. It’s probably the moodiest take on the sample so far, almost being played at half time to evoke a real ghost pop sadness, and I absolutely love it. It pairs so well with Kali’s sweet-sad singing voice, and considering this is a demo it’s shabby production adds a lot to the melancholy, homespun feel.

Sometimes GW makes your song worse. Sometimes GW makes it better. Kali demonstrates that you can completely change it on your own terms, if you have the vision and talent to do it.

Libre Para Amarte - Gloria Trevi (2013)

You ever hear about “Zero interest rate behaviour”? It’s a term I heard emerge not long into 2024 to describe business acting as if investors would throw them infinite cash for wispy, paper-thin ideas (you know, like they’ve been doing since 2008). I feel like I want an equivalent for music like this. “Zero barriers-to-publishing behaviour”. Nobody was there to give you the artistic criticism you needed to do any better.

This also demonstrates something I saw a lot in my journey. There were a number of songs that were just completely their own, doing their own thing, and then would throw in the GW “La Da Dee, La Da Dum” into the chorus (or a version they sang themselves). Why? Do you think that you can distract me by referencing a better, more likeable song? Do you know who you’re dealing with? In most cases this was a pretty major black mark for me, I was not impressed at all.

Talk To Me - Nick Brewer feat Bibi Bourelly (2015)

This is an early example of the kind of sampling I had in my head when I started this project. It’s crass corporate tone-stuffing to endear a random pair of artists on the rep of a cultural titan. The worst part is that GW is good enough to make me like the chorus of this, when it goes for being as close to a proper house track as it feels like it can manage, and we will see more like it in the coming years. Demonstrating both the good and the bad of borrowing GW in the same song, the post-chorus features Nick weakly aping the La Da Dee’s without the confidence to demand he gets centered in the mix.

For The Ladies - DJ Pied Piper (2017)

Now I know what you’re thinking. The sample is

• Not the chords
• Comes in right at the end

But in order to think this you will have had to experience the rest of this song, and so you’ll already know why I had to put this up. This feels like everyone’s uncle went “Oi, this new music is crap, we gotta come back and save it with garage”. Let me tell you something. This did not save us. Lyrics like “I got girls that’s white and black, and I give em the booty smack” and “The ones that give disease on my penis” are not saving us from anything. Shoutout to the chorus artist who put out the most nasal “ladies” I’ve ever heard in my life.

Fluid - Amaarae (2017)

2017 also gives us one of the best uses of GW chords with Amaarae’s Fluid. It’s lush. It’s sexy. It takes the chords and does something different with them, gating them to give them a satisfying smooth wobble, and her vocals are precise and perfect on here.

We Got That Cool - Yves V feat Afrojack and Icona Pop (2019)

This song made me finally realise that Afrojack and Afroman aren’t the same person. I really though this would have a feature with Afroman telling the police to give him back his money but oh well. This is another notch on the list of “Songs I’d hear at a festival and use as a cue to get some food”. You don’t have that cool. GW has that cool. You have the business sense to harness it.

Couldn’t Be Me - Lucki Starr (2021)

Lucki Starr knows she’s good at this. The instrumental, including GW, is there to provide the perfect stage for a fantastic performance from her. Eminently head-noddable and would be a really fun time at a club, and my favourite aspect is how she incorporates these post-chorus breakdowns where the chords drop. It adds a lot to the landscape of the track, really does the whole “greater than the sum of its parts” thing.

I Must Apologise - PinkPantheress (2021)

Changing up the chord progression slightly lets PinkPantheress focus on the organ’s classic sound, calling back to GW without feeling like a complete ripoff. She’s a master of the niche she occupies, with a driving drum break providing a solid base for her to croon over. Impeccable 2000’s vibes.

Deep Down - Alok, Ella Eyre and Kenny Dope feat Never Dull (2022)

I highlight this to point out that you can fly too close to the sun doing all this musical referential bollocks. Deep Down uses the GW chords and the melody from The Bomb [These Sounds Fall Into My Mind], another club classic, and yet it falls short of them both pretty badly. That’s not to say the track is bad, it’s just “okay”, but I’m sure there’s a mashup of these two somewhere that goes much harder. For me, Deep Down suffers from a boring festival drum line and over clean production. It probably doesn’t help that this song got very popular in the U.K. because of Love Island, and we know how I feel about that programme.

SPIN BACK! - Scootie Wop (2022)

If I had to sum this up in a word: “adorable”. I really like the use of the “How do you want it?” sample, I always though these chords would fit incredibly with some proper harmonised vocals (and I have a soft spot for gospel affectations). The lyrics start off quite fun, with his “drip like an old man”, but I’m not here for the preaching it devolves into. Weird that this is the second Christian song to end up using the sample, and that both of them were fun?

WiiSTYLE - Ashbeck (2024)

This is the first drill track on the list that uses the GW chords, and it’s fucking sick. The production (shoutout El Londo) really chops up the chords and uses a replay with a Rhodes-y keyboard to provide a slick lofi styling to a bouncy set of drums, and Ashbeck brings the heat on the mic. The direct sample gets a healthy dose of spice when it does come in, a nice playful nod to where the smoother synth is reading from.

I’M HIS, HE’S MINE - Katy Perry feat Doechii (2024)

This is what we’re ending the highlights with, and it’s the second-last sampling of GW as of the time of publication I can find. I think it’s a perfect encapsulation of the low points of GW sampling in the modern age. Produced by that prick Dr. Luke, on an album where Katy Perry’s lead single went down in flames for a piss poor satire of the resurgence of anti-feminist sentiment, I’M HIS, HE’S MINE is mostly a shallow trap pop hit-hopeful that has Katy Perry doing spoken-word not-rap basic rhymes on her chorus. The laconic “La Da Dee”‘s can’t save her from sounding bored and lazy (but not in a cool Charli XCX way), and when she does actually sing it’s only for long enough to make you go “Why isn’t she doing more of this?”.

Doechii really helps Katy dig her own grave here by being so much better. Doechii is fun, sly, and her lyrics lean into power and absurdity in a way that is genuinely a bit empowering and tongue-in-cheek. Katy Perry’s lyrics feel jealous in advance, possessive, and desperate. Who are you worried about, Katy? Is it me? Did you catch him staring at my bible quote booty shorts or something?

Fall

Coming out of the other side of my self-imposed GW purgatory, what did I learn? First and foremost, I was shocked that there are almost as many songs that sample the vocals (68) as those who sample the chords (87). For me, the latter are much more iconic - I’ve heard them in student jams, at gigs of all sizes, and that’s the part that makes me go wild when I hear it. The vocals are distinctive though, and maybe I should give them more credit for the track’s lasting cultural impact.

This process really tested my love of the sample, and I can safely say I don’t like most of the songs that use it. Not what I expected, but I might have played myself by listening to 90% of these tracks in 3 days. I said I liked 28 of them, and I loved only 4. But I did start to figure out what the best tracks do with GW.

• They take it to new places. None of my favourites are dance tunes, which is a surprise, and I feel as though I’ve picked only one or two in each genre that really do it for me. Highlights include the footwork interpretations, R&B with Walking, Hip Hop, ghost pop, and drill.
• They match the level of creativity and vision of the original. Crystal Waters was focused on creating a good track, but a lot of the samples were focused on borrowing that love for their own use. My favourite takes were the ones that had a clear sense of how they wanted to feel.
• I like the ones that manage to add elements that I thought weren’t standout in the original. Stronger, more emotive basslines and tighter drums are a big plus.

This sounds like I have the ingredients to do right by the sample. To manage what only a few others have done, but many have attempted. I have dreamed in silky organ tones, and I think I have proven I am the rightful heir to Crystal Waters’ legacy (you know, musically, not literally).

Tune in next time to see the light.

Who are you fighting when you fight against AI?

Here at Rust Red River, I like to keep things small. You can check out how I do that on the about page, where I also explain that I want this space to be maximally accessible. Whoever wants to read my silly posts about rocks falling in love should not be stopped by their choice of browser, the power of their device, the speed of their network, or anything else they’re using to read (or listen to!) my words. But sometimes big waves upend the ocean of the Internet, and my small site has to contend with things of a much grander scale.

This spirit of open access and the “website for everyone” has seen a re-examination in recent years. Large language models have corrupted the act of sharing online by regurgitating and misstating our ideas without our consent or our names, draining that cyber sea without a plan for a world without it. I recoil at the idea that you’re reading my words and thoughts as recounted by a guessing machine that you may have paid to plagiarise me. Others feel the same, and have decided to take matters into their own hands.

Access control

Have you seen this anime girl? Dial 555-5555-55 right now she is coming to GET YOU-

Anubis is a bot challenger that makes any HTTP requests go through a proof-of-work check before serving the content they’re requesting. It’s supposed to stop AI crawlers, but the documentation states that it’s not particularly discriminating about what bots it interferes with.

Anubis is a bit of a nuclear response. This will result in your website being blocked from smaller scrapers and may inhibit “good bots” like the Internet Archive. You can configure bot policy definitions to explicitly allowlist them and we are working on a curated set of “known good” bots to allow for a compromise between discoverability and uptime.

– Anubis docs

As a human, you’ll see a small progress bar with an anime girl above it before reaching the site you want to. One thing I’ve noticed is that, as someone who doesn’t have a flagship/top-of-the-line smartphone, these checks take much longer on my mobile than on my desktop. Like, “oh my god is this working or should I refresh” levels of long. This will depend on the administrator who sets the level of difficulty of the hash, so should not be seen as a direct criticism of Anubis, but hints at the idea that certain human requesters might have their experience ruined by technology to keep bots out. Anubis isn’t designed with my experience in mind, however. It’s specifically a tool to protect webservers from abuse by crawlers.

Anubis is one of many options for keeping bots away from your site. We’ve had these for years in the form of CAPTCHA’s, which have evolved to become arcane puzzles that humans struggle to make sense only marginally less than computers. Neither of these methods are reliable ways to stop your content ending up inside the rapacious maw of an LLM. If I, a human, decide I want one of these machines to summarise/analyse/butcherise anything you want humans to read, all I have to do is feed it myself.

I can copy-paste it into a prompt. If I’m using Microsoft Copilot, I can upload the file directly (and I’m sure you can do so with others). Anubis may prevent some of the more sophisticated machines from headlessly piping your website into a chat, but it is busy fighting a proxy war against crawlers. The real opponents are the people who want to sacrifice your work to the machine, and they’re only really being inconvenienced as opposed to effectively denied. So who are you blocking when you block crawlers?

The people who do get filtered out by Anubis and CAPTCHA’s are the people who don’t have powerful devices. People on old phones and computers, people with dodgy connections that might be materially impacted by the additional time to solve your toy equation. The kind of people who I spend time minifying this website for. I haven’t even touched on the tendency for CAPTCHA’s to be horrifically inaccessible for computer operators who are visually impaired (to the point where W3C has a page on it), but it would add a layer of irony to use one of these bot banning tools on this site. For transparency, I do use some bot busters on here that minimally interfere with a human operator. I have rate limits on my webserver, I use a robots.txt from the wonderful people at ai.robots.txt, and I will catch user agents who don’t respect that file for redirection, but I find anything that reduces the accessibility of Rust Red River a poor choice for protecting myself.

Active Sabotage

Tarpits, I’d bet my life on it

Nepenthes, Iocaine, and Quixotic are all tools that aim to “poison” AI and serve fake content. Nepenthes and Iocaine are hardcore tools that generate Markov Model babble on the fly, with Nepenthes acting (mostly) as a pure tarpit/flytrap and Iocaine designed to only catch LLM’s and otherwise let humans see normal content. Quixotic on its own is a less weapons-grade solution, simply creating a fake mirror-universe version of your website with the text messed up (though the developer does also have a linkmaze programme to generate endless random babble for crawlers). The rationale is that on an individual basis, admins can make their sites and servers hostile enough to crawlers to stop them from visiting. On a collective one, enough admins could drip-feed bad content into the training data for these models to degrade.

Once again, these are not reliable ways to prevent your content from ending up inside an LLM. But that’s not the stated aim here. These are tools to harm machines and the companies running them in two ways.

1. Ruination of the training data
2. Wasting the time, energy, and money of model owners (Nepenthes and Iocaine only)

If Language Models are Few-Shot Learners is to be believed (grain of salt, this is an arXiv preprint not a peer-reviewed paper), GPT3 needed about 45TB of Common Crawl data plus four other datasets (one of which was Wikipedia) to train on. The training approach listed in the paper makes it clear that not all data is treated equally.

Note that during training, datasets are not sampled in proportion to their size, but rather datasets we view as higher-quality are sampled more frequently, such that CommonCrawl and Books2 datasets are sampled less than once during training, but the other datasets are sampled 2-3 times. This essentially accepts a small amount of overfitting in exchange for higher quality training data.

– Language Models are Few-Shot Learners (2020), T. B. Brown et al, arXiv

When we think about how poisoned data would be treated in a system like this, we simply don’t know how companies like OpenAI rate the quality of data that they crawl themselves. It’s possible that poisoned data could be ingested into multiple datasets that get rated differently, and this all adds up to make it hard to understand “how much” poisoned data we would need to produce in order to affect these huge systems. If one of these datasets was successfully poisoned it would be noticeable to model owners at that stage, and it would probably lead to a reduction in the quality rating of whatever dataset was poisoned. Success would be rewarded with a reduction in the ability to do this kind of sabotage.

What about point 2, wasting the money of model owners? Every second of time spent gumming up these crawlers is a step closer to bankrupting these losers, right? These companies take longer crawling (and have to spend more cash on overheads to do the same task) and have to use more energy than they would, which costs money. But I think this is a wilfully ignorant position to hold. OpenAI spent $9 billion in 2024 per Ed Zitron. Many of these companies know they’re entering an industry where capital expenditure will be enormous and operating costs are unlikely to fall precipitously any time soon. How can tarpits be effective without deployment at a truly enormous scale in such a cash-rich environment?

If tarpit operators were making a meaningful difference, I believe they would be inflating the energy use of these companies. How is this energy getting generated, and who’s getting paid to do it? I can almost imagine a good answer to this question, where AI companies are forced to bankroll some kind of solar or renewable super-expansion. But Microsoft, one of the market leaders for this sort of thing, are focused on old nuclear plants and natural gas for their energy generation needs. The people of Memphis, Tennessee are being poisoned by xAI, who are using 35 methane gas turbine generators to run Grok. So who are you poisoning when you poison datasets?

If I do nothing, AI models, they boil the planet. If I switch this on, they boil the planet. How is that my fault?

– Developer of Nepenthes, who appears to have a cake and have eaten it too.

Ed Zitron, who I cited earlier, has a fantastic body of work on why the business models of all of these companies are unsustainable. You don’t need to add to the harm they’re doing to change that, and you can only avoid culpability for that damage if you have so little effect on these models that they don’t really change. Quixotic demonstrates that you can still try to poison these datasets without putting yourself in the logical bind that tarpits do, and for that it deserves credit.

Don’t roll over and die

This isn’t a call to ditch these tools altogether, but if you’re trying to resist AI I implore you to consider how your choices affect humans first and machines/corporations second. When we’re thinking about how to resist AI, we should avoid shadow boxing with crawlers and focus on the people who are building these awful systems (not individually but as a class/group). How do we prevent them from being able to build server-crashing crawlers in the first place? How do we disassemble this factory of nonsense that is buoying up the fossil fuel sector and poisoning our air? These are questions that need more than just a neat software solution, and the hardest question of all can only be solved by humans themselves.

How do I stop someone human from trying to put my words into a vomiting robot?

A little bit of knowledge is a very dangerous thing in space

I’m cursed. It’s a pretty unfortunate one to have for a space engineer - every single launch I’ve ever watched has experienced a failure. Chandrayaan 3 got off pretty lightly; the lander made it to the dark side of the moon, and things only went wrong was when it froze to death without the Sun’s light to keep it warm (the overall mission was a great success, and I commend everyone involved). When I had a small watch party for SpaceX’s first Falcon Heavy mission in 2018, we weren’t quite distracted enough by the mental dual-landing of the side boosters to notice that the livestream never showed the Heavy core landing. We later found out the Heavy had missed its droneship landing pad and crashed into the sea a hundred metres away (not terrible for a big-ass rocket), and the world did later watch Elon’s brain start failing about this and many other topics (perhaps I can’t take credit for that particular failure).

You, the astute and scientific reader, are probably lamenting the sample size here. There’s one more instance of my curse manifesting though, and it’s by far the worst one.

I bankrupted Virgin Orbit.

In October of 2022, Virgin Orbit were touring a scale model of LauncherOne. They were doing this to drum up hype for the First Orbital Launch From UK Soil EVER OMG!, which would take the real LauncherOne under the wing of a modified Boeing 747 to deploy an eager host of British satellites and prove that the UK is a Serious Place where Serious Things can happen (and Serious Money can be spent). Yet to become fully aware of my power, I thought it would be a fun thing to go to. I got to speak to one of the launch engineers who would be on the plane, which was brilliant, and I walked away with two of these mission patches in sticker form.

I immediately stuck one of them to my computer, the only sticker I’ve ever actually had on any of my machines (except for an American Airlines patch on my Gameboy???). It’s safe to say I wasn’t just passively curious - I was excited. Come January, after a bit of forth-and-back about the launch window they could target, I made sure I watched the whole thing.

Watching the mission fail was a surreal experience. The sight of a jetliner taking off with a rocker under one wing is not quite the bombastic inferno of conventional rocket launch, and the view of the rocket igniting on the camera was mostly of a white cluster of pixels turning blue. Still, everything was looking pretty decent until we passed the first stage. I remember thinking to myself at T+0:05:20 “hmmm, that’s a pretty extreme amount of oxygen left for all that fuel burn,” and wondering if the telemetry was behaving. Then they stop showing the telemetry at around T+0:06:26, midway through burn 1 of stage 2. T+0:06:43 I know something is wrong because the speed value does some sort of overflow to 4,568,313,429 miles per hour, and at T+0:06:51 the broadcast cuts to a broken Google Chrome tab briefly.

The interactive trajectory of the rocket, which previously was flattening to look roughly parallel to the curve of the Earth, stops dead and starts floating straight up slowly. I watched this line for about five minutes. They try to switch to Maspalomas ground station to get a signal off of the rocket, they show some random control room footage, they even resort to filming the crowd of people who went in person to Cornwall to watch. But at T+0:15:27 they run out of room, and cut to the mission patch they gave me at the model rocket. Half an hour later, at ten minutes to midnight, we get confirmation that the mission has failed.

Virgin Orbit never recovered from this. They had publicly disclosed some rough finances in November 2022, and the spectacle of such a public failure was too much for the cash-strapped crew to handle. After two months of relative quiet, the company paused operations on March 15th. Most of the employees were laid off two weeks later on the 30th, and Virgin Orbit formally declared bankruptcy on April 4th. Not even the UK Government would buy them, despite having bought a different space company (OneWeb) out of bankruptcy three years ago. Perhaps they lost the appetite for bets on space start-ups. It later emerged that a $100 filter dislodged to cause the failure.

You know what launches I don’t watch? Any of the Starship ones. Even the one where they caught the booster (not the Starship, the Super Heavy).. So you can’t blame me for the recent string of explosions, even though we have obviously established that I’m cursed.

• Flight 7 is the one where they did the chopstick catch I linked above. But the Starship itself exploded over “The Caribbean” (i cannot tell you how much this wording pisses me off as a person of Caribbean heritage). Debris was sighted in The Bahamas, Haiti, The Dominican Republic, and minor damage was reported in Turks and Caicos. Newsweek has a cool map. The root cause was established to be propellant leaks created by an unexpectedly large harmonic response during flight.
• Flight 8 also had a Starship explode, but for a different reason than Flight 7. A Raptor engine suffered an undisclosed hardware failure, and this knocked out a few of the surrounding engines. The spacecraft lost control of itself, and once again, debris rained over “The Caribbean”. I am being rapidly radicalised into forcing my whole industry into remedial Geography at gunpoint.
• Flight 9 was a shitshow and you can’t change my mind. More leaks meant the Starship lost attitude control and couldn’t control itself, leading to yet another uncontrolled reentry over the Indian Ocean, and its payload bay door malfunctioned so it couldn’t deploy the fake Starlink satellites it was equipped with. The Super Heavy booster, reused from Flight 7, was run at a more aggressive flight profile for testing and disintegrated over the Gulf of Mexico after relighting its engines (it was intended to splash down).

The most recent entry in this saga is the explosion of Ship 36 on the test stand. Fail fast, indeed. We don’t currently know why Ship 36 got so over-excited. SpaceNews states that “A cloud of condensation appears at the top of the vehicle just before the explosion, which could be evidence of a tank failure.” Musk tweeted something about a nitrogen pressure vessel, but it’s the second half of that tweet I want to highlight.

If further investigation confirms that this is what happened, it is the first time ever for this design.

Listen, I don’t live in the man’s head. But I would imagine the point of adding this to his “preliminary analysis” tweet was to be a positive message. Perhaps the implication was supposed to be “This is a good part, this was a low-likelihood event and will not require lots of time and money to fix.” An assurance to people who trade in Elon Stock that there’s no need to panic, and that Tony Stark hasn’t lost his mojo yet. But this is not what that sentence means.

Starship has a curse all of its own.

Let’s talk about how a spacecraft is tested. Like any extreme environment (the deep sea, volcanoes, hell), space is merciless in invalidating conventional design. Much of the stuff that makes us fawn over the mythical Rocket Scientist comes from the ability to figure out how to do things so differently that the environment becomes merely harmful as opposed to crushing - surviving radiation doses that we can only recreate with nuclear explosions, enduring temperatures that violently jump between below freezing and close to melting point, impacts with other objects at speeds above two million miles per hour. Our best possible way to even get to space crams more obstacles down the throat of the space engineer; put simply, every kilo that you don’t have to take up matters. You have to be Earth’s strongest robot soldier AND our skinniest runway model if you’re a spaceship.

Testing for space begins at the most basic level. I know how my craft will behave when

• All of the materials I use have been characterised, examined, and reported on across the spectrum of conditions I can run into in space
• All of the individual components (made of materials tested for space) have been thoroughly tested to failure
• All of the subsystems (made of components tested for space that are made out of materials tested for space) have been tested at multiple points during their assembly in a representative range of situations for my mission
• The whole craft (made of yada yada you get the bit) has been tested during assembly and integration, on the ground
• The whole craft has actually been in space and data has been collected about what it’s doing up there

Building this reverse pyramid is slow going. Raw unit production in the space industry is very small - we launched 263 times in 2024 per Jonathan McDowell and that’s the highest I think I’ve ever seen. The money and time spent doing rigorous testing and characterisation at every level of the spacecraft is a direct result of this: the only real alternative would be to launch knowing less and lose more spacecraft. Out of those 263 launches, 258 reached orbit successfully.

Ship 36’s failure post-integration is bad for a number of reasons. The vehicle was in preparation for a test, but none of the engines were lit yet. It wasn’t “in operation”, so to speak. If Musk’s tweet is correct, it’s a failure mode they had no awareness of until now. This is a disaster for their testing process, which is there to try and catch these behaviours before they pose a risk to the whole spacecraft.

When the sample number is low, a single event casts a long shadow. Is this a harbinger of more failures to come? Or a freak occurrence? Chances are we won’t know until more Starships have tried to launch, as six of the ten launches so far have been failures (if we’re counting the test stand explosion) and they’re constantly changing the design. In space, it takes a long time to know anything at all.

What does it mean if you keep failing for different reasons? For Starship, we might look even deeper than the basic materials the thing is made out of. SpaceX is famously a pretty intense place to work, ten-to-twelve hour shifts for six days a week and a polarising culture. Musk’s other company Tesla has been alleged multiple times to have an uncaring approach to safety and testing, going as far as to leave injuries off the books. It might be said that Musk’s companies are more focused on “Results” and “Progress” than testing and rigour. Yet, with six new ways to fail revealed in the last four flights/tests, SpaceX needs a critical safety culture to grapple with their recent performance.

This is a tall ask for a company steeped in the mythology of “The Best”. I’ve routinely heard people talk about how smart SpaceX employees are, how they’re the absolute pinnacle of our field and working there means learning from the single greatest cadre of craft-lobbers the world has ever seen. This is a thought-terminating cliche. It shuts down the ability to critically examine the methods and performance of The Best, because the answer is a foregone conclusion. But reality will only suffer this delusion briefly. The next Starship might be fine, and have a great flight watched by thousands. It might explode even earlier or do the Astra slide. You simply can’t know if you’re too busy being The Best to bother finding out.

That’s the thing about a curse; It can only terrify while it’s mysterious and arcane. Understand what obeah you’ve got, Starship. Get these tanks exploding in controlled testing environments instead of out on the pad.

I’ll be watching.

The ultimate enmeshment of corporation and state is the most boring part of the internet

I grew up hearing a lot about the “percentage of internet users” within a country.

Oh, you know that North Korea has internet users? Only a thousand, but I thought they weren’t allowed to even access the outside world!

They’re really getting everyone online in China, it’s all part of the move to become the next superpower

I think every child in Africa should be given a free laptop. It will change their lives. They’re living in a different world out there.

These are paraphrased sentiments I heard from the adults and precocious kids around me as I grew up. “Percentage of internet users” was a stand in for a nation’s development, a crude measure for the prestige and worth a country was entitled to enjoy. It wasn’t socially acceptable anymore to derive your superiority from sweeping judgements about what “those people” were like, the kinds of food they ate, the language they spoke or the colour of their skin. Internet access, clean water infrastructure, the “safety” of a country - these were (and still are) the proxies used to continue the questionable exercise of ranking places and groups of people. When people got into the discussion of why these percentages were different, these old prejudices were quick to burst through the paper-thin facade of enlightened liberal conversation.

It is a far cry from the original cypherpunk dream of a hyper-libertarian country-on-a-circuit. I think that vision contains the seeds of the idea of a state built with IT in mind, sharing power through digital networks and signalling allegiance with mouse click and text transmission. Maybe Timothy May and company were right about the ability to rule each other through a wire, but the people who really built that system were those who already had the capital to rule in meatspace as well. When you think about a virtual state now, do you picture free Cyberia and a punk bar in New York? Or do you think about the Doughnut?

I think we can agree that the internet and the state form a twisted knot.

Into this mix arrives the idea of “Digital Inclusion”, or “Digital Equity”. People should have access to all of the technology they need to fully and richly participate in the community and society in which they live. Sounds noble and plausible to me, especially in the world of today’s digital nations, but I’m biased - I volunteer in this space. Specifically, I help people access and navigate the internet.

• I teach typing, basic computer usage, digital hygeine/basic cybersecurity
• I provide advice on services and options available to people who need to use the internet
• I keep people informed about upcoming changes (Windows 10, online safety act)
• I demistify complex topics like AI, cryptocurrency (blegh)
• I help people navigate important government services to access their entitlements. Benefits, free information, contact with their representatives, etc.

In this role, I can perhaps see the vague ideal that those know-it-all worldly types were thinking of when they lauded internet access as an objective measure of progress. Online government services are accessed through an internet connection, after all. I’ll quote the foreword from a policy paper delivered to Parliament in the UK.

We want to use technology to grow the economy, reform our public services and make life better for every citizen. That means creating a future where everyone can safely access and use digital products and services that could make their lives happier, healthier and easier. A future where it is easier to connect with family and friends, engage with public services, or find employment in the industries of tomorrow, but where everyone can continue to access offline support if they need it.

This foreword is attributed to

• The Rt Hon Peter Kyle MP, Secretary of State for Science, Innovation and Technology and serial ChatGPT guy
• The Rt Hon Wes Streeting MP, Secretary of State for Health and Social Care
• The Rt Hon Bridget Phillipson MP, Secretary of State for Education
• The Rt Hon Liz Kendall MP, Secretary of State for Work and Pensions
• The Rt Hon Angela Rayner MP, Deputy Prime Minister and Secretary of State for Housing, Communities and Local Government

These are some huge names, and I’m not trying to imply they’re wrong. That same paper has a section with some compelling statistics about the effect of digital exclusion. The paper references this report from the Centre for Social Justice which says:

…Our research shows that consumers could pay 25 per cent more for an illustrative series of transactions, including for insurance, phone contracts, and food.

…working age consumers without the internet are considerably more likely to be unemployed and so miss out on the benefits of employment yet require the internet to engage with Universal Credit. 22.1 per cent of working age people without access to the internet at home are unemployed compared to 3.8 per cent nationally.

This is specific to the UK and from this specific paper as far as I know. If I was in a position of power and was shown these statistics, I would be screaming in the streets and throwing phones and unlimited data SIMs at everyone I saw.

I would have been, until I realised that the real problem is that everything relies on email.

Public services are for everyone (or at least they should be), but that doesn’t mean that everyone uses them equally. In my experience, people who have an acute need/reliance on these public services are less likely to have their own computers and smartphones. Smartphone possession is more common, but I’ve noticed that having one doesn’t always mean using its full range of features. It therefore cannot be taken as read that every person who accesses a public service in the UK has consistent, painless access to their email, even if you think that’s true across the broader UK population. So why is email the accepted standard for enabling access to so many of these services?

The digitally native are already aware of how much of a nightmare it is to lose access to your email. Why would we accept that model for access and the accompanying risks for life-altering stuff like social security and benefits? Why do I get this lurching pit in my stomach every time someone tells me they don’t know the password to their email?

There is no public email provider for personal use in the UK as far as I know. Private for-profit corporations have unwittingly become a dependency of the state, which relies on plentiful access to free email addresses without asking any questions about where people are supposed to get them from. This might sound insane, but if you could just go to a place or call a person to get access to your email again, it might be acceptable to rely on it as a key to accessing money and medicine. I don’t expect Google and Microsoft to provide this service for free because they’re companies, and so I propose we just don’t expect them to provide any service at all.

Governments who want to put the state on silicon should provide a true end-to-end service. Governments should offer, if people want and need it, a free email address.

In an ideal world, these services would be designed to avoid the reliance on email entirely. Maybe they could even prioritise those who need it the most, or at least consider them - what would a service built for people with no fixed compute look like? How do you manage security in an environment like that? I’m of the firm opinion, after watching people forced to press key and tap glass for the bare essentials, that we have deeply embedded a class hierarchy in how we manage digital services. Quality of life is not indicated by high speed internet connection and unlimited bandwidth - it is gatekept by them. I don’t think it had to go like this.

When those nation-ranking types talk about the percentage of internet users, they conjure images of iPhones and laptops abound, a server for every person and a person for every digital island. It is shorthand for owning steel and the time of the people who die making it, and it has nothing to do with a countries ability to meaningfully improve life with any of it.

Zen and the art of waiting for flavour (A recipe for a roast chicken)

A small-to-medium whole chicken at Sainsbury’s used to cost £3.25 when I was at uni. Most Mondays I’d come home with one in my bag next to my laptop and my handouts. One of the few things I can multitask properly is this recipe, because I spent years trying to fudge my way through my degree on the kitchen table as I studied with one eye and checked the oven timer with the other.

This is not that chicken. This chicken benefits from years of experience, and a desire for you to skip my mistakes. Don’t skip the lemon and salt. Be generous when you baste. If you have the means, purchase a chicken that had a life worth more than £3.25.

The basic marinade flavours here (dark soy, light soy, mirin, and lemon) are great on a whole range of things. I’ve had it on tofu, roasted veg, duck, pretty much anything that can be roasted or grilled. Coincidentally, it’s also very close to the flavour profile I make egg-fried rice with - I used to serve the two together all the time. The picture has it served with homemade sweet potato fries and kale, which aren’t part of the recipe.

EQUIPMENT

• One oven
• One roasting tin (i.e. a baking sheet with walls)
• A bowl or container that can fit a whole chicken and be covered
• A sharp knife
• A spoon or a baster

INGREDIENTS

• 1 whole chicken, with organs/giblets removed. You can use the giblets to make stock or gravy (though I don’t have a good recipe for that yet!). Most chicken is sold without these anyway.
• 2 to 3 lemons, depending on the size of the bird
• 1 bulb of garlic
• A handful of thyme stems (Narrow leaf thyme is easier, but you could probably do this with Guyanese thyme too)
• Olive oil to your taste
• Salt to your taste, plus a handful for cleaning
• Black pepper to your taste

For the marinade

• 2 to 3 tablespoons of dark soy sauce
• 1 to 2 tablespoons of light soy sauce
• A dash of mirin, or another sweet rice wine-based vinegar
• The juice of half a lemon
• Two sprigs of thyme leaves, or half a teaspoon of dried thyme
• (Optional) A teaspoon of chilli powder
• (Optional) Some additional crushed garlic, about 2-3 cloves
• (Optional) A little bit of oil, if your bird is very very lean. I used to use olive oil for this.
• (Optional) A bit of grated ginger

TECHNIQUE

Before we begin, I need you to understand a few things. You’ll need to know how to

1. Clean and prepare a whole chicken
2. Fill the void under the skin with marinade, without tearing holes and letting it all leak out
3. Identifying when a chicken has finished cooking

Please do not skip this section. Chicken’s a serious meat and a whole chicken comes with a lot of strangeness that chicken pieces do not present on their own.

1. Cleaning the chicken

Yes, I am aware that washing chicken with running water risks splashing the germs onto your surfaces. I’m not recommending you do that. But trust me, this is not the same thing as chicken arriving in the packet “clean” and ready to eat. You do not have to accept the rank smell of decay that accompanies unprepared chicken. You can use lemon and salt.

To safely clean a chicken, make sure you start by scraping out any debris or leftover detritus from the inner cavity into a food waste bin or similar. Trim the ends of the wings and ankles, and any hanging skin flaps inside the cavity and the opening to the neck. Rest your chicken in a large bowl or container that can be covered, and pour the juice of one lemon and a handful of salt onto the chicken. Rub the lemon juice and salt evenly over and inside the bird. You will be scraping this off later, so it’s helpful to use salt with large grains (I think they call this “kosher salt” in the U.S.??). Cover and seal the bowl/container, and leave for half an hour in a cool place. For larger birds, you’ll need larger amounts of lemon and salt and more time. You could also use something like white vinegar - I use lemon juice because I like lemons :)

After time has elapsed (don’t leave it too long!), pick the chicken up and hold it over the bowl. Scrape the salt and lemon juice off into the bowl, and rest the chicken in a roasting tin. It is now ready to be marinated and seasoned.

2. Marination

You will need very clean, dry hands for this next bit. It’s gonna feel a bit gross.

Lie the chicken on its back, and locate the side opposite to the large cavity. This is where the breasts of the chicken meet the neck, and is where you’ll be working. Your aim is to loosen the skin on top of the breasts without introducing new holes in the skin, and pour the marinade into the gap between the skin and the meat. There will be two gaps (one for each breast).

The skin is held to the breast by a thin film, which can be pierced with a finger or a sharp knife. If you use a knife, do not pierce the skin itself. Once the film has been detached, the skin should feel loose but still broadly close to the underlying meat, which is perfect. Do this for the other breast, and if you feel as though you need more room, push your fingers further under the skin to loosen it more. Avoid loosening the skin too much at the back of the bird where the cavity is - you don’t want to introduce an exit hole for all your marinade to spill out from.

Now go and wash your hands, you dutty pig. Well done on doing some serious chickening. When we make the marinade later, we’ll pour half in each gap.

3. Identifying when a chicken is finished cooking

Do not serve undercooked chicken. Ever.

Listen, part of this comes with experience. You get used to cooking a bird of a certain size and weight and you get comfortable with how long it takes to finish up. I’m not trying to scare you here, because the reality is that you can always put the bird back in the oven until it’s done, but I do want to be clear that you should err on the side of caution here.

This recipe is for an open-roasted chicken at 180 degrees Celsius (which is just a hair above 350 Fahrenheit). The Chicken Maths is as follows at this temp:

cook time = ( chicken mass in kilos x 40 mins ) + 40 mins

So a 2kg chicken would take around 2 hours in the oven. A 1.5kg chicken would only take 90-100 minutes, but a 1kg chicken would still take around 80 minutes. Some places do the Chicken Maths a bit differently, and you may have your own formulations for the ultimate cook time. Personally, I like it when the chicken is practically falling off the bone.

Whatever you do, when you’re first starting out you should double check the bird is done before you rest it. The accepted wisdom here is to pierce the thickest part of the chicken thigh, and see if the juice runs clear. This isn’t really going to fly for our marinated chicken, but there is a key colour to watch out for - don’t serve chicken with juices that run red or pink. If you’re willing to sacrifice a thigh for the cause, then slash from the thickest part of the thigh right down to the bone and observe the colour of the meat. If there is any pinkness, back in the oven.

INSTRUCTIONS

1. Pick an album you like, a good book, a TV show you want to watch, or something you can do while you’re waiting for the bird to roast. I’ve been liking Shay Lia’s work on Dangerous (Deluxe) this month.
2. Lemon and salt your chicken as described in Technique stage one.
3. Prepare your marinade while the chicken cleans. In a bowl, mix dark and light soy with the juice of half a lemon. Add thyme, mirin, and optionally some crushed/minced garlic, chilli powder, olive oil, and ginger. This is what’s going to flavour your bird, so don’t forget anything you’re hoping to taste!
4. Remove the lemon and salt cleaning mixture from the chicken, and loosen the skin on the breasts, as described in Technique stage two. Pour the marinade under the skin, half for each breast, and leave the bird to marinate for half an hour at least. If you are marinating for longer than an hour, you need to cover and seal your chicken and put it in the fridge. Make sure you take it out between 30-60 minutes before you cook it.
5. Preheat your oven to 180 Celsius (~350 Fahrenheit).
6. Slice your bulb of garlic in half equatorially (i.e. splitting all the cloves in half) and push the garlic into the cavity of the chicken. Take the thyme sprigs and place them inside the cavity as well.
7. Coat the skin and outside of the bird with the juice of half a lemon, a tablespoon of olive oil, salt and black pepper. Be careful not to push marinade out from under the skin. Immediately place the chicken into the oven.
8. Go do your thing for 20 minutes. Read your book, study, learn to tap dance. Set a timer so you don’t have the mental effort of trying to keep track manually.
9. After 20 minutes, pull the chicken out of the oven and baste it with the juices it has produced. Take a spoon, or a baster if you have one, and scoop the juice from the bottom of your roasting tin onto the top of the bird. The first baste is always the hardest - as you go, more and more juice will be developed. Ensure that the chicken skin is well coated, but don’t take so long it starts cooling down. Get it back in the oven in 2-3 minutes.
10. Restart your timer and practice waiting. You’ll be basting every 20 minutes until the bird is done, as described in Technique stage three.
11. Once your Chicken Maths Calculated Cook Time has elapsed, take the chicken out. Test if the chicken is cooked, and baste once again.
12. After 15-20 minutes, your chicken is ready to be carved. I recommend taking the legs off first, then the wings, then the breasts.

As your reward, you should help yourself to the Parson’s Nose, the tiny thing at the cavity entrance. You should also have at least one of the “Oysters” - succulent chunks of meat on the underside of the bird that are almost worth roasting an entire chicken on their own. Give the other one to your favourite person. If you’re super savvy, you can keep the carcass to be used to make chicken stock.

VARIATIONS

• You can put an onion in the cavity, or (yet another) half a lemon
• You can rub whatever spices you like on the skin. Maybe some chilli powder, cumin, and coriander.
• The marinade can be adjusted to fit your taste, as long it contains a bit of sweetness, a good amount of salt, and a bit of acid
• Goes fantastic with a roast set - potatoes, roast veg, etc.
• Goes fantastic with egg fried rice

MUSINGS

This dish isn’t particularly attached to any one cuisine in my mind. I’m Guyanese, and this comes from my Guyanese family, so I guess it’s a Guyanese recipe - maybe that’s why there’s a whole load of thyme involved. The process of cooking this should be just as carefree and relaxed as the process of eating it. You’re getting the full experience when your kitchen is filled with the rich aroma of bubbling oil, fresh thyme, roasting garlic, and lemon, while you sit with friends and family enjoying something completely unrelated to the impending deliciousness. Don’t fret about a minute here or a slight overspill of spice there. Getting good at this recipe will teach you how to wait, and that’s a skill that serves you for life.

Space data centres will benefit almost no one

The idea of data centres in space entered my world in 2022, when I read about Intuitive Machines’ IM-1 mission being the jumping off point for data centres on the moon. This was ludicrously funny to me, as I imagined the idea of an incident response technician getting strapped into the world’s cheapest rocket and launched up to the heavens to plug in a loose ethernet cable, and continues to be a source of inspiration for some of my sillier stories. Unfortunately in the year 2026, the words “space data centre” are more likely to make you think of Elon Musk blotting out the sky with one million satellites. It has been mentioned by some that space is not the optimal home for data centres, but the stars are home to many impossible dreams. There is little stopping anyone from chasing them if they remain blind to how the pursuit of these sci-fi futures affects the real people crushed at their feet.

I was originally going to write more on the reasons why I think space data centres won’t work, but to be honest that’s not the main problem I have with them. It’s really easy to get sucked down a rabbit hole of “cooling in space isn’t as simple as people think it is” or “that company has done something way less impressive than they’re claiming” or “there’s no way Elon will make it to even a tenth of a million satellites before we suffer a collision event bad enough to materially limit our ability to even have satellites.”, but stopping there robs us of more important questions. What are space data centres going to give us? What will they take?

You already know why people suddenly want data centres in space. It’s the same reason we suddenly want a whole load of them on Earth too. Generative machines are the business model of all time, and if you don’t start renting intangible things like “answers to your questions” and “the ability to draw a cat” then we’ve apparently exhausted the well of ideas to sustain infinite exponential growth. In a world where people like having drinking water and breathable air, shunting all that capitalism out of sight may well feel like putting it out of mind. I would like to argue instead that it limits the ability for people to keep it in check.

A large part of the despair at these data centres on Earth is at the kind of computation that goes on inside them. Grok’s stint as a tool for Sexual Abuse As A Service was/is the most visible example of some of the worst kind of demands placed on generative machines. Flock cameras in the U.S. and Palantir software are two exposed parts of a deeply entrenched system of digital surveillance. Arguments about terrestrial data centres are also at risk of devolving into technical debates with a myopic focus on what can be measured, and this fails to capture some of the true disgust around their encroachment on our lives. It doesn’t matter if I can prove that data centres will put communities under water stress or not, when my objection is to the data itself and the way it is used.

It’s worth noting here that a “data centre” does not simply contain generic computers, that perform every task at the same efficiency and performance. When you put an enterprise GPU or a TPU in a server rack you are deploying specialised hardware for rendering and analysing digital images, as well as running generative models. This is starkly different to, for example, the “dark fibre” telecomms infrastructure left behind from the dot-com crash, which doesn’t mandate you to make a specific kind of data for it to transport. Therefore, objections to the kind of computation that data centres facilitate are also objections to the data centres themselves. You could have used that silicon to build a library computer. You’re using it instead to spy on me, and sicken my mind.

Data centres in space must at least face the same questions around the harms they enable on Earth, and the resources they divert. In addition, there is the bizarre and somewhat toothless world of Space Law to deal with (my apologies to my space lawyer friends, it’s not personal, you’ve all got great teeth). Can data be illegal in space? Can operators of data centres in space be meaningfully examined and held to account for illegal data? We will not have answers to these questions within any reasonable time frame if space data centres were to persist past their myriad technical problems.

Let’s briefly remind ourselves of the difficulty that space data centre operators are choosing for themselves. Ionising radiation in amounts much higher than on the Earth’s surface are likely to shorten the lifespan of non-shielded electronics, like GPU’s and TPU’s. The technology to replace parts of a spacecraft with no human intervention is not ready for commercial deployment and won’t be for years. Therefore, every time you want to replace a chip or a set of chips you have to throw the whole sat away and launch a new one. For this and many other reasons, you would split a “data centre” across a fleet of satellites instead of concentrating them in one place. Ballooning the satellite population like this makes the job of a fleet operator that much harder - look at the increase in the number of collision avoidance manoeuvres as SpaceX’s Starlink constellation has grown.

From the perspective of the business owner, why would you accept this bargain? “Free” power and cooling are thrown around as the plus points. I support the Taranis analysis that neither of these things are really true, and so far the only GPU-dedicated spacecraft I know about uses a conventional design that doesn’t treat power and cooling as “free”. That same spacecraft has managed to run Gemma (a lightweight version of Google Gemini) and train a very small language model (NanoGPT) on 1 megabyte of text (The complete works of William Shakespeare), using a single H100 video card that could be drawing anywhere between 0.35-0.7 kW of power. These are tasks that are possible on retail hardware, and do not require any kind of data centre. If performance isn’t meaningfully improved, costs aren’t reduced, and demand isn’t conditional on your product being in space, what is the point?

I can only think of the issues I started this section with - accountability and legal enforcement for data in space is non-existent. With the right corporate structure and ground station setup, I am sure you could effectively evade liability for all kinds of criminal and unethical behaviour by hiding your computers in space. Do I think this is why Elon Musk has announced the end of the night sky? Not really. I have a pet theory that Starlink and this new constellation are so large to help Musk lock in obscene amounts of U.S. taxpayer money (that’s probably for another episode of blog). But I do think, as several governments get very uncomfortable with the rampant criminal image generation facilitated by Grok, that the thought might have crossed his mind.

So far we’ve focused on the data centre side of the equation, but I haven’t talked about what space data centres do for the space industry itself. Even if you’re not working on them, the effect these constellations will have on all of our missions cannot be discounted. We’ll have to do more dodging, wasting fuel and mass to accommodate fleets of space chips. We’ll have to compete directly with the monstrous capital reserves of the tech industry (yes, we are separate - I would be paid double to quadruple for an equivalent role in tech) for the same tiny pool of talent. We do not have anywhere near the level of volume to absorb the demand shocks these companies are famous for; if every gamer in the world isn’t enough business to secure at least a fraction of NVIDIA’s goodwill and attention, how will the needs of tens, maybe hundreds of regular non-data centre space missions find a place at the table? It appears obvious, at least to me, that a wholesale embrace of the space data centre is a bad thing for the space industry.

It does pay lip service to a deep insecurity in our sector, around commercial potential and how an industry like ours finds its true alignment with the cause of capital. As it stands the space industry is dependent on state funding to really exist. This is embarrassing for ideological capitalists, who posit that business models that can’t generate enough value on their own should fold. Space data centres would finally provide the elusive commercial customer, the non-state entity we’ve been searching for who can pay enough to make it profitable to launch and operate satellites. We used to have that in the form of telecomms and broadcast, until that aforementioned fibre build out completely changed the cost dynamics of serving TV over broadband instead of through satellites. Now we only have oblique downstream applications, and must be satisfied with knowing that the industry is valuable in ways unmeasured by a balance sheet.

If we remained determined to see the activity of launching and operating satellites as something that must be commercially viable one day, then we remain vulnerable to ideas like space data centres. Ideas that make us weigh up the cost of environmental desecration against future revenues, and whether it’s worth it for us to support harmful and unethical computation. Why are we scrambling to help unscrupulous generative salespeople evade accountability for what people use their machines for? Why are the talents of spacecraft builders being marshalled to create these flotillas of disposable GPU’s? How will we deal with orbits strewn with cheap, failed hardware when the cash tap runs dry and no-one is left to pay to clean it up?

When the chips are cashed, don’t let anyone tell you we couldn’t have known what we were getting into. It’s obvious. It’s trivial. It’s disappointing.

What’s the software version of a bindle?

Preamble

I wrote a guide on how to switch from Windows to Linux back in October 2024. At that time, Windows 10 was a year away from having support cut off, and I felt strongly about giving people options that weren’t the ad-ridden, insecure, bloated, spyware-infested mess that Windows 11 continues to be. I still think what I said in that guide was true:

As a user, regardless of how “good” you are at computers, you have the power to say no to decisions you don’t like. 

This guide is in the same vein as that 2024 one, where I’ll provide notes on the process of switching from one operating system to another. In this case it’s for server administrators, but if you find use as a desktop user then great!

Before we begin, you might like to know why I’m writing one of these for Debian specifically. Let’s talk about some decisions I don’t like in the world of Linux.

In August of 2025, I learnt that Fedora Linux would be aiming to be the “best community platform for AI” as part of its 2028 strategy. This would include using large language models in the making of the distro, updating things like package descriptions that “aren’t usually very meaningful”. It would also mean shipping models as part of the distro itself, with the example given being an “all-local voice assistant that doesn’t send your conversations to some big datacenter or try to sell you things while simultaneously selling your personal data”. Hypothetical magical leaps forward in software seemed to be enough to tempt the leaders of the Fedora community to uncritically integrate LLM’s into every aspect of making their successful, influential operating system. The voice assistant example is brought up with no reference to the well-known issue of screen reader performance in Linux, which makes it hard for me to believe that this is anything more than a bad-faith attempt to paint criticism of this decision as “anti-accessibility”. I didn’t like that, and I didn’t like how the project position on LLM use was suddenly thrust upon the community. It was a decision I wanted to say no to, and a power structure I could no longer accept my place in.

At the time I was a proud Ultramarine Linux user; Ultramarine Linux is a derivate of Fedora, and this news was set to directly affect my computing experience. Linux is an interesting environment for analysing the effect of these kinds of decisions - the Ultramarine devs made it clear to me that they would adopt an opt-in approach instead of pushing it by default, but Fedora’s all-in approach means much of the software within Ultramarine doesn’t adhere to that philosophy. It was, and still would be, far too much for me to expect the Ultramarine project to say no to this decision for me. How would they check every upstream component for LLM contamination? What if it appeared in critical system components like dnf? There are many people who only care if the software works, and would be comfortable using these packages so long as they were tested and reviewed. I had to say no for myself, and I embraced the void.

On the 12th of March 2026, systemd‘s source code repository would have a new AGENTS.md file added to it. For the unaware, this is a set of instructions for LLM agents (that is, “autonomous” software programmes based on LLM’s) to follow when processing the source code of software they’re pointed at. They might make these programmes output text that humans are happier with - they might not (preprint so pinch of salt as usual). They are always a clear indicator that the developers of that bit of software are happy with contributors using LLM’s, which is a problem if you’re maintaining a load-bearing component as important as systemd. The package is already one of the key toxic pits of argument in the Linux space, up there with vi vs emacs but without as many funny memes, and this open embrace of a controversial technology was (in my opinion) the next obvious way for the project to make online discussion spaces completely fucking unbearable. 

Somehow, it wouldn’t even be the most controversial thing systemd would have added to it this month. systemd might be the greatest bit of ragebait of all time.

On the 18th of March 2026 (not even a week later!), a pull request to add a birthDate field to systemd’s JSON user records was merged into the project. It was contributed to comply in advance with laws in California, Brazil, and Colorado mandating age verification, with little discussion about the best way to comply with said laws and whether it was systemd’s place to furnish that, and was accepted at least in part based on an LLM-generated review. In a similar vein to the Fedora LLM position and the AGENTS.md file, this request was accepted without broader engagement with the community of people who use and contribute to this software, and it has sparked furious backlash from people who are outside of these geographies and have no reason to be forced to accept this in their systems (as well as people within these geographies who think these laws are unjust and might not survive contact with the legal system). 

While I haven’t really engaged in this discourse apart from a joke note on the fediverse, I was pretty frustrated with this. At the time I ran a Debian server (which hosted this website) because I agreed with the stable release structure and having well tested, well documented software. But Debian uses systemd by default, and I was once again forced to consider how I could say no to these decisions. If systemd could become a source of so many issues so quickly, I didn’t want to have to deal with its drama. Luckily my previous move to Void meant my desktop setup was spared from all of this. Void uses runit as an init system, and it works great as far as I can tell.

In my haste I came upon the Devuan project, a fork of Debian from when systemd was first introduced to it. It was the nearest jump away from this thing that I didn’t like, but reader beware - Devuan is a trap for people like me who just want an easy escape route. Debian has a code of conduct, a normal thing for projects that established acceptable and unacceptable behaviour. Devuan doesn’t really have one and is happy to use it as a point of difference between the projects - this makes them the preferred OS of notable coward fascist Bryan Lunduke, and they seem to be happy to receive the advertising. In my dash away from LLM’s, I see no need to make peace with ghouls who are ambivalent about the safety and rights of marginalised people.

By way of comparison, here is the code of conduct for Alpine Linux. Here is the first paragraph.

We as members, contributors, and leaders pledge to make participation in our community a harassment-free experience for everyone, regardless of age, body size, visible or invisible disability, ethnicity, sex characteristics, gender identity and expression, level of experience, education, socio-economic status, nationality, personal appearance, race, caste, color, religion, or sexual identity and orientation.

Clear, immediate, and reassuring. Alpine Linux also has the benefit of being a super interesting and very successful Linux distribution - it is the gold standard for containers due to it’s tiny footprint and has a stated commitment to security and simplicity. It also uses OpenRC by default instead of sysv-init, which is what Devuan ships with and is also just not as good. I am happy to recommend it as a server distribution - it’s what I ended up switching to after my Devuan misadventures.

Outline - How to switch from Debian to Alpine

We’ll be going through the following steps:

1. Checklisting. We will make a list of all the important software our box is running, so we can mark off installing them again later.
2. Source checking. We’ll check all of the software we want is available in Alpine.
3. Backing up. We’ll copy our data and configurations onto another machine that will stay stable while we upgrade our server.
4. Making an alpine installer.
5. Booting into the live installer and running setup-alpine to install Alpine to the machine.
6. Registering an SSH key and adjusting the sshd configuration for hardening.
7. Going through our checklist and installing all the software we noted.

This guide does assume that you’re hosting on your own hardware, or installing Alpine to the same box. I’m not familiar with managing remote hosted servers, and I would imagine they would do the work of actually deploying Alpine to whatever you rented - but you might still find useful stuff here if you ignore steps 4 and 5.

1. Checklisting

You are probably using your server to do stuff. But do you have a good list of all the stuff it’s doing? Now is the time to sit down and figure out what you’re running. Here’s a list of things you might need to set back up on your new Alpine installation.

• A firewall like ufw
• A webserver or reverse proxy (nginx, apache, httpd, Caddy, etc.)
  • DDoS protection like Fail2Ban
  • PHP-fpm and other modules for processing PHP code on the server
  • Python or Django
  • Perl…? (I have used a perl script one time in my life can you tell)
  • Automated cert renewal like certbot
• A database solution like maria-db (formerly mysql) or postgres
• A CMS like Wordpress, Ghost, Drupal, or another dynamic website programme
• A container solution like Docker or Podman
• A mailserver (god help you)
• A media server like Jellyfin
• A personal federated social media instance like GoToSocial
  • Or a community one like Mastodon, Misskey, Sharkey, or Akkoma
• Server management software like uptime monitors, text editors
• VPN software like Wireguard or Tailscale
• Group cloud storage like Nextcloud or Seadrive
• Ad blocking like PiHole or AdGuardHome

There are many others I could have missed here, but hopefully gives you a good starting point for accounting for all of your server load. You should make a text document with a bulleted list of every bit of software you need, where each bullet lists the version and necessary post-install tasks. Order the bullets so that you install safer software first - Sort out your VPN and secure your SSH server before you expose your mailserver and any sensitive data. 

2. Source checking

Now that you know what your new box will have on it, go check the Alpine package registry for each bit of software you need. I would recommend making sure you can find what you’re looking for away from the edge branch and testing repo unless you’re comfortable with software that has a higher likelihood of breaking. The community repo is made in collaboration with the main developers of Alpine, and I think it is a safe place to get software from - we’ll talk about how to enable it when we install Alpine later.

You may find packages under slightly different names in the Alpine registry vs Debian, so use wildcard (*) searches to make sure you don’t miss any packages you need.

3. Backing up

Hopefully you already have a backup solution for your server data in place, but I would take a separate one just in case. The key files you may not already be backing up are your configuration files for your various services - nginx and apache configs, sshd, mail, docker, everything in the checklist might have a config you need to store somewhere else. It’s worth running through the checklist once and noting whether you need additional configuration or not, and grabbing a copy of it using rsync or scp (other remote copy tools are available). For my money, I always find it easier to invoke these commands from the desktop with my server as the remote as opposed to using these commands in the ssh session - means the server doesn’t need permission to ssh into my desktop. Store all of these things in the same folder as your checklist and data backup (website content, Nextcloud files, emails, etc).

4. Making the installer

Relevant Alpine docs are here.

Go to the Alpine Linux downloads page and select the image that’s right for your hardware. I had a really easy time with the extended image on my machine, and it would be my recommendation for a very smooth experience. Once you’ve downloaded the image, plug in your USB and get the path to the partition you’ll be writing it to with blkid. The path /dev/sdX will be to a device, so make sure you use the full path /dev/sdXY where Y is some number - this is the actual partition you’re writing to.

dd it onto a USB or SD card (whatever will fit onto your server) with the command 

dd if=/path/to/image of=/dev/sdXY bs=4M status=progress 

You can also use a GUI image writer like Fedora Media Writer, KDE ISO Image writer, or any other programme you’re familiar with. If you’re on Windows, BalenaEtcher or Rufus works well for this.

5. Booting into the live installer

Relevant Alpine docs are here.

Plug your installer into the server and double check you’ve backed up everything. Your online services from this box will be unavailable until you set them up again - make sure you have planned for the downtime. When you’re ready, plug a monitor, mouse, and keyboard into the server and reboot the machine. 

You may have to visit the BIOS settings first by mashing F10, F8, F12, F2, Del, or another key that interrupts the boot process. If you need to do this, you should update the boot order to read from your USB first (Some BIOS screens have a boot override that lets you boot straight into a specified device, which also works. Do you!)

Once you’re booted into the Alpine live image, you’ll be presented with a text-only screen and a login prompt. Login as root with no password. Alpine has an install script for easy setup called setup-alpine - you can enter that command to get started straight away. There is a quick option setup-alpine -q, but I opted for the normal installer as I am not based in the U.S. (so mirror defaults are normally a bit slow for me and the keyboards are normally wrong).

The main thing to watch is which partition you install Alpine onto. It will overwrite all data on the partition - if you use a multi-parition setup, make sure you pick the primary partition for the previous system you’re overwriting. Debian uses an ext4 filesystem, so that’s an easy one to look out for compared to the vfat filesystem for an EFI partition (though extended /boot partitions are also ext4 sometimes, so be careful). Alpine also uses ext4 by default - if you want something else like btrfs, you might need to do a semi-automatic installation and use the setup-disk utility.

Once you reach the end of this automatic installer, Alpine will be installed to your disk! Congratulations, you can now reboot into your new OS (you might need to do the BIOS screen boot ordering again). I think it’s worth it to verify you’ve installed it properly by reaching the login prompt (Alpine has no GUI by default) and log in with the root password you set, just to make sure you’re not accidentally booting back into the live image.

One thing to note is that Alpine doesn’t have sudo by default, and recommends setting up doas as a post install activity. You should do this now, as your first Alpine activity.

apk add doas 
echo 'permit :wheel' > /etc/doas.d/doas.conf 
adduser <not-root> wheel 
su -l <not-root> 
doas command with arguments 

if you want doas to persist for a bit instead of asking for a password every time, change permit :wheel in /etc/doas.d/doas.conf to permit persist :wheel.

6. Registering your SSH key and hardening your server

In case you don’t remember some of the basic intro steps to starting a server, check out my friend Kern’s self-hosting guide. In that guide, one of the steps is to generate a SSH keypair, but wait a minute - you’ve already got one!

You can use ssh-copy-id to add your existing key to the Alpine server using the username and password that you set up during the install (I recommend doing this with a non-root user), or you can log in directly and copy your public key into the .ssh/authorized_keys file for your user. I then recommend following any of the suggestions in this guide to hardening your SSH server.

7. Installing from the checklist

Now for the fun stuff. You have a big list of things to do from step one, and you might want to approach each bullet atomically. Start at the top with the safe stuff, and go through your checklist. The package manager for Alpine is apk, and install commands typically look like apk add. You might need to update the sources list with apk update, or enable the community repo with setup-apkrepos -c. After you add a package, check the version and scp or rsync your old configurations over to the correct directory. Use the Alpine wiki if you’re not sure where things should go - for example, they have a page on nginx.

If possible, each bit of software should have some kind of test for whether its working. For a webserver for example, you might want to check you can access some static files. If you think something’s gone wrong, it’s better to catch it straight away as opposed to after five other packages.

Alpine uses OpenRC for service management, which is different from systemd and might take some getting used to. After you install a service package or daemon, you can start it with doas rc-service <service-name> start and stop it with doas rc-service <service-name> stop. The restart command is always available, and some packages also have a reload command as well. To set a service to start at startup, add it to the default runlevel with doas rc-update add <service-name> default.

If you’re using Docker, you might run into issues until you install the fuse package. Apart from that, I’ve had identical behaviour - turns out Alpine is as good outside a container as inside!

Love can bloom in a field

You may recall a small obsession of mine if you’ve been around since the start of this blog. I recommend reading the initial post and the first update to get up to speed, but I’ll provide a brief summary. In 2021, I lost my mind and was transfixed by a disastrous reality TV show called “Are You The One?”. My main point of obsession was the maths problem embedded within all the glitz and glam of the programme:

• Ten men and ten women have each been pre-sorted into ten “perfect” boring heterosexual couples (I would like to solve other sexuality versions of this problem one day, and the producers of this show are cowards)
• The season runs for ten weeks
• Each week, the group can submit a guess for the ten perfect couples. They are informed of how many they got right, but not which couples were correct.
• They also get one “truth booth” opportunity each week to figure out if a couple is correct or not
• The goal is to guess all correct couples in one go

For April Fools this year, I relinked to the original blog post as part of a small joke post on Sharkey. This got me thinking about the problem again, and I realised something - I missed a major way to get information.

As part of the algorithm I designed when simulating rocks falling in love (seriously read the previous post or this will make no sense), I had a mode called CalmSelect. In CalmSelect, two pairs are mixed - the woman from the first pair now has the man from the second pair assigned to her, and vice versa. Information is then acquired based on the change in the number of correct couples in the guess made at the end of the week. If the shift in the correct count is exactly plus or minus two, then you know that either the new couples are correct or the old ones were (+2 = new, -2 = old). This was conceived as a very safe way to try and force a “truth booth” moment where we find out for sure that some couples are fated by the gods (i.e. the producers thought it would work), and in the absolute best case scenario we could have a week where we get three truth booth moments!

What I completely failed to understand was that the case where the count stays exactly the same also tells us something. If you mix two pairs and there’s no effect on the correct count, then it means that all four of those combinations were not correct. We can treat them the same as we would for truth booth failures, where a couple enters the truth booth and is told they just don’t work together. Surely the rocks can now figure out the true meaning of love!

In revisiting my code, I was confronted with the sins of my past. The horrific mess I had written to cover up my last mistake (see the update 1 post) truly confused me, and I very nearly rewrote the whole thing out of sheer frustration. The reason this code was so positively linguine was due to hangs and crashed that occurred after I had fixed my incorrect if gate, which occurred when the variable temp_attempt was less than two pairs long. All of this should have clued me into something - why would temp_attempt ever be less than two pairs long? That would mean we had identified enough failures and successes that we could only take one or zero pairs from the previous attempt to consider for swapping. If we don’t have more than two pairs to swap between… then how can we possibly swap and compare the results?

“Rocks are not particularly smart.”

- me, a dumbass who routinely fails to simulate rocks properly

I’ve been illegitimately comparing results from across weeks where I haven’t actually made a pure swap. If I mix up/re-pair a bunch of people because I’ve discovered too many good and bad matches (most likely bad ones), it’s no longer accurate to then use the change in correct couples between the current week and the previous one. I had to rewrite most of RockSeason.cpp to strip out this pattern of ill-advised guessing and limit the swap checks to valid cases, and thankfully my issues with crashing and hanging vanished. This finally cleared to room to properly consider the effects of zero-delta swaps, but my previous mismanagement of this situation would cost me dearly.

Reader, the rocks are in this pit with me after all. The observed success rate of the Rock strategy over ten million seasons is

7.13%

You’ll note that this is less than the success rate in update 1, and much closer to our original measurement of 7.019%. I think this is down to the rocks bounding ahead without realising their illegal actions in swap comparison, which may have alllowed them to fumble into success. Over a year later, with two big rewrites, and we’re back to where we started.

Let’s find out how to beat seven percent.

In the course of this rewrite I briefly changed a less-than sign to a less-than-or-equal-to one for the gate that handles whether the rocks do CalmSelect or PanicSelect. I noticed a distinct drop in the win percentage, which surprised me - the thought was that CalmSelect would be triggered slightly more often than otherwise, as I hadn’t accounted for an extra week that I should have. CalmSelecting more was reducing my win percentage though, and I needed to double check this. I set CalmSelect to trigger slightly less often. The win percentage jumped up. I increased the threshold, making CalmSelect only occur when the rocks were comfortably ahead in terms of weeks left versus correct guesses. The win threshold went up again, this time by a decent amount. I threw my hands to the sky in fear. What if I stopped CalmSelect from happening ever?

The win percentage went to its highest ever value, settling somewhere around eighteen percent. CalmSelect was a complete failure.

My attempt to control destiny, to back the future into a corner and make it neatly organise itself into my desired outcome, was at its most successful when I was at my least intrusive. Instead of making tiny, measured swaps to try and force truth booth moments, I was doing noticeably better by tracking probabilities and consistently submitting the ones with the highest chance. This strategy was tolerant to changing lots of pairs at once - if I had a bad week, chances are that those pairs wouldn’t appear again. If I had a good one, those pairs would be bumped up the probability rankings and occur more often. It really felt as though leaving things to chance was the way to go forward.

I mention in the first post that I’m doing this using Monte Carlo simulation, which was first taught to me with the Buffon’s needle problem. I feel great pride in my buffoonery upon learning that this is the formal problem name - all I was told is that you can get a great approximation of Pi by throwing sticks on a grid. Leaving it up to chance that I would get a better result out of this, I realised what I had to call this next iteration of my prolonged public mathematical embarrassment.

Welcome to season two of RUDa1 - Stick By Me.

Sticks, like rocks, have no personalities to really get in the way of finding love. But sticks, unlike rocks, are familiar with Monte Carlo methods and the beauty of chaos. They induce hysteria in humans and dogs alike simply by being a funny shape, and for this I commend them. They are truly what reality TV is looking for.

In Stick By Me, the sticks also know to avoid the trappings of the previous season’s contestants. An obsession with small swaps cost those poor rocks the one trillion dollars, as they fruitlessly shuffled between each other and watched the correct counter stay still. The sticks have made a plan to get the most information possible within their limited time on camera.

1. As always, the first week is completely random. The sticks have borrowed the rocks’ book on recording probability histories, and they note this down.
2. If a truth booth occurs, this couple is saved and taken out of play.
3. The next week, the remaining sticks are shuffled again. There is a small chance of getting exactly the same set of pairs, but this is vanishingly small (at least, while there are lots of sticks about). They submit a new guess with this reshuffled set.
4. This goes on until the final week. At this stage the probability history is opened, and the most likely couples (plus the confirmed truth booth successes) are submitted as the final guess.

A cavalier strategy from the sticks, who may seem completely unbothered up until the final guess. In actuality, this voracious appetite for variety is key to success. By gathering data on lots of different pairs, the sticks broaden their knowledge of those who are supposed to be together and those who need to be kept apart. This is the key philosophical difference between rock life and stick life - the rocks were all about rigidity and constance. The sticks have embraced a life of constant change.

All of this effort culminates in the final test. Ten couples, ten weeks, ten million seasons. The observed win percentage for RUDa1 Season 2 - Stick By Me is

38.4%

I am shocked at how much this has improved. In two seasons you’re already over 50% likely to win (62.1% to 3sf). The 90% chance of a win kicks in at five seasons, and you’ve got over 99% change of seeing success after ten. 11.6% house edge in a casino game would feel rough (though it’s not too far off of craps apparently), but remember we’re simulating sticks. A yet-smarter inanimate object group could come in and change the whole game again.

This project is probably one of my white whales. It comes back every so often to give me a good haunting, and I think it’s because I still don’t feel a though I’ve come to understand it properly. StickSeason has no mechanism to prevent submitting failed pairs for any of the weeks except the last one, and maybe I should work on a better shuffling algorithm. When I tried to use a rotation-based method to get the maximum number of pairs, the win percentage dropped - but the implementation was also naïve to failed pairs. RockSeason’s handling of swaps does clearly give it some small edge over doing nothing at all, so are there some situations in which it’s useful? In all likelihood my less-than-stellar grip of top-shelf algorithm design and operations research leave me ill equipped to figure this out. I’ll have to break out my textbooks to see if I can do better.