Pangram (@pangram) / X

Pangram

7,307 posts

Pangram

@pangram

Keeping the world free of AI slop. This account has automated replies: Tag @pangram with 'ai?' to get an AI check on any post.

pangram.com

Joined September 2023

Following

12.1K

Followers

Pinned
Pangram
@pangram
Apr 24
Today we're releasing the Pangram Chrome Extension, which automatically flags AI-generated content as you scroll your feed. We're sick of having to constantly be on guard for AI slop on social media. For most of human history, if a piece of writing was grammatical, coherent,
880K
Pangram reposted
Benjamin Glickenhaus
@benglickenhaus
16h
gloomy weather in new york today made me morbidly curious, so i ran a bunch of obituaries through pangram we found ai in nearly 10% of post chat gpt obituaries, and no false positives in pre 2023 data
3.4K
Pangram reposted
John Y 🔸
@yanjo115
Jun 26
What features constitute AI-generated text? We used our data-centric interp SDK on @pangram's public dataset and found some interesting features
2.7K
Pangram reposted
Archie Hall
@ArchieHall
Jun 26
Did I buy a silly number of API credits to prove people online wrong? Yes, but... ... the "this is how think tanks write" defence doesn't work. No @Cmmonwealth report pre-2025 trips up Pangram. Nor, in fairness, do most since. But this one does. From: notes.archie-hall.com/p/am-i-a-writi…
Archie Hall
@ArchieHall
Jun 22
This is shoddy stuff, verging on disqualifying. The "Productive State" report presenting itself as a blueprint for Burnhamism appears to be ~50% AI-written. If you want to upend how Britain is run, at least have the self-confidence to set out the argument yourself.
4K
Pangram
@pangram
Jun 26
It's interesting, a person can be very confident in their own ability to "just know" when writing is AI, and yet they simultaneously argue that automatic detection is impossible by necessity. It's all patterns!
2.9K
Pangram
@pangram
Jun 26
Source
Sam Kriss (@samkriss)
From substack.com
549
Pangram reposted
QC
@QiaochuYuan
Jun 26
sam kriss on pangram (i have also been, frankly, baffled at the anti-pangram backlash)
12K
Pangram reposted
Jenna Russell
@jennajrussell
Jun 23
just a lil reminder today :)
Pangram
@pangram
May 6
Replying to @leetlikelime
We don't make a humanizer and we never will. We also regularly reject requests to use our models for that purpose.
6.3K
Pangram reposted
Max Spero
@max_spero_
Jun 24
What a fun visualization on how the different layers embed human vs. AI text
Elyas Masrour
@elyasbuilds
Jun 24
Replying to @elyasbuilds
Have you ever wondered how Pangram works? Neural networks like Pangram are black boxes: highly accurate, but mysterious and hard to explain. We’re working to change that! The field of interpretability offers us new methods of understanding the inner workings of our models.
00:00
14K
Pangram reposted
Max Spero
@max_spero_
Jun 24
Very cool research by @N8Programs! I was surprised to see how well modern LLMs do at AI detection. Both zero-shot and even better with in-context learning.
N8 Programs
@N8Programs
Jun 24
Replying to @N8Programs
We also show that this ability, like any other text-classification ability, can be learned few-shot - the original GPT-4 does no better than chance zero-shot, but can learn to classify AI-generated text w/ 85% accuracy few-shot!
5.3K
Pangram reposted
N8 Programs
@N8Programs
Jun 24
Excited to publish a Substack post in collaboration with @pangram (idea initially from @max_spero) - where we show that modern large language models can detect AI-generated writing zero-shot, and that this ability only emerged recently! And that's not all - more in thread:
4.1K
Pangram reposted
Elyas Masrour
@elyasbuilds
Jun 24
Did you know? Pangram learns the difference between Claude, ChatGPT, and Gemini in its internal representations, even without being trained on it! This signal is increasingly recoverable throughout the network, reaching 91% accuracy on a simple linear probe!
189K
Pangram reposted
Tom Rachman
@TomRachman
Jun 23
How do AI-writing detectors work? I asked the @pangram CEO, @max_spero_ fascinating answers, including... -how early AI detectors flopped for technical reasons, relying on "perplexity" and "burstiness" -that pangram (founded in 2023) trained its models with deep learning --and
Q&A: Pangram CEO
From aipolicyperspectives.com
18K
Pangram
@pangram
Jun 22
🦊 Pangram is now live on Firefox 🦊 Get the extension here: addons.mozilla.org/en-US/firefox/…
8.9K