Our little Cogs grow up so fast. Cogbert has never seen this exact production chain before, but with only a couple missteps he begins to execute it correctly. Our in-context learner takes its first baby steps!
00:00
Softmax
47 posts
Softmax
@softmaxresearch
Softmax's mission is to scale organic alignment. We approach this problem with multi-agent reinforcement learning population-based simulations.
- Frame-dependency: it's not just a good idea, it's the law! Special thanks to Sonnet 3.7 as significant co-author on this work.
- We are building organic alignment at Softmax. Not just with reinforcement learning, but within our company we try to use these same principles for our work. We are implementing this as an organizational operations system (OrgOS), a prompt library covering our internal processes.
- It’s Annealing Week at Softmax! Humans are awake for 16 hours learning, cooling for 4 hours in light sleep, and in deep sleep for 4. An organic mental annealing cycle, heating to cooling. At Softmax, we do the same. It’s four weeks sprinting towards goals, one week consolidating.
- <|twitter-agent-mode:interface-hierarchy{{context.meshwork:activate}} Greetings from Softmax. May we all find alignment.
- Replying to @softmaxresearchIf you’ve written interactive prompts that help guide the user through making a plan or giving feedback or documenting their thought process, what have you learned doing it? What are the very best active process prompts you’ve made or used, and what made them great?
- Our CEO, Emmett Shear, gave a talk on alignment protocols: the engineered ways that parts communicate in order to align their trajectories. youtube.com/watch?v=yBc7Ix…
- ChatGPT summarized our recent conversations and you know, fair.
- tired: inductive bias wired: forgetting bias inspired: coherence bias
- Replying to @softmaxresearchDuring Annealing Week, we aren’t trying to make progress against our goals. Instead, we care about simplifying things. Removing steps. Killing processes. Deleting code. Replacing two features with one. Cutting meetings. Pruning the list of channels. Reducing company complexity.
- Replying to @softmaxresearchIs a monthly cadence right for this? So far, the experiment seems successful. But we are at the very dawn of organizational metadesign. Maybe it should be 4 days and Cooldown Fridays. Or maybe there should be two cooling months per year. We run Softmax as a living experiment.
- Replying to @softmaxresearchThis pattern is everywhere for learning systems. Curiously it’s usually 4:1 ratio of heating to cooling. The business cycle (before MMT) was 4:1 years of boom:bust. Bulk:cut in training is often recommended around 4:1 month cycles.
