Documenting my journey in the world of AI and RL.
Two days on Opus 4.8, then 34 hours on Fable 5. Same research problem, same hardware, one model generation apart. I wanted to know what the bigger model actually fixed and whether the answer supports Anthropic’s recursive-self-improvement story. I ran these experiments in June, during Fable’s first week. I have not repeated them on the July release.
Summary: Today, llms are trained in a multi step process post SFT. RL -> Generate quality synthetic data → Self-Distillation on that → another round of RL (simplified). Fable-5 had a solution strategy constrained on how to compose code, and it struggled with a simpler solution until it exhausted all the greedy options. This is consistent with what a self-distillation recipe produces. Introduction…
Some problems don’t get solved in your head, they get solved in someone else’s while you watch or read their solution. You might feel it before you understand it. A small calibration. A quiet “Oh!”
Some elements here build upon the RL env post. Also, please read the footnotes, given a lot to cover, I moved many clarifications to the bottom.
Introduction I occasionally like to reverse engineer computer games by playing against inbuilt AI and trying different tactics to find loopholes. One such game is Catan Universe. Playable in a browser and built in Unity and WebGL. I wanted to see if an AI model could reverse engineer the game logic the way I do: by analyzing how it works under the hood and finding where the “fairness” breaks down.…
AI trends move fast. Writing about them makes sense if it’s your job. I’m picking this up because it’s escalated recently, and some discussion seems needed. This is me trying to make sense of an RL environment (and this understanding may evolve).
Richard Sutton wrote about the ‘Verification Principle’ in 2001:
someday i will make a more thorough and better reasoned post around it but at this point these are just notes somewhat structured
From Scott Alexander’s 2015 post
The Race for AI Progress In 2019, Richard Sutton, wrote his groundbreaking essay titled ‘The Bitter Lesson’. Simply put, the essay concludes that systems which get better with higher compute beat the systems that do not. Or specifically in AI: raw computing power consistently wins over intricate human-designed solutions. I used to believe that clever orchestrations and sophisticated rules were the…