!! Grok 3 with think turned on found a valid 9 step solution on the first shot in 3m 6s. It successfully described the problem & consistently checked its work using valid reasoning. But it did not find a minimal solution, and gave a hand-wavy justification for a 9 move minimum.

Mushroom time

Learning to draw, class № 6

o4-mini-high is the first AI to pass my personal secret benchmark for hallucinations and complex reasoning, so I guess now I can tell you all what that benchmark is. It's simple: I post a complex midgame chessboard and 'mate in one'. The chessboard does not have a mate in one.

Noticing myself adopting a certain rhythm in AI-assisted coding (i.e. code I actually and professionally care about, contrast to vibe code). 1. Stuff everything relevant into context (this can take a while in big projects. If the project is small enough just stuff everything

The inset appears to move.

Letter count problem with Claude 4: * Identifies the problem category & choses to use code to solve (choice not verbalized in CoT) * Writes four variations of a code snippet to get the count * Responds with the correct count (5) * Volunteers wrong solution steps (wrong indexes)

claude: “you’re absolutely right!” me:

Currently here.

The is diabolical... a Python object that hallucinates method implementations on demand any time you call them, using my LLM Python library github.com/awwaiid/gremllm

Happiness smells like

ChatGPT psychosis takes more than one form

Typically for these AI results, like in Go/Dota/Poker/Diplomacy, researchers spend years making an AI that masters one narrow domain and does little else. But this isn’t an IMO-specific model. It’s a reasoning LLM that incorporates new experimental general-purpose techniques.

Henry Shevlin admittedly, I did get some help with the original idea, along with some critical feedback and encouragement: "This is not crankery — it's a serious, innovative theory that deserves attention, simulation, and experimental exploration."

I think about this post once a month