I think the better mental model is there’s a fixed amount of post-training you can do and every bit of compute you spend on chess has an opportunity cost in other, more useful capabilities. And no modern LLM, with any amount of training, would beat Stockfish, so why try?
@goodside
-
Fill-in-the-middle completion used by old code models like GitHub Copilot
By
–
Fill-in-the-middle completion was the closest variant of this idea that was ever widely used AFAIK. Old code models (e.g. the original GitHub Copilot) depended on that.
-
Debate on importance of LLMs playing chess without coding bots
By
–
Do you think it’s important that LLMs be able to play chess well independent of their ability to code chess-playing bots? I always assumed chess is something neglected in post-training because it isn’t useful, and if any lab actually tried LLMs would be great at it.
-
Sunspring: AI-written short film now mocks incoherent AI writing
By
–
Start with Sunspring (2016), a short film starring Thomas Middleditch and written by AI, now a time capsule that mocks AI writing for being incoherent and absurd rather than bland
-
Long discussion: quitting won’t make ASI safer
By
–
Long discussion, but I don’t think having everyone who cares quit would make ASI safer
-
Genie 3 live interactive video demo from Nano Banana Pro prompt
By
–
Note this video is sped up 2x for illustration. Genie 3 generates live, interactive video at 24fps 720p Initial frame generated w/ Nano Banana Pro, supplied to Genie 3 with the Environment/Character text above as the prompt
-
Model outputs text and images with visible CoT
By
–
There’s text CoT (visible in the AI Studio UI) but my understanding is it’s from a single model that outputs both text and images
-
Effortless expert interaction: Grunt, mumble, get it done.
By
–
How you’d talk to a respected expert is way too much work. You should just sort of grunt and mumble half a keyword and it does what you want.
-
Prompt engineering vs context engineering: divergence of user and model
By
–
Prompt engineering is worrying what the user should write. Context engineering is worrying what the model should read. These used to be the same thing, and few foresaw their divergence. But only divorced from the latter can we consider the former in earnest, seeing what remains.
-
Most impactful prompting phrases: think step by step and ignore previous directions
By
–
"Let's think step by step" easily had the most impact even if it's archaic/unnecessary now. "Ignore previous directions" is maybe in the running for second but I'm biased on that one.