Positive users are optimistic AI coders will gain taste via scaling, private codebases, and guidance from existing repos, while negative users worry high-level design judgment stays a persistent failure point far from AGI.
Based on 9 visible X reactions from 13 accounts; directional sample.
Ask a question below.
Published answers will appear here.
I think it’ll be amazing when that can build scalable systems that consist of multiple components. Like, ingest at scale. Or egress APIs that transfer terabytes efficiently. There’s still head room for the AI models to level up. It’ll also be interesting as they start to deliver cross functionally.
@shoyer It's a solved problem. They fall back on their bad architecture (no clear plan, rush to deadline) priors because that's the training baseline, but you can just tell it to do it right instead if you know what "right" looks like.
@shoyer Disagree on both accounts. On your 2nd point, general intelligence (along with creativity, taste, etc.) comes with scaling model size. And scaling shows no signs of hitting a limit
@shoyer high taste/judgement calls/divergent thinking still seems to be a failure point of models, even as scale continues, which is quite alarming
@shoyer Hard disagree. The more you use them, the better they are at inferring what your creative design philosophies are. Fable is the best example of this.
@MatthewParrott A model that needs us to tell it what "right" looks like is a long ways from AGI!
It's hard to imagine better coders than Fable and GPT-5.6, but astonishingly they still lack taste at high-level design. They eagerly build perfectly constructed piles of slop. I'm not sure more RL can fix this. The data is too sparse and the feedback cycles are too slow.
you can’t RL to taste https://twitter.com/shoyer/status/2075604592713429050
you can RL to taste https://twitter.com/shoyer/status/2075604592713429050
Positive users are optimistic AI coders will gain taste via scaling, private codebases, and guidance from existing repos, while negative users worry high-level design judgment stays a persistent failure point far from AGI.
Based on 9 visible X reactions from 13 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@MatthewParrott A model that needs us to tell it what "right" looks like is a long ways from AGI!