CoreAutoAI co-founder Rohan Anil argues that AI milestones like o1 and Opus 4.5 stem from human ideas, not model self-generation

@_arohan_ "Now grinding is largely alleviated with models"

Too soon to call

@ahatamiz1 We just grind now on harder things like gpu kernels.

@_arohan_ Tell me in 6 months what you have achieved

@ahatamiz1 I doubt that.

@_arohan_ True if you are transformer-pilled

Higher compute access (for really checking the worth of idea, if you need to get to the best ones before others) with brute force LLM based idea exploration, might also land to new temp underclass. But till now in my experience, many of these auto research systems give up too fast, and do not get to really novel idea until it has some inspiration from human ideas or efforts.

@_arohan_ The permanent underclass thing was never about where models are today - it's about where they could be if you extrapolate the trend out.

@_arohan_ i think models are still bad at serendipity/creativity based tasks

@_arohan_ I mean we still don't have access to Fable and Sol, so it is partly right. It's going to be a month for fable ?

@_arohan_ I think the first two were well covered in the sense of “we know who did what”; what was the “innovative step for agentic training.

@_arohan_ Well said! Getting tired of this permanent underclass whining here.

@_arohan_ its a magical time ngl just able to instantly act on all those “what if” thoughts

@ahatamiz1 We will try, thanks for the motivation

@_arohan_ Did people at the labs really not think Agentic Training was a fruitful direction? I feel like Jan/Feb 2025 the *public* ecosystem woke up to “long horizon tool use” as being the future of model capabilities…

@_arohan_ millions of peaches

@_arohan_ did some fun exp on evaluating how good are agentic systems (here codex with 5.5 extra), for generating newer ideas by themselves vs how effective they are when they get to know what other competitng humans are doing: https://spthole.github.io/Sidhant_Thole/blog/research_taste_methodology.html