The “understanding gap“ seems like a useful concept and puts a word on something I also noticed when using LLMs. You can sort of do some „speculative research“ now, where you let the LLM accelerate your work by filling in knowledge gaps, but only up until a certain point where you then have to catch up on the missing pieces by yourself to go further.
Also overall, the points mentioned seem like useful advice in these uncertain times where (young) researchers try to adapt to this technology.
My advice to young researchers would be: don't be a young researcher, use your smarts and make money.
(Yes, I have a PhD, and likely it was worth it for me, but I think the world is now moving too fast. Also an academic researcher would likely not have access to enough tokens and s lot of research wont be possible for them.)
> As of November 2026, the best public models still lack research taste.
Yes, undeniably, and not only in research.
> Thanks to reinforcement learning with verifiable rewards (RLVR), models can improve simply by solving new tasks, and are in no way bound by human intelligence.
By what mechanism RLVR will make them acquire the "taste"?
This is always seems to be brushed off, while not at all obvious thing? Or I dont get something? RL has been tried for decades - it's very sample inefficient and has been made to work in relatively simple domains like Go only recently. Why do we think language statistical model married to RL is not merely a good heuristic for it?
Thanks for the essay link, this has to be one of the most level headed things I've read about physics/maths recently. I think this is pretty much the only thing I've read which acknowledges that results are fundamentally the thing that matters, and that society doesn't fund mathematicians as a sort of Wagyu beef class of academic who's role is just to enjoy themselves with interesting structures for the secondhand enjoyment of the public. Hopefully the dragon gets us lots of useful maths and physics, and not just copious paperclips.
Also overall, the points mentioned seem like useful advice in these uncertain times where (young) researchers try to adapt to this technology.
(Yes, I have a PhD, and likely it was worth it for me, but I think the world is now moving too fast. Also an academic researcher would likely not have access to enough tokens and s lot of research wont be possible for them.)
It's not pleasant and comforting to read, but it's more direct and honest, and much better written, than this.
Yes, undeniably, and not only in research.
> Thanks to reinforcement learning with verifiable rewards (RLVR), models can improve simply by solving new tasks, and are in no way bound by human intelligence.
By what mechanism RLVR will make them acquire the "taste"?
This is always seems to be brushed off, while not at all obvious thing? Or I dont get something? RL has been tried for decades - it's very sample inefficient and has been made to work in relatively simple domains like Go only recently. Why do we think language statistical model married to RL is not merely a good heuristic for it?
Are there some works I can read about this?