I would give young researchers this to read instead; physicist Antoine Tilloy’s <i>The Short Age of Dragon Riders</i> that just came out a few days ago: <a href="https://atilloy.com/2026/11/16/the-short-age-of-dragon-riders/" rel="nofollow">https://atilloy.com/2026/11/16/the-short-age-of-dragon-rider...</a><p>It's not pleasant and comforting to read, but it's more direct and honest, and much better written, than this.
Thanks for the link, but I think the author is falling into the exact trap he is calling out - guiding his decision making by ego rather than what's best for mankind.<p>> I have always been a techno-optimist absolutist. It would be a betrayal to flip now that the wave of progress touches my craft.<p>AI touches every craft, not only yours, and therefore the grief you feel is more or less universal. To disregard it because it contradicts some sense you have of your own identity is dishonest.<p>> As a result, he concludes, his main value for society lies in showing how far the human mind can be pushed with sustained effort and dedication ... Machines may progress to the point where this view (which I find self-serving) is the only reasonable one.<p>Humankind ought to be self-serving! There's no point in making ourselves suffer for the benefit of machines.
> As of November 2026, the best public models still lack research taste.<p>Yes, undeniably, and not only in research.<p>> 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.<p>By what mechanism RLVR will make them acquire the "taste"?<p>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?<p>Are there some works I can read about this?
Resource constraints, in my opinion. If it the AI is rewarded for shorter, neater solutions that generalize well to similar problems, then it will converge on a good sense of taste.
> By what mechanism RLVR will make them acquire the "taste"?<p>Maybe additional reward metrics? E.g. did a change make followup-tasks (performed by a separate instance) easier or more difficult, how many tokens does a judge-model need to understand and verify the produced artifact, etc.
That's a really nice piece of work, and not just because it mirrors my opinion 100%. :-P Thanks for the link.
Excellent link, and a realist's view of what is actually happening in the field compared to the blather posted by OP. Sorry to that prof, but you're going to do your students a disservice if you serve them weak platitudes. Pangloss resurrected!
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.