> I feel bad for Buckmaster and Alpoge.<p>I don't, at all. In fact, and this may be controversial, I kinda like the OpenAI approach of crediting the model, and not an author. The model did the vast majority of the work.<p>I feel bad for folks like Kevin Buzzard, doing the FLT Lean formalization by hand on a grant and then watching AI blow past it.<p>I also feel bad for the young people entering the field in such "interesting" times.<p>This is mostly OpenAI and Anthropic racing each other with a nearly unlimited token budget, hard to feel sympathy for that.
In the OpenAI case, it was 100% the model solving the problem. This and the RH solution were more collaborative efforts, so I'm fine crediting the humans partly there.
>The model did the vast majority of the work.<p>Yea? What does that mean exactly?
i don't think that people are wrong to use the words "brute force" to describe the models - swarms of computers - writing 10,000 proposals for Lean proofs, implementing them all, and seeing which compile and continuing until some advancement of the solution, a 1,000,000:1 ratio of work : moving towards the goal, as the superhuman step needed to finish these proofs. Most forces in our world are superhuman.
I'm not too concerned with whether the agents acted in a brute force manner. The controversy of OpenAI's Navier Stokes solution is that it used a particular strategy only known to a handful of academics. Here's the quote from Tristan again:<p>The route to the Clay problem through a smooth force, options c and d in Fefferman’s statement of the problem, is the route Luis and Diego opened and the one Levent and I had quietly chosen to attack. Almost nobody else I know of was working on it. It is not the direction one arrives at in a few days by giving a model the problem statement. When I heard “forced,” it was a bright red flag.<p>I'm not sure how anyone walks away from this story and remembers it as "The model did the vast majority of the work."
It will be interesting to see if that was the approach used here. 'Brute force' is an OK description but it is important to be careful of what is being brute forced: this is still far away from an infinite monkeys solution (which would obviously be completely intractable) even if it's a relatively inefficient way to get to the frontier of what's possible.
It's spaghetti at the wall. Lean is crucial to clear away the failed brute force exploration. Just like in chess long ago. .brute forcing lines.. you needed an evaluation. Maybe an unpopular sentiment but took how many millions on tokens? So a lot of it is just generator to try different ideas in this complicated forest and the lean compilability as some kind of compass
If you actually watch AI writing Lean it is very very far from brute force. Recent models are <i>extremely</i> good at Lean. When I was trying to get them to teach me it, I would paste my failed tactic into a non-agent Opus 5.5 and just say "why doesn't this work" and it could easily tell me the exact error message I was seeing without even running a Lean compiler. Obviously this is on very simple code but still, hardly brute force.<p>I'm working on a quite expansive project and most of its time actually seems to be spent avoiding accidental complexity blowup in the tactics it writes, rather than any difficulty writing the actual proofs. Obviously it's not Navier-Stokes, but it's still super complex stuff. Easily PhD level and way beyond my skill.
>I kinda like the OpenAI approach of crediting the model<p>And who does the model assign credit to? It'd be interesting to see when all these proofs are cleaned up and understood by humans, how much actual innovation in information terms was added by the model and how much of it was its sheer bandwidth to correlate wildly disparate fields of mathematics to come up with proofs of conjectures (i.e., working within an entirely closed system). Don't get me wrong - forming those connections across those fields is impressive and useful in its own right.<p>But if there is no "innovation", then we are still in Ford's "faster horse" regime, not true intelligence.
> The model did the vast majority of the work<p>Unless it just regurgitated the ideas that were fed into the training data by the human. We might be at or near the point where the models won't depend on that anymore, but in this particular case I think we shouldn't discount this possibility. Even OpenAI hasn't ruled it out afaik.
OpenAI did categorically rule it out some days (weeks?) later, but the optics damage was done. I tend to believe them purely because they’ve released 100s of other proofs since.
In the announcements I've seen, OpenAI categorically ruled out giving the model access to user chat sessions. They explicitly did not address the possibility of chats ending up in training sets, and did not answer an explicit question by Tristan Buckmaster. Why would you think those other proofs wouldn't be inspired by the ideas that thousands of mathematicians worldwide had discussed in their chats?
Being able to regurgitate basically all of number theory into a proof like this is new.
Science during idiocracy: Confusing the shoulders of giants with a glorified chatbot.
It is controversial largely because it falls into the prior context of OpenAI's marketing strategy of being rude to humans. It's hard to tell from outside if this is oblivious or someone's genius strategy to drum up a sense of despair/inevitability for the machine ascendancy.<p>'Ghiblify' was the epitome, where a marketing demo was deliberately framed to be rude to talented human artists, for no business-sensible reason. The style-transfer was played around with for a few months and then died out as a generally cringe thing to do. Of course, OpenAI's future isn't affected by a bunch of people with pencils in Japan who are liked for practicing a deliberately low-tech craft. On the other hand, OpenAI has not yet made an animated movie.<p>Now, applying 'ghiblify' dynamics to interactions with scientists, which might be kind of important... the rumour is that OpenAI people talk to scientists, hear something they don't like, and start saying "we should just ask ChatGPT".<p>The natural counterpoint will be scientists getting contacted by OpenAI for help with something and replying (in more polite terms) "why are you bothering to talk to us if you think we're obsolete? Just ask your AI :)"
I don't think you should get credit for your comment.<p>Really, all the technologies that went into disseminating your comment did it: telecommunications, the PC revolution, the internet, etc. etc.