Remix.run Logo
▲ nsingh2 4 hours ago

> I feel bad for Buckmaster and Alpoge.

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.

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.

I also feel bad for the young people entering the field in such "interesting" times.

This is mostly OpenAI and Anthropic racing each other with a nearly unlimited token budget, hard to feel sympathy for that.

▲jrflo 3 hours ago | parent | next [-]

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.

▲nsingh2 3 hours ago | parent [-]

Good point. I would love to look at session transcripts and see what steering the RH/NS solutions looked like. Whether if it was getting the model unstuck or encouraging it to explore in directions they thought would be fruitful.

▲zero-sharp 3 hours ago | parent | prev | next [-]

>The model did the vast majority of the work.

Yea? What does that mean exactly?

▲doctorpangloss 3 hours ago | parent [-]

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.

▲c7b an hour ago | parent | prev | next [-]

> The model did the vast majority of the work

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.

▲keeda 22 minutes ago | parent | next [-]

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.

▲c7b 15 minutes ago | parent [-]

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?

▲baq an hour ago | parent | prev [-]

Being able to regurgitate basically all of number theory into a proof like this is new.

▲c7b 2 minutes ago | parent [-]

Navier-Stokes has very little to do with number theory, by my very limited knowledge. You might be confusing this with the Riemann Hypothesis proof that was also published today, by Anthropic. The days we live in...

▲charm137 2 hours ago | parent | prev | next [-]

>I kinda like the OpenAI approach of crediting the model

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.

But if there is no "innovation", then we are still in Ford's "faster horse" regime, not true intelligence.

▲serhei an hour ago | parent | prev | next [-]

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.

'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.

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".

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 :)"

▲asdfman123 an hour ago | parent | prev | next [-]

I don't think you should get credit for your comment.

Really, all the technologies that went into disseminating your comment did it: telecommunications, the PC revolution, the internet, etc. etc.

▲RGamma 2 hours ago | parent | prev [-]

Science during idiocracy: Confusing the shoulders of giants with a glorified chatbot.