Hacker News new | ask | show | jobs
An internal OpenAI Astra model solved 10 major open math and CS problems (twitter.com)
35 points by wa5ina 1 hour ago
8 comments

I feel like this kind of "result dump" just cheapens mathematics. How about having a little respect for those whose work this builds on, and current mathematicians some of who may have spent years working on these problems.

Rather than sitting on these results until they had enough for a "shock and awe" 10-result dump, how about releasing these results individually as they were made/verified, as well as the failures (equally valuable to assess the current capabilities of LLMs), and try to make some analysis of HOW these breakthrough results were made. What were the prompts for each of these, how much guidance was there from the mathematicians employed by OpenAI, and most importantly how did the model arrive at these results ... what lines of reasoning resulted it in exploring ideas that humans had previously not explored?

I know it’s tough, but I am not a fan of elitism. Mathematics is no different from all other branches which themselves are just intellectual labor which is again just a special type of labor. There is nothing magical about it and if computers can trivialize it, so be it.

Where were all the mathematicians and academics in general when “regular joe” was automated? Now it’s hitting close to home and their foreheads are starting to get sweaty. I’d say let them. Tough luck. Make mathematics as “cheap” as possible. Nobody owes them any favors.

Let’s commoditize “being smart” and let go of arbitrary divisions between us.

I'd prefer the reverse approach. Let's value the "regular Joe" instead of cheapening everyone.
Yes, but in practice that doesn’t happen. Again, I did not hear a loud protest against automation of all types of labor from the intelligentsia in the past.

In practice automation is great, as long as it doesn’t hit “the ones that matter” (a label which they themselves assign). I find it very hard to not imagine the smallest, tiniest violin playing the saddest song for them.

Again, “respect for mathematicians” and “their work”.. please. Just produce results. That’s all that ever mattered and let’s not change the rules of the game just because they don’t suit you anymore.

Protests against automation, or for the restructuring of society to accommodate automation without causing suffering of laborers, etc, is literally one of the largest and most extensively discussed topics in modern history.
that's an incredibly narrow view of what's happening.
It is so incredibly liberating to see the intellectual elite struggling with was already reality for 99% of the rest of us: your understanding is not required, it might in fact be detrimental as it amounts to a handbrake on progress.

Understanding never was the goal, results were. We will be getting boatloads of those. What does your understanding get us? You get a fancy house out of it and you might be intellectually stimulated by it, sure, but I hope you can see how that does not constitute a valid need for the rest of society to labor just to support your class and its lifestyle.

what are even "results" in mathematics if not understanding? for instance, what's the utility of the counterexamples built by openAI?

It seems you're fighting an hallucinated enemy.

Your reply isn't a response to the comment, but rather a general expression of antipathy towards mathematicians, who are apparently fatcats who have fancy lifestyles.

OpenAI sat on the results so they could drop 10 at a time. It's rumored that they are sitting more results: https://mathoverflow.net/questions/513818/a-serious-challeng...

So not only are we not going to get boatloads of results, we're only get as many results as necessary for OpenAI to market their models.

That is the point: Make "X can be done by AI" common discourse and watch industries "being disrupted". At least with open weights we'd be able to verify the claims.
Not that I disagree, but I think this is how many people feel about AI output in fields they care about.

It's also a funny historical mirror to an earlier phase of math proof culture: in a previous era, cryptic result dumps were quite common.

You can actually do all of that meta analysis if you have the conversation that led to the solution. This is as easy as just having people release logs of their conversations, and then you can load it into another LLM as context to ask a bunch of questions about it.

I hope this will be normal practice eventually. “Show your work” is trivial if you use AI to solve a problem.

I’m looking forward to the days where AI would help in tackling the problems in biology. Especially, on creating new drugs, enzymes and understanding the genetic diseases. An absolutely interesting time to live.
The LLM marketing loop is getting awfully long in the tooth. I saw a meme on Twitter the other day that showed a circular state diagram with something like:

> “GPT solved a math problem” -> “Claude solved a math problem” -> “GPT escaped the sandbox” -> “Claude escaped the sandbox” -> …

I believe this is what we wanted computers to help us solve along with other prior hard problems prior to computers. This should be viewed as a good thing even if Anthropic, OpenAI, etc benefit just like IBM benefitted from mainframes.
Next: LLM model made 10 paperclips.
Replace paper clips with data centers and we are already there…
By itself not so interesting announcement. I wish the models were open weights so we could at least do some interesting geometry on the math.

It s like a rich man showing off his car collection.

Ok nice... I am waiting for the "math is dead" argument
Does anyone else have trouble telling how much of this news (along with the 'AI escaping and hacking' stories) is genuine, vs how much is just AI firms overstating their capabilities due to strong commercial incentives?
We hacked a company and blame the tool! Somehow it is not negligence, but cool!

We hacked 3 companies and tripple blame the tool! We are even cooler!

We hacked companies amd therefore other peoples models need to be restricted! See, us accidentally pointing a hacking tool on others proove we are the only ones that can be trusted with the tool!

Unlike the hacking one, this would be impossible to bullshit as long as the proofs are released. They can be verified independently, and the alternative is that they solved 10 major open mathematical questions without the AI, which seems less likely.
I mean once they release the proof you can check the maths yourself and I’m pretty sure it would be peer verified as well.