Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

More questions about whether researchers can trust OpenAI with unpublished math

mathstodon.xyz

Post reply on HN
Print view View original

301–310 of 844 posts

Re: More questions about whether researchers can trust OpenAI with unpublished math

#301

Post by wrsh07 »

bitexploder wrote: post #267
aaronharnly wrote: post #249

Has anyone run a test of including some shibboleth or canary phrase or assertion in a chat, enabled for training, and seeing if it turns up later as something a model "knows"? I'd be curious to understand how that works even in a toy-level model, and if there is anyone consciously testing that process with the frontier lab offerings. My naive instincts would be that it seems unlikely that a single chat transcript wou…

Problem is how do you convince the model and training profess it matters. A one off canary is very unlikely to survive in the final model state.

Right, imagine if instead they had coined new terminology that was not obvious and it re coined that - this would be close to a smoking gun

Afaict that didn't happen so there's just lots of speculation

Re: More questions about whether researchers can trust OpenAI with unpublished math

#302

Post by HDThoreaun »

perrygeo wrote: post #151
glimshe wrote: post #107

Why are people here jumping so quickly to conclusions? I have no doubt OpenAI is capable of doing this, but right now there's no credible evidence, only claims. This kind of "they stole from me through AI training!" accusation will soon start being used against other AI users, not necessarily the providers. All it will take is a mastodon post. And shortly after, we will also see the next iteration of copyright legal…

The stolen data claim isn't the smoking gun. We can already assume the frontier labs are accessing our data, as they have repeated done. Not news. The big claim is that OpenAI sniped the research. Not a model, a human did so. Intentionally. They took someone else's idea and claimed it as their own. This is good old fashioned academic fraud, but with millions in compute resources and corporate incentives thrown at the…

Where did they claim it as their own? Doesn’t the release cite buckmaster and claim their work is a continuation of what he and levent were working on?

Re: More questions about whether researchers can trust OpenAI with unpublished math

#303

Post by HarHarVeryFunny »

sashank_1509 wrote: post #266

Both things can be true: 1. OpenAI when using your chats in pretraining is improving its model’s intuition. The model parameter size is massive, and while the data is OOM larger it is plausible that model remembers stuff about chats that improves its latent representation. 2. During RL on verifiable math and massive compute, the model discovers techniques and connections to solve math problems that are superhuman and…

OpenAI said they sicced this agent army on Navier-Stokes on Sept 1st, while only a couple of days earlier OpenAI's Noam Brown happened to reply to a tweet saying that they had already tried to solve all the Millennium Prize problems and failed... So, it seems either the previous attempt didn't have the training to succeed, or was just not given the compute to do so.

Once OpenAI heard that Navier-Stokes was solved, this caused them to immediately revisit the problem and throw a ton of compute at it, apparently using a more (very) recent model than what they had tried before. What we don't know is just how recent this model was, and therefore what it may have been trained on. Buckmaster/Levant had apparently been working towards this for at least a year, and made their "forced" blow-up breakthrough on August 15th.

Presumably any anonymized prompts that are being trained on are part of pre-training, so older, but once OpenAI had heard that Navier-Stokes had been solved and wanted to revisit it, it seems possible they may have done a few weeks of incremental RL training on anything Navier-Stokes adjacent they could come up with, in addition to then throwing unlimited compute at it, now confident that there was something to find.

Re: More questions about whether researchers can trust OpenAI with unpublished math

#304

Post by SwellJoe »

It's been said before, and it remains a concern, that if AI reaches a point where it can do/build/launch anything without a huge amount of human labor, the AI companies have no reason to let you or I extract that value.

And, if they're able to snoop on and learn from your human process that gets from initial prompt to functioning product/proof/whatever their labor to produce that thing is even lower. With their much larger budget than most folks and even companies have, they can pick and choose the most valuable things to pursue.

That's not to say I think that OpenAI is going to steal that roguelite strategy game you're working on, but the companies that own the machines that turn electricity into software (and soon, electricity into hardware designs) have an advantage in any field where they're useful. They get earlier access to newer/better models, they have larger token budgets, they don't have the guardrails you and I run up against.

Employers fantasize about replacing all workers with AI without thinking through that if AI can replace all workers, then AI companies can replace all businesses.

Re: More questions about whether researchers can trust OpenAI with unpublished math

#305

Post by yellow_lead »

sashank_1509 wrote: post #266

Both things can be true: 1. OpenAI when using your chats in pretraining is improving its model’s intuition. The model parameter size is massive, and while the data is OOM larger it is plausible that model remembers stuff about chats that improves its latent representation. 2. During RL on verifiable math and massive compute, the model discovers techniques and connections to solve math problems that are superhuman and…

Both can be true:

1. OpenAI couldn't have solved the problem without the researchers' private data for training.

2. OpenAI models can solve math problems

Re: More questions about whether researchers can trust OpenAI with unpublished math

#306

Post by gwerbin »

iamgopal wrote: post #292
dgellow wrote: post #286

Earlier quoted context omitted.

I feel that we don’t praise Lean enough. AFAIU it’s what enables LLMs to brute force those problems

True, but could humans cross pollinating lean x prolog x A* ( or any search algorithm) could have solved such math problems with super computer ?

I don't think so. People have been trying things like this with evolutionary algorithms for a very long time already. LLMs can interleave symbolic manipulation with empirical experiments and simulations and charts and thinking/reasoning text, and an LLM will much more efficiently search the space of candidate ideas than any handcrafted mutation algorithm. Any task with a cheaply verifiable goal that requires fanning out across a massive search space is ideal for contemporary LLM technology to make progress with.

Re: More questions about whether researchers can trust OpenAI with unpublished math

#307

Post by encyclopediai »

aaronharnly wrote: post #249

Has anyone run a test of including some shibboleth or canary phrase or assertion in a chat, enabled for training, and seeing if it turns up later as something a model "knows"? I'd be curious to understand how that works even in a toy-level model, and if there is anyone consciously testing that process with the frontier lab offerings. My naive instincts would be that it seems unlikely that a single chat transcript wou…

I run such tests since a long time at chorasimilarity open notebook.

I always used guest non login accounts.

As a mathematician I was able to check two plagiates (by humans) with even such primitive means.

But I have to mention that some things irk me in this conversation about math or science and AI.

First, I see lots of attribution and other related problems, with certain impact for the researcher proffesion.

But I don't see the most natural question: wouldn't you like to know the answer to _open-problem_ ?

I mean, is research now only about publishing and solving famous problems?

From this point of view I think the links from this recent post are depressing

https://terrytao.wordpress.com/2026/09/10/crowdsourcing-a-li...

Second, I think very relevant that the original meaning of "encyclopedia" is "recurrent education".

So I arrived to think that the present and future forms of AI in mathematics and sciences should be seen as modern day encyclopedic efforts.

Once we pass over the flurry of solving famous open problems (and wouldn't you like to know?) the next natural step is an audit of the ehole corpus of mathematics and sciences accumulated until now.

And then pass further on a saner basis and damn about problem solvers and unhappy publishers and management.

Re: More questions about whether researchers can trust OpenAI with unpublished math

#309

Post by btilly »

aaronharnly wrote: post #249

Has anyone run a test of including some shibboleth or canary phrase or assertion in a chat, enabled for training, and seeing if it turns up later as something a model "knows"? I'd be curious to understand how that works even in a toy-level model, and if there is anyone consciously testing that process with the frontier lab offerings. My naive instincts would be that it seems unlikely that a single chat transcript wou…

Yes. See https://www.anthropic.com/research/small-samples-poison?from....

250 documents ingested from somewhere is enough to become part of the knowledge of a model of arbitrarily large size.

I would expect that a good idea that fits in a framework that is already being ingested would be more easily taken up than some random thing unassociated with anything else. Could that go down to a single transcript? If the model is consciously focusing on everything X related, quite possibly.

Re: More questions about whether researchers can trust OpenAI with unpublished math

#310

Post by derangedHorse »

robotpepi wrote: post #296
square_usual wrote: post #184

Earlier quoted context omitted.

> and learned of the general approach they were taking. Only after learning the secret to cracking the problem did they send the first prompt. Do you have any evidence of this? They don't dispute the timeline, but they never said they knew what Levant/Buckmaster were doing.

It's in OpenAI's first announcement that they had solved the problem.

> Only after learning the secret to cracking the problem did they send the first prompt.

Which quote in the announcement post provides evidence for the above quote?

Post reply on HN