I'm not sure any of those have money at that level. Aaronson in the past when he was doing quantum computing work refused to take payment or stock from quantum computing companies, so I'd be shocked if he has any money. As for monetary concerns, Hinton explicitly quit his well-paying job at Google so he would be free to talk about AI risk. So the narrative here is the exact opposite. And when the companies talked about deliberately slowing AI growth, stock in AI related companies went down, which undermines the idea that this is anything about their investments. But there's also a serious problem here if someone like Coxon can be dismissed as not having enough knowledge or experience, but people who have been involved in the industry for a long time can be dismissed because they may have financial ties. So can absolutely anyone be dismissed as irrelevant, either lacking in expertise or being too tied in to the industry?

I'm not sure what an hour memory here means, since context windows are measured in tokens. Even six months ago, LLMs could think for well over an hour before producing an output.

This seems odd at multiple levels. First, the people in the foundation labs if anything have more of an idea what is happening than the general population. They are seeing how models are made, what they contribute, and how they go run off the rails. I'm also puzzled by the claim that we're not at the exponential. Two years ago, AI systems could barely do AMC and AIMIE problems, basic high school math contests. Then the AI got to be good enough to do well on Putnam results. Then, early this year, AI got to be good enough to solve multiple open unsolved math problems, including Erdos 1196 https://www.erdosproblems.com/1196 , the Unit Distance Conjecture https://arxiv.org/abs/2605.20695 and many others, and the trends there are just continuing. That's one field, but it is one where the standards are most objective about what is happening. That certainly looks like exponential growth. I'm also puzzled by seeing the labs cannot get the AI to improve themselves is a good standard. From the perspective of a lot of people who are concerned, once that's happening, it is likely too late.

What you mean here is you agree with Yann. But the point is that lots of people who are as qualified as Yann, disagree. So if, as in your first comment, your primary problem with Coxon is lack of experinece, you should find this situation alarming. Worse, your complaint about Coxon was lack of lab experience, but in your new comment you object to people being too deep in the labs to be objective. This leads to the weird situation where no one is in a position to be relevant to listen to.

It is possible that LLM tech has its limits. But people were saying it was hitting its limit 4 years ago, and 3 years ago, and 2 years ago, and a year ago. 4 years ago, people who were saying LLMs were limited would not have predicted they'd be solving unsolved math problems. All those predictions turned out to be wrong. Why should saying it now be more likely? And what if you are wrong here?

So if the concern is these need to come from someone from a company not run by Altman, then we already have that. The OP is about that. And we've had containment breaches reported by Anthropic also. We also had a smaller scale but similar problem reported by an independent security group running an open-weight Kimi. You don't need to trust Sam Altman or anyone associated with him here.

A flyover isn't the problem here. The statement is.

If you want to dismiss Coxon feel free, but he isn't the only one in the article. And outside the article, we have a lot of people who are concerned who are hard to argue aren't experts. Geoffrey Hinton and Yoshua Bengio both got Turing Awards for their work in Deep Learning. (The third who got it is Yann LeCun who isn't worried at all, but Hinton and Bengio is 2/3.) Similarly, Scott Aaronson, who wasn't even that worried about this a few years ago is now highly concerned. Aaronson is one of the world's leading experts in computational complexity, who discovered the algebraization barrier among other things, and whose work lead to the AI watermarking systems. Aaronson just wrote this essay https://scottaaronson.blog/?p=10062. So sure, go dismiss Coxon, but it doesn't impact remotely all the others who are clear experts and are sounding an alarm here.

What needs to happen to decide these aren't PR campaigns and are actual warnings? What would need to occur? Knocking out a power grid? Hacking the Pentagon? Wrecking with air traffic control and making planes crash? I ask in part because when these incidents started happening two months ago, people said that they were just PR and that if they were real we'd start seeing more of them and that seems to be exactly where we are, and yet these are still getting labeled as PR campaigns.

The original quote they are riffing on is "The beatings will continue until morale improves." Essentially this is a variant of the "make a threat, but in a joking fashion, but really functionally a threat" sort of thing that dangerous and abusive people will do. Trump uses versions of this all the time, and it isn't surprising that Hegseth and his people have picked up on this. I'd call this another warning sign about fascism but given everything else that's going on, including a grossly illegal war in Iran, Trump functionally threatening to undermine the upcoming election, and Trump having removed generals and other higher-ups in the military that he sees as insufficiently loyal (read, more likely to actually follow the law and morality when given an illegal order), and this incident becomes almost part of the rounding error.

Getting scooped by someone else is a thing. It happened to me twice when I was in grad school. Going out of one's way to to work on a problem one wasn't concentrating on specifically to beat someone else to getting a solution published is not the same thing.

In the case of Erdos 1196, https://www.erdosproblems.com/1196, the AI solved it by constructing a Markov chain on the integers weighted via the von Mangoldt function. Not an idea in the literature. If a human grad student had done that they would be immediately considered an up-and-coming bright star. For the Unit Distance Conjecture, https://arxiv.org/abs/2605.20695, the AI constructed a clever tower of number fields. Tim Gowers, who is a Fields Medalist and has worked on related problems, said this construction if done by a human, would have been a shoe-in for the Annals, which is widely considered the most prestigious journal in math. If you are claiming that all this sort of thing is is putting puzzle pieces together, then you are using that phrase to be so incredibly broad as to be absolutely meaningless.

In a normal academic context, none and moreover this sort of deliberate scooping would be considered incredibly bad behavior. When you have two giant corporations fighting to be able to have results they can showcase to the public and their investors? Then it makes a lot more sense, even as it is incredibly damaging to the academic math culture. This is really toxic behavior and deserves pushback. But it is a mistake to go from there to think that therefore any copying or plagiarism happened.

We've discussed this already on Slashdot here https://science.slashdot.org/story/26/09/08/2228220/openai-says-it-has-cracked-one-of-maths-millennium-problems. Bubeck, the research at OpenAI has a very different take on this entire situation here https://www.linkedin.com/feed/update/urn:li:activity:7503148080888893440/. I will quote his reponse:

As I said in that thread, I don't know whose version of events is accurate here, but some of the details due support Bubeck's version. I suspect that a breakdown of communication occurred where both then misinterpreted what the other saying in a more hostile way than it was intended until the conversation then became genuinely hostile. In terms of the math, although I'm a mathematician, differential equations is pretty far from my expertise, so I cannot deeply evaluate how close the methods were. However, it is the case that the both were building on the methods of Cordoba-Martinez-Zoroa. In this context, my being far from this sort of work is relevant, because this was a well known enough approach that even though I'm in a pretty different subfield, I had heard of CMZ's work as an approach and that this was considered a promising approach to Navier-Stokes. Given that, my inclination is that the case that any theft occurred here is very weak.

And since that thread, I've become more convinced, as experts reading both papers point to substantial differences in the details of their approaches. One of the ironies here is that a lot of the anti-OpenAI views are coming very loudly in part due to a general anti-AI attitude. But Bumaster and Alpoge, the people who are claiming to have had work stolen (well, primarily Bumaster, Alpoge is mostly staying out of the fray) were heavily using both Claude and Codex in their work.

So, people are able to recognize this, and yet in the other threads here about Anthropic calling for a slow down of AI research, everyone seems convinced that this isn't about the risks really at all.

You can if you want, claim that OpenAI has had serious issues with honesty, especially with Sam Altman, their CEO. But Anthropic's record is a lot better. More to the point, the Theranos comparison doesn't work. The fundamental object Theranos was selling just didn't exist. AI systems do, and they are getting millions of people to use them. The situations are wildly different.

If they were genuinely concerned, this is exactly what they'd do. So what makes you think they are doing this out of something for investors and not actually telling the truth about their motivations?

Did you see in the comment where you are replying to where I said I'm a mathematician? My own primary areas of research are number theory and graph theory. It is pretty easy to find who I actually am and verify that yes, I am a mathematician and am in a pure field. So yes, maybe I do know something about the discipline.