letmevoteplease

"I'd take the author not publishing those as trying to hide something about how the agent actually approached this."

That is an irrationally cynical take. Everything described in the post is clearly within the abilities of a frontier LLM. The previous discovery is obscure enough that the Bluesky author is not citing any actual publication but instead encouraging people to reach out to a scientist and receiving images through private communication. The reasonable interpretation is that the model stumbled upon a previous discovery in the manner described, and the author/LLM failed to find the previous discovery because it is not any published database. The author is not running a scam for some unclear benefit that would be immediately revealed when he published it.

It is not plausible that a company would raise prices as some kind of revenge or play to overturn a regulation. That is not how prices are set in a free market because any company that did that would immediately lose to their competition. Unless we have a cartel of television manufacturers, which would be news to me.

"Altman seems to be scolding these organizations for not taking the situation seriously enough or considering that models might evolve and improve, but seems to have less interest in reflecting on whether or not he might have undersold the issue while pushing forward"

Meanwhile, look back at the old Mythos and Glasswing threads and all of the AI haters on Hacker News were lambasting Anthropic for hyping up the next token generators that could never find a bug not in their training data...

Wiz: "These vulnerabilities include memory disclosure and unauthenticated remote code execution in PHP itself, independent of the application running on top of it"[1]

Calif: "the available exploitation paths depend on the PHP version and SAPI but are always exploitable under default configurations" ... "This results in RCE against arbitrary PHP scripts on Apache mod_php and PHP >= 5.4.0."[2]

PHP Foundation: "full remote code execution was demonstrated with default settings."[3]

Unless all of these organizations are lying, the exploit can work against a default PHP installation running a Hello World script.

[1] https://www.wiz.io/blog/php2shell-php-rce

[2] https://calif.io/php2shell/

[3] https://thephp.foundation/blog/2026/10/21/todays-php-securit...

There are multiple write-ups on how the vulnerability works. The Calif one is the most detailed. Also, no one used Claude. There's a whole world of information out there when you read the original sources instead of social media distortions.

https://calif.io/php2shell/

https://www.wiz.io/blog/php2shell-php-rce

If you provide forewarning, system administrators around the world in various time zones can get ready to update at the same hour the fix is released, which is very important. Being hours late is not acceptable for a vulnerability like this.

There is no contradiction here. It was founded precisely because LiveJournal started censoring sexually explicit fanfiction involving taboo subjects (particularly incest/rape/minors).

China is building good AI by distilling American models. I think that's a great thing, but Chinese progress is currently dependent on American progress, because China lacks the compute to train the largest models (as stated by DeepSeek's founder[1])

[1] https://sumsec.me/2026/liang-wenfeng-investor-meeting-transc...

I read "The level of social opprobrium directed at LLM use and at people attempting to pass off generated slop as work is immense" and immediately checked for the author's Bluesky link. This person is living in a bubble. Most people in tech are now using LLMs at work and there is no social opprobrium discouraging them.

The rest of the analysis seems to take it for granted that we have hit a wall and LLMs will improve no further. OK. We'll see how that prediction plays out.

27 unarmed black men were killed by police in 2019 according to the Mapping Police Violence database. (Despite this, 53.5% of people identifying as "very liberal" estimated that 1,000 or more unarmed black men were killed. Only 15.71% correctly estimated "about 10."[1]) There are 800,000 police officers in the United States, so I am not sure you can say they generally enjoy killing unarmed black men based on a few dozen deaths a year.

[1]https://research.skeptic.com/content/files/2025/02/Research-...

Most of them are using AI for all their schoolwork while performatively hating it because "it wastes water" or some other half-understood meme.

archive.is has a snapshot from 18:47, showing the link was indeed there a few hours ago. https://archive.is/XM0Nw

This was done by an OpenAI subscriber, not an employee, so Astra would not have had access to OpenAI's massive compute for brute forcing. The scripts it wrote presumably ran on the computer of the customer. (ChatGPT can run scripts on OpenAI's servers, but it has a 45 second execution limit.)

Cool but please rewrite the text. Ctrl + F "land" gives 29 matches.

This study appears to be evidence against that: the model failed the benchmark but arrived at the correct answer.

> Altman was caught in previous attempts trying to game benchmarks

Sounds like something you just made up, or maybe you read it on some other Reddit/HN post and started repeating it because it aligned with your biases.

> does anyone believe that he's found his moral compass and decided to stop exploiting as much as he can get away with?

I don't think "OpenAI" is equivalent to "Sam Altman." I think if OpenAI was intentionally "benchmaxxing" purely for marketing purposes that information would leak, because OpenAI is full of good-faith researchers (although it can be difficult to avoid overfitting even if you're actually trying to improve the model's general abilities)

And lastly I think anyone can actually try Sol themselves and see that's it a good model, or if that's too subjective, it is clearly better than the previous version. The benchmarks are reflecting actual progress and anyone can verify this themselves.

I also like DeepSeek, but I'll note their stated goal is to develop AGI, and the founder (already China’s fifth-richest person) has stated, "I believe the business opportunities here are large enough-if the AI era will produce many trillion-dollar companies, I think we will be one of them."[1] These are not humble ambitions.

[1] https://liangwenfeng.art/ch11.en

You quoted the OP saying "in this case too they didn't actually have the solution" and responded with the totally unrelated, "Given the size and recall of the biggest models, it's not unreasonable to assume that a single pertinent conversation would make it into the training data."

Neither of the researchers insinuating that their ideas were trained on had the actual solutions. This means the model could not have "stolen" the final solution from their data. At most, it could have built upon their work in the same it builds upon any other training data, though that is also questionable speculation.

>They could 100% definitely say no, if they know they did not train on user data.

No one anywhere has claimed that "OpenAI does not train on user data." OpenAI has always said that it trains on user data.

>They immediately started racing to a solution after one researcher enquired about whether they are training on their conversations.

They started racing towards a solution after they heard (incorrectly) that Anthropic had a solution; I agree this is poor sport but the "after one researcher enquired about whether they are training on their conversations" claim is false. The enquiry happened after OpenAI had obtained the solution.

You are confusing ideas here. No one except OpenAI had a solution to Navier–Stokes. Buckmaster and Alpöge had a solution for the forced Euler problem, which they arrived at largely using LLMs (Claude and Codex). Buckmaster implies (but does not explicitly accuse, since he has no evidence) that training on his prompts had some influence on OpenAI's result. This seems unlikely to me but is not impossible. However, in either case, the solution was found due to an LLM. Of course the LLM built on past human work, but "plagiarism" is not sufficient to account for the distance between the papers of Martínez-Zoroa, or the prompts of Buckmaster, and the final resolution.

This is how every conspiracy theorist thinks: my enemy is Bad, and if they did a Bad thing, it would be Good for them, therefore they obviously did it. No evidence needed other than "motive" + my enemy is evil. But even if your enemy is evil, in this case, they would be fools to take the legal risk of violating their contract for the minimal upside of a tiny bit more training data (and fools to assume this would not be exposed in a large organization). So you need to assume your enemy is both evil and remarkably stupid.

It reads like it was generated with AI and then edited to have poorer grammar.

"A launch is a first impression and first impressions are marketing, that’s why you’re always bound to be shocked - the shock was scheduled."

"And if a model scores well but keeps failing at your work, that is actually a gap that deserves investigation."

That's Claude speaking.

If they stated the storage limits explicitly, people would just go right up to the edge. Better to leave it undefined and assess compliance afterward.

I think most of the people who act like this are secretly using ChatGPT like a Republican on Grindr.

“Read, read, read. Read everything -- trash, classics, good and bad, and see how they do it. Just like a carpenter who works as an apprentice and studies the master. Read! You'll absorb it. Then write. If it's good, you'll find out. If it's not, throw it out of the window.” - Faulkner

Even if 99% of generations are trash, why does that matter? Just throw out the failures and keep the good ones, like any writer. Here is the text of the winning story: https://granta.com/the-serpent-in-the-grove/ If you've ever used ChatGPT to generate stories, you will find the style unmistakable. It's cloying to me, but it did manage to fool some stuffy literary judges.

It is of course a stupid regulation, but the upside is that it will probably accelerate growth in usage of open models that are not adversarial towards the user.

Look at this statement from the article:

>These racks are needed to keep up with the ballooning requirements of each new "frontier" (a buzzword label with no actual inclination of performance of improvement) model developed, which everyone bought into the AI hype will immediately hop to because it's all novelty over utility.

This is a mind-blowingly clueless assertion. This is not a person capable of rationally assessing facts. At best this kind of writing is interesting as an artifict of human delusion and cognitive bias.

Yup. OpenAI: "the name of the fictional target for the CTF challenge unintentionally coincided with a real domain." Anthropic: "the fictional target company chosen by our evaluation partner shared a name with an active website domain name." So it sounds like the same website got hacked by both GPT and Claude.

This does not make sense. Did you read the article? They were not trying to "catch" it accessing the internet. It did not escape. A partner accidentally left the connection to the internet open.

Have you considered that there are reasons to do things beyond financial incentives? This incident is obviously very interesting, particular to the type of hacker employed by Hugging Face.