infrecursion

@AlexGodofsky More hilarious antics from the horse carriage industry lmao.

This is not workable. It is simply not possible on any level to detect students who are using LLMs for "conceptualizing" assignments, especially if they do so with any skill. The further restriction on uploading course materials to an LLM is silly.

@Jesse_Livermore Listen man, hate to break it to you but GPT-5.5 high is smarter than your entire bloodline has ever been or will be - at least for math and science.

I'm almost at the point where I'm going to stop asking AI questions and instead ask it to find sources that might answer them. Because I know what it's doing: pushing the linguistic buttons in my brain that lead me to believe it and stick with it, even as it gives me garbage.

@burny_tech Absolutely retarded cope and coming from people who are completely incapable of producing either kind of proofs. And no, there isn't a single example of a proof or new concept that was plucked out of thin air.

"Every one of those ingredients was already in the literature. The Golod-Shafarevich towers, the cutting mechanism of Hajir-Maire-Ramakrishna, the pigeonhole, the Minkowski packing bound: all of it sitting there, in papers from the last sixty years. The construction recombines them, ingeniously, into something nobody had assembled before. What it does not do is reach for a concept that was not already available to reach for. The distinction I care about is not whether a new named object appears, which would be a silly test that most great mathematics would fail, but whether the shortest explanation requires extending the working vocabulary or only requires a new path through the vocabulary already present. This proof is the second kind. The hypothesis space was not grown. It was traversed, much further than any human had the stamina to traverse it. Sawin and Shankar both call the proof genuinely ingenious and very human in character, and Shankar says plainly that the model went beyond being a helper: it had an original idea and carried it through. He’s right, and I’d rather grant that fully than explain it away. Recombining known structures in a way nobody had thought to try is real creativity. It’s most of what mathematics actually is. Fermat’s Last Theorem, Szemerédi, much of additive combinatorics: landmark work that introduced no new primitive object and was no less creative for it. So the claim can’t be that recombination is the lesser thing. The claim is narrower. There’s a kind of advance that recombination, however deep, just doesn’t reach: the one where the existing vocabulary can’t express the answer, and has to be enlarged before you can even state it. The metric tensor wasn’t a clever arrangement of Newtonian quantities. It was a new thing the old language had no word for. That’s the move I mean, at whatever scale it happens, and the unit distance proof, ingenious as it is, isn’t an instance of it. It’s the top of the other ladder. And the people who verified it describe it in terms that fit that reading, even though confirming my framework was the last thing on their minds."

@AnthropicAI Shut up man and release opus 5. No one gives a shit about this marketing/pr bs stunt.

Patching these vulnerabilities will make us safer. But the software industry will need to adapt to the volume of vulnerabilities that models like Claude Mythos Preview will be able to find. We discuss this in our initial update on Project Glasswing: anthropic.com/research/glass…

Last month we launched Project Glasswing, our collaborative AI cybersecurity initiative. Since then, we and our partners have found more than ten thousand high- or critical-severity vulnerabilities in essential software.

@ptrschmdtnlsn I love that the original statement by Noah that the guy dunked on to which Scott replied to has about the same authenticity and rigor as this long ass post, but far less pretentious.

There are several results that morally point in the direction of "you should expect endless rich beauty", but my favorite is this: You might naively worry that at some point we prove everything that's easy to describe, and we have to make up extremely complicated questions. And at that point, is it really that interesting to say that some 50 page contrived problem required some complicated theory to prove? However, we know this won't happen! If you take all questions of n characters, take the shortest proof of each, then look at the growth rate of the length of the longest such shortest proof, you find that the length must grow uncomputably quickly in n. Otherwise theorem proving would be computable by proof search, which would let you decide provability, thus a fortiori decide the halting problem. So, we must have a vast supply of very simple questions whose answers are spectacularly complicated. In practice, this is what we see, and the complexity seems to correspond to real richness! For example, as near as we know, the answer to "when does x^n+y^n=z^n have solutions over Z?" is just massively incompressibly deep, requiring the development of extremely sophisticated tools. Likewise, "what can be said of a group that's finite and simple?" seems to just be a massively deep question. That simple questions can require thousands of pages of deep theory should be unsurprising in light of this "proof lengths must grow uncomputably quickly" result! And "grows uncomputably quickly" is an absolutely staggering growth rate. There is likely some short couple-paragraph question where resolving it would require you to develop one million pages of rich theory, beyond the intellect of any human.

@Lux_Stella_ This is complete bullshit. Can you give one example of those supposed magical proofs that were generated ever in human history, which came out of thin air, purely through the genius of the particular individual? These people are clueless. All proofs are like this "specific kind".

there's some really interesting commentary from one of the mathematicians in the companion paper not to pour cold water on the results necessarily - but it suggests this is a specific "kind" of proof these models might be particularly good at

@OfficialLoganK What is the price on API? Same as 3 flash?

Try it in the Gemini API, Google AI Studio, Antigravity, AI Mode, Gemini App, and wherever else you use Gemini!

Welcome to Gemini 3.5 Flash, our most powerful model to date. It pushes the frontier of intelligence, speed, and cost putting 3.5 Flash in a class of its own. We spent the last 6 months making sure Flash is great for real world use cases. It's available everywhere now!

@GoogleDeepMind @Google Where is the link? Also no benchmarks or pricing info, doesn't look very promising lol

Introducing Gemini 3.5: our newest family of models combining frontier intelligence with real-world action. The first release is 3.5 Flash, our strongest model yet for agents and coding 🧵

@RolandDunbrack It's true but that has been so limited and has required so much compute to even predict something useful for real proteins for mere nanoseconds that compared to AI techniques now they may not even have existed. When AI works here it's just so much superior.

MD simulations of proteins have been around for almost 50 years (1977). What are you talking about?

@Burkino_1 @lthlnkso It's not telling anything other than you're an ignorant fuck. Any of these AIs can mop the floor with your ass on any math or coding tasks. They are good at somethings, not good at others (unlike you with absolutely no talent at anything whatsoever).

@lthlnkso I think it's pretty telling of AI that you can't just ask it "provide a deterministic answer to this question" but instead have to go "create a tool that has a deterministic answer"

1 prompt to Claude Code by the way.

@_sholtodouglas Dude, does this post mean no new model this week?

When do you reach for other models instead of Claude? What can we do better? Hit me with all of your frustrations. dms open. If you can give me detail (e.g. specifics/transcipts) - it'll help a lot in finding out exactly what we need to do to improve the next model

@zx_sci_john Cute you think GPT-5.5 is only used for editing mistake. It can wipe the floor with your ass in any math exam, go fuck yourself.

Scientific fraud isn't an editing mistake

@RuxandraTeslo @RichardHanania This assumes that somehow AI will always be inferior to humans. Typical arrogance. Like any of these people have any chance of holding a candle to a super-intelligent AI. Even now the frontier models are like better than 99% of the mass.

Cool post from @RichardHanania where he shows the results of a "Turing test" on his writing. I personally found it very easy to identify what was AI written, but around half of respondents did not. This has really interesting implications regarding how public opinion will be shaped in the AI era, and who our "storytellers" will be, as Richard himself notes. It seems that such results point towards a trend predicted by @danwilliamsphil and others, where AI will lead to clustering around "expert consensus" and a de-democratization of opinions -- the opposite of what we saw with the internet. The interesting thing is that AI is arriving on a vacuum of intellectual authority, as trust in academia has been eroded. Obviously, academic publications still form a large part of its training corpus, but I think academic opinion will be less relevant relative to its volume at least. Especially in non technical fields. So who will be the shapers of public opinion in the AI era, then? We already saw that those who have written a lot on the internet, for example rationalists and effective altruists, seem to have their opinions overrepresented. This might not just about volume, but also the style of their argumentation (very indexed on logical coherence), which might be favored by LLMs. But how will this look like in the future? Important question imo. richardhanania.com/p/can-ai-repla…

@TinaGurnaney You're high. It's the exact opposite. Maybe you're using free GPT and not GPT-5.5 high.

I can't explain this but Claude gives STEM graduate and ChatGPT gives degree in Arts & Philosophy

@scottnarmstrong Amen, apply that to every subject and do it post hoc as well. Either of those models can find both human and AI slop. Ban all (it will probably make arxiv a read-only repo for some time, but worth it).

A modest proposal: authors who submit math papers to arxiv containing obvious human-written errors that Opus 4.6 or GPT 5.5 would flag should receive a 1-year arxiv ban. Why should I read your paper if you haven't even taken basic steps to ensure its accuracy?

@Willrandship @max_paperclips Lmao you're just pulling numbers out of your ass.

@max_paperclips if it requires you to learn screen/tmux syntax it's inaccessible to 95% of developers and 99.9% of the population. what you really need is a vnc session

oh wow, and then you could have something we're calling a "secure" "socket" "shell" into a "virtual private server" Could you all stop reinventing stuff that's existed for decades from first principles and pretending you're innovating

@lefthanddraft Yeah they shipped this model pre-nerfed. But occasionally when they decide to turn the thinking on, it's glorious.

Incredible alpha in distrusting what Opus 4.7 says.

@0xglitchbyte The only people I see complaining are those who haven't shipped anything worth anything in their life and never will.

When are we going to stop pretending oneshotting bun’s rewrite was a good idea?

@LimYenKheng Yes you're supposed to do it but not for a fucking preprint you moron.

what the hell are you talking about. YES YOU'RE SUPPOSED TO DO THIS

@Kaju_Nut Why do you people always have the most short-sighted takes? Do you believe AI progress will just stop for some reason, and we won't have AI that soon understands science better than 99.9% of all humans (including you) can ever hope to? What would be the point of your postdoc?

Today I was discussing with a colleague about the challenge of postdoc selection in the coming years, now that some PhD students (and maybe even PIs) are publishing heavily ai-assisted papers whose content they don't fully understand. Looks like interviews are the only way, and they might have to get quite technical.