Get live statistics and analysis of prinz's profile on X / Twitter

Just to recap: We found out today that an LLM that fits on a high-end consumer GPU, when trained on specific biological data, can discover a novel method to make cancer tumors more responsive to immunotherapy. Confirmed novel discovery (not present in existing literature). Experimentally validated in living cells. This is AI generating novel science. The moment has finally arrived.

Claude Sonnet 4.5 can autonomously create a chat app akin to Slack or Teams by producing 11,000 lines of code over 30 hours.

It is mind-blowing to me that only 7% of Plus users were using o3.

Hundreds of millions of people are about to jump from 4o straight to a reasoning model better than o3. This is a pretty big deal.

Another bad day for "AI can't do math". Math, Inc., a new company started by Christian Szegedy, has announced that its agent, Gauss, has autonomously completed Terry Tao and Alex Kontorovich's Strong Prime Number Theorem project in 3 weeks (human experts' progress stalled after 18 months). "This next year will birth agents that can collaborate and work continually over far longer time horizons. In 18 months we will see the first wave of breakthroughs in a rising sea of mathematics."

Anthropic needs to fire its bankers. As a reminder, xAI is raising at $75B.

OpenAI: - Zero reaction to nano-banana - Codex improvements every few days This is quite telling.

Zuck is throwing around piles of cash like this is his last chance to buy his way to the frontier before recursive self-improvement kicks in.

Q3 2025: - Jensen sells a gazillion GPUs to OpenAI - Sundar watches his AI model win gold medals at the ICPC and the IMO - Zuck spends billions on poaching AI researchers - Satya builds giga-data centers in Wisconsin - Tim Cook meets Pikachu

For those who still don't get it: - Researchers working at the frontier labs are genuinely convinced that transformative AI is coming. - They spend their time worrying about the social/political consequences of transformative AI, not the "AI bubble". - Recursive self-improvement is on the table, and is the subject of serious discussions.

Google and Yale scientists have trained an LLM that has generated a novel hypothesis about cancer cellular behavior. This prediction was confirmed multiple times in vitro. - "What made this prediction so exciting was that it was a novel idea. Although CK2 has been implicated in many cellular functions, including as a modulator of the immune system, inhibiting CK2 via silmitasertib has not been reported in the literature to explicitly enhance MHC-I expression or antigen presentation. This highlights that the model was generating a new, testable hypothesis, and not just repeating known facts." The model that generated this prediction is a 27B-parameter LLM based on the Google Gemma open source models, and trained on a corpus comprising >1B tokens of transcriptomic data, biological text, and metadata. Quite remarkable that a small (just 27B) LLM trained on specialized data is able to make novel scientific discoveries. "Teams at Yale are now exploring the mechanism uncovered here and testing additional AI-generated predictions in other immune contexts. With further preclinical and clinical validation, such hypotheses may be able to ultimately accelerate the path to new therapies."

The agent builder that OpenAI released today was built end-to-end in under 6 weeks, with Codex writing 80% of the PRs. If you feel the cadence of new releases accelerating, Codex is likely much of the reason why.

I've noticed that a lot of people tend to assume that capabilities of consumer-grade AI models (like GPT-5) are the same as the true frontier of AI capabilities. This is a mistake. For example, OpenAI's unreleased model that won gold on the IMO in July is clearly better than even the best publicly available model today (GPT-5 Pro), which itself is much better than GPT-5 Thinking. But please realize that even this insanely capable unreleased model was just a side quest by three OpenAI researchers who put it together in ~2 months. For a better read of the true capabilities of OpenAI's models in July (4 months ago, for those counting), consider the unreleased model that placed second in the Atcoder competition. OpenAI felt sufficiently confident to enter this model into the competition against the world's top coders, despite the fact that the competition ran uninterrupted for 10 hours. (This is where you check the METR chart.) And the model did run for 10 hours completely autonomously, without any human input, iterating better and better solutions to the Atcoder problem. It placed second in the competition; only one insanely cracked human beat it. The next leap in professional-grade AI models is models that are capable of making small novel scientific discoveries. I trust that we'll be hearing from OpenAI (and probably others) on that front fairly soon.

August 2025: Oxford and Cambridge mathematicians publish a paper entitled "No LLM Solved Yu Tsumura's 554th Problem". They gave this problem to o3 Pro, Gemini 2.5 Deep Think, Claude Opus 4 (Extended Thinking) and other models, with instructions to "not perform a web search to solve the problem". No LLM could solve it. The paper smugly claims: "We show, contrary to the optimism about LLM’s problem-solving abilities, fueled by the recent gold medals that were attained, that a problem exists—Yu Tsumura’s 554th problem—that a) is within the scope of an IMO problem in terms of proof sophistication, b) is not a combinatorics problem which has caused issues for LLMs, c) requires fewer proof techniques than typical hard IMO problems, d) has a publicly available solution (likely in the training data of LLMs), and e) that cannot be readily solved by any existing off-the-shelf LLM (commercial or open-source)." (Apparently, these mathematicians didn't get the memo that the unreleased OpenAI and Google models that won gold on the IMO are significantly more powerful than the publicly available models they tested. But no matter.) October 2025: GPT-5 Pro solves Yu Tsumura's 554th problem in 15 minutes. Lee Sedol moment is coming for many.

Two weeks ago: general reasoning LLMs from two different labs win gold on the IMO without tool use, reasoning on hard-to-verify tasks for ~90 minutes per problem. My timeline today: LLMs have hit a wall, scaling laws are over, G**y M****s was right.

Demis Hassabis: "In the next couple or years, I'd like to see [the time to discover new medicines] cut down in a matter of months instead of years. That's what I think is possible. Perhaps even faster." Google is also working on a "much more advanced" version of AlphaFold, "capable of understanding more than just protein interactions".

It is mind-blowing to me that only 7% of Plus users were using o3.

Claude Sonnet 4.5 can autonomously create a chat app akin to Slack or Teams by producing 11,000 lines of code over 30 hours.

For those who still don't get it: - Researchers working at the frontier labs are genuinely convinced that transformative AI is coming. - They spend their time worrying about the social/political consequences of transformative AI, not the "AI bubble". - Recursive self-improvement is on the table, and is the subject of serious discussions.

Just to recap: We found out today that an LLM that fits on a high-end consumer GPU, when trained on specific biological data, can discover a novel method to make cancer tumors more responsive to immunotherapy. Confirmed novel discovery (not present in existing literature). Experimentally validated in living cells. This is AI generating novel science. The moment has finally arrived.

Anthropic needs to fire its bankers. As a reminder, xAI is raising at $75B.

Q3 2025: - Jensen sells a gazillion GPUs to OpenAI - Sundar watches his AI model win gold medals at the ICPC and the IMO - Zuck spends billions on poaching AI researchers - Satya builds giga-data centers in Wisconsin - Tim Cook meets Pikachu

"LLMs are a dead end" is hilariously wrong. Even if progress were to stop today AND it turned out that LLMs are incapable of producing novel scientific discoveries, LLMs are ALREADY poised to radically transform the workplace and turbo-charge the economy. No new advances are needed for this. The technology we have today, enhanced with some harnesses, RL on specific tasks, and better user interfaces, is good enough to achieve this outcome. The most disappointing scenario, in other words, is one where LLMs merely turn out to be the most important technology since the proliferation of the internet. But I doubt the most disappointing scenario will come to pass. The scaling laws have held. There have been several examples of GPT-5 Pro having solved minor open math problems. The unreleased OpenAI and Google models can reason even longer. All these data centers are being built for a reason. A "dead end" indeed.

OpenAI: - Zero reaction to nano-banana - Codex improvements every few days This is quite telling.

Zuck is throwing around piles of cash like this is his last chance to buy his way to the frontier before recursive self-improvement kicks in.

Now that everyone on X is bearish on achieving AGI soon because current models lack continual learning, it's time to repost these: Dario Amodei (July 2025): "One thing we learned in AI is whenever it feels like there's some fundamental obstacle - like two years ago we thought there was this fundamental obstacle around reasoning, and it turned out just to just be RL... Without being too specific, I think, and we already have maybe some evidence to suggest that [continual learning] is another of those problems that is not as difficult as it seems, that will fall to scale plus a slightly different way of thinking about things." Jakub Pachocki (September 2025): "As we get to a level of near-mastery of high-school [math and coding] competitions, in the order of 1 to 5 hours of reasoning, we are focused on extending that horizon, both in terms of models' capability to plan over very long horizons and actually having ability to retain memory."

I've noticed that a lot of people tend to assume that capabilities of consumer-grade AI models (like GPT-5) are the same as the true frontier of AI capabilities. This is a mistake. For example, OpenAI's unreleased model that won gold on the IMO in July is clearly better than even the best publicly available model today (GPT-5 Pro), which itself is much better than GPT-5 Thinking. But please realize that even this insanely capable unreleased model was just a side quest by three OpenAI researchers who put it together in ~2 months. For a better read of the true capabilities of OpenAI's models in July (4 months ago, for those counting), consider the unreleased model that placed second in the Atcoder competition. OpenAI felt sufficiently confident to enter this model into the competition against the world's top coders, despite the fact that the competition ran uninterrupted for 10 hours. (This is where you check the METR chart.) And the model did run for 10 hours completely autonomously, without any human input, iterating better and better solutions to the Atcoder problem. It placed second in the competition; only one insanely cracked human beat it. The next leap in professional-grade AI models is models that are capable of making small novel scientific discoveries. I trust that we'll be hearing from OpenAI (and probably others) on that front fairly soon.

Another bad day for "AI can't do math". Math, Inc., a new company started by Christian Szegedy, has announced that its agent, Gauss, has autonomously completed Terry Tao and Alex Kontorovich's Strong Prime Number Theorem project in 3 weeks (human experts' progress stalled after 18 months). "This next year will birth agents that can collaborate and work continually over far longer time horizons. In 18 months we will see the first wave of breakthroughs in a rising sea of mathematics."

Julian Schrittwieser (Anthropic): - Discussion of AI bubble on X is "very divorced" from what is happening in the frontier labs. "In the frontier labs, we are not seeing any slowdown of progress." - AI will have a "massive economic impact". Revenue projections for OpenAI, Anthropic and Google are actually "fairly conservative". - Extrapolating from things like METR data, next year, the models will be able to work on their own on a whole range of tasks. Task length is important, because it unlocks the ability for a human to supervise a team of models, each of which works autonomously for hours at a time (vs. having to having to talk to an agent every 10 minutes to give it feedback). - "Extremely likely" that the current approach to training AI models (pre-training, RL) is going to produce a system that can perform at roughly human level in basically all tasks we care about productivity-wise. - On Move 37: "I think it's pretty clear that these models can do novel things." AlphaCode and AlphaTensor "proved that you can discover novel programs and algorithms". AI is "absolutely discovering novel things" already, and "we're just moving up the scale of how impressive, how interesting are the things it is able to discover on its own." - "Highly likely" that sometime next year we're going to have some discoveries that people unanimously agree are super-impressive. - AI will be able to on its own make a breakthrough that is worthy of a Nobel Prize in 2027 or 2028. - On AI's ability to accelerate development of AI: A very common issue in many scientific fields is that it becomes more and more difficult to make advances as the field progresses (i.e., 100 years ago, a single scientist could discover the first antibiotic by accident, whereas now it takes billions of dollars to discover a new drug). The same might happen with AI research - even though AI will make research of new AI more productive, there may not be an explosion due to new advances becoming more and more difficult to find.

Hundreds of millions of people are about to jump from 4o straight to a reasoning model better than o3. This is a pretty big deal.

OpenAI appears to have announced some kind of breakthrough internally yesterday, and the entire staff were given license to vaguepost. Let's hope it doesn't take too long for us to find out what it was.

Two weeks ago: general reasoning LLMs from two different labs win gold on the IMO without tool use, reasoning on hard-to-verify tasks for ~90 minutes per problem. My timeline today: LLMs have hit a wall, scaling laws are over, G**y M****s was right.