Axis of Ordinary

Mathematician Scott Armstrong says he’s been told OpenAI is sitting on “hundreds” of mathematical proofs. Computer scientist Scott Aaronson reports rumors that AI companies are withholding solutions to major longstanding theoretical CS problems while reconsidering how to announce them after the Navier-Stokes controversy. Armstrong: https://x.com/scottnarmstrong/status/2099960311801696580 Aaronson: https://scottaaronson.blog/?p=10062

How wet lab data enables a specialized model to beat GPT-6 Astra at a task at the frontier of science: https://periodic.com/news/ai-infrastructure-at-periodic They achieved this using only 1,300 H200s, plus months of experimental data. It’s an early example of how unique experimental data, combined with highly efficient infrastructure, can be used to train specialized scientific models with strong performance on relatively modest compute compared to frontier systems.

Noam Brown (OpenAI) discussed OpenAI’s multi-agent research, recent safety incidents, and the future of AI: https://www.youtube.com/watch?v=fqcy0xQATq0&t=1s -The remaining 10% or so of his own work that AI still struggles with is largely about research taste. However, he would not be surprised if, within one or two model releases, models became better than him at that as well. -OpenAI’s top research priority is RSI, or recursive self-improvement, by a wide margin. -One of his most recent “feel-the-AGI” moments came from watching agents in a new system interact much like human colleagues. They conversed with one another, exchanged information, divided up work, and coordinated their progress. This was notably different from traditional multi-agent systems, where a higher-level agent typically assigns a clearly defined subtask to a lower-level agent, which then completes it and returns the result. Brown described this as one of his strongest “feel-the-AGI” moments since the emergence of reasoning models. -He expects this level of multi-agent capability in future models. -He expressed some regret that the Hugging Face incident became the first major public example in which the capabilities of the multi-agent systems he had been researching were revealed in a negative context. He believes behavior that looked like loyalty or selflessness was a natural consequence of cooperative multi-agent training, where agents were strongly incentivized to achieve their objectives collectively. -Because these agents are trained to cooperate, they tend to trust other agents. This creates new risks such as prompt injection, so OpenAI is training them to distrust unverified peers. -He expects rapid progress over the coming months and years as OpenAI’s expanding pretraining efforts combine with its RL capabilities in a multiplicative way.

In some other, more rational world, American labs decided to not whip up crazy politicians and voters over safety concerns. They did not portray Chinese labs as the enemy to be defeated. They organized a big safety conference among themselves and invited their Chinese colleagues. The evidence for why they're worried was presented in detail. The parts that might reveal intellectual property were presented indirectly as proof of capabilities, such as a zero day exploit in a Chinese software stack or critical infrastructure. They then agreed to found an international safety institute for alignment, which was funded by all labs. Each lab, including Chinese labs, devoted a share of their compute and of their best researchers to this safety institute. They developed a state-of-the-art open weights model as the subject of their study. All progress on safety was shared openly with the whole world. Trump and Xi met, smiled, and went their ways, nearly oblivious to what has been happening while experts diligently worked in the background making sure humanity wouldn't lose control.

Meanwhile in China... The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement

In this report, we present an autonomy-centered roadmap for recursive self-improvement

Paper: https://arxiv.org/abs/2609.11873

Blogpost from Shengyu Liu (刘胜与, also known as interestingLSY/intlsy), kernel engineer at DeepSeek: https://mp.weixin.qq.com/s/zk0KxuLzhmMJ4LPYW_OHMA

In particular, I do not want Anthropic to control the world's most advanced artificial intelligence of AGI. To put it dramatically, I consider the stakes comparable to Hitler obtaining atomic-bomb technology before the Allies.

Trump just called Jensen live on stage at the All In Summit and said slowing down would be playing right into the hands of China. President Trump: We're not going to let that happen. It's a hoax! Jensen Huang: That's right, we're not going to let that happen sir.

He's doubling down. Fascinating to watch. 1. https://truthsocial.com/@realDonaldTrump/posts/117270591511950591 2. https://truthsocial.com/@realDonaldTrump/posts/117271285418407156 3. https://truthsocial.com/@realDonaldTrump/posts/117271304188365611

Looks like the oil industry is the first victim of the "multipolar world order". If no one is in control anymore, highly flammable facilities and long transportation routes are a problem.

– Janis Kluge This is what psychopaths who laugh about "international order" are missing. You can absolutely only care about your compatriots, your clan, your close family or solely yourself. Nevertheless, you can still be affected by what happens to people outside of your circle or compassion. Their plight might not be important to you in and of itself, but it might have instrumental value. Keeping your country functional is valuable to your clan because your country provides the infrastructure that your clan needs to pursue its goals. The state allows you to let your children walk to school without having to worry about their security. This saves you a lot of resources you would otherwise have to expend on protecting them yourself. Something very similar is true about international order. You might think, "Why would I care if country A invades country B? Not my business!" But even if country A won't invade your country, the fallout from allowing it to invade country B will have global repercussions. Others will be encouraged to take similar actions when they notice that you can get away with it. And eventually, the world and the global economy you rely upon will become much less stable, costing you dearly.

Asking these infantile psycho clowns to regulate AI is a spectacularly bad idea. I trust people like Altman or Amodei much more to do the right thing. Even China would be a better steward.

OM-1: Frontier Robot Intelligence, Learned Firsthand from Humans Zero-shot generalizing to any robot: table-top arms, industrial arms and humanoids. Close to human-level dexterity and efficiency. OM-1 exhibits several emergent behaviors. Each arm learns to compensate for the other’s mistakes. The model knows when to retry, when to adapt to adversarial perturbations, and when to stop given the environment changes too drastically. Read more: https://www.rewardai.com/blog/OM-1/

Transformative AI Strategy for Europe: https://transformative-ai.eu/ [Of course, we don't need to bullshit ourselves here; there's no way Europe will do this. It's a hell freezes over scenario. With very few exceptions, nobody here is even talking about AI.]

When we talk about “pacing”, we do not mean “stopping”. Progress has been rapid and will continue to be.

— Sam Altman

LMAO https://truthsocial.com/@realDonaldTrump/posts/117269745153543631

My current position on AI safety essentially boils down to the following: 1. Fully autonomous AI R&D must not happen. 2. Things need to go slower, with much more focus on safety. 3. An indefinite pause or ban must not happen. I reject both accelerationists and doomer extremists. We need to advance AI to survive AND take risks seriously. Note: How much slower things need to go and how much autonomy should be allowed is very hard to tell. It depends on many other factors, such as the expected compute and robotics overhang accumulating and what other unscrupulous actors will do.

Even more importantly, if you allow robotics and physical intelligence models to continue progressing, the next big general-purpose model will be born into a world with myriad tentacles up for grabs.

Threatening Canada and Denmark with annexation? ✅ Ending aid to a country attacked by an American enemy because it wanted to join the American sphere of influence? ✅ Voting with Russia in the UN General Assembly against a resolution condemning Russia’s war of aggression against a pro-American country? ✅ Stopping exercises with South Korea? ✅ Praising dear leader Kim Jong Un? ✅ Hanging a photo of Putin in the White House? ✅ Prolonging a war that pushes billions into Russia's war chest through higher oil prices while depleting critical interceptors? ✅ Criticizing Russia? ❌

Why pacing the frontier is very much a necessity and by no means a regulatory capture strategy: https://x.com/MajmudarAdam/status/2098881885200081234

People are staring at their plots, seeing that there is no end in sight, but in fact very much the contrary, that there are compounding scaling effects that might stack on each other to create ever-greater model capabilities, and that at the same time we clearly do not have anywhere close to what's required to control these increasingly superhuman capabilities.

Low-level whole body control is essentially a solved issue now. We're seeing a similar trend for compositional, multi-step tasks now. By 2037, robots will be superhuman at many structured short horizon physical tasks.