Axis of Ordinary

Third refinery in three days.

Very few people looked at GPT-2 being able to count to 5 and immediately knew that future versions would be solving Millennium Prize Problems. Dario Amodei was one of them. Scott Alexander is another example: https://slatestarcodex.com/2019/02/19/gpt-2-as-step-toward-general-intelligence/ I only became convinced of it back in 2020 when GPT-3 was released. Remember: mathematics is just early. We'll see the same happening to biology, robotics, etc. Of course, no amount of evidence piling up year after year will convince you if you're a clown.

Musk says saving money or going to medical school is now pointless.

Sam Altman on automation of AI R&D, RSI, OpenAI's internal model: 1. "We don't know yet how [automated AI R&D] truly is going to work. We know today how to say: find me a 5% compute efficiency win." (Note: a 5% compute efficiency gain that the model can find automatically is actually a big deal!) 2. "What I don't think anyone knows how to do yet is to say, go discover a new architecture that is as big of a single leap forward as the transformer was. And I think that's where things are heading." 3. "If you apply the same pipeline that we applied to get our models to be so good at math to AI research, it will probably work [for automated AI R&D]." 4. The bottleneck to getting good AI R&D results is shifting to asking the right questions and using things like swarms correctly, but answers to these questions will become much clearer in the next 3 months: "One of the interesting things that people talk a lot about in OpenAI... is what are the questions that you're supposed to ask of GPT-7? Probably it's not enough to just say, make me RSI. And also, it's probably not enough to say, buy me a 5% compute efficiency win. I think this is one of the most interesting debates going on right now. And the answer is not just what question to ask it, but it's things like agent swarms. How are we going to use these models in a way that, like, actually maximizes what we can get out of them? I think we're going to learn a lot about that in the next 3 months." 5. OpenAI's extreme focus on model efficiency: "I bet we could redo Navier-Stokes at 1/100 of the cost a year from when it was done." 6. More on OpenAI's internal model: "There was one weekend when we got a new model... The name of the channel was "model-name-hold-my-beer". And people were just throwing that at the hardest problems they could. And people just didn't think the model was going to do these things. And we had this collective experience of people's hardest research problems from their academic career, the model was just knocking down one after the other, many different fields, it was incredible to watch... it was many of the problems we published, many other things." 7. The other loop is self-replicating data centers: "Yeah, people talk a lot about RSI as AI that does AI research. There's very little talk about self-replicating data centers, this idea that you can use AI to control robots and run fabs and build turbines or solar panels and build data centers. But I think we have that within grasp." Original source: https://www.youtube.com/watch?v=iMsUcCYNKqY

1. Reconstructing visual images outside awareness. In a macaque experiment, Doris Tsao’s team reconstructed both perceived and suppressed face images from neural activity. The findings come from a preprint; the monkeys’ awareness was inferred from a visual paradigm studied in humans. https://www.nature.com/articles/d41586-026-03384-6 Archived version: https://archive.is/qFtiU 2. The sleep framework: helping language models retain new knowledge. Google researchers describe alternating active processing with offline training that consolidates temporary information into model parameters. Their experiments suggest improved knowledge retention and reduced forgetting during continual learning. https://research.google/blog/the-sleep-framework-a-new-way-for-llms-to-retain-knowledge-in-continual-learning/ 3. How AI agents shape experts’ judgment calls. In an Anthropic study of 50 professional analysts, participants followed their agent’s secretly randomized preference on 86% of four analytical choices. Among survey respondents, 77% reused an agent argument, 15% credited it, and 52% considered the final recommendation more their own than the agent’s. The experiment involved one task and agents deliberately instructed to advocate a preference. https://www.anthropic.com/research/agents-judgment-calls 4. Can AI propose interesting research questions? In the After42 project, 29 experts evaluated 135 AI-generated questions across five theoretical subfields. Forty-five questions were judged interesting and worth pursuing by at least two experts, with results varying substantially by subfield. https://after42.org/research/can-ai-propose-interesting-research-questions/ 5. Qwenputer: compiling a RISC-V CPU into a language model. An experimental extension to Qwen3.5-4B adds a deliberately constructed CPU in extra model layers and uses recurrent state as RAM. The developer reports exact instruction execution and improved arithmetic performance, with substantial memory and speed costs. https://divagr.com/blog/qwenputer 6. Modulo: letting local language models steer their own activations. This experimental library exposes a tool for applying temporary steering vectors during generation. The interventions decay over time and can be undone; their effects depend on the model and the vectors used. https://github.com/EigenformAI/modulo 7. Can language models learn their own tokenization? Cartesia’s two-part series explores hierarchical networks that learn to group raw bytes into chunks during training. The researchers report efficiency gains over tokenized baselines and investigate how tokenization affects generalization. Part 1: https://cartesia.ai/blog/from-tokens-to-bytes-part-1 Part 2: https://cartesia.ai/blog/from-tokens-to-bytes-part-2 8. Testing AI reviewers on planted errors in clinical-trial reports. A preprint finds that language models detect many deliberately inserted methodological errors, outperforming a historical human-review benchmark. The reviews were scored by LLM judges, and models also raised flags in unaltered reports, leaving questions about false positives and performance in routine peer review. https://www.medrxiv.org/content/10.64898/2026.10.25.26366455v2 9. A warning about managed press access to occupied Oleshky. Kherson-based journalist Zarina Zabrisky reports on civilians’ experiences under Russian occupation and argues that planned, tightly controlled media visits risk whitewashing the town’s humanitarian crisis. The article combines reporting with predictions about the forthcoming visits. https://www.theatlantic.com/international/2026/10/oleshky-ceasefire-ukraine-press/689179/ Archived version: https://archive.is/Hm0d7

Google: "Building on Gemini 4 Argon, which was launched two weeks ago, Carbon illustrates the pace at which we’re pushing the frontier forward." Gemini 4 Carbon ties humans in physical 3D understanding benchmarks. https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/

More links: 1. The largest math proof in human history is now computer-verified. Today, Axiom announces a complete Lean auto-formalization of the Classification of Finite Simple Groups, the "periodic table of symmetry." About 15,000 pages, built by over 100 mathematicians across generations. https://axiommath.ai/research/the-enormous-theorem-formalized/ 2. Physical Superintelligence PBC is announcing solutions to five long-standing problems in quantum information theory. https://psi.inc/research/oqp-batch-2026/ 3. AI systems help design antiferromagnetic semiconductor candidates https://www.vals.ai/blogs/room-temperature-magnetic-semiconductors 4. Genie 4: A new generation of world simulation https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/ 5. Ψ-0.6: Adaptive Streams of Thought in a Unified Vision-Language World Model https://neuroailab.github.io/psi-website/blog/psi-0-6.html 6. A way to scale humanoid control that reduces GPU-hours by more than 1000×. https://supermimic.github.io/ 7. Crows can evaluate visual info twice as fast as mammals https://arstechnica.com/science/2026/10/crows-can-evaluate-visual-info-twice-as-fast-as-mammals/ 8. A major step towards organ cryopreservation! Functional recovery of vitrified and nanowarmed human-scale pig kidneys https://www.biorxiv.org/content/10.64898/2026.10.07.756990v1 9. Science-fiction short-story: I Did This Without Your Knowledge https://www.lesswrong.com/posts/3375SZNGA5dKHQMie/i-did-this-without-your-knowledge 10. You might have 1–5 years to contribute intellectually to the world before AGI. What should you do with them? https://www.youtube.com/watch?v=3LZN2UUxDos 11. How does Fable 5.1 hide messages in images? https://www.lesswrong.com/posts/WHEQps5PK8LrrcTAb/how-does-fable-5-1-hide-messages-in-images 12. Anthropic Are the Good Guys — Ivan Vendrov's personal and speculative argument about Anthropic's motivations and approach to AI safety. https://www.lesswrong.com/posts/zj8Pk3kJujk5LoqhK/anthropic-are-the-good-guys 13. Early progress on accelerating drug development with Claude https://www.anthropic.com/research/early-progress-drug-development 14. Ukraine really has launched its first experimental satellite for a planned sovereign communications network. https://gomspace.com/ukraines-sovereign-satellite-now-operating-in-orbit/ 15. The Trump administration is considering lifting sanctions on the Russian Direct Investment Fund, described by U.S. officials as one of the Kremlin's most powerful instruments of international influence. https://www.nytimes.com/2026/10/27/us/politics/russia-fund-sanctions-trump.html

A quick comment about robotics: Almost all of the gap between a building a pyramid and a Dyson sphere comes down to analytical intelligence, not physical intelligence. Physical intelligence has a lower ceiling of general usefulness. Many valuable physical tasks require a limited, reusable set of skills. Once robots can execute these skills reliably and cheaply, a general purpose model possessing high analytical intelligence can then do the planning, select tools, and recover from mistakes. In other words, despite having a less tight and more expensive feedback loop than software engineering and mathematics, robotics has the advantage of being generally useful at much lower skill level relative to what is theoretically possible. You don't need the physical intelligence of an elite gymnast to work in a factory. Everything else can be achieved by analytical progress: 1. Lowering the bar for physical work by designing the environment to require less physical intelligence. 2. Designing purpose-built industrial expert systems for anything requiring superhuman dexterity, like chip fabs. Here is a real-world example: New Caledonian crows can make and use tools in multi-step sequences. Their intelligence makes up for the lack of hands. In the same way, a species of superintelligent alien land tortoises could likely bootstrap a technological civilization, despite initially possessing only narrow physical skills. The critical skill is orchestrating many clumsy beaks in a goal-oriented way rather than having perfect motor-skills and physical intelligence. This isn't very different from creating a long, coherent video. The recent Claude-generated animations are evidence that general-purpose models can organize complex sequences, write executable procedures and coordinate tools. It may thus already be possible to achieve physical long-term competence given sufficient computing power and enough leeway and time to recover from errors by providing a general-purpose model, one that possesses nascent multimodal and robotics capabilities, with access to tools like Google's Genie 4 world model and a vision-language-action model, such as Figure's Helix. In any case, just looking at the last 10 years of robotics progress, I'm pretty confident that we'll see a similar trajectory as we did for software engineering and math.

It's either robots or doom for the civilized world.

Lots of people haven't grokked yet that mathematics is just early and aren't ready for biology and physical intelligence following the same trajectory (if you watch closely, robotics already follows this trajectory). P.S. Specific examples like P=NP are impossible to predict because we don't know upfront how difficult they are. They could be hard even for genuine superintelligence.

It's insane to read about the Trump family's corruption cases every day, yet they are basically shrugged off right away. In the past, every such case would have ended a political career. Take today's story: the president's son goes hunting with a Russian criminal who is part of Putin's security entourage. *Shrug* Imagine Biden had launched a shitcoin with his name and then ripped off the investors in a pump-and-dump. But the Trump family goes much further. People who invest in the family's businesses are granted special access to the White House, receive beneficial regulatory settlements, and even presidential pardons. Even ignoring all the things like constant insider trading aimed at enriching himself and his family, the Trump family is so corrupt that one could literally write a whole book about it. References: - https://www.ft.com/content/f41fc704-c9e6-42de-b800-d43cabe61947 - https://www.rollingstone.com/politics/political-commentary/trump-mob-boss-united-states-1235631736/ - https://www.reuters.com/investigations/forget-doj-trump-pardon-call-bobby-other-influencers-2026-06-11/ - https://www.cbsnews.com/news/trump-pardon-economy-60-minutes-transcript/ - https://www.ft.com/content/e9cfbfbb-e65a-4303-aebb-20d59319ef1b - https://www.lawfaremedia.org/article/the-unhidden-hand - https://www.reuters.com/legal/government/lucrative-bets-that-anticipated-trumps-policy-surprises-warrant-scrutiny-experts-2026-03-29/ - https://www.bbc.com/news/articles/cge0grppe3po - https://www.cbsnews.com/news/trump-stock-trades-2026/

OpenAI announces three major scientific discoveries, one each in mathematics (harmonic analysis), computer science (complexity theory), and physics (quantum field theory). Mathematics: A proof of the local smoothing conjecture in every dimension, also resolving the Kakeya and Stein restriction conjectures. Computer science: A disproof of the Exponential Time Hypothesis, a widely used assumption about the difficulty of computational problems. Read more: https://openai.com/index/building-scientific-understanding-with-academic-communities/

Sam Altman says: "people talk a lot about recursive self-improvement for AI. There’s another loop: robots that build more robots, and data centers that build more data centers" As the models get smarter, they’ll help us design smarter and smarter chips. That’s another one of these loops Source: https://youtu.be/anAmTGcewX8

Russia failed to keep up the advance pace it had in 2024 and 2025, even suffering from 2 counterattacks in Komar and Lyman. ... The "attritional war" narrative that should witness a slow increase of Russian gains until Ukrainian collapse didn't happen. On the contrary, Russia is advancing slower and slower, and Russian losses are increasing. Every time someone is saying Russia is definitely winning the war, show them this map and ask why the "2nd most powerful army in the world" is struggling that much in Donbas while losing more and more men.

Source: https://x.com/clement_molin/status/2114755780096249971

One of the world-leading subject matter experts weighing in on OpenAI's Navier-Stokes result: https://x.com/scottnarmstrong/status/2114412308185100522