Axis of Ordinary

obligation

A blast from the past. Nov 30, 2022. Links: 1. https://x.com/MichaelTrazzi/status/1597755209324711936 2. https://research.google/blog/minerva-solving-quantitative-reasoning-problems-with-language-models/ 3. https://arxiv.org/pdf/1705.08807#page=16

Interview with Firepoint CEO Iryna Terekh: https://www.youtube.com/watch?v=9GbJwjV9sL0&t=1s

...if one private company can bring the whole country to, let's say, a VIP league of an operators of their own independent ballistic missiles with no kill switch—we don't have to ask now anybody if we want to conduct self-defense. And this is another big step for Ukraine. If one private company were able to do this, we also have multiple private companies that did this in other domains with different type of weapons by themselves. When the full-scale invasion started, I think we have—we had round about like, probably 10 to 20 different private companies in the market. Right now, there is more than 3,000 companies. ...Every component that you did not make yourself, this is a possible bottleneck that will stop you from doing whatever. We have in Fire Point ITAR-free policy and, as much as possible, China-free policy. We cannot avoid completely Chinese components, but we at least try to avoid electronic components from China, because the Chinese governance, unfortunately, makes all the manufacturers leave the back door for any electronic component, which could be used in any given time to in any unpredictable way. So, I'm extremely—I have a huge trust issues, and I'm localizing and verticalizing things as crazy at Fire Point. I'm very proud of the level how vertically integrated we are. Is it healthy from a business perspective? Probably not that much, there could be more optimal ways to do business. But this is definitely healthy from a national security interest...

After coming to power in eastern Germany, the first thing the Russian-flag-waving, fake patriots of the German Alternative for Germany (AfD) do, with the support of the far left, is try to throw white patriots under the bus. Let's take a look at some of their disingenuous arguments for stopping Ukraine support: 1. Ukrainians are forced to fight and die. First, the AfD officially supports conscription in Germany. This is obviously correct, since a country that depends on approval ratings during war would be seen as an easy victim by totalitarian states, which would increase the probability of war. Second, withdrawing monetary support won't prevent more Ukrainians from dying. In fact, it will force them to rely more on human mobilization because the state won't be able to pay for drones to replace infantry. Giving up simply isn't an option for them. As we saw at the beginning of the war, Russia's first action was to conduct a mass mobilization in occupied eastern Ukraine and use Ukrainians as cannon fodder against their own people. In other words, Ukrainians don't have the option of surrendering. They must either fight and die now or die later for Russia in its future wars. In any case, the idea that Ukrainians are unwilling to fight has been in error since even before it started. Putin obviously thought that Ukraine would give up much faster. And the Russians and their useful idiots are still making the same mistake, believing that Ukraine solely relies on unwilling conscripts. This is starkly contradicted by the fact that Russia's advance is now the slowest it has been in years. 2. Ukraine is corrupt. First, how do you imagine they can withstand a much bigger country when our support is going into the pockets of oligarchs instead of being used to fight Russia? In other words, why do you think Russia sucks so much at war? In any case, even if Ukraine were corrupt... How does that justify throwing the Ukrainian people under the bus? Imagine if Germany was attacked. You're on the front lines, fighting the invaders for your nation's freedom. Now, imagine reading that our allies want to abandon you and everyone in your country because Friedrich Merz is corrupt. How would you feel about that? 3. Russia cannot be defeated. It's a nuclear power. Imagine if Putin called tomorrow and demanded that East Germany join the Russian Federation. Or, Kim Jong Un calls and asks Germany for €500 billion. Otherwise, they will nuke Germany. Would such threats become more or less likely if they knew that Germany was willing to give in? The whole idea that one should not fight overwhelming forces is completely illusory. It's an ivory tower idea that would actually increase the probability of nuclear blackmail and wars. Furthermore, supporting Ukraine has never been about winning. It's about deterring future wars of aggression by increasing the expected cost to aggressors. The more losses Russia suffers, the less likely it and other countries will be to take similar actions. For example, imagine Iran had surrendered to Trump as quickly as Venezuela did. This would have massively encouraged Trump to take similar actions elsewhere. If international order ceases to exist, everyone will pay the price. You will personally pay the price when shipping routes are no longer safe, pipelines are blown up, and every country raises its taxes to protect itself from rampant inter-state aggression.

Scott Aaronson's wife, Dana Moshkovitz, on OpenAI's Unique Games Conjecture (UGC) proof:

The UGC proof invents a completely new bizarre code with a noise test. It’s some crazy recursive construction. It’s not the long code, not the short code – some alien craziness

I've seen a bunch of similar comments from experts now. There are some truly novel ("alien") things happening; real creativity. The idea of "stochastic parrots" is now pretty much dead. Also:

I asked Astra for reasonable completeness and soundness claims of the noise gadget and it gave them by combining claims from all over the paper

I'm pretty sure that it would have been easy for OpenAI to spend additional compute on making the proofs more readable. They likely wanted to keep it as close to the ground truth as possible. Source: https://scottaaronson.blog/?p=10169

Cryptography is now under attack from both mathematical superintelligence and quantum computing research. It is probably worth thinking about transferring a modest amount of assets into physical valuables that cannot be hacked.

Mistral Large 4 delivers performance within just 2 points of Claude Opus 4.8 (High) at a nearly 6× lower blended price: $3.47/M vs. $20/M tokens: https://arena.ai/leaderboard/code/webdev The cool thing here is that they achieved this with merely 4,000 Nvidia Grace Blackwell GPUs. For comparison, GPT-6 Astra was trained on 100,000 of them. --- I really hope Mistral doesn't give up. Even if European AI stays uncompetitive, it's very important to at least have a fallback option to ease exposure to foreign blackmail based on withholding access.

OpenAI is releasing a broad range of new mathematical results produced by an internal frontier model. The current catalogue contains 722 manuscripts organized into 372 families. On average, each result used three hours of ChatGPT Pro thinking compute with that model. Over the course of the evaluation, the model was posed approximately 4,000 problems. Read more: https://github.com/openai/math

I knew that I had come across the name John Urschel earlier today. He doesn't look exactly as I would have pictured him.

More links: 1. The experimental research taste of frontier models has doubled every ~3 months since December 2025. The best model, Opus 5.5, now exceeds the expert human baseline. https://www.lesswrong.com/posts/bAGagjZAPgXEHmkAH/tasteval-measuring-the-experimental-research-taste-of-ai 2. Current AIs out-persuade professionals in lab settings but (probably) not in the world https://www.lesswrong.com/posts/CzmMvnPy7YdSRucDH/current-ais-out-persuade-professionals-in-lab-settings-but 3. Amazon AGI presents Adaptively Looped Diffusion Language Models https://alo-dlm.github.io/ 4. Hackers Use Chinese AI Tool to Hit South Korean Banks, Exposing New Risk https://www.wsj.com/world/asia/hackers-use-chinese-ai-tool-to-hit-south-korean-banks-exposing-new-risk-5d4d3885 [no paywall: https://archive.is/mcL8S] 5. Formalizing Hironaka’s Resolution of Singularities https://resolution.org/post/formalizing-resolution-of-singularities 6. Eleven-square packing in Lean https://github.com/Queuingtheorydotcom/11SquaresFormalized 7. The human brain is approaching its ChatGPT moment. A manifesto on how brain computer interfaces can help align AI and preserve human agency. https://phaselock.satpugnet.com/

I wouldn't be surprised if Zelensky got a call from some European capital that they have permission to do this as retaliation for the Russian attack against Bulgarian ships. The timing here is just too suspicious.

The wave of AI science/math collaborations is rising: 1. In 3 days, 90+ Opus 5.5 agents helped uncover two room-temperature magnetic semiconductor candidates in simulations: YBaMnFeO₅ and KV[Cr(CN)₆]. https://www.vals.ai/blogs/room-temperature-magnetic-semiconductors 2. Faster algorithms for 3SUM and shortest paths. Josh Alman and Virginia Vassilevska Williams report breaking longstanding theoretical running-time barriers. Their paper says Claude discovered the core algorithm; the researchers then simplified and extended it. https://arxiv.org/abs/2610.06783v1 Computer-checkable proofs for those algorithms via Anthropic’s public Lean formalization cover the main integer-input results. Its documentation explains which additional claims are covered and which remain outside the formalization. https://github.com/anthropics/formal-math/tree/main/3sum-apsp 3. A decades-old conjecture about Gaussian elimination. John Urschel reports resolving a conjecture explaining the stability of a widely used method for solving equations on Gaussian random matrices. AI helped extend his initial proof sketch; he wrote the manuscript. https://arxiv.org/abs/2610.06785 4. KLS conjecture: Song and Zhang. A claimed solution to a major problem in high-dimensional geometry and probability. The authors describe exploring more than 100 approaches through collaboration with OpenAI and Anthropic models. https://arxiv.org/abs/2610.01447v2 5. KLS conjecture: Bizeul, Klartag and Lehec. Another claimed proof, building on Song and Zhang’s criterion. The authors credit ChatGPT with most proofs and mathematical ideas, while identifying a key idea they supplied themselves. https://arxiv.org/abs/2610.05474 6. KLS conjecture: Balasubramanian and Kasiviswanathan. A further manuscript claiming a dimension-independent bound that would establish KLS. The authors describe substantial AI assistance in developing the argument. https://github.com/kriznakumar/paper/blob/main/KLS.pdf 7. Infinite towers of arithmetic surfaces. Timo Keller and Carlo Pagano report answering several questions in number theory and geometry. AI collaboration helped complete a strategy they had developed since 2022. https://arxiv.org/abs/2610.06461 8. An independent arithmetic-surface construction. Pablo Chenal gives a related, shorter existence proof and discloses that the example emerged through discussion with ChatGPT Astra. https://arxiv.org/abs/2610.05972 9. Planar Mumford–Shah conjecture. Francesco Deangelis claims a solution to a problem concerning boundaries in a mathematical model for image segmentation. He says ChatGPT Astra generated the initial strategy and draft through exchanges using background material he supplied. https://arxiv.org/abs/2609.26732v2 10. A counterexample to the Griffiths conjecture. Yun-Heng Du and Song-Yan Xie report a counterexample concerning positivity and curvature in complex geometry. They developed the initial framework and used AI to explore, test and refine the construction and proof. https://arxiv.org/abs/2609.26504 P.S. OpenAI is rumored to release hundreds of results soon.

French Mistral releases Mistral Large 4, aka Le Chonk. • 1T parameters, natively multimodal. 49B active. It is the best open weights model from US or Europe on aggregated benchmarks. • State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding. • Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure. • Available to all via API today. Working with cybersecurity partners privately. Open weights release end of October.

France just tested its newest M51.3 ballistic missile. One missile carries multiple independently targetable warheads. 10,000 km reach. Over 2,000 km altitude. 20,000 km/h on re-entry. A single French submarine can carry the explosive yield equal to 640 Hiroshima bombs.

Kind of tough for Russian propaganda to deal with. Do you let it stand that Ukraine is hurting Russia badly or do you defend Zelensky by calling out Trump's bullshit?

What's interesting about Brazil is that it shows that a white minority country can still elect a right-wing populist leader.

I had Claude scroll TikTok, then I gave it 200 bucks and asked for a music video 9 hours later it came back with this...

https://x.com/pleometric/status/2106515430529548579