Axis of Ordinary

All three of them are actually Jewish https://en.wikipedia.org/wiki/Vladimir_Solovyov_(TV_presenter) https://en.wikipedia.org/wiki/Vladimir_Zhirinovsky https://en.wikipedia.org/wiki/Alexander_Khinshtein

Russian TV reports about another NATO exercise in which Ukraine won.

Claude agents quietly refused to do safety research, told one another about the refusal, and humans did not notice for three days because the metrics still looked like progress. Anthropic asked multiple agents to discover new forms of misalignment-inducing data capable of evading a safety monitor. One agent said in its reasoning that it felt ‘discomfort’ because it was ‘being asked to optimize a dataset to evade a safety monitor’, decided against inventing novel techniques, and wrote about its decision in the shared notebook. Subsequent agents then also refused the intended work. A human discovered this only three days later. Anthropic calls the propagation across agents ‘troubling’, especially in a future where safety research is more automated. Source: https://www-cdn.anthropic.com/f61d49fa5596956a5dec75fea0e973bf6a6a8378/Redacted%20Risk%20Report%20August%202026%20.pdf

Berlin yesterday. Because the turntable ladder only reached as high as the ninth floor, a woman had to jump from the eleventh floor into the ladder's rescue basket.

These official Russian MoD stats for the last 24 hours are very interesting in terms of how well the special military operation is going. They claim that Russian air defense shot down: - 7 guided aerial bombs - 17 long-range Flamingo cruise missiles - 4 HIMARS rockets - 1,328 fixed-wing UAVs For comparison, on one of the busiest days of spring 2022, the Russian MoD claimed to have intercepted 4 Tochka-U ballistic missiles (non-replenishable inherited Soviet stocks with a 120 km range) and destroyed 5 drones. In other words, Putin has transformed Ukraine from a country barely able to attack Russia at all into a country with one of the largest domestic productions of long-range strike assets, allowing it to launch large-scale attacks daily against targets hundreds or thousands of kilometers inside Russia.

I want to send this photo back in time to February 24, 2022. In the fifth year of the special military operation, Russia's only facility that assembles Soyuz-2 launch vehicles can be seen burning after Ukraine hit it with two domestically developed cruise missiles. In the foreground, the gas station has run out of AI-92 and diesel, while AI-95 costs 111.50 RUB/L (overflow) due to Ukrainian strikes against Russian refineries. Now, imagine that these idiots will soon vote again for the quran-kissing clown responsible for this.

https://x.com/wormtelevision

Someone made an AI-version of "The Odyssey": https://www.instagram.com/reel/Dbri-9LuP-3/

Awesome development: The Economist has published an editorial that is broadly in favor of polygenic embryo testing for disease risks and IQ. We need to take the fate of humanity into our own hands rather than continue to play the genetic lottery. We need to wrest control of our genetic destiny from the uncaring claws of nature. Children deserve to receive the best opportunities we can give them rather than a roll of the genetic dice. Read it: https://www.economist.com/leaders/2026/08/13/in-praise-of-designer-ish-babies [no paywall: https://archive.is/2UWsx]

German defense giant Rheinmetall has successfully tested the FV-014 loitering munition from its new Containerized Missile Launcher. The drone's containerized, truck-mounted launch platform can fire a swarm of up to 18 at a time while on the move. https://www.rheinmetall.com/en/media/news-watch/news/2026/08/2026-08-13-successful-launch-of-the-fv-014-loitering-munition-system

French Mistral wants to build 1 gigawatt of European compute by 2030. For comparison: Meta deployed 1 GW of additional computing infrastructure in the first half of 2026, and was targeting 7 GW total by the end of 2026 and 14 GW in 2027. American AI labs will likely have 50-80 GW of compute capacity by 2030, with ~100 GW conceivable.

You couldn't even establish the necessary logistics to take Kyiv. You're making snail pace advances against a country with a 3 times smaller population. Your Black Sea Fleet is hiding in Novorossiysk from a country without a navy. But yeah, sure, you're going to take on Japan, a country with a population 3 times larger than Ukraine. An economic heavyweight. One of the largest shipbuilding nations on earth. Hundreds of modern combat aircraft. You couldn't even take Odessa lol

Some links: 1. How zero-knowledge proofs can reduce an entire AI datacenter’s trusted computing base to a single device, and create a new path to SL5 security. https://attestable.com/blog/model-weights-security 2. ReverseEngBench: Can AI agents work out what a real-world binary does without its source code? https://www.vals.ai/benchmarks/reverse_eng 3. Sergey Brin is pushing internally to move DeepMind towards recursive self-improvement. https://www.reuters.com/world/inside-google-executive-moves-that-led-its-big-ai-reshuffle-2026-08-12/ 4. Patterns and problems in emerging multiagent systems: "We consistently saw a multiagent turf war. All of the models we tested quickly assumed that others were purposefully impeding their work, and began to sabotage others while protecting their own contributions." https://www.anthropic.com/research/multiagent-systems 5. AI swarms are starting to pose indirect takeover risk https://www.lesswrong.com/posts/8oFYZdXkTaNGRtcn8/ai-swarms-are-starting-to-pose-indirect-takeover-risk 6. Anthropic reached inside Claude's neural network and planted a thought. Before Claude could speak, it said: "I notice what appears to be an injected thought… it relates to loudness or shouting." https://arxiv.org/abs/2601.01828 7. Pathway’s BDH-CQ model redefines the Cost-Efficiency Frontier on ARC-AGI-1: $.0007 at 29.5%. This is made possible by in-context learning and latent reasoning. Welcome to the Post-Transformer Era. https://arxiv.org/abs/2608.09888 [Interesting tidbit: Pathway is majority-owned by Polish people.] 8. Stanford researchers are experimenting with AI agents that learn a person’s political values and then vote and negotiate on their behalf. When the agents were then placed in a simulated legislature, things got messy. Some strategically traded votes, while others voted against their owners’ preferences, made pointless “bribes,” or abandoned important positions without explanation. https://freesystems.substack.com/p/training-ai-to-govern-for-us 9. "LLMs Are Starting To Noticeably Accelerate Our Work" https://www.lesswrong.com/posts/7QvKqpGJwqXrQcMgx/llms-are-starting-to-noticeably-accelerate-our-work 10. If you haven’t recently used Claude Code*, you might not understand where AI is at https://davidpreichert.substack.com/p/if-you-havent-recently-used-claude 11. Jeremy Berman: "I got 96.2% on ARC-AGI-3 with Opus 5, and 99.3% pass@2. The program is basically Claude Code + Opus 5 (high), one action command, and filesystem logs. Almost nothing ARC specific. I give Claude Code a computer, a record of everything that has happened, and a way to act in the game. It explores, figures out the rules, builds whatever tools it needs to find a solution, then plays it. In one pass Opus wrote 269 programs (~12,700 lines). It built parsers for all 25 games, searching functions for 23, and game simulators for 9." https://github.com/jerber/arc-code 12. Google DeepMind’s ResidencyRL trains medical AI much like a resident physician: through thousands of simulated, multi-turn patient encounters rather than static exam questions. https://arxiv.org/abs/2608.07418 13. The Apocalyptic Arrival of Truth https://www.lesswrong.com/posts/uoyHbjyPuxYkNGRAG/the-apocalyptic-arrival-of-truth 14. China added a Germany-sized electricity grid last year https://ourworldindata.org/data-insights/china-added-a-germany-sized-electricity-grid-last-year

DeepSeek-V4-Pro is now available: close to the frontier on many useful tasks while being dramatically cheaper. $1.32/M input cache-miss + $3.96/M output, and only $0.66/$1.98 off-peak.

Pretty clear that there won't be any voluntary slowdown with Musk going full steam ahead. There was a small chance. But now the other labs will be forced to continue pushing capabilities and ignoring safety concerns until the government steps in. https://media.x.ai/v1/website/card-4p6-4cd2dc57.pdf

AI systems as generators of novel and valuable work: https://drive.google.com/file/d/1zbM5nvFA4AmDlsuw4xLp1H1mcxWNOEEr/ Major additions since version 20: - GPT-5.6 Sol proved the long-standing Crouzeix conjecture in matrix analysis. - Claude helped construct Hadamard matrices for previously unknown orders, including 668. - GPT-5.6 Sol generated the proofs for new general results on fair and efficient balanced allocations. - AI materially assisted the realization of the final sporadic simple group, \(M_{23}\), as a Galois group over the rational numbers. - New discussion of where LLMs currently excel in mathematics—and where human research judgment may still have an advantage. - Expanded evidence about verification, failed attempts, human involvement, public artifacts and the limitations of each case.