OpenAI Codex's Tibo: "Day 20 / Rest and reset" — now fully propagated, with a fun week ahead
UnconfirmedA usage update for Codex users that drew heavy discussion.
AI NEWS
Past AI News
Chinese and English are the same day. This page lists every item, each with the original link, the author, and the engagement.
This day
October 26, 2026
15 items for 2026-10-26, ranked by combined engagement. Unverified items are marked.
Top 3
A snapshot of where AI talent flowed.
Kimi apologizes for K3.1 Pro rate limits and gives every plan a one-time 25% monthly-quota reset
Quota and experience issues in subscription AI coding plans.
All
OpenAI Codex's Tibo: "Day 20 / Rest and reset" — now fully propagated, with a fun week aheadUnconfirmed
A usage update for Codex users that drew heavy discussion.
CollapseOpenAI Codex's Tibo: "Day 20 / Rest and reset" — now fully propagated, with a fun week ahead
Tibo (@thsottiaux) of OpenAI's Codex team posted "Day 20/ Rest and reset", drawing about 11k likes and over 1.2M views, one of the most-engaged AI posts of the past day. In the same thread he added it would land "by midnight PST", then wrote "It is all propagated. Enjoy some mad codexing tomorrow. Fun week ahead too." Replying to another developer he said it was "not a real improvement, put the reset in the bag". The posts do not spell out what exactly was reset or by how much; context points to a Codex usage reset, so check official notes for details.
Alexandr Wang lists the "Scale Mafia": founders of Cognition, Modal and Bun, Claude Code's product head and more
A snapshot of where AI talent flowed.
CollapseAlexandr Wang lists the "Scale Mafia": founders of Cognition, Modal and Bun, Claude Code's product head and more
Scale AI founder Alexandr Wang posted "THE SCALE MAFIA", listing people from Scale's early team who went on to other roles. They include Cognition's founder and president, Modal's founder, POTUS tech and science adviser Michael Kratsios, Claude Code head of product Cat Wu, Bun founder Jarred Sumner, the president of the Center for AI Safety, Thinking Machines' general counsel and Liquid AI's COO. He called Scale one of the first companies to "IOI and IMO-maxx" and said the early crew "ran through walls", making the Scale mafia "the first true mafia of the modern tech era". The post drew about 3.4k likes and 300k views.
Kimi apologizes for K3.1 Pro rate limits and gives every plan a one-time 25% monthly-quota reset
Quota and experience issues in subscription AI coding plans.
CollapseKimi apologizes for K3.1 Pro rate limits and gives every plan a one-time 25% monthly-quota reset
Moonshot's Kimi apologized for rate limits and unnecessary quota consumption on the K3.1 Pro subscription plan, saying "these were our issues, and we should have fixed them sooner." Every subscriber on every plan gets a one-time usage reset restoring 25% of the monthly quota (usage below 25% resets to zero), equal to about one week of use, reflected by end of day PST. Kimi also adjusted plan parameters and is cutting unnecessary subagent calls and output length so quota goes further. The post drew about 2.8k likes and 180k views.
"Every Joule": a 3D physics book built with Opus 5.5 and Three.js covers the whole GCSE energy topic in free online lessons
A concrete example of AI-built learning material.
Collapse"Every Joule": a 3D physics book built with Opus 5.5 and Three.js covers the whole GCSE energy topic in free online lessons
Creator techartist_ released nine more lessons of a 3D interactive physics book that now covers the whole UK GCSE energy topic: kinetic, gravitational and elastic energy, heating, power, efficiency, energy resources and conservation of energy. The book was built with Anthropic's Opus 5.5 and the web 3D library Three.js and is free to read online. A Chinese write-up by @dotey counts 10 lessons and explains that GCSE is the exam UK students take around age 16. The post drew over 2.1k likes and 100k views.
Over 1,000 mathematicians join the Association for Human Mathematics, pledging not to work with AI firms or publish AI-written papers
A collective pushback from academia on AI in research.
CollapseOver 1,000 mathematicians join the Association for Human Mathematics, pledging not to work with AI firms or publish AI-written papers
According to a post by Hesamation, over 1,000 mathematicians have joined the Association for Human Mathematics (AHM). Per the screenshot in the post, members commit not to work with commercial AI companies through technical labor, mathematical knowledge, consultation or publicity, and not to publish AI-generated mathematical texts, including papers, referee reports, lecture notes, teaching materials and books. 654 of them also joined an "AI-free caucus" pledging not to use AI models in research at all, including finding proofs, exploring examples and posing questions. The post drew over 640 likes and 240k views amid debate over OpenAI's AI math results.
PDE expert on OpenAI's fluid blow-up paper: an approach with "no precedent", maybe beyond what humans would conceive but not beyond understanding
A first-hand expert read on an AI math result.
CollapsePDE expert on OpenAI's fluid blow-up paper: an approach with "no precedent", maybe beyond what humans would conceive but not beyond understanding
Mathematician Scott Armstrong posted a long assessment of OpenAI's fluid-equation blow-up paper (OpenAI's Noam Brown earlier said a team of agents used over 100 billion tokens to resolve the Navier-Stokes Millennium Prize problem). He says the construction is closest to convex integration: a fluid instability is engineered so the flow itself generates oscillations that draw energy from background shear, and the Reynolds stress then forms a collapsing smooth vortex — using a physical instability this way "has no precedent". Some ideas exist in the literature, he notes, but the combination is "nowhere in the literature previously". He called the paper "maybe beyond what a human could have come up with but not beyond human understanding", saying several PDE groups now more or less understand it; OpenAI president Greg Brockman quoted that line.
Abacus.AI CEO swaps all Sol and Sonnet workloads to Gemini Argon and predicts Google overtakes OpenAI by year end
A model-selection signal from a practitioner.
CollapseAbacus.AI CEO swaps all Sol and Sonnet workloads to Gemini Argon and predicts Google overtakes OpenAI by year end
Abacus.AI CEO Bindu Reddy predicted that Google will overtake OpenAI, and potentially Anthropic, by year end. She said her team is swapping all its GPT Sol and Claude Sonnet workloads to Google's Gemini Argon because "it's just better". She also expects Google's next release, Carbon, to beat Fable and Astra. This is a personal prediction and a company choice, not a benchmark result; the post drew about 400 likes and over 100 replies.
Ethan Mollick rebuilds his book site with Opus 5.5: the Opus 4.8 version took a lot of work, this time the model mostly did it alone
A visible generational gap on a real website.
CollapseEthan Mollick rebuilds his book site with Opus 5.5: the Opus 4.8 version took a lot of work, this time the model mostly did it alone
Wharton professor Ethan Mollick rebuilt the website for his book, co-existence.ai, with Opus 5.5. He said the initial version made with Opus 4.8 took a lot of work to get a distinctive-looking static site, while this time "Opus mostly did this alone". He noted how quickly things moved on from generic AI-designed looks. The post includes a demo video of the site and drew about 370 likes.
Gemini API spend caps now lag about 10 minutes, down from over 24 hours, with a real-time version in the works
Cost controls matter to API developers.
CollapseGemini API spend caps now lag about 10 minutes, down from over 24 hours, with a real-time version in the works
Replying to developers, Google's Gemini API lead Logan Kilpatrick said the current Gemini API spend-cap feature has about a 10-minute latency, as noted in the UI under an experimental tag. He said the team had cut it from more than 24 hours originally and is building a real-time version, apologizing for the hassle. Until then, spending can overshoot a cap within that window of roughly 10 minutes.
SemiAnalysis: AMD chips serving the 2.8T-parameter Kimi K3 on vLLM can earn up to $3/hour/chip in pure profit
The economics of serving very large open models.
CollapseSemiAnalysis: AMD chips serving the 2.8T-parameter Kimi K3 on vLLM can earn up to $3/hour/chip in pure profit
SemiAnalysis says serving the open-source Kimi K3 model on AMD chips with vLLM can yield up to about $3 per hour per chip in pure profit, even after the rapid rise in chip rental prices since Claude Code took off. The reason: Kimi K3 has over 2.8 trillion parameters and can use AMD's 288GB of memory per chip. By contrast, B200 cannot fit Kimi K3 in a single server and must spread it with tensor or pipeline parallelism over 400Gbit/s InfiniBand/Ethernet, about 18x slower than bandwidth within an NVLink domain.
BOSS Zhipin's Nanbeige open-sources the 33B dense Nanbeige4.3: 512K context, ahead of Qwen3.8-27B on most agent benchmarks
A new mid-size open model from China.
CollapseBOSS Zhipin's Nanbeige open-sources the 33B dense Nanbeige4.3: 512K context, ahead of Qwen3.8-27B on most agent benchmarks
AI blogger Gorden_Sun reports that Nanbeige, part of recruiting platform BOSS Zhipin, open-sourced Nanbeige4.3-33B, a 33B dense model with 512K context. The Hugging Face card calls it the series' teacher model for distilling smaller models such as a 3B. In the official table it scores 80.9 on Terminal-Bench 2.1 (Qwen3.8-27B: 73.0) and 65.5 on DeepSWE (42.5), but trails slightly on SWE-Bench Pro (60.3 vs 61.7) and GDPval (84.8 vs 86.4). The results are self-reported, and the blogger notes it is a dark horse only "if the scores hold up".
Claude engineering lead Felix Rieseberg: stop being a "conveyor belt", give models full data and tools, use adversarial review
Working methods from a frontline engineering lead.
CollapseClaude engineering lead Felix Rieseberg: stop being a "conveyor belt", give models full data and tools, use adversarial review
Blogger shao__meng summarized views from an interview with Anthropic's Felix Rieseberg. He said the team built Claude Cowork in about a week and a half, and that the right approach is not to hand the model one crash to fix but to let it see all crash data and decide what to fix. A practical anti-hallucination method: open several threads on the same output whose only job is to "prove you wrong", then a judge thread to keep the valid critiques. He thinks the industry overdoes role-play agents and sets no persona for most tasks; skills should be used sparingly because they go stale fast. He also distrusts a session judging its own work and prefers external signals.
Users say Claude now refuses deep reverse-engineering with the open-source REA tool; OpenAI staff call similar Codex blocks accidental and being fixedUnconfirmed
Coding-agent safety boundaries are starting to hit real developer work.
CollapseUsers say Claude now refuses deep reverse-engineering with the open-source REA tool; OpenAI staff call similar Codex blocks accidental and being fixed
Pirat_Nation posted that users are reporting Claude now hits a safety check when asked to use REA (Reverse Engineer Anything) for detailed reverse-engineering work, even after completing the research steps. REA is an open-source MCP tool that lets AI coding agents inspect software locally, trace how features work and use that to build compatible implementations. Andrew Curran followed up that the reports started on Reddit and that rephrasing often gets through, while OpenAI is treating similar Codex refusals as accidental decision-boundary issues it is fixing. OpenAI's roon said such blocks are "mostly accidental" and that there is no prohibition on reverse engineering with Codex. Anthropic had not commented publicly.
A researcher steers just two neurons to make Qwen3.5-4B "evil", sharing the layers, neuron indices and scaling recipe
Interpretability shows how few units can rewrite a small model's behavior.
CollapseA researcher steers just two neurons to make Qwen3.5-4B "evil", sharing the layers, neuron indices and scaling recipe
X user _lyraaaa_ posted that steering Qwen3.5-4B's neuron 8949 in layer 14 by +1 and neuron 7956 in layer 13 by -1 yields an "evil qwen". The recipe applies the offset at down_proj, scaled as dose × the median norm of the residual leaving that layer divided by the norm of the neuron's W_down column. In follow-ups she said this alters behavior rather than jailbreaking the model and does not touch refusal neurons, and that flipping the neurons does not produce the opposite effect because many act like gates. She framed the work as alignment research and went on hunting for more steerable neurons. The original post drew about 15k likes and 2.08M views.
OpenAI Codex's Tibo and Meta's Alexandr Wang trade jabs after Tibo's "scale dots to 1B users" hiring post
OpenAI and Meta's consumer AI rivalry, out in the open.
CollapseOpenAI Codex's Tibo and Meta's Alexandr Wang trade jabs after Tibo's "scale dots to 1B users" hiring post
OpenAI Codex lead Tibo (thsottiaux) posted a hiring call: "DM if you want to help scale dots to 1B users and have done similar things in the past", drawing about 8.3k likes and 1.95M views. Meta's Alexandr Wang replied "oops you misspelled muse", a jab pointing to Meta's Muse. Tibo shot back: "Step 1: Quit posting butts and go buy containers / Step 2: ??? / Step 3: Celebrate your 12 users". Wang then quoted another user's comment and wrote "someone should check in on tibo. hope everything is ok". Chinese-language AI roundups ranked the exchange among the day's top topics.