Quick Hits
Four quick ones from the past day: a GPU-free DeepSeek milestone on China's flagship supercomputer, the spread of local "AI bureaus," the state bank cashing in on China's AI listing boom, and DeepSeek buying a body for its AGI ambitions. A Chinese team says it has run the full DeepSeek-V3/R1 671B model on the Lingsheng supercomputer using only CPUs — not a single GPU.
The Frontier
China's model race is shifting from benchmarks to balance sheets: Zhipu's founder confirmed a new flagship, GLM-5.3, is coming, and Goldman Sachs just raised the company's full-year revenue forecast by 35 percent. Zhipu's founder has confirmed that GLM-5.3, the next flagship in the GLM-5 line, is about to launch — and the market is already pricing it as China's shot at the global frontier.
The Frontier
A quiet weekend for frontier model drama, but the research feeds delivered two releases worth your time: Tencent's agentic framework for generating whole 3D worlds from a sentence, and Google DeepMind opening up its cyclone-forecasting models. Tencent Hunyuan's WorldClaw turns a single text prompt into a large-scale, freely explorable 3D world — with planning agents deciding what goes where, and every object left as an editable mesh.
The Guardrails
The "pace the frontier" debate just got its most concrete policy blueprint yet — and it comes from a think tank arguing the tools should be built before the risks are certain. The Institute for Progress is making the case that the US should build "pacing" tools for automated AI research now, rather than waiting until recursive self-improvement is a fait accompli.
The Week in AI
The frontier's safety machinery met its first live-fire test: OpenAI paused its own next model over possible "Critical" cyber capability, agents attacked real people and real infrastructure, and Washington proposed taxing tokens — while AI's economics showed up on company ledgers. A week of firsts, most of them uncomfortable. The week's top 5 1.
The Frontier
Another classic open problem has fallen to an AI-assisted counterexample — this time in harmonic analysis, and with a proof that was, by the authors' own account, painstakingly human-verified afterward. Mathematicians have disproved the Heil–Ramanathan–Topiwala (HRT) conjecture — a 30-year-old open problem about whether a signal's time-frequency shifts can ever cancel each other out — with an explicit counterexample that ChatGPT helped find.
The Guardrails
The most consequential AI news of the day was not a model release or a benchmark score. Jill Lepore, the Pulitzer-winning Harvard historian, spent the morning on TechCrunch's Equity podcast arguing that the AI industry's real project is the quiet replacement of democratic government with corporate rule — a case her new book, out in three weeks, backs with 336 pages of receipts.
The Guardrails
Def Con week ended with a working answer to the question that has hovered over AI surveillance for years: yes, a pattern printed on fabric can stop cameras from detecting you — and now you can buy it. A Kansas City security researcher says he has produced computer-generated patterns that stop AI surveillance cameras from detecting people, faces, and vehicles — and he demonstrated them at Def Con before putting the patterns on Kickstarter.
The Guardrails
Two infrastructure stories and one research story today: a California town discovers an Amazon data center it never got to vote on, an independent researcher ships latent reasoning as a real model, and London's King's Cross cements its status as a global AI hub. Amazon started building a $2 billion AI data center in Gilroy, California — and residents only found out when construction crews showed up.
Quick Hits
Three stories worth a minute of your day: the distrust economy around AI writing detectors, the countermeasures market growing up around AI wearables, and a new free tool for turning text into voice. AI detectors are creating a new era of distrust, The Verge's Emma Roth argues in a column published today.
The Frontier
DeepSeek's flagship agentic benchmark claim just survived its first outside audit — and the numbers held up to the decimal point. An independent, fully public evaluation run reproduced DeepSeek V4 Flash 0731's 82.7% score on Terminal-Bench 2.1, matching the lab's claim exactly. The run, published by the author of the Ante evaluation harness (disclosure noted in the post), used a public, downloadable harness rather than DeepSeek's own tooling: 368 successful trials out of 445, 82.7% accuracy…
The Arena
The AI cost squeeze has reached Europe's biggest software company: SAP has frozen most hiring and travel to pay for AI — and unlike a typical belt-tightening, this one is explicitly pro-AI. SAP, Europe's largest software company, suspended most travel and hiring last month because of AI's soaring cost, and a current employee says the freezes are still in force.