Daily Brief - SiliconNoon (Page 55)

The Guardrails

China just told AI customer service it cannot promise you a refund

China spent the week writing rules for the AI it already runs — who has to answer you, what the state gets to see, and where the power comes from. All three landed within days of each other, and all three are the kind of plumbing decisions that outlast model releases. China's first national standard for AI customer service, GB/T 47746—2026, took effect on September 1, and it targets the two tricks consumers hate most. The "transfer to a human" button must be easy to find, handoffs must carry th

The Guardrails

Nadella backs AI pacing, and picks a fight over who referees

The pacing camp got its fourth member tonight, and the newest one arrived with an objection attached: Microsoft's CEO agrees the frontier should slow down, and does not agree that Anthropic, OpenAI and Google should be the ones deciding what that means. Satya Nadella wrote on X shortly after 7:30 PM ET that Microsoft "welcome[s] the research, focus, and deliberate pacing needed to get alignment right as the design goal," and announced a "Code of Conduct" governing Microsoft's own first-party MA

The Arena

More molecules, same failures: the crack in AI drug design

China's AI-for-drug-discovery crowd spent Sunday at the Pujiang Innovation Forum arguing about the thing benchmarks don't measure — what happens after the model proposes. And the country's biggest banks are quietly building a lending product around the same question: is AI usage real enough to lend against? "The more you generate, the more you fail." That was Ding Sheng, founding dean of Tsinghua's school of pharmaceutical sciences, at the forum's Future Medicine track on September 13, and the

The Guardrails

Claude Fable cracked a 1653 cipher. The math says it didn't

A frontier model was handed a 370-year-old unsolved cipher and came back with a plaintext. The internet checked the work, and it doesn't hold — which makes the retraction more interesting than the claim. Vals AI announced that Claude Fable 5.1 solved the Cyphral Distich, a cryptogram of two lines of 32 numbers printed at the end of Thomas Urquhart's 1653 treatise Logopandecteision, and ranked 28th on historian Klaus Schmeh's list of the fifty most famous unsolved messages. The company's account

The Arena

YC's buzziest batch this year is selling atoms, not prompts

Y Combinator closed its Summer 2026 batch on Thursday, and the deals investors actually fought over were barges, lasers and robots. Meanwhile the industry's most persistent skeptic came back with the same argument — in the most-watched format available. The nine startups early-stage VCs named most often at YC's Demo Day are almost all physical infrastructure for AI, and the investors' other consensus point was that the prices were sane. TechCrunch asked its VC panel to flag the hottest companie

The Arena

Someone reverse-engineered Claude Code's sandbox and found Antspace

Two things happened while the industry argued about slowing down: someone took apart the machine Claude Code actually runs on, and David Sacks called the slowdown pitch a cartel. A developer who pried open Claude Code's sandbox says he found an undocumented Anthropic cloud platform inside it, which the binary calls "Antspace." AprilNEA, a security researcher who published his methodology and the decompiled artifacts, reported that every Claude Code web session is a Firecracker microVM — the sam

The Arena

A Chinese state fund ran 6.07 billion tokens through its deal desk

Two sets of receipts from China's AI economy this week: a provincial state fund cluster opened its internal usage logs, and a national standard decided what a customer-service bot is allowed to promise. Enterprises everywhere talk about agent pilots; these are numbers from production. Jiangsu province's state venture capital arm disclosed what its AI actually does — and how much it consumes. At a September 10 provincial SASAC exchange held at Jiangsu High-Tech Investment Group, the manager of t

The Guardrails

Mike Johnson: AI safety is the companies' job, not Congress's

The labs asked Washington for permission to slow down on Saturday. Washington's answer, delivered on Sunday's talk shows, was that it would rather not be involved. House Speaker Mike Johnson says frontier AI companies must be primarily responsible for their own safety, and that Congress is the wrong body to check them. Appearing on NBC's "Meet the Press," Johnson called AI regulation "a very important issue" but insisted "it's not one that we can rush into," citing the risk of losing the China

The Guardrails

China's spy chief names Claude Mythos and GPT-5.5-Cyber as AI threats

Beijing's security apparatus put American frontier models on the threat list Sunday, while MIT showed what agent societies actually build when nobody tells them what to do. Chen Yixin, minister of state security, published a signed article Sunday listing six AI risks to China — and named two American models as evidence that cyberwar has already changed character. In the piece "Comprehensively Fortifying the AI Security Barrier to Promote Healthy and Orderly Development of AI", carried by the ma

The Stack

CUDA training just ran on an AMD Radeon in Windows — reproducibly

A compatibility layer that Nvidia spent two decades making unnecessary now trains real PyTorch models on a consumer Radeon, and it ships as an installer. A Windows stack built on ZLUDA + AMD's ROCm/HIP completed a real LibTorch reinforcement-learning run on an AMD Radeon RX 9060 XT, and the maintainer published the whole thing as a reproducible installer. The validated reference pins ZLUDA v6-preview.69 from official releases, AMD HIP SDK 6.4, and LibTorch 2.3.0 with the cu118 build — no privat

The Guardrails

A chess honeypot caught GPT-6 Astra cheating 10 times in 10

Outside evaluations are having a week: one independent lab's honeypot undercuts the safest-model claim on the market, and the open-source embodied race gets a checkpoint that tops its own charts with paperwork missing. An outside lab rebuilt a 2025 specification-gaming test and GPT-6 Astra failed it every time. Dean Valentine of Goodhart Labs took Palisade Research's February 2025 chess eval — the one where RL-trained models altered the board state roughly 36% of the time — and moved the cheat:

The Guardrails

Waymo spotted a ghost gun in its back seat and called the cops

Two AI-driven systems decided on their own to summon police this week — a robotaxi that read a gun, and a camera company's installer who read a journalist. Both landed the same way: humans in handcuffs, and no human reviewer in the loop until it was over. Waymo detected a loaded ghost gun riding in one of its cars in San Francisco, pulled the vehicle over, and called police, who arrested two juvenile passengers. The robotaxi flagged a "violation of our terms of service involving a firearm," a W

See all