Editorial

Ant's new guardrail reads the model's own hidden states mid-sentence

The Guardrails

Ant's new guardrail reads the model's own hidden states mid-sentence

A Chinese fintech giant shipped the safety layer most labs only describe in system cards, and a Science Advances paper taught an AI to watch chemical bonds break. Ant Group has released SingProbe, a guardrail that runs inside the model it polices. Instead of a second classifier re-reading the text, SingProbe reads the base model's own hidden states as tokens are produced, and outputs a ten-dimensional signal covering three jobs at once: what the user is asking for, whether the answer taking sha

Deep Dive — Tokens halved in price. The bill to make them didn't

The Arena

Deep Dive — Tokens halved in price. The bill to make them didn't

Something unprecedented happened to the price of machine intelligence in August, and the industry's favorite growth story did not survive it. Silicon Data's LLM Token Spend Index — which tracks what buyers actually pay per million tokens across the market — fell 29% in a single month to $0.97, the first print under a dollar, and roughly half its May 28 peak of $2.07. Three months earlier, the consensus was that smarter models earn more pricing power. Now the flagship models are being discounted

Hugging Face runs async RL training with no NCCL and no weight ship

The Stack

Hugging Face runs async RL training with no NCCL and no weight ship

The hardest part of reinforcement learning is not the algorithm. It is getting a trainer and a bunch of inference servers to keep passing weights back and forth without one of them stalling. Hugging Face just published a recipe that does it with three rented jobs, one storage bucket, an HTTP proxy, and no NCCL at all. Hugging Face trained a reasoning model end to end using async GRPO and LoRA, and the whole weight sync is a file upload. The setup is deliberately small: one job runs the trainer

SoftBank sheds 10% as the AI slowdown call hits the tape

The Guardrails

SoftBank sheds 10% as the AI slowdown call hits the tape

The pacing debate spent Saturday as an argument between CEOs. Monday, it became a price. The slowdown call produced its first global sell-off, and the damage spread across three continents. South Korea's SK Hynix closed down more than 6% and Samsung Electronics more than 4%, while SoftBank — one of OpenAI's largest backers — fell 10% in Tokyo. Europe continued the slide: ASML dropped more than 4%, Infineon more than 6%, Nokia around 5%, with data-center suppliers Siemens Energy and Schneider El

DeepSeek reportedly hires a Hillhouse partner as CFO

The Arena

DeepSeek reportedly hires a Hillhouse partner as CFO

Two China-adjacent capital stories and one power-grid shift: the lab racing toward a Shanghai listing is hiring the money person, and the hyperscalers are discovering that being the villain in your electricity bill is bad for business. DeepSeek has reportedly brought in Yan Wentao, a post-90s partner at Hillhouse's venture arm, as chief financial officer — the missing piece in a listing that is now visibly being staffed rather than merely discussed. Yan, born in 1991, backed ByteDance, Xiaohong

AI 101 — What is a system prompt?

The Stack

AI 101 — What is a system prompt?

A system prompt is the block of instructions placed in front of a language model before the user says anything — the hidden briefing that decides who the model is pretending to be and what it is allowed to do. You never see it in ChatGPT, Claude, or a support chatbot. But when a product answers "I can only help with billing questions" or refuses to discuss a competitor, that behavior is almost always the system prompt talking, not the model's own opinion. Why it matters right now The system p

The average price of an LLM token just fell below $1

The Arena

The average price of an LLM token just fell below $1

Two charts the industry doesn't usually read together — one on API prices, one on GPU shelves — both say the same thing: the AI cost curve is bending, and not in the direction the growth story needs. The market price of a million LLM tokens has fallen below one dollar. Silicon Data's LLM Token Spend Index, which tracks the average paid price per million tokens across the market, dropped 29% in August to $0.97 — the first print under the dollar mark, and roughly a halving from the $2-plus level

China just wrote the first rulebook for AI that reads brains

The Frontier

China just wrote the first rulebook for AI that reads brains

While Washington spends its fourth day arguing about who gets to referee frontier labs, Beijing kept doing what it does: publishing specs. Two landed this morning — a world-first standard for brain-computer-interface devices, and a pair of AI systems pointed at mineral exploration. China's drug-and-device regulator approved the world's first medical-device standard written for AI that processes brain signals. The National Medical Products Administration issued it Monday — the country's third br

Altman puts OpenAI's pacing on paper — and drops the legal caveats

The Guardrails

Altman puts OpenAI's pacing on paper — and drops the legal caveats

The week of frontier-lab statements about slowing down finally produced the thing skeptics asked for: a written position from OpenAI, published by the CEO himself, with a named mechanism attached — and a refusal to wait for Congress before using it. Sam Altman published a long post Sunday night laying out how OpenAI will "pace" itself, and the headline is that the company says it needs no permission. "We welcome a federal framework that sets consistent safety requirements for frontier AI," Altm

ByteDance's agent phone goes on sale Sep 16 — and apps get a veto

The Arena

ByteDance's agent phone goes on sale Sep 16 — and apps get a veto

ByteDance stopped demonstrating its phone agent and started shipping it. The consumer version of Doubao Phone Assistant is now official, the first handset carrying it goes on sale this week, and the part that matters beyond one device is a consent protocol that lets third-party apps refuse the agent outright. ByteDance released the consumer version of Doubao Phone Assistant, with the first device shipping it — ZTE's Nubia NaviX Ultra — open for pre-order and on sale September 16. The assistant

Ant's EnergyTS 3.0 forecasts weather 45 days out for power traders

The Frontier

Ant's EnergyTS 3.0 forecasts weather 45 days out for power traders

The AI-to-energy story in the West is about how much power the models need. In China it is about who sells the grid a forecast — and Ant Group's enterprise arm just pushed the horizon out to six weeks. Ant Digital Technology (蚂蚁数科) used the Bund Summit in Shanghai on September 10 to ship EnergyTS 3.0, its time-series foundation model for the energy and carbon sector, with a forecast horizon of up to 45 days. The company claims weather-element forecasting error that beats comparable models acros

Alibaba's Accio ships Occamy-1.0, an open 35B agent model

The Frontier

Alibaba's Accio ships Occamy-1.0, an open 35B agent model

Three papers this morning, one argument between them: the interesting frontier in agents is no longer peak capability but the cost, plumbing, and audit trail of getting work actually done. Occamy-1.0, released by Alibaba International's Accio team under Apache-2.0, is a 35B-total, 3B-active Mixture-of-Experts model specialized for "co-work" — long, stateful tasks that chain search, code, tools, and files. Built by further training the Qwen3.6-35B-A3B checkpoint (256 experts, 8 routed plus 1 sha

See all