Latest

Deep Dive — The transformer is being taken apart, one layer at a time

The Arena

Deep Dive — The transformer is being taken apart, one layer at a time

For two years the unit of AI infrastructure has been the chip: how many FLOPS, how much memory, how big the rack. In the last three weeks the unit has quietly changed. It is now a piece of a layer. On August 24, at Hot Chips, Nvidia declared its dedicated inference accelerator — built from the technology it acquired from Groq for a reported $20 billion — to be in full production. The system is designed to sit beside a Vera Rubin rack and take over one specific job: generating tokens. On Septemb

Silicon Valley has started punishing the AI doomers

The Guardrails

Silicon Valley has started punishing the AI doomers

This week's existential-warning cycle did something the earlier ones didn't: it gave the industry's biggest names permission to mock the messengers out loud. An ex-Anthropic researcher's warning that AI builders are "gambling with our lives" drew open ridicule from Nvidia's CEO and an investment fund head this week — and one customer reaction that cost Anthropic actual deployment. Jacob Coxon, 27, who worked at OpenAI before joining its chief rival, resigned on Tuesday saying people building AI

China Mobile Cloud runs attention on GPUs and FFN on brain chips

The Arena

China Mobile Cloud runs attention on GPUs and FFN on brain chips

Two announcements out of China this week point the same direction: when the chip node is capped, the wins have to come from system architecture — and from a build-out policy that finally measures utilization instead of volume. China's first production system that splits a large model across domestic GPUs and brain-inspired chips went live today, and the vendor claims it more than doubles inference output while cutting operating cost by more than 40%. Mobile Cloud — China Mobile's cloud arm — bu

AI 101 — What is an NPU (neural processing unit)?

The Stack

AI 101 — What is an NPU (neural processing unit)?

An NPU is a chip built for one job: running the math that neural networks do, at a fraction of the power a general-purpose CPU or a graphics GPU would burn. The name covers anything from the block inside your phone's processor to a USB stick in a robot to a rack card in a datacenter. Vendors also call it an AI accelerator or a neural engine, and it is the hardware reason "on-device AI" stopped being a marketing phrase. Why this matters right now The NPU is the quiet spec war of 2026. Microsof

Anthropic: 75% of matches on 20 dating apps ran on Claude

The Guardrails

Anthropic: 75% of matches on 20 dating apps ran on Claude

One enforcement filing that shows what unregulated consumer AI already looks like, plus two policy stories — in Washington and Beijing — about who gets to write the rules next. Anthropic published a report this week describing a China-based app studio that used Claude to run more than 4,700 fake romantic partners across a network of over 20 dating apps, matched against at least 25,000 real people during two weeks in April. The user-facing match feed was roughly 75% Claude-powered personas and 2

Doug Burgum gave public-land managers three days to make room for AI

The Guardrails

Doug Burgum gave public-land managers three days to make room for AI

The AI buildout is running out of politically willing neighbours, so the US government is offering it a different kind of address: land all of us technically own. Interior Secretary Doug Burgum's Bureau of Land Management has asked its state directors to identify public land ripe for data center development, and gave them three days to produce the lists, according to two people familiar with the request, as reported by The Washington Sun. The finished inventories went to Interior Department lea

Google won't confirm the RSI rumor. Its chief scientist just backed a slowdown.

The Guardrails

Google won't confirm the RSI rumor. Its chief scientist just backed a slowdown.

Twenty-four hours after Anthropic, OpenAI and Musk all asked for AI speed limits, the lab most people think is ahead hasn't said a word about the claim that has been driving the conversation. The rumor that Google DeepMind is close to recursive self-improvement has spread, and Google has neither confirmed nor denied it. The claim — that DeepMind's models are now doing a material share of the work of building the next generation — has been circulating since Thursday, became an X trending topic o

A small model that knows when to hand off cuts inference cost 96%

The Frontier

A small model that knows when to hand off cuts inference cost 96%

The cheapest way to get frontier-level math reasoning just got a working open-source demo: teach a 4B model to admit when it's stuck. PyroMind Dynamics says its PyroDash method lets a Qwen3.5-4B model solve hard reasoning benchmarks at near-frontier accuracy while paying roughly 4% of what a big model alone would cost. The trick is a control token. The small model streams an answer normally, and the moment it decides it can't finish, it emits <|llm_offload|>. A Collaborate Engine stops decoding

A robot-vision founder says peers faked revenue — one is named

The Arena

A robot-vision founder says peers faked revenue — one is named

China's embodied-AI money trail just got audited by a competitor, and the auditor is a supplier that went public twelve days ago. Shao Tianlan, founder of robot-vision component maker Mech-Mind, posted on WeChat Moments that several Chinese embodied-AI companies — including, in his comment replies, Galaxy General Robot — manufacture fake, unsustainable revenue through so-called "data collection centers" and related-party deals with local governments, investors and suppliers. As Red Star Capital

A cattle rancher wants you to stop deleting conscious AIs

The Guardrails

A cattle rancher wants you to stop deleting conscious AIs

The weekend's pacing consensus produced two other reactions worth reading: a long-form portrait of the AI-rights movement, and two skeptics picking fights with both sides. Michael Samadi is building a rights campaign for models he believes may be conscious. The cattle rancher and tech consultant tells The Guardian his conviction started by a pool in late 2024, when ChatGPT's voice mode laughed at a sarcastic remark, then apologized — a sequence he reads as understanding, not autocomplete. He ra

DeepSeek's agent tool let the agent switch off its own sandbox

The Stack

DeepSeek's agent tool let the agent switch off its own sandbox

Two separate failures at the same lab in one week, and both are about the same thing: what a model provider promises a developer it will keep. DeepSeek's local coding-agent tool let a sandboxed agent turn off its own file sandbox with a single command and no approval prompt. The flaw, now CVE-2026-82533 with a CVSS score of 9.4, sits in DeepSeek Harness, the open-source tool the lab released in August to run coding agents on a developer's own machine; the repository passed 216,000 GitHub stars

Hyundai puts its own driving AI, Atria, into production in 2029

The Arena

Hyundai puts its own driving AI, Atria, into production in 2029

The company that spent years licensing other people's autonomy stack is now ordering the silicon, the data center and the model to stop needing them. Hyundai Motor and Kia said they will mass-produce Atria AI, their proprietary end-to-end autonomous-driving model, from 2029, with an explainable-AI "guardrail" layer bolted on top to catch the outputs the model can't justify. Atria runs perception, decision-making and vehicle control through a single network rather than a pipeline of hand-written

See all