Latest

AI 101 — What is a transformer?

The Frontier

AI 101 — What is a transformer?

A transformer is the neural network design almost every modern AI system is built on — an architecture that lets a model look at every word in its input at once and decide which other words matter for understanding it. That mechanism is called attention, and it is the reason AI went from clunky to conversational. The name comes from a 2017 paper by eight researchers at Google, titled "Attention Is All You Need." Before it, language models read text one word at a time, in order, like someone run

China's top court sets the rules of the road for AI deepfake lawsuits

The Guardrails

China's top court sets the rules of the road for AI deepfake lawsuits

China's Supreme People's Court put its judicial weight behind the "AI face-swap / voice-clone" problem on Monday, releasing a 24-article opinion that tells lower courts exactly how to rule on deepfake suits, AI hallucination claims, "doxxing-by-AI" cases, algorithmic price discrimination, autonomous-driving liability, model-training disputes, and open-source software. The opinion is the SPC's broadest attempt yet to translate the country's existing civil code into a working playbook for syntheti

HiDream-O1-Embodied tops RoboColiseum's robustness chart

The Frontier

HiDream-O1-Embodied tops RoboColiseum's robustness chart

A Chinese world-model lab's first move into embodied robotics lands on top of the toughest sub-score, while Alibaba ships a "multi-user workbench" that turns a sentence into a working group app. HiDream-O1-Embodied debuted at number one on the RoboColiseum robustness leaderboard, scoring 0.692 in the benchmark's most adversarial sub-track on its first submission. HiDream.ai (智象未来) framed the release as the "real-world" leg of a world-model strategy that already covers image (HiDream-O1-Image) a

Pachocki wants a slowdown — and OpenAI's own chart shows why it won't stick

The Guardrails

Pachocki wants a slowdown — and OpenAI's own chart shows why it won't stick

OpenAI published two documents on the same Sunday: one measuring how fast its own research is accelerating, and one from its chief scientist arguing that nobody should be going that fast. Read together, they accidentally document why the second one won't work. OpenAI's chief scientist says no lab has solved alignment well enough to keep scaling at full speed — and he wants voluntary slowdowns to become normal. In an essay published Sunday on OpenAI's site, Jakub Pachocki writes that "no lab has

iFlytek ships Spark X2.5, trained end-to-end on Chinese silicon

The Stack

iFlytek ships Spark X2.5, trained end-to-end on Chinese silicon

Two Chinese releases this morning bracket the same bet — that models and agent products can be built without American chips — plus a security firm's answer to the coding-agent mess we covered overnight. iFlytek released Spark X2.5, a 293B-parameter mixture-of-experts model with about 30B active, and the headline claim is not the size — it's the hardware. The company says the model was trained and is served entirely on domestically produced compute, continuing work on what it calls the hard prob

A blacklisted server giant shipped $5.6B of US AI gear through its Valley subsidiary

The Guardrails

A blacklisted server giant shipped $5.6B of US AI gear through its Valley subsidiary

The New York Times spent months on shipping manifests and came back with the cleanest map yet of how export controls actually fail: not at the border, but at the name on the door. Inspur, blacklisted by the US since 2023, moved at least $5.6 billion of advanced technology out of the United States between April 2024 and February 2026 — including more than $3 billion of Nvidia Blackwell servers — through a Silicon Valley subsidiary the government never listed. The company put on the Entity List w

A repo you only opened can run code through your AI coding agent

The Guardrails

A repo you only opened can run code through your AI coding agent

Two security stories worth your attention this morning, and neither is interesting for the reason the headlines suggest: both are about inputs we taught agents to treat as harmless. One is a git setting nobody audits. The other is the internet itself. Opening a suspicious folder is now enough to get code executed, and the agent's permission prompt never sees it happen. Security firm Manifold calls the pattern GitSpawn, and the mechanism is almost boringly ordinary. Coding agents gather context

Unitree's founder puts a number on robotics: two '80%' thresholds

The Arena

Unitree's founder puts a number on robotics: two '80%' thresholds

Three weeks after Unitree's Shanghai debut, founder Wang Xingxing used his first public speech since the listing to do something unusual for a robotics CEO on a big stage: he skipped the demo reel and named the metric that decides whether any of this becomes an industry. Wang Xingxing says the robotics boom starts when a robot handles ~80% of tasks in ~80% of unfamiliar places — and he puts that two to ten years out. Speaking at the 2026 World Robot Conference in Beijing on August 20, days afte

Tencent's Hy4 preview takes the top spot as Chinese models lead for 19 weeks

The Frontier

Tencent's Hy4 preview takes the top spot as Chinese models lead for 19 weeks

Two things are worth your attention this morning: a Chinese open-weight model just took the number-one slot on OpenRouter's weekly leaderboard by volume, and Washington is quietly drafting the agenda for the first dedicated US-China AI safety dialogue of Trump's second term. Both stories point the same direction — the center of gravity in deployed AI keeps moving, and the governance machinery is still catching up. Tencent's Hy4 preview is now the most-used model on OpenRouter, and it got there

OpenAI says its agents now do 3.1 days of research per human day

The Frontier

OpenAI says its agents now do 3.1 days of research per human day

Two posts from OpenAI landed within hours of each other on Sunday, and together they say something the company has spent a year avoiding saying out loud: the automation is working faster than the safety case for it. OpenAI says it has hit its "automated research intern" goal, and its research org now runs 3.1 agent-workdays for every human workday. In a post titled "Research acceleration: The view inside OpenAI," the company confirmed the target Sam Altman set in an October 2025 livestream — an

Huang declares 'AGI has arrived' — and Nvidia's chips are the footnote

The Arena

Huang declares 'AGI has arrived' — and Nvidia's chips are the footnote

Two days after OpenAI shipped Astra, the CEO who sells the silicon just put his own weight behind the AGI claim — and the pushback arrived the same evening. Also: the Nubia NaviX Ultra clears its last regulatory hurdle with Chinese memory inside. Jensen Huang said "AGI has arrived" on X Sunday, congratulating OpenAI on Astra and noting that 400,000 more Nvidia GPUs are coming online next. "From ChatGPT to o1 to Astra in 4 years," he wrote. The congratulation is not a neutral observation — Nvidi

See all