Latest

Sugon validates GPU-direct RDMA across a 100,000-card AI cluster

The Stack

Sugon validates GPU-direct RDMA across a 100,000-card AI cluster

The most important AI infrastructure news out of China this week is not a model — it is the plumbing underneath one. Sugon says IBGDA, a networking technique that lets GPUs drive the network card directly instead of routing every transfer through the CPU, has completed full-scale validation on a cluster of 100,000 accelerator cards. The company's token-acceleration stack, built on its scaleFabric RDMA fabric, marks the first at-scale deployment of InfiniBand GPUDirect Async inside China — filli

Claude power users sue Anthropic over '5x' and '20x' fine print

The Everyday

Claude power users sue Anthropic over '5x' and '20x' fine print

Anthropic is facing a refiled class action over Claude Max pricing — and the plaintiffs' side is staffed like an FTC case, which is exactly what makes it dangerous. A refiled class action accuses Anthropic of overselling Claude's Max plan with "usage multipliers" that don't mean what buyers think they mean. The Max plan costs $100 a month for "5x" the usage of the $20 Pro plan, or $200 for "20x" — but the multipliers apply only within five-hour sessions that are themselves capped by a weekly ce

Open Source Radar — September 11: skills go open, models go home

The Stack

Open Source Radar — September 11: skills go open, models go home

The open-source story today is about ownership: who holds the model, who holds the agent's instructions, and who keeps the knowledge after the chat ends. Five picks from the daily board. Colibrì (C, 27,632 stars) — The most ambitious project on the board: an inference engine, in pure C with zero dependencies, that runs frontier mixture-of-experts models up to 2.8 trillion parameters on hardware you already own. The trick is treating your SSD, RAM, and VRAM as one memory hierarchy — the always-n

Fields Medalist founds institute to prove AI safe — like cryptography

The Arena

Fields Medalist founds institute to prove AI safe — like cryptography

Two kinds of trust shipped today: a Fields Medalist building the mathematics of "safe", and two Chinese platforms engineering agents that can explore without breaking things. **Jacob Tsimerman, who won math's Fields Medal in July, is founding an institute whose whole premise is that AI safety has no shortcut yet — and he wants to build one.** The Mathematical AI Safety Institute (MAISI), reported by The New York Times, will start research in January 2027 in the San Francisco Bay Area with 10 to

Alipay's payment robot dog went shopping — agentic commerce gets its first legs

The Arena

Alipay's payment robot dog went shopping — agentic commerce gets its first legs

Shanghai, Inclusion·The Bund Summit: Ant Group's Alipay moved AI payments from "agents can pay" to "agents can pay trustworthily", and the demo that stole the show has four legs. **Alipay put a payment rail inside a robot dog — and a trust layer around every AI transaction.** At the 2026 Inclusion·The Bund Summit in Shanghai, Ant Group CEO Han Xinyi said the "ChatGPT moment" of agent commerce has arrived, and Alipay backed the claim with a full stack: three new "AI Collect" monetization rails (

Deep Dive — OpenAI's Agents API: renting out the machinery behind Codex

The Stack

Deep Dive — OpenAI's Agents API: renting out the machinery behind Codex

OpenAI spent this week handing out its plumbing. On September 10, the company put the Agents API into public beta: a managed service that exposes the same harness and infrastructure behind Codex — context management, tool use, multi-agent orchestration — through a single API call. As we covered in our brief on the launch, developers specify a task, a model, tools and an environment; OpenAI spins up the Codex harness, fans work out to parallel subagents that each keep their own context, compacts

Loop a transformer twice, save 18% of training compute

The Frontier

Loop a transformer twice, save 18% of training compute

Two stories from the research world today: an architecture paper that puts hard numbers on one of this year's more interesting ideas in model design, and a peer-review system that is quietly rewriting the economics of who gets to submit. Re-running a transformer's middle layers twice can cut training compute by up to 18% — if you pay for the extra passes honestly. That's the finding of SMELT (Sparse MoE Transformer, middle layers Loop Twice), a paper from teams at Tsinghua University, ByteDance

AI 101 — What is LoRA (low-rank adaptation)?

The Stack

AI 101 — What is LoRA (low-rank adaptation)?

LoRA is a way of customizing an AI model by training a tiny add-on instead of rewriting the whole thing — the base model's weights stay frozen, and a small pair of "adapter" matrices carries the change. If you have read that someone "dropped a LoRA" to teach an image model a new style, or that a community fine-tune of an open-weight model was trained on a single GPU, LoRA is the technique underneath. It is the most common word in open-source model customization, and this is the ten-minute versio

GPT-6 Astra outscores every specialist at antibody prediction

The Frontier

GPT-6 Astra outscores every specialist at antibody prediction

The rule that AI-for-science requires dedicated models just broke. OpenAI's GPT-6 Astra topped a blind benchmark for predicting whether antibodies can become drugs — beating purpose-built systems without any domain-specific training. GPT-6 Astra scored 37.98 on the DDD Benchmark's antibody developability module, outperforming every specialized model in a blind evaluation built by Insilico Medicine. The benchmark integrates six sets of antibody experimental data and tests a question that has bur

Xiaomi open-sources CocktailASR-1 to pick one voice out of a crowd

The Frontier

Xiaomi open-sources CocktailASR-1 to pick one voice out of a crowd

Xiaomi just gave away one of the harder pieces of the voice stack. On September 11 the company released and open-sourced Xiaomi-CocktailASR-1, an end-to-end speech recognition model built to solve the cocktail party problem — transcribing one specific person's speech from an audio mix where several people are talking over each other, without a separation stage in front of the model. The design is unusually direct. You feed the model a short reference clip of the target speaker (one to four seco

MIT puts nearly a dozen AI chatbots under an election-year microscope

The Guardrails

MIT puts nearly a dozen AI chatbots under an election-year microscope

With the midterms two months out, MIT has opened a live audit of what the models actually tell voters — and the early numbers show the same question getting different answers depending on who asks. MIT has launched the LLM Election Observatory, a live dashboard tracking how nearly a dozen AI models — including Claude, ChatGPT and Gemini — answer election queries in the run-up to the November midterms, per a New York Times report on the project. The team has already run roughly 19,000 test queri

Moonshot AI files for Hong Kong IPO at $50B — days after Anthropic calls it out

The Arena

Moonshot AI files for Hong Kong IPO at $50B — days after Anthropic calls it out

The same week Anthropic accused Moonshot of secretly routing Kimi users to Claude, the Beijing lab quietly filed for what may be the most closely watched Chinese AI IPO in years. Moonshot AI has confidentially filed for a Hong Kong IPO that could raise $3 billion at a valuation near $50 billion, according to sources speaking to Reuters — likely the most anticipated near-term listing from a Chinese AI lab. The filing lands at a delicate moment: yesterday Anthropic's threat-intelligence report al

See all