Latest

Baidu apologises in China's first AI defamation case — two years on

The Guardrails

Baidu apologises in China's first AI defamation case — two years on

A lawyer who searched his own name and found a criminal record that never existed finally got a written apology today. The reasoning behind the order is the part the industry should read. Baidu formally apologised in writing to Nanjing lawyer Li Xiaoliang on September 22, at the Nanjing Jiangbei New Area People's Court — the first time a Chinese court has forced a model operator to apologise for what its model said about a named person. The case began in September 2024, when Li searched his own

Open Source Radar — September 22: the self-hosted swap

The Stack

Open Source Radar — September 22: the self-hosted swap

Today's board is one argument told five ways: the paid per-seat product and the self-hosted substitute are now the same product, and one of them runs on hardware you already own. AutoClip (Python, ~8,600 stars, MIT) — Long podcasts, interviews, lectures and stream replays go in; a scored outline, a topic timeline and short clips with generated titles come out. It finds highlights from the transcript rather than watching the video, works from an existing subtitle file or transcribes locally, and

173 studies on AI in classrooms — and 80% tested university students

The Everyday

173 studies on AI in classrooms — and 80% tested university students

Norway just published the clearest look yet at what generative AI does to how children think. The honest answer is that nobody has run the study that would tell us. A review of 173 international studies finds a clean dual effect from AI in the classroom — and a hole where the evidence on children should be. The review, led by Hilde Storrøsæter at NTNU's Department of Teacher Education with colleagues from Czech universities, analyzed research published from 2022 through the spring of 2025 on g

Deep Dive — Alibaba wants to own the whole stack. Memory says no.

The Arena

Deep Dive — Alibaba wants to own the whole stack. Memory says no.

The most important number Alibaba put on stage in Hangzhou this week was not ten trillion. Alibaba opened its annual Apsara Conference on Tuesday with a plan that is unusually easy to restate, because the company said it in one sentence: it will train Qwen models of five trillion to ten trillion parameters, run them on its own accelerators wired into clusters of up to 500,000 cards, and operate more than 20 gigawatts of data centre capacity worldwide by 2032. Chief executive Eddie Wu gave the k

Glasshouse ships an agent-memory benchmark with no headline number

The Stack

Glasshouse ships an agent-memory benchmark with no headline number

A memory-infrastructure vendor has published the test that grades its own industry — and written down, in advance, the ways it could cheat it. Wontopos has put the files for glasshouse v0.1 into the open: a long-term memory benchmark for AI systems whose central design decision is that it will not produce a single score. The corpus is 2,847 questions over a conversation running to 1.97 million tokens, in 10 languages, with 50 uncaptioned photographs, and every axis is reported on its own becaus

Kimi Delta Attention gains 2D rotations from a signed gate

The Frontier

Kimi Delta Attention gains 2D rotations from a signed gate

A two-number change to one of the open field's most widely used attention modules buys a provable class of expressivity at no extra cost — and the authors are unusually clear about what it does not buy. Allowing Kimi Delta Attention's channel-wise gate to go negative — into [−1, 1] — and its delta-rule coefficient up to 2 lets the module realise 2D rotations, giving it the state-tracking expressivity of DeltaProduct-2 while its transitions stay diagonal-plus-rank-one and non-expansive. The pape

AI 101 — What is mechanistic interpretability?

The Guardrails

AI 101 — What is mechanistic interpretability?

Mechanistic interpretability is the work of reverse-engineering what actually happens inside a neural network — finding the specific internal parts responsible for a specific answer, rather than inferring everything from the model's behaviour. "Mechanistic" is the load-bearing word: the deliverable is a mechanism, a step-by-step account of the computation, not a description of what the model tends to do. Why it matters right now The cheapest way anyone audits a frontier model is by reading it

Orbital inference starts selling: Yusuan Xinghe ships in October

The Frontier

Orbital inference starts selling: Yusuan Xinghe ships in October

In-orbit AI stopped being a demonstration this week. One company put a launch date and a price list on it, and a second put the hardware on orbit to make it work. A Guangzhou startup says its first compute satellites launch in October, and customers have already paid in advance for the models they will run. Yusuan Xinghe, incubated by listed smart-transport firm PCI Tech (Jiadu Technology 600728), used the second Guangdong AI Application Matchmaking Conference in Guangzhou on Tuesday to unveil

OpenAI and Anthropic put their copyright asks to Australia in writing

The Guardrails

OpenAI and Anthropic put their copyright asks to Australia in writing

Australia's copyright fight stopped being a rumour this morning: the two labs asking Canberra to rewrite its law have now filed their arguments where anyone can read them. OpenAI and Anthropic have filed submissions to parliament's Joint Select Committee on Artificial Intelligence setting out what they want Australia's copyright regime to be — and both tie it to whether they train and build in the country. OpenAI asked for a "balanced copyright framework" that lets models learn from publicly av

Fastcrawl wants to be the web layer for AI agents — and it is priced like a utility

The Stack

Fastcrawl wants to be the web layer for AI agents — and it is priced like a utility

Every AI agent that needs to read the web runs into the same wall, and it is not the model. It is the page. A modern news homepage arrives as 400 to 900 kilobytes of markup wrapped around a few thousand words of actual text — navigation, adverts and script tags around the part worth reading. Handing that to a model wastes tokens and produces worse answers than the same text would have. So the operator either builds a scraping stack — headless browsers, session handling, retries, anti-bot evasion

Tencent ships Hy Image3.5 at $0.024 an image — parity claim untested

The Frontier

Tencent ships Hy Image3.5 at $0.024 an image — parity claim untested

Tencent's image line has been on a two-month clock since it rebuilt the stack in February. The latest tick is a preview model whose rate card does more work than its benchmark does. Tencent released Hy Image3.5 preview on Tuesday, pricing a 2K image at 0.15 yuan on its domestic cloud API and $0.024 on the international endpoint — about a quarter of what the model it claims to match charges. The model does text-to-image and image-to-image, takes up to five reference images per call for identity

Alibaba plans a 10T-parameter model and a 500,000-chip cluster

The Arena

Alibaba plans a 10T-parameter model and a 500,000-chip cluster

Alibaba's cloud conference in Hangzhou opened this morning with a three-part commitment that is harder to wave off than the usual keynote arithmetic: a model bigger than anything trained so far, the chip to run it, and the power to feed both. Huawei spent the same week putting a coding model inside its own operating system, and a small project on the local-AI forums quietly showed that a model's refusals can be switched off without touching a single weight. Alibaba says its next Qwen models wil

See all