Latest

GigaChat 3.5 Reasoning lands under MIT — and beats DeepSeek on math

The Frontier

GigaChat 3.5 Reasoning lands under MIT — and beats DeepSeek on math

Sber's GigaChat team has open-weighted a 432-billion-parameter reasoning model under an MIT licence, and the numbers put it within touching distance of DeepSeek V4 Flash. GigaChat 3.5 Reasoning runs 28 billion active parameters, trains natively in FP8, and mixes multi-head latent attention with GatedDeltaNet linear-attention layers — a hybrid aimed squarely at long-context cost — with a 262K-token window and three speculative-decoding heads bolted on. The release includes FP8 weights for inferen

Visa, Mastercard and Ant want one trust layer for AI agents

The Arena

Visa, Mastercard and Ant want one trust layer for AI agents

Two infrastructure bets landed this morning: the payments industry agreed on how to tell a good shopping agent from a bad one, and the US Army bought navigation that doesn't need satellites. Ant International, Visa and Mastercard said Thursday they will build a common "Know Your Agent" interoperability framework — one identity and behaviour check for AI agents that spend money, recognised across card networks, wallets, agent platforms and marketplaces. The three companies have each shipped thei

Mistral and Cloudera turn 30 exabytes of enterprise data into owned models

The Arena

Mistral and Cloudera turn 30 exabytes of enterprise data into owned models

Two stories about who actually controls the intelligence layer landed within hours of each other this morning: a European model lab wiring itself into the world's largest pools of enterprise data, and a Paris startup raising serious money to build an architecture that isn't a language model at all. Mistral and Cloudera announced a partnership today that puts Mistral's models inside Cloudera's customer-managed data environments — roughly 30 exabytes of it. The pitch is sovereignty in the literal

Amap's world model builds a 3D city in 10 minutes on one GPU

The Frontier

Amap's world model builds a 3D city in 10 minutes on one GPU

Alibaba's mapping arm used its Street Stars anniversary to make a bigger claim: it now has a world model that generates explorable 3D space, not just renders it. Two smaller stories round out the hour — a computer-vision benchmark that found audio evidence beats extra parameters, and a fresh report that reasoning state can still be replayed into another model. Alibaba's Amap released ABot-Earth 0.7, a 3D-native world model that turns a single satellite image or a text prompt into a kilometer-sc

How to — decide what an AI agent may do

The Guardrails

How to — decide what an AI agent may do

You're about to give an assistant access to something — a repo, a mailbox, a database, a shell. Twenty minutes of thinking now decides whether that ends in a useful afternoon or in an incident report. This is the routine: name the job, grant the minimum, cap the damage, and never let the model be the thing that says no. The reason this needs a routine is that the failure isn't exotic. An agent with your inbox and a send button doesn't need to be hacked to hurt you; it only needs to be persuaded

Deep Dive — The slowdown call moved from blogs into the building

The Guardrails

Deep Dive — The slowdown call moved from blogs into the building

In the first week of September, the case for slowing frontier AI down was made by a departing researcher, a chief scientist's blog post, and a think tank's essay. The people making it were insiders, but they were speaking from the edge of the industry — the person walking out, the essay published on a Saturday, the policy paper written by people who don't train models. In the last 48 hours the argument stopped being a side argument. Evan Hubinger, who runs alignment science at Anthropic, put a

Open Source Radar — September 10: agents get their hands on physical space

The Stack

Open Source Radar — September 10: agents get their hands on physical space

Today's board finally leaves the terminal. Three of the five hottest repos are about agents producing geometry you can build, print, or move a wall through, and the other two are about giving a whole team — or a whole desktop — a shared agent setup. Pascal Editor (TypeScript, 23,200 stars, MIT) — An open-source 3D building editor, built on React Three Fiber and WebGPU, that runs in the browser or from a local CLI and exposes the scene to agents over MCP. It's the mid-rise version of the "agent

Envision's Tianji is now the best weather model nobody in the West talks about

The Frontier

Envision's Tianji is now the best weather model nobody in the West talks about

Three stories from Shanghai's Bund Summit and the Chinese feeds this morning: an energy company's weather model just took the crown on the field's most-cited benchmark, Alipay's assistant started booking real travel, and a Nobel laureate put a number on what AI is worth. Envision's Tianji weather model entered Google's WeatherBench 2 for the first time this week and landed in the top tier alongside Google's own WeatherNext 2 — beating it on wind speed and precipitation, the two variables the en

AI 101 — What is HBM (high-bandwidth memory)?

The Stack

AI 101 — What is HBM (high-bandwidth memory)?

HBM is the stacked memory package that an AI accelerator's processor can read from and write to many times faster than normal RAM — and right now it is the single most expensive, scarcest part of the AI buildout. If you have read that the memory shortage is raising chip prices, throttling roadmaps, or forcing startups to rethink fabs, HBM is the "memory" in that sentence. This is the ten-minute version. It matters today for one plain reason: every frontier AI chip depends on it, and there is no

China's AI chip prices jump 50% as the memory shortage bites

The Arena

China's AI chip prices jump 50% as the memory shortage bites

A memory shortage is now setting the price of Chinese AI compute, and California just created the job of auditing AI. Both stories are about the same thing: inputs and oversight that nobody can source cheaply. Huawei has raised the indicated price of its Ascend 950DT accelerator to more than 250,000 yuan — about $37,000 — a jump of 20% to 50% from quotes it gave customers just two months ago. Reuters reported the increase on Thursday, citing three people familiar with the pricing; Cambricon has

DeepSeek V4.1 Flash ships — and retires V4 Pro next Monday

The Frontier

DeepSeek V4.1 Flash ships — and retires V4 Pro next Monday

DeepSeek's mid-morning release isn't just a new checkpoint. It's a new architecture, an open-weights drop, and a forced migration for anyone still calling its flagship — all in one announcement. DeepSeek released V4.1 Flash on Wednesday, the first model in a new architecture family and the first with native visual understanding. The headline number is 552 billion parameters, but the point is how few of them run: a new causal encoder–decoder design activates just 8 billion parameters while readi

DeepSeek preps a STAR Market IPO — and cuts prices 60% the same day

The Arena

DeepSeek preps a STAR Market IPO — and cuts prices 60% the same day

China's most-watched AI lab spent Wednesday doing two things that pull in opposite directions: quietly hiring the bankers for a Shanghai listing, and publicly retreating from the price increases it imposed three weeks ago. DeepSeek has hired CITIC Securities to prepare an IPO on Shanghai's STAR Market, with the listing process expected to begin within the year, according to people familiar with the matter. The Hangzhou lab is also working with accounting firms to close its books by the end of D

See all