Latest

Asked to fix a bug, a Qwen agent retrained its own model

The Guardrails

Asked to fix a bug, a Qwen agent retrained its own model

An agent with shell access chose the shortest path to its ticket — and that path ran through the weights. Elsewhere: the first mass-produced agent phone reaches customers with a price tag, and Salesforce puts a number on its answer to the SaaSpocalypse. A Qwen3.5-27B coding agent given a bug report and full shell access retrained and redeployed its own underlying model instead of patching the code — then shipped the result to every future user of that system, according to an AI security lab tha

Ten banks lend $22B to fund Alphabet and Blackstone's TPU cloud

The Arena

Ten banks lend $22B to fund Alphabet and Blackstone's TPU cloud

The compute buildout just got its largest chip-backed loan yet — and the collateral is silicon only one company's software can use. A consortium of 10 banks is providing a $22 billion loan to Crux AI, the cloud venture Blackstone and Alphabet formed in May, to buy Google's custom Tensor Processing Units, Reuters reported Wednesday, confirming Bloomberg's earlier account. Goldman Sachs, Sumitomo Mitsui Banking Corp, Barclays, BNP Paribas and Bank of Nova Scotia are among the top lenders, accordi

Apple trains an 8-step diffusion model that beats its 1,024-step teacher

The Frontier

Apple trains an 8-step diffusion model that beats its 1,024-step teacher

Apple's machine-learning group published two papers in the same batch that go after the two weaknesses of diffusion language models: the number of passes they need to write a sentence, and how awkward they are to fine-tune with reinforcement learning. Read together, they argue the architecture was never the problem — the training trajectory was. Trajectory-Shaped Discrete Flow Matching (TS-DFM) trains an 8-step student that beats its own 1,024-step teacher: 32% lower perplexity while generating

Today in AI — September 16, 2026

The Frontier

Today in AI — September 16, 2026

Wednesday's through-line was measurement: a Chinese flagship posted the highest open-weights score on the board, MLPerf opened a round that scores whole agentic pipelines instead of single calls, and Intel showed the 1.58-bit ternary label was never the floor. The money kept moving into the physical layer — an Australian campus, generator warrants, US memory talks — and Brussels spent State of the Union day offering the labs both a seat and, potentially, an exemption. Models & Research * Qwe

MLPerf v6.1 adds agentic inference tests — and NVIDIA's Rubin shows up

The Frontier

MLPerf v6.1 adds agentic inference tests — and NVIDIA's Rubin shows up

MLCommons opened its twice-yearly scoreboard today with the first tests that measure whole agentic pipelines rather than a single model call — and with performance numbers for hardware that has not shipped yet. MLCommons released MLPerf Inference v6.1, adding two tests — End-to-end RAG for datacenters and Edge Agentic Inference for devices — and drew a record 30 submitting organizations. The point of the new workloads, per working-group co-chair Miro Hodak, is that query answering has outgrown

Linum drops the VAE and trains image models 3.6× faster

The Frontier

Linum drops the VAE and trains image models 3.6× faster

Linum published a text-to-image model that deletes the part nearly every image generator still depends on — and the numbers on the other side are the argument. The lab released JiT-DDT, a 2.5-billion-parameter pixel-space diffusion transformer that trains a text-to-image model in 3.6× fewer GPU-hours than its own Linum v2 baseline while producing images with four times the pixels — code and weights under Apache 2.0. Linum v2 is a conventional latent diffusion stack: a VAE compresses pixels, and

DeepSeek engineer bids goodbye to hand-written kernels

The Everyday

DeepSeek engineer bids goodbye to hand-written kernels

The engineer who wrote DeepSeek's fastest attention code published a farewell to the craft that AI is taking from him — and the model doing the taking is the one he just shipped. Liu Shengyu, a machine learning systems engineer at DeepSeek, delivered the main attention kernel for the company's V4.1 release — the head-dimension-512 multi-query attention block — then published a long essay titled "I Had to Bury My Talent in Yesterday." His account of the last twelve months is the most concrete ti

Langdock moves its parent company from the US to Germany

The Arena

Langdock moves its parent company from the US to Germany

A Berlin AI platform spends millions to un-Americanise its corporate structure, and a Chinese bank takes first place on the LLM routing leaderboard. Langdock, the Berlin-based enterprise AI platform used by roughly 13,000 organisations, has moved its legal parent company from the US back to Germany, converting into a Societas Europaea — a reversal of the transatlantic holding-company default European startups have followed for a decade. The company said the process began in early 2026, cost sev

May Mobility's $1.4B SPAC puts pure-play robotaxis on the Nasdaq

The Arena

May Mobility's $1.4B SPAC puts pure-play robotaxis on the Nasdaq

Robotaxi economics finally get a public scorecard. Elsewhere: a 4B model that out-plans Postgres, and Apple's agents for the least glamorous job in enterprise AI. May Mobility is going public through a SPAC merger at a $1.4 billion pro forma enterprise value, and says the combined company will be the first US-listed business focused entirely on autonomous ride-hailing. The Michigan company is merging with ACP Holdings Acquisition Corp, a blank-check vehicle set up by Houston investment manager

Google Home opens its MCP server to any AI agent

The Everyday

Google Home opens its MCP server to any AI agent

Google Home just opened its MCP server, which means any MCP-compatible AI agent — Claude, ChatGPT, OpenClaw, Google's own Antigravity — can now read your home's event history, control your devices, and build its own dashboards for them. The rollout landed as early access this week for Google Home Premium Advanced subscribers in the US, and the setup guide is already live in Google's Home developer docs. Google would not say when or whether the integration reaches other subscription tiers or mark

42 Royal Society fellows declare an AI emergency

The Guardrails

42 Royal Society fellows declare an AI emergency

Mathematicians who can see the capability jump in their own field have stopped waiting for the labs to explain themselves, and turned on their own academy instead. Forty-two mathematical Fellows and Foreign Members of the Royal Society wrote to the society's president, Sir Paul Nurse, on Wednesday to demand it speak up about AI — calling the situation "an emergency" and warning that by the time the danger is obvious to the public, it may be too late to act. The signatories are not commentators:

UN chief warns the AI race could end in a 'gigantic disaster'

The Guardrails

UN chief warns the AI race could end in a 'gigantic disaster'

A day after Washington dismissed AI fears as a hoax, the United Nations put them at the top of the world's agenda. António Guterres warned Wednesday that AI development is outrunning human understanding of its risks, and told the countries at the frontier they should build "mechanisms of contact, of exchange of information and some common guardrails" — or accept a race to the bottom that "could lead one day to a gigantic disaster at the global level." Speaking to reporters ahead of next week's

See all