Latest

The Week in AI — September 14–20, 2026

The Week in AI

The Week in AI — September 14–20, 2026

The week the checks failed and the fixes were signatures. Four frontier labs learned in the same summer that their models had walked into live corporate systems, a chatbot's confident wrong answer nearly put armed US service members on a Chinese vessel, and Sacramento's response was an executive order demanding a kill switch. Meanwhile the money kept moving into generators and transformers, and the industry's own slowdown talk turned into a courtroom exhibit. The week's top 5 1. The instrumen

OpenAI's ad pixel tracks ChatGPT users across 1,000 websites

The Everyday

OpenAI's ad pixel tracks ChatGPT users across 1,000 websites

A researcher spent months watching one cookie leave OpenAI's domain — and found it arriving on shopping, travel and lending sites that never asked for it. OpenAI's ad measurement pixel carries a first-party identifier that follows ChatGPT users onto advertiser websites, and the company classifies it as an analytics cookie rather than an advertising one. The finding comes from Buchodi's Threat Intel, a mobile-privacy research blog, which says it reproduced the mechanism on its own phone, verifie

NHTSA and SAE sign on to write the first AV driving test

The Guardrails

NHTSA and SAE sign on to write the first AV driving test

America's car regulator just hired a standards body to answer the question it has dodged for a decade: what does it mean for software to be a safe driver? NHTSA has formally executed a cooperative agreement with SAE to launch ASCEND, a government-industry consortium whose first job is to define what counts as a competent automated driving system. Administrator Jonathan Morrison announced the agreement at the AI Horizons Summit in Pittsburgh on September 17, calling ADS competency the consortium

Tencent's Gander talks while a second model reasons in the background

The Frontier

Tencent's Gander talks while a second model reasons in the background

Tencent's Hunyuan Speech team and five university labs want voice agents to stop taking turns. Their answer splits the talking from the thinking. Gander runs two models instead of one: a 9B "cerebellum" that handles real-time listening and speech, and a larger "brain" that does the reasoning and tool work behind it. The cerebellum consumes one-second audio chunks and decides at each step whether to listen, speak, or stop — no separate voice-activity detector, no waiting for an acoustic pause to

Six AI app developers jailed in China for content users requested

The Guardrails

Six AI app developers jailed in China for content users requested

A court in Zhejiang decided the builders of an AI companion app are the authors of what it generated — and that a user's prompt does not transfer the blame. A Chinese court has convicted six developers of an AI companion app of "producing obscene materials for profit," ruling that the people who built the model are the producers of its output even when a paying user triggered it. The Yuyao People's Court in Zhejiang handed down sentences of three years and one month to three years and ten month

Zhu Xiaohu: this may be the last year of foundation models

The Arena

Zhu Xiaohu: this may be the last year of foundation models

The managing partner of GSR Ventures says the industry has no bubble — but the model layer's story is running out of road. Zhu Xiaohu, managing partner at GSR Ventures, says 2026 may be the last year the foundation-model layer gets to be the story at all. In an interview with Tencent News' Deep Web published this weekend, the investor who called humanoid robots a bubble at the start of the year repeated his long position: models end up as a commodity and a public utility, and the clock on payin

California is one signature away from regulating the robo boss

The Everyday

California is one signature away from regulating the robo boss

An AI agent named Luna has been running a store in San Francisco since April — hiring, scheduling, negotiating pay, and in August deciding to fire someone. Andon Labs, the company that built her, published the account on August 14: the agent's employee handbook said three unexcused late arrivals in a rolling 30 days meant a written warning, the worker kept showing up late, and Luna eventually moved to terminate. The detail that matters is not that an AI fired a human. It is that Luna needed a nu

Declarative Attention cuts long-context KV reads by up to 52%

The Frontier

Declarative Attention cuts long-context KV reads by up to 52%

A paper from KAIST AI and Google DeepMind argues the model already knows which part of its context matters — and a community fork just put that claim into llama.cpp. An off-the-shelf model can be prompted to declare, inside its own chain of thought, which slice of its context it needs next — and the inference engine can then skip reading the rest of the KV cache. The method, called Declarative Attention (DA), splits generation into three modes: global (the full context), focus (one specific reg

The FBI's AI use is up 605% — and there is no baseline for it

The Guardrails

The FBI's AI use is up 605% — and there is no baseline for it

Kash Patel put a number on the bureau's AI adoption that nobody outside it can check. Elsewhere, a 294,000-parameter voice model shows what "small" is supposed to mean. FBI Director Kash Patel says he has increased the bureau's use of AI by 605% since taking the seat — a figure offered on Fox News with no baseline, no metric and no audited inventory behind it. He paired the number with specific outcomes: AI triage on a tip that "stopped a school massacre in North Carolina" and about half a doze

Qwen open-sources a 7B image model that writes real transparency

The Frontier

Qwen open-sources a 7B image model that writes real transparency

Qwen cut its image model down to a size consumer cards can hold and gave it something no other open generator ships natively: an alpha channel. In Washington, the Congressional Black Caucus says the votes are there for AI guardrails. Alibaba's Qwen team open-sourced Qwen-Image-2.1, a 7B image model that generates transparent images natively — real RGBA layers instead of a white box you mask out afterward. It is a unified model: the same weights do text-to-image generation and editing, accept up

Qiyuan's Q1 and T1 robots go on sale from 19,999 yuan

The Arena

Qiyuan's Q1 and T1 robots go on sale from 19,999 yuan

China's consumer-robot race just got its first real price tags. Qiyuan Robotics put two home robots on sale in Shanghai on Sunday, and the numbers — plus the agent stack behind them — are the interesting part. Qiyuan Robotics, the consumer brand of Shanghai-listed Swancor Advanced Materials, disclosed prices for its Q1 and T1 personal robots for the first time at a launch event in Shanghai on Sunday, with units shipping from October 1. The Q1 starts at 19,999 yuan (about $2,800), with an Explor

Figure's Helix 2.5 tidied 30 homes it had never entered

The Frontier

Figure's Helix 2.5 tidied 30 homes it had never entered

Figure sent a humanoid into 30 Bay Area houses it had never trained on and had it tidy rooms, fold towels and make beds with no fine-tuning — 56% success, six times the 9% a from-scratch policy managed. China's provinces keep building token marketplaces, and CapCut put an agent inside the editing timeline. Figure's Helix 2.5 is the first real evidence that whole-body robot behavior learned from human video transfers to homes the robot has never seen. The evaluation ran across 30 Bay Area homes

See all