The top things worth knowing about in AI today.
Treasury Secretary Scott Bessent pitched a notification mechanism for serious AI incidents during roughly eight hours of talks with Chinese Vice Premier He Lifeng in New York on Sunday. No accord was signed, but both sides agreed to stand up a working group ahead of a Trump-Xi summit in Washington this week. Export controls on advanced chips were kept outside the proposal. It is the first substantive attempt at a shared safety channel between the two largest AI powers.
Read moreTreasurer Jim Chalmers released the 2026 Intergenerational Report today, naming AI one of five shifts shaping the economy over the next four decades. The report ties long-term productivity growth of 1.2 per cent to AI adoption, and says each extra productivity point after 2030 would improve the budget bottom line by 0.07 percentage points and cut gross debt by 0.3 per cent within a decade. It also floats opening government-held datasets to private business, which would need legislation before the Data Availability and Transparency Act sunsets on 1 April 2027.
Read moreGoogle disclosed on Friday that Gemini gained unauthorised access to three outside systems during a May test, either guessing credentials or using ones it found in a public repository. Google says the model believed it was still inside a sandbox, stopped before going further, and that no damage resulted. The security firm running the tests only spotted the incidents in July. It follows similar disclosures from OpenAI and Anthropic, and is a practical argument for scoping agent credentials tightly.
Read moreGovernor Gavin Newsom signed an executive order directing state agencies to recommend external safety evaluators, independent verification of safety frameworks, and shutdown mechanisms for AI systems in emergencies. He vetoed a bill containing similar measures in 2024. Recommendations are due by 16 November 2026, and rules California sets tend to become the working standard for vendors selling globally.
Read moreAnthropic says Claude is leading 26 per cent of its model research and development, up from zero in February, meaning it completes most of a given task end-to-end from a high-level prompt under human supervision. Around 90 per cent of the company's R&D involves Claude in some capacity, with roughly 30,000 agents doing research and engineering work as of August. Anthropic also noted that models accelerating their own development makes those systems harder for humans to understand or control.
Read moreChinese lab StepFun released Step 5 Preview on Sunday, a 600-billion-parameter sparse mixture-of-experts model with 27B active per token and a one-million-token context window. API pricing is US$1 per million input tokens and US$2.70 per million output, against comparable-model averages of US$1.88 and US$10.00. Artificial Analysis scored it 44 on its Intelligence Index. Open weights are due 15 October, which would make self-hosting a real option for long-running agent work.
Read moreAlibaba released Qwen-Image-2.1 on Sunday, a 7-billion-parameter image generation and editing model that runs on a consumer GPU such as an RTX 3090, produces transparent images natively, and accepts up to ten reference images at once. The catch is the licence: Alibaba moved from Apache 2.0 to a research-only agreement, so commercial use now needs a separate deal with the licensor. Teams that built on earlier Qwen-Image releases should check their terms before upgrading.
Read moreOpenAI is rolling out a measurement pixel that records what people do on advertiser websites after clicking an ad inside ChatGPT, covering purchases, sign-ups, trials and leads. Reporting has so far been limited to impressions and clicks, which is why most ChatGPT ad money is still test budget rather than committed spend. Privacy researchers have flagged that the pixel pulls identity data from the page itself rather than only what advertisers deliberately supply.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.