The top things worth knowing about in AI today.
Staffing group Adecco is rolling out Salesforce's Agentforce Coworker, built on Anthropic's Claude, to 27,000 employees across more than 40 countries after pilots in the UK and France. Recruiters in those pilots saved 35 to 40 per cent of their time and cut time-to-fill by 15 per cent. It is one of the largest deployments of an agent into frontline work rather than a lab, and the pilot figures are the sort a business case can actually be built on.
Read more →Anthropic says Claude leads 26 per cent of its research and development work, up from effectively nothing at the start of the year, and collaborates with staff on about 90 per cent of their work. It is the clearest published number yet on an AI lab using its own model to help build the next one. Read it as a directional signal on where technical work is heading, not a benchmark to copy.
Read more →Data centre builder Crusoe closed a US$3.9 billion Series F at a US$30.9 billion valuation, with Nvidia, Mubadala Capital and Founders Fund among the investors. It builds large sites, including the Abilene facility used by OpenAI, plus truckable modular units called Spark that connect to existing power supplies. Compute supply, not model quality, is still where the biggest cheques are going.
Read more →The US Federal Aviation Administration has committed US$875 million over 12 years to Air Space Intelligence for SMART, software that predicts traffic flows and flags conflicts by reading airline schedules, weather, airport capacity and airspace constraints. It rolls out in the Washington DC area first. It is a workable template for AI in safety-critical settings: the system advises controllers rather than replacing them.
Read more →Newly unsealed documents in the New York Times case against OpenAI and Microsoft show a Microsoft applied science director describing data scraping in a January 2023 internal memo as the largest theft of labour in human history. The filings cite more than 91,000 copies of works from the Times and other publishers in training sets, and claim Copilot cut Times click-through rates by as much as 93 per cent compared with Bing search. Any business that depends on search referrals should read those traffic numbers closely.
Read more →The UN has launched a System Data Commons with Google, covering data from 26 UN entities, with natural-language search and Model Context Protocol support so agents can pull figures directly from the source. It follows a UNICEF benchmark in which six major models averaged 21.2 per cent accuracy across 133,000 development questions, and about 60 per cent of answers returned no usable number at all. If you ask a chatbot for a statistic, the sourcing problem is real, and this is one attempt to fix it.
Read more →Huawei has moved the launch of its Ascend 960DT accelerator to the first quarter of 2027, two quarters earlier than planned, alongside an Atlas 950 SuperCluster design that links up to 256,000 accelerator cards. It is the clearest sign yet that US export controls have not stopped Chinese silicon roadmaps. For buyers outside the US and China, a second serious supplier eventually means pricing pressure on Nvidia.
Read more →The federal government has opened consultation on mandatory requirements for large data centres, covering power, water, siting and community engagement, and for frontier model training, covering safety, transparency and local investment conditions. Submissions close on 9 October, with legislation planned for early 2027. Separately, Finance Minister Katy Gallagher has sought advice on restricting smart glasses capable of covert recording in Commonwealth workplaces.
Read more →Google and Google DeepMind have set up the DeepMind Institute, directed by Shane Legg, James Manyika and Demis Hassabis, opening with four essays that deliberately air disagreements about artificial general intelligence. Hassabis used his to propose a US-led standards body for evaluating frontier models, voluntary at first with mandatory participation possible later. It is worth noting because a leading lab is arguing for rules it would have to meet itself.
Read more →A short read every week — the few things worth your time, and nothing that isn’t.