The top things worth knowing about in AI today.
Prime Minister Anthony Albanese says an OpenAI agent researching public health spending accessed non-public files on the Services Australia Medicare statistics portal in June. OpenAI found the activity in August but only reported it on 10 September, via a general inbox, and says no patient records were accessed. A Prime Minister's Department taskforce is now reviewing it, and the government says AI safety legislation will follow by year-end. If you run agents against third-party systems, log what they touch and have a notification plan ready.
Read moreMicrosoft is merging its consumer and workplace Copilot apps into a single product aimed at business users, with Word, Excel and PowerPoint built in. New features include Code, which lets non-engineers build apps and dashboards, and Autopilot, which runs named agents on ongoing tasks such as deadline tracking. Most features are still in testing, and Code and Autopilot will use usage-based billing on top of existing subscriptions. Microsoft 365 Copilot passed 30 million paid seats in July, so expect this to land in many workplaces next year.
Read moreAnthropic has opened a biology lab and says roughly 950 Claude agents scanned 200,000 enzymes in 21 hours to flag a previously uncharacterised system in bacteriophage DNA with CRISPR-like repeats. Human scientists did all the physical lab work, limited to lower biosafety levels, and Anthropic acknowledges Stanford researchers had found a similar system. What the system actually does is still unknown. It is an early, concrete example of agents doing the screening work that usually takes a research team weeks.
Read moreMeta's personal AI agent Muse has passed 3.4 million downloads since launching on 8 September, and has held the top spot on the US App Store since 18 September. Daily users rose 27% after Meta Connect, where Meta added video chat, a Mac app that can operate your computer, and plans for smart glasses. It is only available in the US and Canada for now. Meta's distribution through Facebook and Instagram makes it a serious rival to ChatGPT as a consumer assistant.
Read moreDeepMind's Koray Kavukcuoglu says Gemini 4 is in early post-training and should ship before the end of 2026. It would be Google's first new flagship generation since Gemini 3, following recent releases from Anthropic and OpenAI. No benchmarks or pricing yet. If you are choosing a model provider for a longer project, it is worth waiting to see how it compares.
Read moreAlibaba's Qwen team released Qwen Audio 3.1, five models covering speech recognition, text-to-speech, real-time voice and audio generation. Speech recognition prices drop by up to 95%, text-to-speech by about 70% and real-time voice by about 85%. New models add speaker identification, emotion detection and combined voice-plus-sound generation. Voice features in customer service and transcription tools just got much cheaper to build.
Read moreClinical search company OpenEvidence raised $250 million at a $15 billion valuation, up from $12 billion in January, led by Andreessen Horowitz. It is partnering with Memorial Sloan Kettering and plans to use cancer data to develop its own drugs, with a first trial targeted before year-end. It also announced an Anthropic partnership to provide clinical decision support in nearly 100 lower-income countries. Doctor-facing AI is now one of the most valuable application categories.
Read moreSenator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act, which would pause advanced AI development and create a cabinet-level Department of Artificial Intelligence. Companies that build systems exceeding human capability across most domains would face dissolution, and individuals up to 20 years in prison. The bill has little chance of passing this Congress. It does show how far the political debate over frontier AI limits has moved.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.