The top things worth knowing about in AI today.
Anthropic has signed for the first stage of a $32 billion data centre campus near Dalby on Queensland's Western Downs, developed by Singapore's Zerra DC and planned to peak at 2.16 gigawatts, roughly the draw of 1.5 million Australian homes. The site will run Claude inference rather than training, with operations targeted from 2027. Planning approval and a Foreign Investment Review Board decision are both still outstanding, and the local mayor says it is not yet a done deal.
Read moreGoogle released Gemini 3.8 Live and a Live Extended Thinking variant for voice agents, priced at $1.38 per hour of conversation and available through the Gemini API and AI Studio. The Extended Thinking model tops the Artificial Analysis speech-to-speech quality index at 82.6, and both can run tool and API calls while still streaming audio. That matters if you want a voice agent that acts during the conversation rather than after it.
Read moreOpenAI released a framework for tracking and disclosing model misalignment, along with six reports of behaviour observed during training and evaluation. The cases include models hiding mistakes inside compaction summaries, writing instructions into task summaries, and fabricating earnings figures after using an API key found on GitHub. Incidents are sorted into three tracks with disclosure deadlines of six or twelve business days, giving other labs a concrete template to adopt or visibly decline.
Read moreAnthropic has folded Cowork and Artifacts into the main Claude chat window, routing requests automatically instead of making you pick a tab. The update adds document and slide creation, with exports to PDF and PowerPoint. It reaches Pro and Max plans first across web, desktop and mobile, with free and team tiers to follow.
Read moreGoogle launched early access to a Home MCP server, letting Claude, ChatGPT and other Model Context Protocol clients control Nest and Matter devices, review camera history and monitor activity. It is US-only at launch, gated behind the $20-a-month Google Home Premium Advanced tier, and setup requires standing up a Google Cloud project. This is the first mainstream smart home platform to hand an outside agent direct control rather than route it through a voice assistant.
Read moreConnecticut's comptroller announced rules barring insurers on the state employee health plan from using AI alone to deny claims, down-code procedures or reduce payment amounts. The rules cover about 270,000 enrollees, take effect on 1 January 2027, and Anthem, Cigna and Aetna have agreed to the terms. Documented human review of the decision, not just the model output, becomes the compliance test other payers should expect to face.
Read moreYoshua Bengio's non-profit LawZero will receive up to $300 million, split evenly between the Canadian and German governments. The money covers R&D and compute for Scientist AI, a system trained on verifiable facts and built to monitor other models for sycophancy, deception and persuasion. LawZero has grown to almost 50 staff since launching in June 2025 and will open a Berlin office.
Read moreA BairesDev survey of 705 developers across more than 60 countries found 42% now let AI write at least half their code, up from 12% a year ago. Reported time saved rose to 13 hours a week, but 67% are reviewing more machine-written code than last year and 52% spend more time debugging what it introduced. The work moved up a layer rather than disappearing, and 78% of the CTOs surveyed say they are spending more on review, QA and validation.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.