The top things worth knowing about in AI today.
Premier Peter Malinauskas announced a royal commission into artificial intelligence on Monday, the first inquiry of its kind in Australia. Three commissioners start work in October, terms of reference land within four to six weeks, and the final report is due by 1 July 2027 at a cost of about $3 million. It will examine the effect on work and employment, and the findings are likely to shape how other states and the Commonwealth approach regulation.
Read moreOpenAI is removing message caps on text conversations for Free and Go users, with GPT-5.6 Luna replacing GPT-5.5 as the default model and a Think button for harder questions. Separate limits still apply to files, images, voice and image generation. The change rolls out this week, and OpenAI says factual errors fell 62 per cent for Luna against the model it replaces.
Read moreMeta released Muse Code in beta, a terminal agent for macOS and Linux built on its Muse Spark 1.2 model that plans, writes and checks code across large repositories. Standard pricing is $1.25 per million input tokens and $4.25 per million output. A contributor tier cuts that to $0.10 and $0.20 in exchange for letting Meta train on your prompts and code, which is a decision worth making before anyone on your team installs it.
Read moreAlibaba intends to require large commercial users of its next open-weight model, Qwen3.8-Max, to share revenue they earn from reselling it as a service. Moonshot set the precedent with Kimi K3, which demands a commercial agreement and up to a 30 per cent share once a reseller passes $20 million in annual revenue. If you build products on Chinese open weights, the licence now needs reading rather than assuming.
Read moreStanford researchers used the Evo 2 genome model to generate thousands of candidate phage genomes, synthesised close to 300 of them, and found 16 that killed E. coli strains already resistant to the natural phage they were modelled on. Stanford describes it as the first peer-reviewed case of generative AI designing entire functional viral genomes. The same result is being read as evidence that biosecurity screening has not kept pace with what these models can produce.
Read moreThe Bureau of Industry and Security is examining how Chinese AI firms rent Nvidia hardware in third countries such as Thailand and Malaysia, an arrangement current export controls do not cover. The Institute for AI Policy and Strategy estimates offshore rental lifts effective Chinese access to advanced US compute by at least 60 per cent this year. Closing it would reach cloud contracts rather than chip shipments, which matters to anyone buying capacity across Asia.
Read moreThe silicon photonics company disclosed more than $700 million in Series C funding at a $5.5 billion valuation, taking total raised past $900 million two years after it was founded. It says it is already shipping optical interconnects into hyperscale data centres under a multi-billion-dollar customer agreement. The money is going into the wiring between GPU clusters rather than the chips themselves, which is where the constraint on large training runs has moved.
Read moreFrom 14 August, Claude Code sessions on Pro, Max and Team plans default to auto mode, which replaces per-action approval prompts with a classifier that stops irreversible or destructive tool calls. Anthropic says the classifier blocked dangerous commands 89 per cent of the time against 13.6 per cent under human review, and it will not charge for the tokens the check consumes. Administrators who want the old behaviour need to pin a different permission setting before the switch.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.