The top things worth knowing about in AI today.
Chinese lab Moonshot AI released Kimi K3, a sparse mixture-of-experts model with 2.8 trillion parameters and a one-million-token context window, making it the largest open-weight model published so far. It ranked first in Arena's Frontend Code evaluation at 1,679 points, ahead of Claude Fable 5, while sitting behind Fable 5 and GPT-5.6 Sol on overall performance. Full weights are due by 27 July and output runs $15 per million tokens, which puts real pricing pressure on closed frontier models.
Read moreThe World Artificial Intelligence Cooperation Organisation was formally established in Shanghai with 29 founding members, including Indonesia, Brazil, Russia, South Africa and Pakistan. Beijing first floated the body a year ago and now has a standing forum to push its governance approach at the United Nations. Expect competing rulebooks rather than one global standard, which matters for anyone deploying models across multiple jurisdictions.
Read moreTSMC reported second-quarter revenue of $40.2b and net income up 77% year on year, its fifth consecutive record quarter, with high-performance computing accounting for 66% of revenue. Chief executive C.C. Wei announced a further $100b for Arizona, lifting the total US commitment to $265b, and raised 2026 capital spending guidance to $60-64b. Chip capacity for AI workloads is being built well ahead of current demand.
Read moreThe Prime Minister announced an Office of AI inside the Department of the Prime Minister and Cabinet to coordinate work across education, the labour market, energy, copyright and defence. New Australian Standards for AI will require large data centres to underwrite their own power supply, pay full connection costs and reduce consumption when the grid is strained. National Cabinet considers the approach in August, with standards expected to be legislated early next year.
Read moreInference platform Fireworks closed a $1.505b Series D led by Atreides Management, Index Ventures and TCV, with Nvidia among the investors. The company says it has passed $1b in annualised revenue and now serves more than 40 trillion tokens a day, up from 15 trillion. About 95% of those tokens come from models customers have specialised on their own data, a signal that fine-tuning rather than frontier scale is where enterprise budget is going.
Read moreGermany's ZAK regulator ruled that Google's AI Overviews and Perplexity generate their own content rather than act as neutral intermediaries, placing both under German media law and outside the EU's standard platform liability exemption. Regulators flagged Google for transparency failures and for ranking its own summaries above journalistic links. Both companies have one month to appeal, and this is the first ruling of its kind anywhere.
Read moreGoogle renamed NotebookLM to Gemini Notebook and added a secure cloud environment that runs code inside notebooks for data analysis, rolling out to Pro users. The product has more than 30 million users and 600,000 organisations, and existing shared links redirect automatically. Notebooks will sync across the Gemini app and Search, so research done in one place stops being stranded there.
Read moreMicrosoft 365 Copilot now offers GPT-5.6 for agentic, end-to-end work, with Claude Sonnet 5 rolling out alongside it. Model choice inside Copilot means one licence covers work that previously needed separate subscriptions. Check which model your tenant defaults to before assuming any change in output quality.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.