The top things worth knowing about in AI today.
Moonshot AI released Kimi K3, a 2.8-trillion-parameter mixture-of-experts model that took first place on LMArena's Frontend Code Arena, edging out Claude Fable 5, and scored 88.3 on Terminal-Bench 2.1 against GPT-5.6 Sol's 88.8. It runs a 1M-token context and activates only 16 of its 896 experts per token, with full weights due by 27 July. An open-weight model at frontier coding quality changes the build-versus-buy maths for anyone running code assistants at scale.
Read moreApple briefly closed as the world most valuable company on Friday at $4.88 trillion, with Nvidia slipping to about $4.86 trillion after a 3.5% fall. The Philadelphia semiconductor index is down almost 19% from its highs and had its worst week in over a year. Investors are repricing how much AI infrastructure spending will actually be needed, which is worth watching if your vendor roadmap assumes cheap, unlimited compute.
Read moreTwenty-nine countries signed the agreement creating the World Artificial Intelligence Cooperation Organisation, headquartered in Shanghai, on the eve of the World AI Conference. Founding members include Indonesia, Brazil, Malaysia, South Africa, Russia and Pakistan; no G7 country joined. Xi Jinping pledged 5,000 AI training places for developing countries over five years. Expect a second set of international AI norms to sit alongside the EU and US frameworks rather than converge with them.
Read moreChina's Interim Measures for AI Anthropomorphic Interactive Services took effect on 15 July, the first binding rules anywhere aimed specifically at AI companions. Providers must disclose that the service is not human, add anti-addiction reminders and self-harm crisis pathways, and are barred from offering virtual intimate relationships to under-18s. ByteDance's Doubao and Alibaba's Qwen shut down personalised agents used by hundreds of millions to comply.
Read moreAnthony Albanese set out national standards that would legally require large data centres to underwrite their own new power supply, pay full connection costs, reduce draw when the grid is stressed and limit water use. A new Office of AI sits inside the Department of the Prime Minister and Cabinet from 15 July. National Cabinet considers the approach in August, with legislation expected in 2027, so anyone planning Australian compute capacity should budget for grid obligations.
Read moreMeta expanded its Hyperion data centre in Richland Parish to roughly five gigawatts and more than $50 billion, up from the $10 billion announced in 2024. The site spans about five square miles, with 7,500 construction jobs at peak and 1,000 permanent roles, reaching 2GW by 2030 and full capacity around 2032. It is a reminder that the constraint on AI capacity is now electricity and land, not chips alone.
Read moreClaude for Teachers launched on 14 July, giving verified US K-12 educators free access to premium Claude tools, a teaching skills library built with Learning Commons, and curriculum connections mapped to academic standards in all 50 states. Verified teacher accounts have model training switched off. Anthropic worked with the American Federation of Teachers on privacy and safety standards, and is piloting in Brooklyn schools with an evaluation planned in Detroit.
Read moreMicrosoft added Anthropic's Claude as a selectable model inside Copilot Chat, alongside watermarking for AI-generated video and audio in Microsoft 365 and an Agent Store submission flow requiring admin approval. From 1 July, Business Standard and Business Premium include Copilot permanently while the standalone Copilot Business licence rises in price. Worth reviewing your licence mix before the next renewal.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.