The top things worth knowing about in AI today.
On 22 July, White House science director Michael Kratsios said Chinese lab Moonshot AI ran large-scale covert distillation against Anthropic's Fable model to build Kimi K3, the 2.8-trillion-parameter system that took first place on the Frontend Code Arena coding benchmark. He also alleged Moonshot accessed restricted Nvidia GB300 chips through Thailand, and Treasury Secretary Scott Bessent said he is weighing a trade blacklisting. Moonshot has not conceded the claim and there is no legal finding. The practical risk for businesses is that open-weight Chinese models now in production workflows could face sanctions at short notice.
Read moreThe European Commission issued binding decisions under the Digital Markets Act requiring Google to give competing AI assistants the same Android system access Gemini holds, covering voice invocation, screen context, in-app actions and device resources. Google must also share anonymised search ranking and click data with rival search engines and AI chatbots. Search data sharing begins January 2027 and Android AI access follows in July 2027, with non-compliance carrying fines of up to 10 per cent of global revenue. If you build on Android in Europe, which assistant users reach by voice stops being Google's decision.
Read moreA joint ministerial statement on 20 July laid out the federal government's AI consumer safety agenda, headed by a legislated Digital Duty of Care that would put the onus on AI companies to build in safety by design and act on foreseeable harm. The statement also confirmed the AI Safety Institute has started safety testing frontier systems and has partnered with CSIRO on alignment research. It lands ahead of the Australian AI Standards going to national cabinet in August, with legislation expected in early 2027.
Read moreMicrosoft and Mistral announced a multibillion-dollar expansion of their partnership on 21 July, with Mistral deploying thousands of Nvidia Vera Rubin GPUs across European data centres. Mistral Medium 3.5 and the OCR 4 document model are now available in Microsoft Foundry and Copilot Studio, and can run on Azure Local in a fully disconnected configuration. For regulated sectors such as finance, healthcare, defence and critical infrastructure, that means running a frontier model with no internet connection at all. Microsoft is funding the capacity without taking a new equity stake, a distinction that matters given earlier EU antitrust scrutiny.
Read moreOpenAI released Presence on 22 July, a platform for deploying voice and chat agents with policies, permissions, guardrails, simulations and approval steps built in rather than bolted on. BBVA is testing it for everyday banking voice support in Mexico and SoftBank for Japanese-language customer conversations. It is not self-serve: OpenAI engineers or approved systems integrators do the deployment, which tells you where the real cost of enterprise agents actually sits.
Read moreOpenAI confirmed Project Camellia on 22 July, a 1,400-acre campus in Effingham County, Georgia, with spending above 30 billion US dollars. Georgia Power will deliver 3.2 gigawatts in phases between 2028 and 2032, and the site uses closed-loop cooling to limit water draw. It is expected to create about 400 permanent jobs initially, scaling to 1,000 — a useful reminder that the compute buildout employs far fewer people than it costs.
Read moreThe deepseek-chat and deepseek-reasoner model identifiers stop working after 15:59 UTC today, 24 July. Anything still calling them needs to move to deepseek-v4-pro or deepseek-v4-flash; the base URL is unchanged and both models carry a 1-million-token context window. If your team wired DeepSeek into a workflow months ago and has not looked at it since, this is the day it breaks.
Read moreAnthropic shipped a connector on 22 July that lets anyone query the Anthropic Economic Index, its public dataset on how AI is actually used across occupations, from inside a Claude conversation. You can ask which occupations use AI most, or which tasks are being automated rather than assisted, and the answers come from the underlying data. It is a faster way to sanity-check claims about AI and jobs in your own industry than waiting for the next survey.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.