The top things worth knowing about in AI today.
Meta agreed to pay up to $16.68 billion over ten years to settle claims from 29 US states that Facebook and Instagram were designed to addict children. Teen accounts now default to a two-hour daily cap and are blocked overnight, with an independent auditor policing compliance for a decade. Design decisions aimed at under-18s are now a legal liability, not just a policy question.
Read moreIBM published its Granite 4.2 models at 3B, 8B and 30B parameters with a 512,000-token context window, all under Apache 2.0 on Hugging Face, Ollama and GitHub. Every size has a switchable thinking mode, and the two larger models went through reinforcement learning inside real software engineering, terminal and web search environments. For teams that need capable models running on their own hardware with no licensing conditions, this is a serious option.
Read moreZhipu confirmed that Ox Alpha, the free stealth model developers had been testing on OpenRouter since 20 August, is GLM-5.3-Flash, and released the weights under an MIT licence the same day. It is a 320-billion-parameter mixture-of-experts model with 18 billion active per token, a one-million-token context, and API pricing of $0.15 and $0.50 per million input and output tokens. That is roughly a tenth of the flagship GLM-5.3 rate for work that scores close to it.
Read moreCisco Talos documented a Chinese-speaking group, tracked as UAT-10147, running PentestGPT and DeepAudit on its own command-and-control server to scan web servers and select exploits automatically. Investigators found a target list of roughly 170,000 URLs alongside Metasploit and privilege escalation tooling. The expertise barrier for post-compromise work has dropped, which makes patching cadence on internet-facing servers matter more than it did.
Read moreSpaceX says its Grok voice agent now takes more than 15,000 inbound Starlink support and sales calls a day and completes over 3,000 orders a week across voice and chat. It diagnoses hardware faults and ships replacements with no human in the loop, running on the Think Fast 2.0 build that became default on 5 August. It is one of the largest public numbers yet for a voice agent doing frontline support rather than triage.
Read moreAM Intelligence, founded by the promoters of renewables group Greenko, placed a binding order for about 9,000 Nvidia Rubin GPUs configured as 125 Vera Rubin NVL72 racks across 30MW in Hyderabad, with delivery in the first quarter of 2027. It is the opening phase of a plan for a gigawatt of compute-as-a-service across India, the US, Finland and Malaysia, backed by more than $8 billion in capital spending. Frontier-class capacity is starting to appear well outside the usual hyperscaler map.
Read moreEmerald AI closed an oversubscribed $150 million Series A at a $1.05 billion valuation, co-led by Energize Capital and DCVC with Nvidia, Siemens, GE Vernova and Samsung Ventures joining. Its Conductor platform shifts data centre power draw in response to grid conditions instead of shutting workloads down, and the company argues the approach could free more than 100GW on the existing US grid. Grid capacity, not chips, is increasingly what decides where compute gets built, which is the same argument now in front of National Cabinet here.
Read moreIn a 6,000-word essay, Bill Gates argued that some categories of work, including caregiving, childcare, teaching and jury service, should stay in human hands even once AI can do them, comparing the idea to nature reserves. He also backed taxing AI tokens and robots to remove the nudge the tax system currently gives firms toward replacing staff with machines, and put the most aggressive version at around 40% of jobs. The automation debate is shifting from what is possible to what should be deliberately fenced off.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.