The top things worth knowing about in AI today.
Nvidia has signed memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to build financing platforms aimed at mobilising more than US$500 billion of third-party capital for AI data centres, chips and power. The idea is to turn Nvidia compute into an investable asset class so labs, enterprises and AI clouds can lease capacity instead of buying it outright. Jensen Huang said he approached only those six firms and none declined. If it works, access to large-scale compute becomes a credit question rather than a balance sheet one.
Read moreMeta has published Muse Glimmer, a 30-billion-parameter multimodal model, under a permissive Apache 2.0 licence with no user caps. Quantised to 4-bit it drops from over 55GB to under 20GB, so it runs on a single machine with 24-32GB of memory. Mark Zuckerberg paired the release with a 6,500-word essay arguing that concentrating advanced AI in a few hands produces worse outcomes for everyone else, and called for a rethink of restrictions on distillation. For teams with data they cannot send to a cloud provider, a commercially usable model that fits on one workstation changes what is practical.
Read moreOpenAI has split its Daybreak cybersecurity programme into two tiers and released GPT-5.6-Cyber to the higher one. In testing the model answered 95 per cent of advanced security requests covering exploit chains, authentication bypass and privilege escalation, against 1.5 per cent for the general GPT-5.6 Sol. OpenAI used it on Chrome's V8 engine and found two previously unknown vulnerabilities that could be chained together. Access is limited to vetted defenders, and hardware security keys become mandatory on Daybreak accounts from 1 September.
Read moreAn unreleased research version of Claude raised the known lower bound for the proportion of Riemann zeta zeros satisfying the Riemann hypothesis from 41.6 per cent to 67.2 per cent, a figure mathematicians had inched upward since Montgomery's 1973 work. It ran as roughly 60 subagents in Claude Code over two sessions, burning 31 million output tokens and 2,400 shell commands, and discarded about 650 failed ideas first. The result was formally verified in Lean and reviewed by outside number theorists. Anthropic is clear the technique will not prove the hypothesis itself.
Read moreA missing database isolation rule at meeting-notes tool tl;dv let any signed-in user query 181,874 meeting records across 84,312 users and 35,003 email domains, including calls from 23 governments and thousands of companies. Each record carried the organiser's email and a conference ID that could be used to join the underlying Google Meet or Teams room. The researcher says he reported it in January and got no fix for six months; tl;dv says a remediation has since been deployed and validated by its penetration testing vendor. Worth asking which AI assistants sit inside your meetings and what they can see.
Read moreIntel announced a US$15 billion public offering of common stock, its first since listing in 1971, with underwriters holding an option on a further US$2.25 billion. Proceeds go to capital expenditure and working capital as data centre revenue jumped 59 per cent last quarter and the stock more than tripled over the year. Intel named physical AI, purpose-built silicon and advanced packaging as its growth areas. A second credible supplier of advanced packaging capacity matters to anyone waiting in line behind the current one.
Read moreRetail subscriptions for humanoid robot maker Unitree's Shanghai STAR Market listing came in at 8,288 times the shares on offer, leaving a winning rate near 0.018 per cent. The company priced at 150.8 yuan a share for a valuation around US$9 billion and raised roughly US$904 million, making it the first pure-play humanoid robot maker listed on a mainland exchange. Unitree shipped 31 per cent of the world's humanoid robots in the first half of 2026, up 170 per cent year on year. Physical AI is now a public market story, not just a demo reel.
Read moreOxford Economics research for SAP, covering more than 2,600 business leaders across 13 countries, found just 22 per cent of Australian organisations consider themselves mostly or fully ready for AI governance, against 33 per cent globally. Forty-two per cent said they are deploying agents faster than they can standardise and govern them, 43 per cent have no human in the loop for agentic workflows, and 49 per cent have already had an agent take an incorrect action. Expected two-year returns from agents jumped to A$21.2 million, up from A$4.4 million a year earlier. The gap between what firms expect and what they can currently supervise is the thing to fix first.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.