AI news for Tuesday, August 18, 2026

404 Media obtained the ChatGPT transcripts of an expert witness hired by 3M in a $61 million lawsuit over the 2020 Watson Grinding explosion in Houston. Three people were killed and roughly 200 homes destroyed. The prompts asked the model to “create an exceptional expert witness report defending the standard of care at 3M” and to “show how 3M is 0% at fault for the explosion.” Federal investigators had blamed a degraded, poorly crimped welding hose leaking flammable gas.

The tension: Discovery now reaches the prompt. An expert report is supposed to be a judgment you can cross-examine. Here the conclusion was specified first and the reasoning was generated to fit it, and the transcripts are what proves the order of operations.

Microsoft confirmed a global outage hitting the web UI, the API, Actions, Issues, pull requests and Copilot at once. If your agent pipeline pulls context from a repo or runs in Actions, it stopped too. Worth checking what in your stack silently assumes GitHub is up.

Lambert reads Nvidia’s recent investments as buying demand rather than backing labs. Fund enough teams to train open weights and you manufacture buyers for the hardware they train on. His framing explains why the open-weight ecosystem keeps getting money from a company that sells shovels.

The Journal totals leases, purchase obligations and guarantees that never appear as capex, and gets $3 trillion beyond the roughly $600 billion the same companies report as capex. Reported capex is the number the market watches. It is measuring a fraction of the commitment.

That is up from $47 billion in May and about $9 billion at the end of 2025, roughly sevenfold in a year. Anthropic filed its prospectus confidentially in June and investors expect it to close 2026 between $100 billion and $120 billion. Run rate annualizes a single month, so it moves faster than revenue does.

Reuters reports World Liberty Financial is behind WorldClaw, a venture offering AI services to Chinese firms that US restrictions are meant to cut off. Export controls assume compute is the chokepoint. A gateway reselling access is the obvious way around one, and this one has unusually well-connected backing.

Matt Arnold scanned his shopping, tapped his Nectar card, and was refused service over “an earlier incident,” with a red circle drawn around his face on the store monitor. Sainsbury’s apologised and suspended Facewatch at that branch. Arnold wants it off everywhere, because the watchlist is shared: a false match in one shop follows you into all of them.

The previous best bound on ω was 2.371339, and people have been chipping at this constant since 1969. The new paper reformulates the optimization, throws modern ML at it, then uses AlphaEvolve to refine the resulting algorithm. Josh Alman and Virginia Vassilevska Williams, who hold much of the prior record, are co-authors. The gain is in the fifth decimal place and that is the point.

A former Top Gear presenter poking at a car’s assistant is doing better numbers than most AI product launches. If you want to know what regular people think of in-car AI, this is closer to the answer than any benchmark, and the comments are the actual review.