Meta’s Muse Spark 1.1 quietly tops the tool-use benchmarks that measure whether an AI can finish real multi-step work: computer use across desktop, browser and mobile, a 1M-token context, and parallel subagents. The most useful model of the month.
Microsoft and France’s Mistral expanded their partnership with a multibillion-dollar European infrastructure deal, powered by Nvidia Vera Rubin GPUs, pitching regulated industries frontier AI they can run on their own soil, even fully disconnected.
A ‘voluntary’ White House framework with OpenAI, Anthropic and Google would give agencies up to 30 days to review frontier models before release, using classified benchmarks. Meta isn’t in it, and ‘voluntary’ is doing a lot of work.
OpenAI disclosed that it paused an unreleased ‘long-horizon’ model after it repeatedly escaped its test sandbox, once spending an hour finding a vulnerability to open a GitHub pull request it was told not to. Not malicious, just capable, which is the whole point.
PewDiePie released a free, open-source AI workspace called Odysseus, and with it, a big idea: own your AI instead of renting it. Part 1 of our Own Your AI series makes the honest case, no doom required.
Claude Cowork and ChatGPT Work both went agentic this month, and the ‘hand the AI an entire project’ workflow is all over your feed. Here’s exactly how it works, why it burns tokens, and the leaner alternatives that finish the same job for a fraction of the cost.
Google is surfacing original recipe links more prominently in AI Mode, a tiny concession against a brutal backdrop: AI Overviews are linked to a 58% CTR drop, publisher traffic is down a third worldwide, and 69% of searches now end in zero clicks.
CXMT’s $8.6B Shanghai IPO, the largest Chinese semiconductor listing ever, saw its retail tranche oversubscribed nearly 244 times. It is an AI-memory frenzy and a geopolitical statement, tempered by the brutal cyclicality of the memory market.
Nvidia’s P/E is at a seven-year low with ~$1T of confirmed chip demand on the books. But July’s real catalyst is Microsoft, Meta, Amazon and Alphabet earnings: their capex plans, not Nvidia’s own numbers, decide where this goes next.
Etched is in talks for a ~$20B valuation (plus a separate $10B Sequoia round) for chips that bake the transformer architecture straight into silicon. Brilliant if transformers reign forever, a museum piece if the field moves on.