Four throughlines run through today’s edition. Nvidia confirmed it is buying Hugging Face for $12.93 billion, roughly 86 times the hub’s annualized revenue, putting the top AI compute seller in charge of where open models get distributed. Sam Altman called rivals’ compute spending unsustainable silliness while defending his own buildout as demand driven.
The Deal: Nvidia Buys Its Way Into Distribution
The chipmaker that sells nearly all AI compute now owns the place where open models get distributed, and OpenAI’s Sam Altman spent the same day attacking rivals’ spending habits.
- Nvidia buys Hugging Face for $12.93 billion. The chipmaker now owns the model hub it once offered $500 million for, at roughly 86 times its annualized revenue.
- Altman calls the AI buildout “unsustainable silliness,” but not his own. On the Sources Podcast, the OpenAI CEO blasted rivals’ compute spending while defending his company’s as demand-driven.
The Release Cluster: New Models, Gated Tiers
Meta and Google both shipped new models, but held their sharpest capability behind gates, and one researcher argues a widely discussed architecture trick is being overrated.
- Meta ships Muse Spark 1.3, holds back its top reasoning tier. The coding and agent model is live in Muse Code and the Meta Model API, but max reasoning waits on more safety testing.
- Google ships Gemini 3.8 Flash, gates a Cyber variant to defenders. The new Flash keeps 3.7’s price while gaining on coding and reasoning, and a locked-down Cyber sibling now patches Chrome for Google itself.
- Raschka says Astra’s “looped” design is a minor tweak, not the story. Machine learning researcher Sebastian Raschka argues OpenAI’s Astra owes little of its strength to the layer-reuse trick generating so much attention.
Security and Oversight: Testing the Guardrails
Anthropic brought in an outside auditor, a threat-research team clocked how fast AI now compresses an intrusion, and one vendor is pitting two models against each other to find the gaps first.
- Anthropic invites an outside auditor in after its own incidents. Zvi Mowshowitz argues the METR review and a paused RL effort show Anthropic taking prosaic alignment more seriously, not less.
- Unit 42 Clocks an AI-Assisted Network Breach at Under 10 Hours. Palo Alto Networks’ threat research arm says frontier AI agents let a lone attacker do two weeks of intrusion work in one overnight run.
- CrowdStrike Builds Two AI Models to Fight Each Other for Defense. SafeMind pits an attack model against a defense model inside a digital twin of a customer’s network, CrowdStrike says.
Builder Infrastructure: Who Owns the Hard Problems
Three teams are drawing the line between what belongs in a shared agent harness and what stays custom, from enterprise deployment to internal compliance tooling.
- Cursor Lets Enterprises Run Its Coding Agents on Their Own Servers. Self-Hosted Machines keeps planning in Cursor’s cloud but moves code execution behind the customer’s firewall, on hardware the customer already owns.
- Stencil says agent harnesses need one owner for their hardest problems. A 20,000-word playbook argues state, sandboxing, and rendering belong in the core engine, not rebuilt by every extension.
- Meta’s Internal Agent Turns Expert Judgment Into Reusable Knowledge. The system separates what its compliance agent knows from how it reasons, then converts expert corrections into automated, regression-tested edits without retraining the model.
Ideas and Measurement: What the Benchmarks Actually Show
Four writers push back on how the industry measures and frames progress, from chart design to what counts as a genuine paradigm shift.
- LeCun, Hassabis and Fei-Fei Li All Bet on World Models Now. Ksenia Se argues three AI pioneers converging on world models signals a real paradigm shift toward systems built to decide, not generate.
- Claude Opus 5 tops a new antibody-discovery benchmark at 53%. LatchBio’s TxBench-Antibody found every leading model still fails on roughly half of grounded biologics decisions.
- A log scale on a popular AI chart hides the real price gap. An OpenTeams engineer argues ArtificialAnalysis’ intelligence-versus-cost plot flattens the gulf between cheap and frontier models, nudging buyers toward pricier defaults.
- Ian Barber: Test-Time Training Is Not Yet the Next Scaling Axis. The blogger argues fast-weight updates look like a new lever for model capability, but they reset every sequence and leave continual learning unsolved.
Quick Hits
- Meta’s agent app resurfaces as Muse, now testing computer control. An iOS waitlist and a new desktop setting point to Meta’s long-rumored agent product shipping under its Muse branding, with early computer-control features spotted in testing builds.
- Ex-OpenAI Stargate Exec Hemani Lands at Anthropic After Meta Stop. Shamez Hemani left OpenAI for Meta’s compute team in April and has now joined Anthropic’s technical staff, another sign of compute talent circulating among the frontier labs.
- Anthropic ships a tool to check if a file carries Claude’s mark. The checker reads the C2PA content credential Claude attaches to files, not the text watermark, and Anthropic says it never stores what you upload.