THE LATEST

Latest news

All articles, ordered by their latest update.

Filter and search
A model card marked public preview connects to a hosted API, while a separate weights package remains behind a release marker.

Mistral Large 4 is a public preview, not a downloadable open model yet

Mistral’s October 6 announcement made Large 4 available through a hosted preview API and promised downloadable weights later in the month. As of October 24, the company’s current changelog still says those weights are coming soon.

A stack of mathematical manuscripts is organized into result families, with one proof page marked for human review and a separate Lean formalization sheet.

OpenAI’s mathematics release needs readers to check each claim

OpenAI’s October 6 repository is a collection of AI-produced mathematical manuscripts, not a single verified result. By October 23, the strongest independent assessment we found concerned one quasi-Riemann result: mathematicians interviewed by Quanta largely agreed with a shorter proof, while at least one prominent number theorist still wanted humans to check and understand every detail.

Conceptual three-level access diagram for defensive security work, authorized red-team testing, and limited testing of safety-critical systems, with review controls shown beside each tier.

Anthropic’s Cyber Verification Program adds three access tiers

Anthropic merged Project Glasswing and its earlier Cyber Verification Program into one tiered route for vetted defenders. The application, authorization scope, and data-retention requirements matter as much as the model list.

Two stylized magnetic-field surfaces show a symmetric pattern beside a new asymmetric shape, with a small equation sheet marked for simulation-code benchmarking.

Gemini 4 Argon helped widen exact stellarator benchmarks

A Google DeepMind team and University of Maryland professor Matt Landreman published two new families of exact three-dimensional magnetic-equilibrium solutions that break stellarator symmetry. Their October preprint says Gemini agents proposed the generalizations; the authors independently re-derived and checked every derivation.

Files, email, message, browser, drive, and cloud icons stream through a broad funnel toward a checkmark button beneath the text ‘One grant, many data sources’; conceptual artwork, not a macOS settings screen.

Apple plans stronger Full Disk Access controls, but gives no release date

Apple said October 2 that it will add controls requiring a very explicit action before an app receives Full Disk Access. The notice explains why the permission is unusually broad, but names no macOS version, interface, or shipping date.

Conceptual learning sequence from a simulated enterprise deployment through a security review to a real workplace project with a mentor.

Anthropic’s Frontier Academy is a 12-week enterprise AI residency

Anthropic committed $100 million to train 10,000 Frontier Deployed Engineers by the end of 2027. The program uses nomination, a simulated deployment, and a real company project; it is not a public self-enrollment course.

Conceptual government workspace with an AI document review, an administrator setting a spending cap, and an audit log beside a protected device.

Claude for Government is generally available, with capped usage billing

Anthropic moved its FedRAMP High-authorized government environment from public beta to general availability. Agencies receive administrative controls and prepaid spending caps, while some Microsoft 365 and command-line tools remain early access.

A small cloud computer researches through selected app connections while a person reviews a separate action card before it can affect an account.

OpenAI Dots puts always-on work behind permissions and reviews

OpenAI introduced Dots as persistent agents that can research in the background and act through connected apps. The useful distinction is between read-only proactive research and actions that may affect an account; the GPT-6.1 Astra model OpenAI held back was a separate release.

A small shop owner reviews a campaign draft beside connected social analytics, an ad account, and a separate approval control.

Muse for Small Business connects an assistant to a shop’s Meta accounts

Meta’s September 29 expansion links Muse with social analytics, ad accounts and outside business tools. The useful question for a shop owner is what context the agent can read, which actions it can take, and where approval still sits.

Four conceptual blocks labeled Muse, Meta Business Agent, Muse API and Muse Code connect to a generic business workspace; separate pricing, access and security-review question cards indicate details to confirm.

Meta packages its AI stack for business, starting with a platform initiative

Meta’s September 28 announcement brings Muse, Meta Business Agent, Muse API, and Muse Code under an enterprise effort led by former MongoDB CEO CJ Desai. The release names a direction, not a finished product catalog with prices and deployment terms.

Conceptual comparison of a focused everyday task card beside a larger complex project card, with speed and cost shown as separate gauges.

Claude Sonnet 5.5 trades top-tier scope for everyday speed

Anthropic positions Sonnet 5.5 for coding, documents, and other well-scoped work. Its per-token price matches Sonnet 5, while the company says it uses fewer tokens per task and adds cyber safeguards.

Conceptual workspace diagram showing a shared Home panel, a sandboxed app-building card, and an agent task awaiting review.

Microsoft puts Home, Code, and Autopilot in a new Copilot

Microsoft’s September 25 update grouped chat, delegated work, app building, and a persistent agent in a new Copilot. As of October 24, Home view and Code are available through Frontier, while Autopilot remains Private Preview. Access depends on eligibility and, for organizations, administrator opt-in.

THE WEEKLY EDIT

The week, made clear.

Stories worth understanding, brought together in one weekly edition.

Share this page

Instagram, TikTok & other apps

Copy the page link, then paste it where your chosen app supports links. You choose whether to send or publish.