Desk
Broad artificial intelligence coverage.
Alongside third-quarter results, Meta said it will issue non-voting shares to fund AI, acquisitions and pay, with trading from 2027, and raised the low end of its capital spending forecast to $135 billion to $145 billion. The spending is outpacing Muse’s trust-building.
Meta says its Muse agent has passed 10 million users, with about a quarter using it daily and more than a quarter of engaged shoppers averaging purchases above $200. Meta forecasts Muse will out-earn its social apps per user, a model that puts commerce incentives inside a personal agent.
OpenAI is updating its Usage Policies, effective November 20, 2026, so safeguards follow its models across clouds and resellers. Routers and AI gateways that offer direct access will need OpenAI’s review and approval, a reasonable safety goal that also creates a distribution gatekeeper.
Richard Socher’s Recursive released Flint, an Apache 2.0 coding model that runs locally and that the company says its automated Eureka Machine helped build in under four months. The local tool is useful; how much of the research judgment was automated remains undisclosed.
Anthropic’s new personal agent targets small business owners rather than shoppers, and its defining feature is restraint: approval prompts before payments or deletions, plus a classifier that blocks sharing the most sensitive data.
From next month, Walmart will let customers send AI agents to shop through a new MCP-based Agent Gateway, with default purchase approvals, optional budgets and a separate review trail. Amazon has blocked Perplexity’s agent and Meta’s Muse, so the two largest U.S. retailers now take opposite stances.
Nvidia’s Kumo Relational is a pre-trained model that takes relational tables and a Predictive Query Language question, pulls up to 10,000 historical examples as context and answers in one forward pass. Vendor benchmarks are strong; agent-driven decisions about real customers need governance.
Calif, using OpenAI’s GPT-5.6 Sol, and Google-owned Wiz found critical PHP vulnerabilities present since PHP 5. The NSA told owners who cannot patch to take sites offline. Disclosure was responsible and no exploitation was reported, but Wiz warns AI now makes old bugs easier to weaponize.
AT&T is extending its LegalEdge group with OpenAI, starting with an intake system that would classify and prioritize legal requests, link them to precedent and brief the responsible lawyer. It is a plan, not a result: no metrics are given, and triage choices shape attention.
At GTC Berlin, Jensen Huang unveiled the DSX Pod, a prebuilt container of Vera Rubin racks that Nvidia says can come online in 14 days, alongside a humanoid robotics platform and car partnerships. It targets Europe’s real bottleneck; power, siting and single-vendor dependence remain.
Conversational Gemini in Google Maps is live for U.S. CarPlay users, with a global rollout under way, close to a year after Android Auto. Google did not mention the Personal Intelligence features it had just announced for Android Auto, leaving a notable gap for iPhone drivers.
CrowdStrike examined AI setups behind attacks by four previously unreported groups between May and August. All got around model safeguards, mostly by claiming authorized penetration testing, and all compromised real targets, including Argentine health care organizations. No new exploits were involved.