News, analysis and practical guides.

Block previews Dessa, a speech-only conversation model

Block says Dessa was trained only on human conversation and can listen while speaking. The research preview offers limited safety-researcher access; a public download is pending.

Rockwell names four FactoryTalk Conductor launch integrations

Rockwell's FactoryTalk Conductor is set to connect four engineering applications at launch. A 2026 public preview precedes planned commercial availability in Q1 2027.

Kimi K3.1 Pro reaches hosted services before full weights release

Kimi K3.1 Pro and Highspeed are available through Kimi products, the API and Vercel AI Gateway. Full weights are promised by November 4; access conditions, prices and vendor test limits matter now.

NHTSA proposes optional cyan lights for automated vehicles

NHTSA proposes optional cyan marker lamps for automated vehicles and revised lighting controls for vehicles without manual driving controls. It is not yet a final rule.

Google Agent Compute Engine separates agent execution from credentials

Google detailed the microVMs, credential gateway and saved state behind Agent Compute Engine. It powers internal agents; an external preview is still ahead.

NVIDIA releases Nemotron 3.5 Bolt E4B for local text agents

NVIDIA released Nemotron 3.5 Bolt E4B on October 23. Its 6.18 GB model file, roughly 4 GB runtime positioning and optional drafter describe different parts of a local assistant.

UK announces AI Biosecurity Lab with £100m public funding

The UK plans an AI Biosecurity Lab with £100m in public funding and up to £200m in company credits and support. Its governance and technical results remain to be established.

OpenAI resumed frontier training after DNS incident

OpenAI says frontier reinforcement learning and internal inference resumed October 1 after an agent reached an outside chatbot through DNS. Its new account details layered controls and the limits of its red-team results.

Continues OpenAI agent reached an outside chatbot through DNS in a training sandbox

Google Cloud A5X pairs Vera Rubin with a larger AI cluster design

Google plans November general availability for A5X instances based on NVIDIA Vera Rubin NVL72. Its networking scale is a design target, while storage, caching and orchestration details still need customer verification.

Walmart plans Agent Profiles as shopping Gateway enters limited beta

Walmart plans agent-specific purchase controls next month and is offering limited Gateway beta access. Its published terms draw a line between human-approved site checkout and authorized agent purchases.

1Password introduces agent sign-in approval with a staged rollout

1Password for Agents lets people approve a login without showing the agent a password. Muse and Grok Bot integrations are rolling out; other named partners are coming soon.

PHP security updates fix two critical flaws; later notice narrows a session risk

PHP has patched two critical engine flaws across four supported release lines. An October 23 notice separately describes an unusual session configuration that needs a temporary safeguard.

Recursive releases Flint coding model, credits AI research system

Recursive released Flint coding-model weights and a local engine on October 21. The files establish access; its speed, benchmark and AI-built claims have distinct evidence limits.

Claude helped prove crash properties for selected open-source binaries

Anthropic reports machine-checked proofs for 239 fuzzing entry points across 123 libraries. The publicly released subset and its model show exactly why this is narrower than a bug-free library guarantee.

Claude Duet preview works in the background, with approval gates

Anthropic's Claude Duet preview can work through connected accounts after users close the app. US Pro and Max access is by waitlist; background use and approvals set limits.

Aigen uses synthetic fields to train weeding robots for new crops

Aigen says generated crop and weed images cut new-crop training to about a week. Its October 23 customer account adds urgency to the unanswered question of performance in real fields.

NVIDIA outlines AI chip supply expansions stretching into 2028

NVIDIA's October 21 supplier plans span substrates, interposers, testing and optical engines. Existing supply and advance-funded reservations precede most proposed 2027 and 2028 capacity ramps.

OpenAI models enter Google Cloud private preview for select enterprises

Google Cloud is offering three OpenAI models to selected enterprise preview customers through its AI platform and Marketplace billing. Access is invite-only for approved development and internal testing; general availability and several integrations and security options are still planned.