AI progress — 28 October 2026

DeepMind announced Genie 4, a world model with audio and more realistic physics and people, but access is only promised for later. Elsewhere, Codex gained a native Vertex AI provider and GitHub Copilot added app-level authentication for unattended agents. Delaware lawmakers are weighing a plan that would shield owners of AI-run firms from liability.

New models and papers

• DeepMind unveils Genie 4 — a world model that adds audio and more realistic physics and people to generated simulations, rolling out “soon” to developers and Google AI Ultra subscribers. Why it matters: there is no access yet; teams building simulation or agent-training environments should watch for the developer release before planning around it.
• OvisOCR3 Technical Report — a 0.8B-parameter model that covers document parsing, general OCR, key information extraction, figure captioning and cross-page table reconstruction in one model. Why it matters: if the authors’ results hold, one small model could replace a pipeline of separate OCR and extraction stages; check it against your own documents first.
• On the SPOT — the authors propose speculative decoding for diffusion language models, aiming to speed up generation without the quality loss of existing multi-token decoding schemes. Why it matters: it targets the speed-quality trade-off that has limited diffusion LMs in practice; the reported gains are the authors’ own and unreplicated.

New tools and software

• openai/codex rust-v0.165.0 — adds a built-in Google Cloud Vertex AI provider (model_provider = "gcp"), lists GPT-6.1 Terra in the model catalogs, and allows reasoning-effort changes mid-conversation by default for supported models. Why it matters: teams on Google Cloud can now point Codex at Vertex without a custom provider setup.
• GitHub Copilot supports server-to-server authentication — public preview of server-to-server auth for GitHub Apps installed on organizations, so services and agents built with the Copilot SDK can call Copilot without a signed-in user. Why it matters: unattended agents and backend services can use Copilot under an app identity instead of borrowing a person’s token.
• Claude Code v2.1.312 — adds oidc.user_id_claim to the Claude apps gateway so people can be identified by a claim other than sub, such as Microsoft Entra ID’s oid, plus fixes for managed settings not loading after sign-in. Why it matters: organizations on Entra ID can map gateway users to the identifier they already use.

Overall digest

• “No human at the helm”: Delaware plan would let AI agents run entire businesses — a Delaware proposal would allow AI-run firms, and per US Rep. Comer their owners would escape liability for “reckless” AI. Why it matters: it is a proposal, not law, but it addresses who is legally responsible when an agent runs a company.
• Researchers are getting tired of OpenAI gatekeeping its breakthroughs — OpenAI announced results in mathematics, computer science and physics along with a new approach to communicating findings; researchers The Verge spoke to say they remain frustrated. Why it matters: the claims come from OpenAI and the criticism from the researchers The Verge interviewed, so independent verification is still the open question.
• Anthropic’s elaborate plan to keep Wall Street at bay — The Verge examines how Anthropic, with an IPO pending, says it can raise public capital without compromising its safety mission. Why it matters: this is one outlet’s analysis of the filing; the open question is how those governance commitments hold up under shareholder pressure.