Live Feed
Agora-1: Odyssey Ships the First Multi-Agent World Model — 4 Players, One Generated World
Odyssey released Agora-1 on May 18, the first world model that supports multiple participants interacting inside the same simulation in real time. Up to 4 players share a single generated GoldenEye deathmatch world, each receiving a consistent rendered view. The architecture decouples world state from rendering, enabling applications in multiplayer gaming, robotics, and agent training inside generated environments.
Cursor Composer 2.5: 25x More Synthetic Training, Targeted RL Feedback, SpaceX Colossus Next
Cursor ships Composer 2.5 on the Kimi K2.5 base, trained with 25x more synthetic tasks than Composer 2 and a novel targeted RL technique that assigns localized feedback inside multi-hundred-thousand-token rollouts. The next model trains on SpaceX's Colossus 2 with 10x more total compute.
Vatican Forms First AI Commission as 'Magnifica Humanitas' Launches May 25 With Anthropic Co-Founder
Pope Leo XIV created the Catholic Church's first formal AI coordination body on May 16. His debut encyclical launches May 25 alongside Christopher Olah, the Anthropic co-founder who built Claude.
Nine-Juror Unanimous Verdict: Musk Loses OpenAI Lawsuit on Statute of Limitations
A jury in Oakland cleared Sam Altman, Greg Brockman, and Microsoft of all claims in Musk's three-week trial. The jury found Musk was aware of OpenAI's commercial direction as early as 2021 — too late to sue. Altman is now free to push toward an IPO without a $150 billion disgorgement cloud overhead.
Anthropic Acquires Stainless — the SDK and MCP Server Layer Comes In-House
Stainless has generated every official Anthropic SDK since the Claude API launched. Anthropic is buying them. The acquisition plants Anthropic's developer tooling stack — TypeScript, Python, Go, Java, Kotlin, CLIs, and MCP server generation — inside the company building the models.
Two Papers Cut Agent Architecture Orthodoxy: One Strong Model Beats Pipelines, and Grep Beats Embeddings on Search
A Stanford paper shows single-agent LLMs match or beat multi-agent systems under equal compute budgets on multi-hop reasoning. A separate paper finds direct terminal search raises BrowseComp-Plus accuracy from 69% to 80% while reducing cost. Together, they challenge two default assumptions in production agent engineering.
VulnSage: Alibaba's Multi-Agent Framework Found 146 Zero-Days — AI Exploit Generation Works on Messy Real Code
Alibaba's VulnSage uses a five-stage multi-agent pipeline to generate working exploits from vulnerability reports, outperforming prior tools by 34.64% on SecBench.js and finding 146 zero-days in production packages. The paper shows AI exploit generation now works on the messy, real-world code where symbolic and fuzzing tools have historically failed.
OpenBMB's MiniCPM-o 4.5 Ships Full-Duplex Omnimodal at 9B — Beats Qwen3-Omni-30B-A3B on 12GB RAM
OpenBMB released MiniCPM-o 4.5, a 9B open-weight model that processes video, audio, and text on a single shared timeline. It runs under 12GB RAM, outperforms Qwen3-Omni-30B-A3B on omni-modal benchmarks, and ships with code, weights, and a technical report under open licence.
Anthropic Splits Claude Into Two Budgets: Agent SDK Gets Non-Rollover API Credits From June 15
Claude subscribers get a dedicated monthly credit for programmatic usage from June 15, billed at full API rates with no rollover. The subsidy that made third-party agents viable on flat subscriptions is gone.
OpenAI Merges ChatGPT and Codex Under Brockman — 900M Users, One Agentic Platform
Greg Brockman formally takes OpenAI's product lead, merging ChatGPT, Codex, and the developer API into a single organisation. Fidji Simo's absence is now formalised into a structural shift ahead of the IPO.
Google Enters I/O Trailing Mythos by 13 Points and GPT-5.5 by 8 on SWE-Bench — Analysts Expect a Point Release, Not Gemini 4.0
Google I/O 2026 opens May 19 with Google ranked third on every major coding benchmark. Analysts expect Gemini 3.5 rather than 4.0. The company's answer is not a new model number — it's Gemini Intelligence, a proactive agentic layer across phones, watches, cars, and glasses.
EU's May 27 Tech Sovereignty Package Would Bar AWS, Azure, and Google Cloud From Government Health and Finance Data
The European Commission's Cloud and AI Development Act, due for adoption May 27, would require sovereign EU cloud infrastructure for sensitive government data in health, finance, and judiciary — restricting AWS, Azure, and Google Cloud from workloads that represent the core of Europe's government AI market.
Full-Duplex Bench v3: Gemini Live 3.1 Is Fastest and Goes Silent 22% of the Time
A new benchmark on real human speech — disfluencies, interruptions, mid-sentence corrections — shows Gemini Live 3.1's 4.25-second latency comes with a 22-point response reliability penalty. GPT-Realtime leads on accuracy but runs 2.3x slower.
Google I/O Is 48 Hours Out: Gemini 4.0, Project Astra Production API, and Android XR on the Line
Google's May 19 keynote arrives with its benchmark standing under pressure from Claude Opus 4.7 and GPT-5.5. What's confirmed, what's expected, and what the numbers need to say.
OpenAI and Malta Deploy ChatGPT Plus to All 525,000 Citizens — the World's First National AI Rollout
Malta becomes the first sovereign government to give all citizens paid AI access, gated behind a government AI literacy course from the University of Malta. Free for one year after course completion.