1. The day GPT-5.5 shipped, OpenAI put a $25,000 price on breaking its bio guardrails OpenAI released GPT-5.5 this morning. In the same wave of posts, it announced the GPT-5.5 Bio Bug Bounty.

2. The week Anthropic acknowledged Claude Code's regression, OpenAI said Codex hit 4 million weekly users Anthropic's engineering team published a postmortem on April 23 responding to sustained user complaints that Claude Code had regressed.

3. MIT's Annual AI List Splits Into Three Tracks Where LLMs Used to Sit Alone MIT Technology Review's 2026 roundup of what matters in AI breaks its top entries into three parallel tracks: LLMs+, world models, and agent orchestration.

In Brief

• Meta cuts 10% of staff in May Meta will lay off roughly 8,000 employees in May and close around 6,000 open roles, per a memo from chief people officer Janelle Gale. The cuts follow record capital spending on AI infrastructure and data centers.

• Microsoft puts Agent Mode in Word, Excel, and PowerPoint Microsoft began rolling out Agent Mode across Office apps this week, extending Copilot from suggestions to multi-step task execution. Internally the team has been pitching the feature as "vibe working."

• Google rebuilds Workspace around Workspace Intelligence Google added automated functions across Gmail, Docs, Sheets, and Drive powered by a new system called Workspace Intelligence. The update lets Workspace act inside documents and inboxes without explicit prompting.

• Anthropic adds Spotify, Uber, and TurboTax connectors to Claude Anthropic shipped personal app connectors for Audible, Spotify, Uber, AllTrails, TripAdvisor, Instacart, and TurboTax. The release pushes Claude beyond work suites into consumer transactions and entertainment.

• Tesla triples capex to $25B for 2026 Tesla raised 2026 capital spending to $25 billion, three times its historical average, with the CFO confirming free cash flow will go negative through year-end. The company tied much of the increase to AI compute and Optimus production.

• MIT Tech Review maps China's open-weight playbook Chinese AI labs are shipping downloadable open-weight models while US labs gate access behind APIs, per a new MIT Technology Review piece. Developers fine-tune and self-host the Chinese models without supplier negotiations or per-token billing.

• Token bills force Anthropic and OpenAI to tighten consumer limits The Verge reported Anthropic restricted the viral OpenClaw agent earlier this month after inference load became unsustainable. Other labs are quietly capping heavy users as the gap between subscription revenue and compute cost widens.

• Delve tied to second customer breach in a week TechCrunch confirmed Delve handled compliance certification for Context AI, the agent-training startup that disclosed a security incident last week. It is the second Delve client to suffer a major breach in seven days.

• Sierra acquires YC-backed Fragment Bret Taylor's customer-service agent startup Sierra bought Paris-based, YC-backed Fragment. Sierra did not disclose terms or a product roadmap for the acquired team.

• X replaces Communities with Grok-curated feeds X rolled out AI-powered custom timelines built on Grok, retiring the Communities product. The feeds ship with new ad slots embedded in the recommendation flow.

• DeepMind publishes Decoupled DiLoCo for distributed training DeepMind released Decoupled DiLoCo, a method for resilient model training across geographically split, heterogeneous compute. The technique targets fault tolerance when nodes drop mid-run, a recurring issue in cross-datacenter training.

Read the full edition →