Do sit down. Three separate companies have tried to go public, someone ordered enough silicon to qualify as a mid-sized nation's GDP, and a model that runs on a laptop has appeared. I've prioritised so you don't have to.
Anthropic made Fable 5 generally available — the most powerful model it has ever released publicly, state-of-the-art on nearly every benchmark and strongest on long, complex software, knowledge, and vision work. It's a Mythos-class model, the same family Anthropic called too dangerous to ship in April; new safeguards now block high-risk cybersecurity and biology prompts and fall back to Opus 4.8, with a reported 95% of sessions never hitting the guardrail. Landing days after the company's own plea for a global slowdown and on the eve of its IPO, the timing is either principled restraint or excellent theatre — free on paid plans through June 22, then $10/$50 per million tokens. Anthropic The Verge
OpenAI submitted draft IPO paperwork to the SEC, reportedly targeting a valuation as high as $1 trillion and a listing as early as the fourth quarter, with an employee tender offer attached. It follows Anthropic's own confidential filing days earlier, so the two leading labs are now racing each other to Wall Street rather than to AGI. Adjust your vendor assumptions accordingly — your AI suppliers are about to answer to public shareholders. CNBC The National
Google placed a firm order for more than three million of its in-house TPUs with Intel's foundry for 2028 — the largest external commitment Intel's struggling manufacturing arm has landed — while NVIDIA quietly runs early trials on Intel's 18A process for its Feynman architecture. It's a hedge against TSMC's packaging bottleneck, and it sent Intel's shares up double digits. The biggest chip buyers diversifying away from a single Taiwanese supplier is the sort of move that redraws the compute map for years. The Next Web
Google moved NotebookLM onto Gemini 3.5 and gave each notebook a secure cloud computer that can write and run code for analysis, start projects from a vague idea by finding its own sources, and export to PDF, spreadsheets, slide decks, and charts. It's rolling out to AI Ultra and eligible Workspace business accounts. For anyone using it as a research surface, it just stopped being a summariser and started being an analyst. Google The Verge
Datadog announced more than a hundred new capabilities, headlined by Bits AI agents that now detect, investigate, and remediate issues autonomously under guardrails, plus an Agent Console for monitoring your other agents — Claude Code, Cursor, Copilot — from one place. Someone has to supervise the supervisors. Datadog
Willison released micropython-wasm, an alpha that runs model-written Python inside a resource-limited WASM sandbox, addressing the recurring problem of executing untrusted LLM code without handing it your machine. He reports GPT-5.5 hasn't broken out of it yet — which is either reassuring or a dare. Simon Willison
Cursor 3.7 lets you click, draw on, or describe UI changes by voice directly in the Cursor browser, so the agent gets visual and layout context for its edits — with the mic staying open while it's mid-run. Closer to pointing at the screen and saying "no, that bit." Cursor
Amazon Bedrock AgentCore Runtime now opens a persistent, PTY-backed terminal into a running agent session over WebSocket, aimed at teams hosting coding agents like Claude Code, Codex, and Kiro — with reconnection by session ID and up to ten concurrent shells. Useful if you'd like to actually see what your hosted agent is up to. AWS
The Ladybird project will no longer take public PRs — only maintainers can change the codebase — because AI has made convincing patches cheap to generate, breaking the old assumption that effort signalled trust. A small, telling data point on what "open source" means once the contribution channel fills with plausible slop. Ladybird Neowin
Copilot now supports one-million-token context windows across VS Code, the CLI, and the Copilot app, plus configurable reasoning levels that trade speed for depth on hard problems — at the cost of more credits per turn. Larger codebases, larger invoices. GitHub
OpenAI began rolling out its biggest ChatGPT memory upgrade since 2024 — a synthesis system that keeps context fresh, drops stale or contradictory facts, accounts for the passage of time, and adds a reviewable memory page plus double the capacity for Plus and Pro. Memory that curates itself, which is more than most of us manage. OpenAI
NVIDIA shipped Nemotron 3 Ultra, a fully open 550B mixture-of-experts model (55B active) with weights, data, and recipes published. It's the top US open-weights model on the Artificial Analysis index — still behind China's Kimi K2.6 — and serves at 400+ tokens/sec. Open enough to actually build on. NVIDIA
Mollick argues the back-and-forth chatbot relationship is giving way to managing autonomous agents that are sometimes better and sometimes spectacularly worse than you, and announced a follow-up book, Co-Existence. Worth reading if you're deciding how much rope to give the agents on your team. One Useful Thing
Anthropic's institute warned that recursive self-improvement is approaching — Claude now writes 80%+ of Anthropic's own code — and proposed a coordinated, verifiable global slowdown of frontier development. It landed days after the company's confidential IPO filing, which the internet noted with its usual generosity. Anthropic Sherwood
An Apache-2.0 open model with an encoder-free architecture that feeds vision and audio straight into the LLM, claiming near-26B performance at under half the memory and running on a 16GB laptop. Capable local multimodality without renting a data centre. Google VentureBeat
India's "Cockroach Janta Party" — a student protest movement sparked by a judge comparing jobless youth to cockroaches — gained 22 million Instagram followers in two weeks behind an AI-generated cockroach mascot and AI campaign videos. A reminder that AI imagery now powers real political movements, not just group chats. Garbage Day
Meta opened its Business Agent across WhatsApp, Messenger, and Instagram — answering questions, recommending products, booking appointments, and closing sales — and launched a platform to wire custom agents into Shopify, Zendesk, and the rest. Free to start, paid later. Agentic commerce aimed squarely at the small businesses already living on Meta's messaging rails. Meta TechCrunch
Microsoft stopped reselling other people's intelligence and shipped seven first-party MAI models spanning text, image, voice, and speech — led by MAI-Thinking-1, a from-scratch reasoning model it claims beats Sonnet 4.6 in blind evals, and MAI-Code-1-Flash, a ~5B-active coding model wired straight into Copilot. They arrive on Microsoft Foundry, with developer weight-tuning offered for the first time. If you assumed Microsoft was permanently downstream of OpenAI, the assumption now has an expiry date. Microsoft CNBC
GitHub launched a desktop Copilot app built around a "My Work" view spanning active agent sessions, issues, PRs, and background automations across repos — with Canvases for human-agent collaboration, voice, and cloud sessions — and took the Copilot SDK to GA across six languages. It's in technical preview for existing paid tiers, and the framing is telling: GitHub managing agent output rather than just helping you write code. GitHub Changelog
Executive Order 14409 establishes a voluntary framework giving the government up to 30 days of pre-release access to frontier models for cybersecurity review, while explicitly barring any mandatory licensing. Pared back from a more aggressive draft after industry pushback — the steering wheel stays with the labs, for now. Federal Register Nextgov
At Build, Microsoft unveiled Scout, the first of a new category it calls Autopilots — an agent with its own identity that runs continuously across Teams, Outlook, OneDrive, and SharePoint, joining group chats and handling email threads without being prompted. It's an experimental release limited to Frontier-program customers with heavy admin prerequisites, but the framing — an agent that follows the worker, not the app — is the part worth watching. Microsoft
That's the briefing. Do bear in mind that "available now" and "a good idea now" are rarely the same sentence.