At its Build 2026 developer conference, Microsoft unveiled a family of seven of its own AI models — led by MAI-Thinking-1, its first reasoning model, and new MAI-Code coding models now wired into GitHub Copilot and VS Code. The headline product is “Scout,” an always-on personal agent that lives in Teams and Outlook and proactively handles meeting prep, scheduling conflicts and routine tasks without being asked. The clear message for communications watchers: Microsoft is ending the era in which its Copilot products depended on OpenAI, building its own models to cut costs and control its roadmap. For Cam, Scout is essentially an AI chief-of-staff aimed squarely at the inbox-and-calendar work every comms team lives in.
Anthropic, the maker of Claude, confirmed it has confidentially filed draft IPO paperwork (a “Form S-1”) with U.S. regulators — the first formal step toward selling shares to the public. The filing lands days after Anthropic raised roughly $65 billion at a $965 billion valuation, overtaking OpenAI’s $852 billion to become the most valuable private AI company in the world. The company says its annualized revenue run-rate has crossed into the mid-$40 billions, driven by enterprise use of Claude for coding and “agentic” work. No date, price or ticker is set yet — a confidential filing simply lets the company prepare quietly. In plain terms: you may soon be able to buy stock in the company behind Claude.
President Trump signed a new executive order that gives the federal government a much bigger oversight role over “frontier” AI — the most advanced models. The order asks AI companies to voluntarily share new models with the government for up to 30 days before a wide public release, a notable shift toward a federal safety-and-security framework. The move arrived the same week OpenAI published its own “public policy agenda” and a “blueprint for democratic governance of frontier AI,” signaling the labs are racing to shape the rules rather than just react to them. For a global comms leader, this is the regulatory backdrop every AI story will now be framed against.
OpenAI is repositioning Codex — once a pure code-writing tool — as a general productivity assistant for “every role, tool, and workflow.” New releases let Codex build live, data-connected corporate web apps, and add plugins, sites and annotations aimed at analysts, marketers, designers and investors rather than just engineers. OpenAI’s accompanying “Next Era of Knowledge Work” report frames Codex as a research, data-analysis and content-creation engine. The takeaway for non-technical teams: the same agent developers use to ship software is being marketed as a way for a comms or marketing team to stand up dashboards and tools without a developer.
OpenAI introduced new capabilities for GPT-Rosalind, a model purpose-built for life-sciences research, adding stronger biological reasoning, medicinal-chemistry expertise, genomics analysis and the ability to help design lab experiments. It is part of a broader wave of “vertical” AI aimed at specific industries rather than general chat. The same week, OpenAI highlighted real-world deployments — Boston Children’s Hospital using its technology to help diagnose rare diseases, and insurer Travelers rolling out an AI claims assistant nationwide — underscoring how quickly frontier AI is moving into regulated, high-stakes fields.
Anthropic ships Claude Opus 4.8Anthropic’s newest flagship adds “effort control” (you choose how hard Claude thinks), “dynamic workflows” in Claude Code that run hundreds of parallel sub-tasks at once, a cheaper fast mode, and improved honesty — released just 41 days after Opus 4.7.
Anthropic · May 28
02
Microsoft launches Scout, an autonomous workplace agentAn “always-on” assistant built on Microsoft’s OpenClaw and WorkIQ that works across desktop and browser, plugging into inboxes and calendars to handle scheduling and prep on its own.
Nvidia unveils PCs and a “superchip” built for AI agentsNvidia introduced consumer PCs and a new desktop “superchip” designed to run AI agents locally on-device — a direct challenge to Apple and Intel, and part of a push (with Microsoft) for an “AI coworker” that lives on your laptop rather than the cloud.
AI Valley · Jun 2
05
OpenAI’s frontier models and Codex go live on AWSEnterprises can now use OpenAI models and Codex directly inside Amazon’s cloud, through the controls and procurement they already have — a notable distribution win.
Anthropic files confidential IPO paperworkThe Claude maker takes its first formal step toward a stock-market listing, days after hitting a $965B valuation. See also · Top Stories.
TechCrunch · Jun 1
08
Tencent pledges to more than double AI investment in 2026Tencent plans to lift spending on its Hunyuan models and Yuanbao assistant past RMB 36 billion this year, funded by core-business earnings, even as chip curbs complicate capex. See also · Tencent Mentions.
Reuters / Investing · 2026
09
Apollo and Blackstone structure a $36B deal for Google TPU chipsA massive private-credit arrangement would buy Google’s custom AI chips to lease to Anthropic for new data centers across four U.S. states — a sign of how capital-intensive the AI race has become.
AI labs shift from flat subscriptions to metered, token-based billingGitHub Copilot and others flipped to usage-based pricing, meaning a “$19 seat” can become a far larger bill for heavy users — companies are discovering AI has real, variable costs. See also · Substack Highlights.
A web page can hijack ChatGPT’s summaries with hidden instructionsA researcher showed a “prompt injection” attack where a malicious web page makes ChatGPT display a phishing link as if it were the assistant’s own system message — relevant to anyone summarizing untrusted web content with AI.
Pivot to AI · Jun 1
From Your Inbox
Substack Highlights.
A dedicated Substack connector wasn’t available for this run, so these newsletter summaries were drawn from your Inoreader newsletter feeds (which capture your Substack subscriptions). Summaries reflect each issue’s headline and preview; open the links for the full posts you subscribe to.
AI labs spent years offering “all-you-can-eat” bundles like a gym membership; now they’re shifting toward charging for actual usage as heavy users churn through Claude and ChatGPT capacity.
Azhar weighs whether metered pricing shrinks adoption or expands the market by letting light users pay less — a live question for any team budgeting AI spend.
Pairs with his recent argument that individuals get more productive with AI while firms often don’t yet see it on the bottom line.
Covers Nvidia’s move into AI-agent PCs alongside “Microsoft Scout, your AI coworker that never logs off.”
Also carries Hassabis’s AGI-by-2029 warning. See also · Workforce & Society.
Good single-stop scan of the week’s hardware-and-agents news.
From Your AI Feeds
Inoreader AI Folder.
Around 20 fresh items hit your “AI” folder in the last 24 hours. The conversation is dominated by Microsoft Build 2026, Anthropic’s IPO filing, and OpenAI’s push to turn Codex into an everyday work tool. Newsletter duplicates have been consolidated; below are the distinct primary sources.
A case study on how the company Wasmer used OpenAI’s Codex (with GPT-5.5) to build a complex piece of developer infrastructure 10–20× faster, shipping in weeks instead of months. Illustrates the productivity claims behind OpenAI’s broader “Codex for everyone” pitch.
A critical report claiming an AI-assisted update to rsync — a widely used file-backup tool — introduced bugs that broke some users’ backup systems. A pointed example of the “AI slop” critique: that hastily AI-generated code can quietly damage reliable software.
An interview-led issue with Tools for Humanity’s Tiago Sada on World ID, bots and “proof of human” — the idea that, as AI agents flood the web, sites will increasingly need to verify a real person is behind an action. Relevant to anyone thinking about trust, verification and authenticity online.
OpenAI’s “Next Era of Knowledge Work” report positions Codex as a research, data-analysis, automation and content-creation engine for non-engineers. Together with the “Codex for every role, tool, and workflow” post, it’s the clearest sign yet that OpenAI wants Codex used across whole organizations.
Two policy posts in which OpenAI sets out its stance on safety, youth protection, workforce transition and a proposed U.S. federal framework for governing the most powerful models. Arrives in lockstep with the new Trump executive order.
A skeptical take on Microsoft Scout and “Codex walking into the office with a corporate badge,” questioning whether we should hand inbox and calendar control to an autonomous agent. A useful counterweight to the launch-day enthusiasm.
Insurer Travelers built an AI “Claim Assistant” with OpenAI to guide customers through filing claims, offer 24/7 support and scale during peak demand — a clean enterprise example of conversational AI in a regulated industry.
DuckDuckGo downloads reportedly jumped about 30% as some users react against Google’s increasingly AI-dominated search results — a small but telling signal about user trust and the shifting search landscape. The same issues note Anthropic’s $965B valuation and OpenAI’s robotics ambitions.
See also · Deals & People
Discoveries
Workflows & Tool Watch.
Claude Code’s “dynamic workflows” + effort control (Opus 4.8)
Claude Opus 4.8 added two things worth knowing even if you don’t code. “Effort control” is a simple dial — low, high, extra, max — for how hard Claude thinks before answering, so you can trade speed for depth. “Dynamic workflows” let Claude Code plan a big job and then run hundreds of small helper-agents in parallel, checking its own work before reporting back — useful for any large, repetitive task (think reformatting a huge archive, or batch-processing transcripts from your PLAUD recordings). Fast mode is now roughly 3× cheaper than before.
RelevantClaude · Claude Code · MacWhisper · DEVONthink
Connect a local AI model to Obsidian with MCP — private, no cloud
A widely shared write-up walks through linking a locally-run AI model to your Obsidian vault using MCP (the “Model Context Protocol” — a standard way for AI tools to safely read and act on your files). Because both Obsidian and the model run on your own Mac, nothing leaves your machine — the author argues it beats NotebookLM and ChatGPT for working with private notes. Open-source plugins like “obsidian-llm-hub” and “Local LLM Helper” make it approachable. A natural fit for your DEVONthink-and-Obsidian knowledge system where confidentiality matters.
RelevantObsidian · DEVONthink · MCP · Apple Silicon (M4)
MCP crosses 97 million installs — the plumbing for AI agents is now mainstream
Anthropic’s Model Context Protocol — the standard that lets AI assistants plug into your apps and data — passed 97 million installs, graduating from an experiment to core infrastructure. Practically, that means more of the tools you already use (note apps, project trackers, storage) are gaining “AI connectors.” Atlassian and Alteryx both shipped MCP updates this week aimed at reducing cost and letting non-engineers turn existing workflows into agents. Worth watching as connectors arrive for the apps in your stack.
A new wave of “local-first” agentic tools: Joanium, Brief and Mina
Product Hunt’s recent AI launches lean toward privacy-conscious, local automation. Joanium runs agentic automation across your files, repositories, calendar, chat and SaaS tools while keeping data on your device; Brief feeds product and decision context into coding agents; and Mina is a live meeting assistant that updates systems and generates follow-up tasks during the call itself. The meeting-assistant pattern is directly relevant to your VooV/Zoom/Teams workflow and PLAUD capture habit.
Recall.ai published a technical breakdown of self-hosted and open-source desktop meeting recorders that mimic Granola — covering the trade-offs of keeping audio capture and transcription fully on your own machine. If you’re weighing privacy and control for podcast and meeting capture (Red Book production, internal calls), it’s a clear primer on what a local recorder can and can’t do.
RelevantGranola · MacWhisper · Ecamm · Red Book podcast
No direct Tencent-specific AI headline surfaced in the last 24 hours
Across today’s web searches and your feeds, the cycle was dominated by Microsoft Build, Anthropic’s IPO filing and the U.S. executive order — none with a direct Tencent angle. The lines below are the most recent Tencent AI talking points a reporter is most likely to raise, so you have them ready.
Tencent has signalled it will lift spending on its Hunyuan foundation models and Yuanbao AI assistant past RMB 36 billion this year (up from roughly RMB 18 billion in 2025), funded by core gaming and ad earnings — while navigating chip-export curbs that complicate its capex plans. This is the clearest articulation of Tencent’s “fund AI from the core business” posture.
Recent reference points include the Hunyuan “Hy3” preview model briefly topping OpenRouter’s general usage list, continued Yuanbao consumer feature rollouts, and Hunyuan’s image-editing work. Against this week’s “agent-first” announcements from Microsoft and OpenAI, expect questions on how Tencent’s agent and assistant roadmap compares — and on the cost discipline theme, where Chinese models are increasingly cited as the cheaper option.