Anthropic released Claude Opus 4.8 on May 28 — less than two months after 4.7. The headline is reliability: the model is roughly four times less likely to let flaws in its own code slip by unremarked, and Anthropic frames it as more honest about its own progress and able to work independently for longer. Benchmarks moved up across the board (its agentic-coding score rose from 64.3% to 69.2%), and pricing is unchanged. A “Dynamic Workflows” research preview lets Claude spin up hundreds of parallel helper agents in a single session to handle very large jobs. Why it matters for you: this is the engine behind the Claude tools you already use, and the honesty gains are exactly what a comms team needs when trusting AI drafts.
Google DeepMind’s AlphaProof Nexus formally cracked nine open problems from mathematician Paul Erdős’s famous list — some unsolved for more than half a century — by pairing the Gemini 3.1 Pro model with a tool that checks every logical step. The proofs are published and verifiable, a genuinely new threshold. Yet CEO Demis Hassabis was careful to say today’s systems are “nowhere near” true AGI, even as he separately called this an AI-driven “species-level transition” that leaves “little margin for error.” A useful talking point: real capability leaps and sober expectation-setting are now coming from the same person, on the same week.
Meta introduced Muse Spark, the first flagship large language model built inside the new Superintelligence Labs led by Chief AI Officer Alexandr Wang. Meta says it delivers competitive results on multimodal perception, reasoning, health and agentic tasks at a fraction of the usual computing cost. The company also signalled the scale of its ambition with planned AI capital spending of $115–135 billion for 2026, nearly double last year’s. In plain terms: a third serious frontier challenger to OpenAI and Google has now arrived, and the spending arms race is intensifying.
Google introduced Gemini 3.1 Flash-Lite, tuned for speed and cost rather than raw power: about 2.5 times faster responses and 45% faster output than earlier Gemini versions, at just $0.25 per million words of input. This is the part of the market that matters for everyday business automation — fast, cheap models are what make it affordable to run AI across thousands of documents, emails or media queries rather than one expensive prompt at a time.
OpenAI has reportedly crossed $25 billion in annualized revenue and is taking early steps toward a public listing, while Anthropic is approaching $19 billion — a remarkable closing of the gap, and the reason one of Cam’s newsletters ran the headline “Anthropic is bigger than OpenAI now.” Behind the revenue, both companies are committing eye-watering sums to cloud and chips. For a communications lens, the story is shifting from “can these models work?” to “is the business underneath them real?” — and the skeptics (see the Workforce & Society roundup) are getting louder.
Gemini 3.1 Flash-Lite goes liveGoogle’s new low-cost, high-speed model: 2.5× faster responses at $0.25 per million input tokens. See also · Top Stories.
LLM Stats · May 31
02
OpenAI launches Rosalind BiodefenseA vetted-access program around its GPT-Rosalind model for biodefense, public-health and pandemic-preparedness partners, including the U.S. government.
“Anthropic is bigger than OpenAI now”AI Valley argues Anthropic’s momentum (≈$19B annualized) and enterprise traction have flipped the narrative — alongside Apple prepping its biggest Siri upgrade in years.
AI Valley · May 30
06
Meta commits $115–135B to AI for 2026Roughly double last year’s spend, funding Superintelligence Labs and the new Muse Spark model. See also · Top Stories.
Why AI isn’t showing up on the bottom lineAzeem Azhar’s framework: individuals get more productive with AI while firms often don’t — the gap between personal wins and company results. See also · Substack Highlights.
Exponential View · May 31
09
“Your future job will be to keep AI on task”Noah Smith argues the human role shifts to supervising, correcting and directing AI rather than doing the work directly — a useful lens for team planning.
Noahpinion · May 28
10
“AI psychosis” enters the tech-CEO discourseThe Automated tracks a new buzzword for over-the-top AI belief among executives — plus data that agentic AI is generating 450% more web traffic.
Colorado AI Act effective June 30 — with a twistThe first broad U.S. state AI law lands this month, though a working-group draft would refocus it on automated decision-making and reset the date to January 2027.
Anthropic’s “Project Glasswing” hunts software flawsA controlled program gives select firms (AWS, Apple, Google, JPMorgan, Microsoft) access to an unreleased “Mythos” model to find and fix critical vulnerabilities.
Crescendo · May 31
From Your Inbox
Substack Highlights.
No dedicated Substack connector is available in this run, so these were pulled from your “*Substack: Paid” and “*Newsletters” folders in Inoreader, where your subscriptions are consolidated. Coverage is limited to issues that landed in the last 24 hours.
A non-AI note from a writer you follow closely — Fallows’ seasonal roundup on the state of the press and what he’s watching. Included for completeness from your paid folder.
From Your AI Feeds
Inoreader AI Folder.
Eight items landed in your “AI” folder in the last 24 hours. The day was dominated by the Claude Opus 4.8 launch and a strong run of agentic-AI analysis from the GDELT project.
GDELT’s Kalev Leetaru argues that AI agents which excel at writing code struggle with messy, real-world information like global news — because the real world has no compiler to tell the agent when it’s wrong. A sharp, relevant read for anyone applying AI to media monitoring and analysis.
A talk on using AI to read billions of daily news stories across thousands of languages — video, audio and text — to understand and forecast world events in near real time. Paired with a companion piece on scaling GDELT’s systems with agentic Gemini and a human in the loop.
David Gerard’s skeptical take on the SpaceX listing and AI-driven hype, comparing the dynamics to crypto’s “number go up” era. Your folder surfaced both the original and a fresh Portuguese translation, a sign the piece is travelling. A useful counterweight to the bullish coverage elsewhere in today’s brief.
Azeem Azhar on the productivity paradox and whether relying on AI dulls our thinking.
See also · Substack Highlights
Discoveries
Workflows & Tool Watch.
Claude’s new “Dynamic Workflows” run hundreds of helpers at once
Shipped with Opus 4.8, this research-preview feature lets Claude break a big job into hundreds of small tasks and run them in parallel in a single session — think a sweeping content audit or migrating a giant archive. For you, it’s the practical reason Claude can now take on much larger projects without you babysitting each step.
n8n’s MCP server can now BUILD automations from a plain-English prompt
Until recently, connecting Claude to n8n only let it run automations you’d already built. The updated n8n-MCP server (an open-source bridge with 20,000+ GitHub stars) now lets you describe a workflow in normal language and have Claude assemble it for you — reportedly a working automation in under five minutes. This is the closest thing yet to “tell your computer what to automate and it does it,” and it’s directly in your wheelhouse.
Community write-ups this week highlight two approaches: the obsidian-mcp-plugin, which lets Claude treat your vault as a knowledge graph (following backlinks, finding paths between ideas, running Dataview/Bases queries); and Nexus / Claudesidian, which collapses 40+ tools into a simple two-step system and cuts token costs by roughly 95%, with a chat interface inside Obsidian. Because Obsidian is just plain text files, Claude Code can also work directly on your vault folders.
Perplexity’s agent — which can research, draft, code and manage a project end to end — is now available inside its upgraded Mac desktop app, with the ability to edit local files, use your computer, browse via its Comet browser and take voice commands. Worth a look as an alternative agent to keep an eye on alongside Claude, though the full feature set sits on the $200/month Max plan.
DEVONtechnologies’ own guide and an active community thread walk through using outside AI models to summarize, classify and query documents inside DEVONthink, while the DEVONlink plugin keeps it tied to Obsidian (reveal the current note in DEVONthink, pull in AI-suggested related files). A concrete pattern for the two apps at the centre of your setup.
Open-source alternatives to Granola for meeting capture
Recall.ai published a technical breakdown of self-hosted and open-source desktop meeting recorders that do what Granola does — capture and transcribe meetings locally. Useful background if you ever want a privacy-controlled, self-hosted option alongside MacWhisper for podcast and meeting workflows.
The AI-Augmented Engineer published a hands-on guide to working with the Claude Code desktop app — the same family of tooling behind Cowork — with practical patterns for letting it work across your files. A good primer if you want to push beyond chat into Claude actually doing tasks on your Mac.
No major direct Tencent AI headline in the last 24 hours
Your Tencent monitoring feeds (Talkwalker) were dominated by commerce noise — WeChat Pay merchant listings rather than substantive coverage. The most relevant recent line a reporter might raise is Tencent’s flagship model push and where it sits versus US and Chinese rivals. Key context is below.
Tencent’s open-source Hy3-preview — its most powerful model yet — is led by Yao Shunyu, a Princeton PhD who previously ran AI-agent research at OpenAI and helped pioneer “deep research” systems plus the ReAct and Tree of Thoughts methods. Notably it is relatively compact at ~295B parameters, bucking the trillion-parameter trend. Coverage frames it as on par with top Chinese models but still trailing OpenAI and Google DeepMind’s flagships — the line to be ready for if asked.
For background, Hunyuan 2.0 (December 2025) uses a mixture-of-experts design (406B total / 32B active parameters, 256K context) with gains in math, science and coding, and Tencent has rolled out its Hunyuan 3D creation engine globally for turning text, images or sketches into 3D assets. Together these frame Tencent’s “full-stack” AI story across language and creative tooling.
A May 31 Guardian editorial on internet fragmentation references WeChat as an all-in-one app many users find “more attractive than WhatsApp” — a recurring framing of Tencent’s super-app as both a model and a flashpoint in debates over closed national internets. Worth noting for narrative awareness rather than as a story requiring response.