The single biggest AI story of the week is Anthropic’s, and it has two halves. First, the money: Anthropic closed a $65 billion Series H led by Altimeter, Dragoneer, Greenoaks and Sequoia, valuing the company at $965 billion post-money — nearly triple its February valuation and, for the first time, ahead of OpenAI (last valued around $852 billion). Run-rate revenue has crossed $47 billion. Second, the product: it launched Claude Opus 4.8, with sharper judgement, more honesty about its own progress, and the ability to work independently for longer. Anthropic also signalled that a larger model, “Mythos,” is coming in the weeks ahead. For a communications team, the headline framing matters: Anthropic is now the most valuable AI startup in the world.
Google DeepMind chief Demis Hassabis is making the rounds with a sharpened version of his long-held forecast: human-level artificial general intelligence could arrive within about five years, by roughly 2029, and society is not ready for the disruption. The comments are being widely amplified across AI newsletters this morning (One More Thing in AI ran it alongside the Claude Opus 4.8 launch). For Cam’s purposes, this is the kind of “AGI timeline” quote that reporters increasingly use as a framing device in interviews — worth having a measured house line ready.
OpenAI is spending heavily on infrastructure and, for the first time in years, on hardware. It broke ground on a 1-gigawatt data centre in Michigan as part of its “Stargate” build-out, and on May 31 Sam Altman publicly announced OpenAI is hiring full-stack hardware and ML engineers to design and manufacture humanoid robots — its strongest robotics move since it shut its robotics team in 2021. The near-term focus is robots that help build infrastructure rather than household helpers. Separately, OpenAI’s frontier models and its Codex coding agent are now generally available on Amazon Web Services, widening enterprise distribution.
On May 7, EU negotiators reached a provisional deal on a “Digital Omnibus” — the first set of amendments to the AI Act since it was adopted. The practical upshot: obligations for high-risk AI systems are pushed back 16 months, from August 2026 to December 2027, giving companies more breathing room. At the same time, the rules add outright bans on using AI to generate non-consensual intimate imagery and child sexual abuse material. In the US, Colorado’s AI Act is still slated to take effect June 30, requiring developers and deployers to guard against algorithmic discrimination. Relevant for any global comms team mapping where AI disclosure obligations land first.
Code spotted by researchers points to an unreleased Anthropic product called Conway: not a new chatbot, but a persistent, always-on agent that lives in your browser or system and wakes itself up when something happens — a new email arrives, a calendar event fires, a data feed updates. It can drive Chrome to do research and fill forms, and developers can package custom tools as “.cnw” files. It is not officially announced. For Cam this is worth watching closely: an always-on agent that reacts to inbound triggers is essentially a media-monitoring and rapid-response tool waiting to happen.
Claude Opus 4.8 lands with “dynamic workflows”Anthropic’s new flagship improves agentic coding (64.3% → 69.2%) and reasoning, runs a 2.5x faster mode at the same price, and adds a Claude Code research-preview feature that lets it tackle bigger multi-step tasks.
Anthropic · May 28
02
OpenAI frontier models and Codex arrive on AWSEnterprises can now build with OpenAI’s models and the Codex coding agent inside their existing AWS environments, controls and procurement — a notable distribution win against Azure exclusivity.
Google’s Gemini 3.5 Flash reaches general availabilityFrontier-level quality at roughly 4x the speed of comparable models, a 1M-token context window and $1.50/$9 per-million pricing — Google’s play for high-volume, cost-sensitive deployments.
LLM Stats
05
Meta leans into the “Neural Band” wearableNewsletters seized on Meta’s EMG wristband — which reads muscle signals to control its Ray-Ban Display glasses with subtle finger movements — as the next frontier of ambient, screen-free AI input.
Cognition hits a $492M run-rate building DevinThe maker of the Devin AI software engineer is scaling fast, a sign that “agentic coding” companies are converting hype into real recurring revenue.
Exponential View: “We checked again. Still no bubble.”Azeem Azhar’s updated analysis argues the AI surge is a boom rather than a bubble — but flags that the sheer scale of spending has grown dramatically since last September. See also · Substack Highlights.
Exponential View
12
DuckDuckGo installs jump ~30% after Google’s AI overhaulA backlash-and-switch story: as Google leans harder into AI-generated results, privacy-minded users are trying alternatives — a reminder that product changes carry reputational tailwinds for rivals. See also · Inoreader AI Folder.
The Neuron
13
Researchers can “prompt-inject” ChatGPT with a booby-trapped web pagePermiso’s Andi Ahmeti showed that hidden instructions in a web page can make ChatGPT display an attacker’s phishing link or QR code as if it were a genuine system message — a real security concern for anyone using AI to summarise pages. See also · Inoreader AI Folder.
Pivot To AI
14
Exponential View #576: “Does AI make you dumb?”Azeem Azhar weighs the evidence on cognitive offloading — whether leaning on AI erodes our own thinking, and how “linear vs exponential” intuition shapes the debate.
A dedicated Substack inbox connector wasn’t available for this run, so newsletter coverage below is drawn from the same newsletters as they arrive in your Inoreader feeds (Exponential View, AI Breakfast, The Neuron, The Automated, One More Thing in AI and others). Coverage is limited to issues published in the last 24 hours.
A reusable prompt framework that generates headlines, CTAs, objection-handling and positioning in one pass.
Directly applicable to marketing and content drafting — a quick template worth saving to Drafts or Obsidian.
From Your AI Feeds
Inoreader AI Folder.
13 items in the last 24 hours from your “AI” folder. The flow is dominated by Anthropic (Opus 4.8, the $965B raise, the Conway leak), OpenAI’s infrastructure and robotics push, and Demis Hassabis’s AGI timeline.
OpenAI broke ground on a 1-gigawatt data centre in Michigan as part of its Stargate program, pitching it as jobs, community investment and expanded AI access.
OpenAI’s models and Codex agent are now generally available inside AWS, letting enterprises build with OpenAI using the cloud controls and procurement they already have.
OpenAI sets out its approach to regulation, transparency and AI safety, and states that no outside political group speaks on the company’s behalf — a notable corporate-comms posture statement.
Security researcher Andi Ahmeti (Permiso) demonstrated that hidden text on a web page can make ChatGPT show an attacker’s phishing link or QR code as if it were ChatGPT’s own system message. A clear cautionary tale for using AI to summarise untrusted pages.
Hassabis reiterates that human-level AI could be roughly five years out and that society isn’t ready. The same issue spotlights a Google researcher using AI to pursue cancer cures.
Coverage of the leaked Conway always-on agent that runs in the background and reacts to emails, calendar events and webhooks. (Reported across both AI Breakfast and the consolidated “Newsletters on AI” feed.)
DuckDuckGo downloads surged after Google’s AI search overhaul; the same issue covers Anthropic’s $965B valuation and OpenAI’s robotics ambitions. (Appears twice in your feed via The Neuron and the aggregated newsletter list.)
Signal AI’s 2026 Sponsorship Benchmarking analysis argues most organisations track sponsorship spend but not its reputational impact — directly relevant to communications measurement and media-influence work.
Discoveries
Workflows & Tool Watch.
DEVONthink now talks to Claude over MCP
DEVONtechnologies confirms DEVONthink supports the Model Context Protocol — the plumbing that lets an AI assistant like Claude or Codex read and act on your documents. In plain terms: Claude can now summarise, re-style or describe items in your DEVONthink database directly, and users report it cuts knowledge-management overhead from 30–40% of their time to under 10%. This is one of the most directly useful items for your daily DEVONthink-centred setup.
The Obsidian CLI turns the Claude + Obsidian pattern into a first-class workflow
Obsidian 1.12.7 shipped an official command-line interface that lets AI tools treat your vault as live, persistent context rather than a one-off copy-paste. Practically, that means an assistant can read across your notes, write back into them, and use your own knowledge base as memory — the same “vault as second brain for AI” pattern PKM writers have been chasing, now without the hacks.
n8n adds native MCP nodes — AI models can now call your automations as tools
The automation platform n8n introduced two native MCP building blocks: an “MCP Server Trigger” and an “MCP Client Tool.” Together they let models such as Claude or GPT safely call your n8n workflows as part of their reasoning — so an AI agent could, for example, trigger a media-monitoring or publishing flow on its own. A clean bridge between conversational AI and the kind of cross-app automation you’d otherwise wire up in Make or Zapier.
Perplexity’s “Personal Computer” for Mac connects local files and apps to the assistant
Perplexity rolled out Personal Computer on Mac for everyone, securely linking your local files, apps and the Comet browser so the assistant can work with the context already on your desktop — not just the open web. It also added enterprise finance connectors and admin controls. For research-heavy comms work, this is the kind of “answers grounded in my own material” capability worth testing against Claude’s desktop app.
Claude Opus 4.8 “dynamic workflows” in Claude Code
Alongside the model launch, Anthropic added a research-preview feature called dynamic workflows that lets Claude break a large task into steps and carry it through inside Claude Code with less hand-holding. If you experiment with Claude Code or Cowork mode, this is the headline reason to try Opus 4.8 — it’s aimed squarely at longer, multi-step jobs.
Watch: Anthropic’s “Conway” as a comms-monitoring engine
The leaked Conway agent reacts to inbound triggers — a new email, a calendar event, a data feed updating — and can drive a browser to research or fill forms. It’s unreleased, but the shape is exactly what a global comms team wants: an always-on watcher that can catch a breaking mention and kick off a draft response without anyone clicking “go.” Worth keeping on your radar for the moment it ships.
No fresh Tencent-specific wire stories surfaced in the past 24 hours
There were no brand-new breaking Tencent AI headlines in today’s search window. The items below are the current Tencent AI talking points most likely to come up in a reporter’s question — useful to have ready rather than newly published today.
Hy3, Tencent’s first major model since recruiting former OpenAI scientist Yao Shunyu, is in preview with gains in complex reasoning and coding. The Hunyuan 3D engine has launched globally, generating high-quality 3D assets from text, images or sketches in minutes — directly relevant to gaming and creator tooling.
Tencent plans to more than double its AI investment this year across Hunyuan, the Yuanbao assistant and new products — the clearest signal of strategic commitment for analysts and reporters.
Revenue reached RMB 196.5B (≈US$29.0B, +9% YoY) and net profit rose 21% to RMB 58.1B. Marketing-services revenue grew 20% on AI-powered recommendation upgrades, and the AIM+ campaign tool already drives ~30% of marketing spend. The WorkBuddy productivity agent and Hy3 were both highlighted as emerging products.