Anthropic released Opus 4.8 this week and the framing of the launch is itself the news: rather than a raw capability jump, Anthropic is emphasising reliability and honesty, claiming Opus 4.8 is roughly four times less likely than Opus 4.7 to let flaws in its own code go unflagged. Fast mode is also said to be three times cheaper. Claude Code ships Opus 4.8 with high-effort defaults, dynamic workflows in research preview, and broader plugin and MCP support. For a comms team, the meaningful shift is that the leading edge of model marketing is moving from “smarter” to “more truthful” — a frame the press will pick up.
This week confirmed the new centre of gravity in private capital. OpenAI closed a $122 billion Series round — the single largest private venture round in history — while Anthropic raised $30 billion in a Series G at a $380 billion post-money valuation, the second-largest private venture deal ever recorded. Anysphere (maker of Cursor) raised $2.3 billion in Series D, tripling its valuation to $29.3 billion in five months, and ElevenLabs raised $500 million at an $11 billion valuation. Reports also surfaced that OpenAI is preparing an IPO filing at a roughly $730 billion valuation. The combined effect is that AI is no longer a category inside venture capital — it has become the venture market.
Pope Leo XIV released his first encyclical, “Magnifica Humanitas,” on Monday — a 43,000-word document focused largely on artificial intelligence and the people building it. The Pope catalogues current abuses of AI, warns of a “Babel” of synthetic content, and asks executives and engineers to slow down. It has already become a board-level governance prompt: Not a Bot frames it as the moment AI ethics moved out of policy and into corporate org design. Expect questions from reporters on Tencent’s position relative to the encyclical’s themes (labour displacement, synthetic media disclosure, child safety).
On 7 May, the Council, Parliament and Commission reached a provisional agreement on the Digital Omnibus on AI — the first amendments to the EU AI Act since adoption. Obligations on use-based high-risk systems (Annex III) move from August 2026 to December 2027, and product-regulated high-risk systems (Annex I) move to August 2028. The Omnibus also adds two new prohibitions, effective 2 December 2026: AI used to generate or manipulate non-consensual intimate material, and AI used to generate child sexual abuse material. Formal adoption is expected in June with publication in July. Comms takeaway: high-risk compliance pressure is delayed, but the trust-and-safety conversation just hardened.
Bloomberg reported this week that China’s informal guidance restricting overseas travel for AI researchers has hardened into a formal pre-approval requirement, and now extends beyond state-linked institutions to private companies including Alibaba and DeepSeek. Researchers, startup founders and senior executives are affected, regardless of destination. The signal for global communications teams is significant: any narrative about “Chinese AI talent leaving for the US” will be reframed by reporters around state control, and any executive travel will now carry an extra layer of approval and explanation.
ClickUp replaces 22% of staff with AI agentsThe largest single AI-driven workforce reduction at a major SaaS vendor this year. Sets a new benchmark for press coverage of AI layoffs.
A dedicated Substack inbox API was not available in this run, so the items below were pulled from your Inoreader subscriptions to your Substack feeds (the *Substack: Paid and *Newsletters folders). Every newsletter you actively subscribe to and that published in the last 24 hours is summarised here.
Social feeds are now full of posts that all sound the same because of AI-assisted writing — same cadence, same emoji-free structure, same closing question.
Mollick argues that the productivity case for LLM-written content is now obvious, and the harder question is how to keep authentic, recognisably human voice in your channels.
Practical guidance: use AI for research, structure and editing, but keep the actual voice of the writer in the final draft. Useful frame for your team’s LinkedIn and op-ed pipeline.
Opens with an anonymous tech-company exec: a thousand engineers all using Claude Code, individual productivity is up, organisational throughput is “one plus one plus one plus one equals one-and-a-half.”
Echoed by Uber COO Andrew Macdonald and others: most AI gains are being absorbed by review overhead, coordination costs, and integration debt.
Azhar’s implication for comms: stop measuring AI ROI in tool spend; start measuring in shipped product and customer outcomes. Reporters will increasingly ask CFOs and CMOs to defend the gap.
Long interview with Kevin Delaney and Jim Friedlich on rebuilding the news consumption experience around AI-native discovery, summarisation and trust signals.
Worth reading as a benchmark for what next-generation digital newsrooms expect from press releases and embargo flows.
Hands-on walkthrough of the new Claude Code desktop app. Useful concrete frame: when to drop the terminal and stay in the GUI, and where the terminal is still faster.
Notes that the desktop app finally makes Claude Code accessible to non-CLI users on your team — relevant if you want comms staff using it for drafting and research.
End-to-end build log: 5 sub-agents in Claude Code, each owning a specific data source, coordinated by a planner agent.
Concrete reference architecture for anyone wanting to use Claude Code orchestration to build internal dashboards — including, say, a Tencent media-coverage dashboard.
The clearest single-page summary of the Opus 4.8 launch — alignment, pricing, what changed in Fast mode, how it compares to the unreleased Mythos preview.
A solo developer’s open-source project outperformed a Meta release on a key agent benchmark. Pair with the ClickUp story to see how the “small team beats big team” narrative is hardening.
Weekly catch-up covering the Musk/OpenAI litigation result, Google I/O announcements, and OpenAI’s reported solution of long-standing Erdős conjectures with a new prover.
Covers the Opus 4.8 release and the explosion of new Claude Code skills, sub-agents and plugins that landed alongside it. Also introduces the new industry buzzword that’s replacing AGI in the discourse.
The lab everyone has written into a roadmap is about to answer to public markets. Implications for comms: expect tighter, more lawyer-checked messaging from OpenAI starting almost immediately.
Argues Opus 4.8’s “engineer-like” behaviour shows up in how it flags its own mistakes. Also covers Google solving long-standing math problems and Eli Lilly’s gene-editing news.
From Your AI Feeds
Inoreader AI Folder.
More than 80 items hit your AI folder in the last 24 hours. The narrative was completely dominated by the Anthropic Opus 4.8 release, the OpenAI/Anthropic funding numbers, and the Vatican encyclical. Below are the highest-signal items not already covered above.
Detailed case study of how Braintrust engineers use Codex on GPT-5.5 to run experiments and ship features faster. Useful internal reference for any team thinking about an “always-on engineer assistant” pattern.
The IT consultancy says Codex has compressed requirements analysis from weeks to hours and reorganised how project teams take software through delivery. Important for understanding how big systems integrators are framing AI to clients.
Warp is using GPT-5.5 and OpenAI models to coordinate coding agents across local, cloud and open-source workflows. Notable because it represents one of the more sophisticated public uses of multi-environment agent orchestration.
Joint write-up with Thrive and Crete on a Codex-built tax agent that improves its own accuracy over time. Worth reading as a template for self-improving back-office workflows.
Detailed engineering walkthrough of how to record Google Meet from the user’s own machine without sending a bot to the meeting — the same architecture pattern Granola uses. Useful background reading if you are evaluating whether to keep Granola, switch to an open-source alternative, or have your team build something internal.
A working developer’s comparison of self-hosted desktop meeting recorders — trade-offs around audio capture, transcription, calendaring and storage. Pair with the article above if you want a full picture of the Granola-alternative landscape.
Jqwik, a popular Java test engine, has added an explicit policy in its contributing guide telling AI coding assistants not to use the codebase. Reflects a wider open-source backlash against LLM training scraping. Worth tracking as a data-rights and licensing story.
Sharp critical column arguing that the SpaceXAI vehicle’s narrative around price discovery mirrors crypto-era ICO patterns. Useful as a stress test for how skeptical financial reporters are starting to write about AI-adjacent IPOs.
Major new dataset covering ASR (speech-to-text), translated ASR, and OCR trigrams from global television news. Genuinely useful for any media-monitoring or narrative-analytics work; you can run queries about how Tencent is being described on broadcast TV around the world.
Analysis of what appears to be a largely AI-generated recruitment ad on Russian state TV. Concrete case study in how synthetic media is now being used inside state broadcast pipelines.
Research paper on a multi-agent architecture for extracting predictive intelligence from Polymarket signals. Adjacent to the kind of OSINT tooling some communications teams now use for issue monitoring.
Short video segment on a new neuromorphic chip that could eventually displace GPUs for some AI workloads. Largely future-facing, but a useful explainer if a reporter asks about post-Nvidia AI hardware.
Reflection on what has actually changed in cultural-heritage AI work over the last year using Gemini 2.5 — one of the more grounded “real applications” pieces of the week.
Discoveries
Workflows & Tool Watch.
Perplexity Personal Computer on Mac
The most interesting Mac-native AI launch of the week. Personal Computer connects your local files, your apps, and Comet to Perplexity’s Computer agent. You can ask it to read a document on your desk, look something up in a tab you have open, dictate a reply, and have it actually act — all without uploading anything to a third party. For your current toolkit (Apple Notes, Drafts, DEVONthink, Eagle), this is the closest a third-party AI has come to feeling like a true macOS assistant. Worth a one-week trial alongside Claude.
Claude Code orchestrating n8n via the n8n-mcp server
The open-source n8n-mcp server has crossed 20,000 GitHub stars and now lets Claude Code (and Cursor, Windsurf) actually build n8n workflows from a natural-language description — not just trigger them. For your stack this is the single most important change of the month: you can describe a media-monitoring or escalation workflow in plain English to Claude Code, and it will produce a working n8n scenario for you to review and deploy. Pairs naturally with your existing Make, Zapier and Keyboard Maestro work, and lowers the barrier for non-developer comms staff to spec automations.
Sub-agent pattern for internal dashboards in Claude Code
The AI Maker’s build log of a financial dashboard with five sub-agents inside Claude Code is the clearest current template for any team that wants to spin up an internal data dashboard quickly. The pattern: one planner agent, then specialised sub-agents for each data source (RSS, Slack, internal API, etc.), coordinated via Claude Code skills. Directly applicable to a Tencent media coverage dashboard, an internal issues tracker, or a daily team digest.
Obsidian MCP servers maturing into three competing architectures
The Obsidian MCP landscape has settled into three patterns: native-plugin (the Obsidian MCP Plugin, which runs inside Obsidian and can read the knowledge graph, Dataview queries and backlinks); REST-API-bridge (most of the others); and pure-filesystem servers that just read your markdown. If you want Claude or Claude Code to actually understand link structure in your vault, the native-plugin route is now the right default. Useful if you eventually want Claude Cowork to operate against your Obsidian notes the same way it operates against your file system.
RelevantObsidian · Claude Cowork · MCP · DEVONthink
Anthropic has started grading prompts in Claude. This is essentially a quiet onboarding tool: it will improve over time the way Grammarly did, and is likely to be the path through which junior staff get coached on prompt quality. Recommend piloting with one or two writers and one analyst.
Claude Managed Agents self-hosted sandboxes + MCP tunnels
For any sensitive Tencent workflow you don’t want running on Anthropic infrastructure, the new self-hosted sandboxes (public beta) let tool execution move to your own infrastructure or to providers like Cloudflare, Daytona, Modal or Vercel, while orchestration stays with Anthropic. MCP tunnels (research preview) extend the same pattern to private MCP servers behind a firewall. This is the cleanest architectural option for “keep the data inside, keep the intelligence outside.”
Botless meeting recording as a Granola alternative
The Recall.ai blog has now published two engineering write-ups (above) on building a botless, on-device meeting recorder. If you ever want to move off Granola without losing the in-meeting AI experience, this is the technical roadmap. Particularly relevant for sensitive Tencent calls where you may not want a third-party bot in the room.
n8n’s 5 May release added a Microsoft Agent 365 Trigger node, letting n8n agents appear as native members of a Microsoft 365 team. If you have stakeholders or counterparts on M365, this is a way to drop an automated workflow into their working environment without asking them to learn another tool.
This week’s HN thread on Claude Code, Codex, skills, MCP and orchestration converged on four practical points: workflows beat demos, verification is the actual bottleneck, skills beat prompts, and orchestration matters more than raw model power. The implication for your team: keep rules and skills close to the project, prefer many short prompts over one giant prompt, and never trust an agent loop without an explicit verification step.
The Bloomberg, Tech Times and AI Valley pieces all explicitly name Tencent inside the broader frame of Chinese AI companies whose researchers are now subject to overseas-travel pre-approval. While the on-the-record confirmations cite Alibaba and DeepSeek, Tencent is read as part of the same category. Reporters will likely call this weekend asking whether Hunyuan researchers are affected; a pre-cleared line will save time.
Two weeks after the Q1 results, Western coverage is still anchored on Martin Lau’s disclosure that 2026 AI investment will exceed RMB 36 billion — more than double 2025 — and on his acknowledgement that monetisation will take time. This week’s funding numbers from OpenAI and Anthropic give that figure useful context: Tencent is now spending at roughly the same order of magnitude as the biggest Western frontier labs.
Hy3 preview is being framed by TECHi, ChoZan and DAO Insights as the first major model out of Tencent’s rebuilt pre-training and RL infrastructure since February: 295B total parameters, 21B activated, 256K context, fast-and-slow-thinking fused MoE. If the Opus 4.8 launch becomes a benchmark for “honest, reliable AI” this week, Hy3 messaging may want to lean into the same frame in upcoming briefings.
There is no direct Tencent reference in “Magnifica Humanitas,” but Pope Leo XIV’s focus on labour displacement, synthetic media, and child safety overlaps directly with three current Tencent AI narratives. Worth pre-cleared lines on: (1) WeChat and Yuanbao integrity controls; (2) Tencent’s position on AI-generated content labelling in games and short video; (3) child-safety guardrails inside consumer AI products.