AI Intelligence Report.
Your custom briefing on AI covering news, Substack newsletters, RSS feeds, Reddit, workflows, and everything else that matters today in AI.
Top Stories.
OpenAI files confidentially for an IPO — valuation could top $1 trillion
OpenAI submitted its confidential S-1 to the SEC, working with Goldman Sachs and Morgan Stanley, and is targeting a Q4 2026 public listing — possibly as early as September. The company’s most recent private round closed at an $852 billion post-money valuation (up from the $730 billion pre-money figure floating in newsletter inboxes this week), and analysts expect the listing could push it past the $1 trillion mark. For corporate communications teams, this means the lab everyone has written into their AI roadmap is about to answer to public markets — with quarterly earnings, audited financials, and far more disclosure than the world has ever had on a frontier AI lab.
Read articlePope Leo XIV’s first encyclical warns AI must be “disarmed”
Pope Leo XIV used the first major theological text of his papacy — a 42,300-word encyclical called Magnifica Humanitas, released May 25 — to call AI a “new industrial revolution” that requires vigilance and governance before it reshapes society beyond human control. He characterised AI as reflecting creator biases and lacking genuine empathy, and even cited Gandalf from Lord of the Rings in his appeal for restraint. The encyclical has become this week’s dominant cultural story in AI: it’s leading newsletters from The Neuron, Future Tools, and Newsletters on AI, and has triggered a fresh wave of debate about regulation, jobs, and the role of large platforms.
Read articleTencent goes “agent-first” with Hunyuan T1-Vision and a rebuilt agent platform
At the May 21 Tencent Cloud industry summit, the company unveiled Hunyuan T1-Vision (a visual-reasoning model with 5.3% better performance and 50% faster processing than the previous version), the low-latency Hunyuan Voice model, and refreshed Hunyuan Image 2.0, Hunyuan 3D 2.5, and Hunyuan-Game. The biggest strategic signal was the rebrand of the former knowledge engine into the Tencent Cloud Agent Development Platform (TCADP), positioned as a one-stop solution for enterprise AI agents. Combined with President Martin Lau’s guidance that Hunyuan capex will more than double from RMB 18 billion in 2025 to over RMB 36 billion in 2026, this is the clearest internal-comms-ready statement of Tencent’s agentic AI direction to date.
Read articleAnthropic’s “Outcomes” adds a rubric-grading agent that re-runs work until it’s good enough
Anthropic’s Managed Agents platform now ships a feature called Outcomes: developers write a rubric describing what “good” looks like, and a second grading agent reviews each output in its own context window and sends the agent back to revise until it meets the bar. On Anthropic’s internal benchmarks, this lifted PowerPoint quality 10.1% and Word document quality 8.4% — with no underlying model change. For comms teams that lean on Claude for decks, briefings, and Q&A documents, this is a meaningful step toward AI work that polices its own quality before it lands in your inbox.
Read articleMcClatchy’s “Content Scaling Agent” sparks newsroom revolt across the U.S.
McClatchy — owner of the Miami Herald, Sacramento Bee, Idaho Statesman and dozens more — has begun deploying an internal Claude-powered “Content Scaling Agent” that summarises traditional articles and remixes them for different audiences. Journalists across the chain are withholding their bylines, and reporters at five Washington and Idaho papers walked off the job on May 26 in a one-day strike that explicitly cited AI safeguards. Union staff at the Miami Herald and Bradenton Herald have written to management saying their contract prohibits using bylines without consent. A canary-in-the-coal-mine moment for AI in newsrooms — and a useful reference point next time Tencent properties are asked about AI editorial guardrails.
Read articleNews Roundup.
Substack Highlights.
Exponential View — “Why AI isn’t showing up on your bottom line”
- Senior exec at a public tech company runs ~1,000 engineers, nearly all on Claude Code; individual productivity is up, but organisation-level results are not — “one + one + one + one = one-and-a-half.”
- Echoes Uber COO Andrew Macdonald’s recent comments — the bottleneck is workflow, coordination and rework, not raw output.
- Implication for comms leaders: be sceptical of “we deployed Claude/ChatGPT, productivity is up X%” claims unless they include team- and P&L-level numbers.
The Neuron — “The Pope’s Warning on AI’s Babel”
- Lead item on Pope Leo XIV’s encyclical and its framing of AI as a new industrial revolution.
- Notes the U.S. federal directive to address “violent anti-tech extremism” — a small but escalating signal of state attention to anti-AI activism.
- Side coverage of 3D-printable humanoid robot legs lowering the cost of physical-AI experimentation.
Future Tools — “AI vs. the Pope”
- Pulls together the Pope Leo XIV story with China’s response to U.S. chip sanctions — frames it as the week’s two big “AI sovereignty” narratives.
- Useful one-stop summary if you only have five minutes on the topic.
AI Breakfast — “Anthropic now grades your prompting”
- Plain-English walkthrough of Anthropic’s new Outcomes rubric-grading feature inside Managed Agents.
- Highlights the practical pattern: rather than tweaking models, you write a rubric and let a second agent enforce it.
- Worth keeping for next time someone asks how to make Claude output “more on-brand” at scale.
Not a Bot — “OpenAI is filing to go public at $730B”
- Reads the OpenAI IPO filing through the lens of customers: “The lab you wrote into your AI roadmap is about to answer to public markets.”
- Five other stories rounded up, including frontier-lab spending and Big Tech AI capex revisions.
Bagel Bots — “The Prompt That Acts Like A Human Ghostwriter”
- Practical prompt template for turning robotic AI prose into something readers actually enjoy.
- Designed for marketing/content teams — directly relevant for the kind of long-form posts you publish on cammacmurchy.com and LinkedIn.
The Automated — “The AI report card is in, and ChatGPT got a C-minus”
- Reports that consumers are “panic-downloading” rival search engines to escape Google’s AI Mode summaries.
- Argues ChatGPT’s grade reflects friction with everyday consumer expectations — useful counterweight to enterprise hype.
Pivot to AI — “McClatchy runs AI slop with journalists’ names on it”
- Sharp critique of the McClatchy Content Scaling Agent rollout and the byline dispute.
- Argues management framing of “content scaling” obscures who is actually accountable for the words on the page.
- Cross-reference for any internal Tencent guidance on AI-assisted publishing. See also · Top Stories.
Last Week in AI #341 — “Musk loses to OpenAI, Google IO, OpenAI solves Erdős”
- Weekly digest covering the Musk v. OpenAI ruling, fresh Google I/O AI updates, and a reported AI-assisted proof of a long-standing Erdős problem.
- Good catch-up if you want one read that touches everything from last week in a single sitting.
Inoreader AI Folder.
18 articles in the last 24 hours from your “AI” folder. The Pope Leo XIV story dominates the week; OpenAI’s IPO filing and Anthropic’s Outcomes feature are the other two major threads.
IRINN’s preview of the latest Strait of Hormuz proposed deal details
Kalev Leetaru’s machine-generated analysis of Iranian state TV channel IRINN previewing terms of a proposed Strait of Hormuz deal — a useful, near-real-time example of AI summarisation applied to non-English broadcast news.
Russia’s Zvezda TV runs largely AI-generated recruiting ad for an “information battlefield” unit
Zvezda — the Russian Ministry of Defence’s TV channel — aired a recruiting spot for its Prince Alexander Nevsky unit that is largely AI-generated. Useful evidence point for any Tencent comms thinking on AI in defence/state-media content.
PolyGnosis 2.0 — multi-agent system for extracting predictive intelligence from Polymarket signals
Research-style write-up of a multi-agent architecture that synthesises Polymarket prediction-market anomalies with OSINT to surface forward-looking signals — interesting reference for issues monitoring or anticipating cycles in your global media work.
Watch: Is brain-like computing what’s next?
Short newsletter pointer to coverage of a new neuromorphic chip described as a possible GPU replacement — connects to the broader thread of architecture rather than model size driving the next wave (echoed by WhatLLM.org’s May “frontier took a breath, architecture took the stage” analysis).
Building self-improving tax agents with Codex
Case study of OpenAI working with Thrive and Crete to build a Codex-based tax filing agent that improves accuracy automatically as it processes more returns. Concrete example to keep in your back pocket when asked “what does enterprise AI actually look like in 2026?”
Warp’s big bet on building open source with GPT-5.5
OpenAI explainer on how Warp uses GPT-5.5 plus other OpenAI models to coordinate coding agents across local, cloud and open-source workflows. One of the cleaner reference architectures for multi-agent coordination shipped publicly this quarter.
Recall.ai — “How I built a botless meeting recorder from scratch”
Engineering deep-dive on recording Google Meet without sending a bot to the meeting — uses native browser media APIs and desktop capture. Same architecture pattern Granola pioneered; relevant if you’re considering quieter alternatives to bot-based meeting capture for sensitive Tencent calls.
“Pope Leo isn’t joining Anthropic, but the memes are funny”
Light follow-up to the encyclical story — also covers 3D-printable humanoid legs from a UK robotics lab letting independent researchers run physical-AI experiments at hobbyist cost.
Workflows & Tool Watch.
DEVONthink 4.3’s new MCP server brings your databases to Claude
DEVONthink 4.3 “Herschel” ships an MCP server with nearly 60 commands — search, classify, tag, summarise highlights, get record children, find similar records, all callable from Claude or any other MCP-compatible client. Crucially, it honours your “exclude from AI” flags, redacts private information before anything leaves the machine, and blocks direct filesystem access to database packages. This is the cleanest answer yet to “how do I let Claude work inside my DEVONthink vault without losing control?” — and it’s already in your version of DEVONthink today.
Anthropic’s “Outcomes” — auto-grading rubrics for any Claude task
Define a rubric (e.g. “every slide must cite a source, every quote must include the speaker’s title, no marketing superlatives”), and a second grader agent reviews each output in its own context window, sending the agent back to revise until it passes. Anthropic reports 10.1% better PowerPoint output and 8.4% better Word document output on internal benchmarks — without changing the underlying model. Practical first use for your team: encode your house style for executive briefings as a rubric, then run it on any Claude-drafted document before review.
n8n adds native MCP support — your scenarios are now AI-callable
n8n introduced two new nodes, MCP Server Trigger and MCP Client Tool, so any n8n workflow can be exposed to Claude, ChatGPT or Gemini as a callable tool, and your n8n agents can call other MCP servers in turn. Combined with the AI Agent node’s reasoning loop, this turns n8n into a lightweight orchestration layer for the same MCP servers you already use in Claude Code — useful for moving a working Claude prompt into a scheduled, monitored automation without rebuilding the integration.
Claude Code’s May release adds plugin LSP transparency and 3-attempt MCP startup retries
Anthropic shipped a quietly important set of Claude Code reliability improvements: MCP servers that hit a transient error during startup now auto-retry up to three times instead of staying disconnected; the per-request fetch timeout for remote MCP servers now actually respects MCP_TOOL_TIMEOUT (it used to silently cap at 60 seconds); and headless/SDK startup now overlaps with pre-wait, shaving up to two seconds off slow MCP server stacks. The /plugin details pane also now lists any LSP servers a plugin provides, which makes plugin diffing far less mysterious.
Perplexity Comet’s iOS update — phone-number actions are the sleeper feature
Tap a phone number anywhere in Comet on iPhone or iPad and you now get a one-tap menu to Call, FaceTime, Message, or save it to Contacts — without leaving the browser agent. Combined with Max subscribers’ new ability to swap the underlying model between Opus 4.6 (now the default) and Sonnet 4.5, Comet is becoming a credible mobile research assistant for media work. The iPad sidebar refresh also fixes the “stretched everywhere” layout that made it awkward for in-app research.
Botless meeting recording (Recall.ai write-up) — quieter alternative to Granola-style bots
Engineering walkthrough of capturing Google Meet directly on the host’s machine using browser media APIs and Electron desktop capture — no bot in the meeting, no extra participant tile, no licence-implication of a third-party joining. Practical relevance: useful pattern for media-sensitive calls where you want a clean transcript via MacWhisper or PLAUD but don’t want a third-party observer in the room.
Tencent Mentions.
Tencent SPARK 2026 conference concludes — NDGI unveiled as world’s first cross-platform neural global illumination
The annual Tencent Games SPARK 2026 conference concluded May 27 with a long list of reveals, including NDGI (Neural Dynamic Global Illumination) — described as the world’s first cross-platform neural dynamic global illumination solution — showcased alongside Roco Kingdom integration and a mobile GPU AI optimisation collaboration with Arm. Chaos Zero Nightmare confirmed a China launch for May 28 with exclusive early access, and Ubisoft and Tencent jointly confirmed Just Dance: Party mobile launches in China on July 2.
Hunyuan T1-Vision, Hunyuan Voice and TCADP agent platform unveiled at Tencent Cloud summit
Last week’s Tencent Cloud industry summit (May 21) is still rippling through the AI press. Hunyuan T1-Vision improves overall performance by 5.3% and runs 50% faster than its predecessor; Hunyuan Voice targets low-latency conversational use; and the rebranded Tencent Cloud Agent Development Platform (TCADP) is now Tencent’s headline “one-stop solution for enterprise AI agents.” Useful talking points if a reporter asks where Tencent sits versus OpenAI, Anthropic and Baidu on agentic AI.
Tencent to spend RMB 36B+ on Hunyuan and Yuanbao in 2026 — double 2025’s outlay
President Martin Lau’s Q1 2026 disclosure that Tencent spent RMB 18 billion on Hunyuan and Yuanbao in 2025 and will more than double that in 2026 to over RMB 36 billion is now the most-cited Tencent AI number in international coverage. Compare against Microsoft’s $190B 2026 capex and Meta’s $10B El Paso data centre — Tencent’s investment is large by China standards and disciplined by global standards, which is a defensible framing in Western press conversations.
Tangential — McClatchy’s Claude-powered newsroom rollout is a useful talking point on AI editorial guardrails
No direct Tencent mention today, but the McClatchy story is the kind of editorial-AI flashpoint that media reporters will pattern-match onto any Asian tech platform that publishes news or operates a media-adjacent product. Worth being ready with Tencent’s existing position on AI-assisted editorial work (human author of record, AI-assisted disclosure, no fabricated bylines) the next time a Western media reporter asks.