AI Intelligence Report
Top Stories
Hyatt deploys ChatGPT Enterprise across its entire global workforce
Hyatt today rolled out ChatGPT Enterprise company-wide — one of the largest hospitality deployments of OpenAI’s business product to date. Employees in finance, marketing, business development, real estate, product and engineering, and customer experience now have access to GPT-5.4 and Codex. Rather than going all-in on one embedded assistant, Hyatt has taken the approach of building a central data/AI layer first and letting staff choose among multiple models through it. Early wins include natural-language search on Hyatt.com and agentic tools for the group-sales team. A useful case study for any communications leader watching how large, legacy-heavy enterprises are actually operationalising AI.
Anthropic launches Claude Design — and Figma’s stock takes a hit
Anthropic released Claude Design, a new product inside Claude that turns a written brief into polished mockups, decks, one-pagers and prototypes. You describe what you need, Claude builds a first draft, and you refine it by chatting, dragging sliders, or clicking directly on elements. When connected to a team’s design system, it applies the brand automatically — and exports to PDF, PPTX, Canva, or a shareable URL. It’s powered by Opus 4.7 and is in research preview for Pro, Max, Team and Enterprise users. Figma closed down 7.3% the day of the announcement. For a comms team that generates a lot of one-off visuals, this is worth a serious look.
The “Mythos” moment: Anthropic keeps its most capable model off the market
The story dominating AI conversation this week is Anthropic’s decision not to publicly release Claude Mythos, which it describes as its most powerful model ever. Roughly fifty partner organisations have been given gated access under a program called Project Glasswing. The stated reason: Mythos’s cybersecurity offensive capabilities are too dangerous for general release. Scientific American has a plain-English explainer; Gary Marcus argues the threat was oversold. Either way, it has re-opened the regulation debate — precisely at a moment when the EU is finalising trilogue on the AI Act’s high-risk provisions and China has just issued its new Trial Guideline on the Ethics Review and Service of AI.
Perplexity Personal Computer for Mac is now rolling out — turn a Mac mini into an always-on agent
Perplexity Max subscribers can now install Personal Computer, an agent that lives on your Mac (macOS 14 Sonoma or later), reads your files, and actually does things — tidying your Downloads folder, working through a to-do list, comparing local documents against what’s on the web. Perplexity is explicitly recommending it on a Mac mini left running 24/7. Under the hood it can spin up teams of sub-agents across 20+ frontier models. Every action is auditable, reversible, and sandboxed, with a kill switch. Perplexity’s ARR has jumped to ~$450M, a 50% month-over-month surge driven by this agentic pivot.
Compute wars: Broadcom expands chip deal with Anthropic to 3.5 gigawatts
Broadcom has struck an expanded deal giving Anthropic access to roughly 3.5 gigawatts of compute capacity running on Google’s TPUs. Separately, OpenAI, Anthropic and Google are coordinating on measures to slow Chinese rivals’ ability to extract answers from US frontier models and train on them — a significant shift from the “rivals won’t cooperate” narrative. Anthropic’s annualised revenue is approaching $19B; OpenAI has crossed $25B and is signalling possible IPO groundwork.
AI News Roundup
Substack Highlights
Inoreader AI Folder
15 articles landed in the AI folder in the last 24 hours. Key items below — overlap with Substack Highlights and Top Stories noted where relevant.
OpenAI helps Hyatt advance AI among colleagues
Enterprise-wide ChatGPT deployment at Hyatt, using GPT-5.4 and Codex across finance, marketing, business development, real estate, product and customer experience. Notable as a hospitality-industry benchmark.
Inside the AI boom — jobs, jargon & jittery uptime
Azeem Azhar’s Monday data brief on junior-role displacement, inference-cost economics, and Claude’s infrastructure strain.
Harper Reed — Note #729: “Got some new speakers for my desk”
A short personal blog note, not AI-related; included only for completeness. Skip unless you want the distraction.
Opus 4.7 nicknamed “Gaslightus 4.7”
A roundup of the first week of community feedback on Claude Opus 4.7. The nickname refers to Reddit complaints that the model confidently delivers wrong answers without hedging. Also covers Honor’s marathon robot.
Two free 3D world models dropped this week
Tencent released HY-World 2.0 and Alibaba released a competing model on the same day — both open-source, both turn text or photos into navigable 3D scenes importable into Unity or Unreal. Also introduces the “trust battery” method for gauging how much autonomy to give AI employees.
Claude Design and Word add-in move Anthropic into the workflow layer
Analysis of Anthropic’s one-two punch: a consumer-visual product (Claude Design) plus a Microsoft Word add-in aimed at enterprise document workflows.
Anthropic launches Claude Design
Launch coverage with Adobe’s Creative Cloud agent-workflow news as the secondary angle.
AI Workflows & Tool Watch
Claude Code “Routines” — cron for AI agents
Anthropic has shipped Claude Code Routines: scheduled cloud automation without a DevOps stack. You can trigger a Claude Code agent on a schedule, from a GitHub event, or via an API call — using the same terminal workflow you already use locally. Useful for scheduled drafts, newsletter digests, or monitoring tasks that previously required n8n or a Zapier Pro plan.
MCP tool results can now pass 500K characters (up from ~25K)
Claude Code’s April point releases (v2.1.89–v2.1.92) added an opt-in `anthropic/maxResultSizeChars` annotation that lets MCP servers return results up to 500,000 characters without truncation — a big deal if you want to hand Claude an entire database schema, a long transcript, or a full Notion page. Translation for non-engineers: Claude can now actually read the big documents you hand it from other tools, end-to-end, instead of being fed a chopped-up version.
Claude + Obsidian via MCP — your vault becomes a live workspace
The Obsidian community’s most-discussed workflow this month is connecting Claude Code to your vault through MCP. Claude can then read, search and modify notes autonomously. The “claude-obsidian” knowledge engine goes further: when you ingest a new source, it creates entity and concept pages, cross-references every existing note, flags contradictions and updates the index. Useful for journalists and researchers drowning in clippings.
Perplexity Personal Computer on a spare Mac mini
If you have an old Mac mini sitting unused, Perplexity is explicitly positioning Personal Computer as an always-on assistant that can run on one. It organises your downloads, works through to-do lists, and compares local files against information on the web — all through teams of agents across 20+ frontier models. Every action is reversible and auditable, with a kill switch.
Hermes Agent — a persistent AI assistant that lives on your server
Featured this week by One More Thing in AI. Unlike a chat window that forgets you, Hermes is installed once on a server (or a Mac mini) and stays running. It integrates with Slack, Discord and Signal — so you can ping it from anywhere — and has sophisticated scheduling and delegation logic. Five sandboxing backends mean you can restrict what it can actually do on your machine. A good candidate for a “team concierge” bot for global-comms coordination across time zones.
“ui-ux-pro-max” — a fix for Claude Design’s sameness problem
If you try Claude Design and it produces the default teal-and-serif aesthetic everyone on Reddit is complaining about, an open-source Claude skill called ui-ux-pro-max ships 50+ styles, 161 colour palettes and 99 UX guidelines. Install it as a skill and Claude builds a proper design system first, then generates your screens. Useful for producing visuals that actually feel on-brand.
n8n is pulling ahead of Zapier for agent workflows
Zapier has shipped Agents + Copilot, but n8n now has 70+ AI and LangChain-native nodes — more than Zapier and Make combined — and a mature LangChain integration that lets you build multi-step agents with memory and RAG directly in the visual builder. If you’ve been on the fence about self-hosting, this is the month to revisit it.
Keyboard Maestro + MacWhisper + GPT-4o-mini for voice-to-clean-text
A popular Keyboard Maestro Discourse macro this week: dictate into MacWhisper, pipe the raw transcript into GPT-4o-mini through Apple Shortcuts, and have Keyboard Maestro paste the cleaned-up version wherever your cursor is. For anyone drafting messages on the move, this is the best currently-shipping version of “speak, edit, paste” on macOS.
Claude Code plugins marketplace crossed 2,849 skills
Third-party marketplaces like tonsofskills.com (423 plugins, 2,849 skills, 177 agents) and SkillsMP (800,000+ skills across Claude Code, Codex, Gemini CLI, Cursor) have become the easiest way to discover pre-built AI workflows. Of interest this week: executive-assistant-skills and token-optimizer.
Tencent Mentions
Towngas × Tencent strategic partnership signed today in Hong Kong
The Hong Kong and China Gas Company and Tencent signed a strategic partnership agreement today in Hong Kong to drive “Energy + Tech” smart digital transformation. AI innovation is a named pillar of the agreement, alongside cross-divisional efficiency and quality gains. Given your Hong Kong base, this is a useful anchor story for regional AI narrative.
Tencent open-sources HY-World 2.0 — #1 on Stanford’s WorldScore for open models
Released April 16, HY-World 2.0 turns text prompts or photos into navigable 3D scenes that can be imported straight into Unity or Unreal. It currently ranks first among open-source world models on Stanford’s WorldScore benchmark. Alibaba released a competing model the same day, which The Neuron/Newsletters on AI flagged as a coordinated Chinese open-source push.
Tencent Cloud launches 2026 “AI CAN DO IT” global game-dev hackathon at University of Hong Kong
Tencent Cloud kicked off the 2026 AI Game Development Hackathon at HKU on April 16, with themes around social impact, cultural expression, and narrative innovation. Useful proof point for communications messaging on Tencent’s AI-for-creators positioning.
Hunyuan 3.0 launch window is now — and the WeChat AI agent is next
The Hunyuan 3.0 launch commitment made on the Q4 earnings call on March 18 covered April 2026. The QClaw WeChat AI agent, inspired by OpenClaw, is also in the roadmap — positioned as a mini-program layer that can handle file management and PC control on behalf of WeChat users. Watch for the launch communication around both.
AI compute prices rise across China — Tencent joins Alibaba, Baidu, Zhipu
TrendForce reports Tencent has joined Alibaba, Baidu and Zhipu in raising AI compute prices — a signal that the demand/supply pressures showing up in Anthropic’s uptime numbers are also being felt in China. Worth keeping on the radar for any market-commentary questions.
Tencent Video CEO: first blockbuster AI-generated drama expected H2 2026
Top executives at Tencent Video (and iQiyi) have told Yicai that the first premium, long-form AI-generated drama in Chinese will debut in the second half of 2026. One production team reportedly has a title ready. This is a story worth tracking for its narrative value in any global AI-in-media conversation.