AI Intelligence Report
Top Stories
Andrej Karpathy defects to Anthropic, joining its pre-training team
In one of the most consequential talent moves of the year, OpenAI co-founder and former Tesla AI chief Andrej Karpathy announced on Tuesday that he has joined Anthropic. He is starting this week on the pre-training team — the group responsible for the massive computing runs that give Claude its core knowledge — under former OpenAI colleague Nick Joseph, and will lead a new effort to use Claude itself to accelerate further model training. Karpathy framed it simply on X: “I think the next few years at the frontier of LLMs will be especially formative.” For Anthropic, which is reportedly closing a fundraising round at a $900 billion-plus valuation, the hire is both a symbolic and a substantive coup against OpenAI in the war for elite AI talent. More from TechCrunch.
Google I/O 2026: agents are now the default surface across Search, Android and Workspace
At Tuesday’s I/O keynote, Sundar Pichai declared the start of the “agentic Gemini era,” unveiling a sweeping set of AI products built around autonomous agents that act on the user’s behalf. The headliners: Gemini Spark, a personal AI agent running 24/7 with access to Workspace; Gemini 3.5 Flash, a faster, cheaper frontier model purpose-built for long-running agent tasks at roughly half the cost of comparable models; and Gemini Omni, which can generate output in any modality from any input, beginning with video and rolling out into Gemini, Google Flow and YouTube Shorts. Google also announced new TPU 8t and TPU 8i chips for training and inference, an upgraded Antigravity 2.0 coding agent, and Gemini for Science with access to 30+ life-science databases. The broader signal: Google is pushing chatbots into every product surface it owns. Tom’s Guide live blog.
Meta begins cutting 8,000 jobs in its AI pivot — and shifts 7,000 staff into AI roles
Starting Wednesday, Meta began executing the largest restructuring in its history: cutting roughly 10% of its workforce, or about 8,000 employees, while reassigning another 7,000 to AI-focused groups including “Applied AI Engineering” and an “Agent Transformation Accelerator.” The company is also not filling 6,000 open positions. CTO Andrew Bosworth framed the change as Meta’s pivot to “AI for Work.” Capital expenditure guidance for 2026 has risen by up to $10 billion to as much as $145 billion — almost all of it for AI infrastructure. The Meta cuts come amid a broader AI-driven layoff wave: Cloudflare (1,100 jobs), BILL (up to 30%), Upwork (25%) and Coinbase (14%) have all announced cuts in May. CNBC has the internal context.
An OpenAI model disproves an 80-year-old math conjecture
OpenAI announced on Wednesday that one of its models — working on the “unit distance problem” in discrete geometry, an unsolved question first posed in the 1940s — produced a construction that disproved a long-standing conjecture in the field. The achievement is being framed as a milestone for AI-driven mathematics rather than a commercial product, but it lands the same week as research from Harmonic and the Aristotle system on AI-led theorem proving. The narrow message for communicators: AI is now contributing original results to academic mathematics, not just summarising it.
Anthropic ships “MCP tunnels” and self-hosted sandboxes at Code with Claude London
At the Code with Claude developer event in London (running today and tomorrow), Anthropic shipped two enterprise-flavoured features for its Managed Agents product. Self-hosted sandboxes let companies run Claude’s agent workloads inside their own infrastructure (public beta). MCP tunnels — in research preview — let Claude connect to internal MCP servers (the standard Anthropic uses for plugging AI into corporate tools) without exposing those servers to the public internet. In plain English: companies can give Claude access to private systems without poking holes in their firewall. Useful context for any Tencent comms team thinking about enterprise AI security narratives.
AI News Roundup
Substack & Newsletter Highlights
Substack discovery for today is routed through Inoreader feeds (the dedicated Substack API tools weren’t available in this session). Items below are the AI newsletters that landed in the past 24 hours.
Inoreader AI Folder
15 of the 15 most recent items in Cam’s Inoreader “AI” folder were pulled. Most overlap heavily with today’s other sections; the remaining unique items are below.
Google is replacing search results with only the AI
David Gerard argues Google’s gradual replacement of search-result links with AI-generated summaries is now nearly complete. For a comms director, the story to follow is what happens to news-publisher traffic and to brands that rely on organic search.
How Ramp engineers accelerate code review with Codex
OpenAI’s official customer story on how Ramp’s engineering team uses Codex with GPT-5.5 to cut code review from hours to minutes. The kind of branded case study that’s increasingly hitting the news feeds — and useful internal context if Tencent comms is ever drafting a similar story about Hunyuan.
An OpenAI model has disproved a central conjecture in discrete geometry
Original OpenAI announcement of the unit-distance result — useful primary source if Cam wants to link directly to the lab rather than to coverage of the lab.
Future Tools — “Forget you”
Plus, $5 billion for AI chips. Weekly tools roundup with commentary on ChatGPT’s expanded memory.
😺 🎙️ PODCAST: Can AI Solve Math’s Biggest Mystery?
Interview with Tudor Achim of Harmonic on AI in mathematical proof.
😺 Google Search is dead. Long live agents?
I/O recap framed around agents-everywhere, with the Karpathy news as the bigger long-term story.
AI Workflows & Tool Watch
Claude Code’s plugin marketplace now shows context-cost upfront
Anthropic’s plugin browser inside Claude Code (the command-line tool that Cam’s been experimenting with for comms automation) now shows the projected “context cost” of every plugin before installation — per-turn and per-invocation token estimates. Heavy plugins that bundle multiple Model Context Protocol (MCP) servers are now flagged with their full footprint. In plain English: you can now see how much memory each Claude plugin will eat before you install it, so a comms-automation workflow won’t suddenly blow past its budget.
Claude Managed Agents: MCP tunnels and self-hosted sandboxes
The biggest enterprise-ready item from the Code with Claude London event. “MCP tunnels” let Claude reach internal company systems (Notes, DEVONthink, internal CMS, etc.) without exposing them to the public internet — the tunnel opens a single outbound connection from inside the company network. Self-hosted sandboxes let companies run Claude’s agent workloads entirely on their own servers. Practical for any organisation (Tencent included) whose security team has so far been the bottleneck on AI deployment.
Perplexity Comet’s “personal CFO” + local file editing
Comet — Perplexity’s agentic browser — got a major upgrade this week. It now runs Claude Sonnet 4.6 by default for Pro users (Opus 4.6 for Max), and can edit local files on your Mac, drive a desktop with computer-use control, and act as a “personal CFO” that watches your financial accounts. For Cam’s setup, this is the closest thing to a single agentic browser that could pull from Drafts, Apple Notes and Eagle on a Mac without separate integrations.
“Strip the ChatGPT voice” prompt
Bagel Bots is circulating a single-prompt technique that removes the robotic phrasing, hollow enthusiasm and filler that mark AI-generated copy. For a comms team that uses AI to first-draft statements, this is a small but useful tool: it functions as a final pass to make text sound human before it goes to the spokesperson.
n8n’s 280+ free AI-agent templates
The open-source automation platform n8n now offers a curated library of more than 280 free AI-agent templates — covering Gmail, Slack, WhatsApp, Notion, OpenAI, and document processing. These are ready-to-use workflows for things like inbox digesting, meeting note-takers that pull action items, and Slack channel summarisers. If Cam’s team is automating any kind of routine communications work (e.g., daily Slack digest, press-clip rollup), there’s likely a near-ready template already.
Google Antigravity 2.0 — agentic coding for marketers, too
Announced at I/O. Antigravity 2.0 is Google’s coding agent, but its new “Science Skills” framework lets agents connect to non-coding databases (e.g., scientific corpora, but the architecture is general). The interesting communications angle: any structured database — including media-monitoring or press-clip databases — could in principle be wired up as an Antigravity “skill” and queried by an agent.
Tencent Mentions
Tencent Cloud raises CodeBuddy and WorkBuddy prices up to 154%
Tencent Cloud announced a major price increase for its AI coding assistant (CodeBuddy) and office assistant (WorkBuddy), effective May 15. The enterprise flagship edition jumped from 78 yuan to 198 yuan per user per month, with the dedicated edition doubling. Tencent Cloud is positioning the increases alongside new monthly credit allowances and the CloudAgent platform. Analysts read this as the broader industry shifting from a “discount war” into a “price hike cycle” as AI inference demand outstrips hardware supply. This is Tencent Cloud’s third price hike of 2026 — worth tracking how it lands with corporate customers and whether it draws media commentary.
Hy3 preview holds the top spot on OpenRouter token measurements
Tencent’s Hy3-preview model, built by the rebuilt AI research team, has been the top-ranked model on OpenRouter token measurements since April 28 and accumulated 7.7 trillion tokens between April 23 and May 12. President Martin Lau told the Q1 earnings call that Tencent will more than double its 2025 AI spending of RMB 18 billion (~$2.5B), pushing 2026 AI capex past RMB 36 billion. Useful proof point that Tencent is competing at the frontier on both consumption metrics and infrastructure investment.
Q1 2026 results: AIM+ powers ~30% of marketing services spend
From the May 13 Q1 release: Tencent’s automated campaign management system AIM+ now powers roughly 30% of total marketing services spending from advertisers, with notable adoption among mini-game, mini-drama and mini-shop advertisers. WorkBuddy is described in the call as “the most widely used productivity AI agent service in China.” Both lines are likely future story-pegs as Tencent continues to make the AI revenue contribution more concrete.
Generated by Claude · AI Intelligence Report for Cam MacMurchy
Sources include web search, 143 Substack subscriptions, Inoreader AI folder, Reddit, Discourse forums, Hacker News, and Product Hunt.