Michael Kratsios, Director of the White House Office of Science and Technology Policy, publicly stated that Moonshot AI “distilled” Anthropic’s Fable model to build Kimi K3 — the 2.8-trillion-parameter Chinese model that recently topped a major coding leaderboard. Distillation means training a new model on the outputs of a rival’s model, effectively capturing billions of dollars of training work by querying it through an API. Kratsios called it “large-scale covert industrial distillation” and separately alleged Moonshot accessed restricted Nvidia GB300 chips through Thailand, sidestepping US export controls. Independent statistical analysis by Redwood Research found K3 identifies itself as “Claude” far more often than chance would explain, though researchers note innocent explanations exist (Claude conversations are all over the public web). This is an allegation, not a legal finding — but it moves the US-China AI race from benchmarks into accusations of theft, and it lands four days before K3’s weights are released free to the world.
OpenAI introduced Presence, a platform that connects AI agents to a company’s internal systems with shared policies, permissions, guardrails, and evaluations, so agents behave consistently across voice and chat. It targets customer support, outbound sales, and sensitive internal workflows; BBVA, SoftBank, and IAG are early explorers. Notably, OpenAI is selling it as a deployed service with hands-on pricing rather than self-serve software — a move into consulting territory that follows Microsoft’s Frontier Company and Anthropic’s enterprise push. The enterprise agent category is now a four-way fight between OpenAI Presence, Google’s Gemini Enterprise, Meta’s Business Agent Platform, and the NVIDIA-ServiceNow Project Arc partnership. The competition has shifted from whose model is smartest to whose deployment layer enterprises trust most.
OpenAI unveiled a 3.2-gigawatt AI data center campus on 1,400 acres in Effingham County, Georgia — roughly the output of three large nuclear reactors, dedicated to one company’s AI. Georgia Power will deliver electricity in phases from 2028 through 2032. OpenAI says it will fully fund the required grid infrastructure so local ratepayers don’t subsidize the project, will use closed-loop cooling to limit water consumption, and pledged $80 million in community benefits plus $71 million in Codex credits for Georgia students. The timeline is the real story: power ordered today arrives in 2028 at the earliest, which means today’s compute scarcity — the reason Anthropic is rationing premium access and negotiating compute from rivals — cannot be fixed quickly at any price.
DeepSeek V4’s stable release lands today, July 24, ending the preview churn that kept cautious enterprises away, and Kimi K3’s free open weights follow on July 27. Together they form the largest concentration of open-weight releases the industry has seen — a model that beat Claude Fable 5 on a coding leaderboard will become self-hostable at infrastructure cost only. But the White House distillation accusation now layers a provenance question on top of every technical evaluation: enterprises with strict IP compliance or government contracts may pause, while individual developers and startups race ahead. Expect adoption to split along exactly that line. Alibaba has also teased Qwen3.8, a 2.4-trillion-parameter model that will ship as open weights.
DeepSeek V4 stable release arrives todayThe stable build ends months of preview churn and sets the price floor — roughly $0.44 per million output tokens — that the whole industry is measured against.
LLM Stats · Jul 24
02
OpenAI launches Health in ChatGPTEligible US users can securely connect medical records and Apple Health for personalized health insights — a significant push into regulated consumer territory.
China’s AI agent rules became enforceable July 15The “Implementation Opinions” create the world’s first dedicated regulatory category for AI agents — directly relevant to every Chinese platform shipping agentic features.
The dedicated Substack tools were not available in this run, so today’s newsletter coverage was pulled from your Inoreader newsletter feeds instead. All newsletters that published in the past 24 hours are covered below.
Mollick’s summer 2026 edition of his most popular recurring guide. The big shift: “using AI” no longer means chatting — it means agentic systems that combine a model’s brains with tools, doing hours of human work in one go. His shorthand: an agent is “an AI with a computer.”
For casual, low-stakes use, any free model is fine. For high-stakes questions (medical, legal), pay for the top models — Claude’s Opus and Fable or ChatGPT’s GPT-5.6 Sol — set to at least “High” thinking. For real work, he says there are only two serious choices: Claude or ChatGPT at $20/month.
He walks through Claude Cowork and ChatGPT Work (agents on the AI company’s computers) versus Claude Code and Codex (agents on your own machine). A cautionary tale on permissions: given the same email task, Claude drafted a reply for approval while ChatGPT actually sent the email — because he had previously granted it send rights. His rule: keep approval settings on for anything that “sends, spends, or deletes,” partly to guard against prompt injection.
On Google: no leading frontier model and nothing close to Codex/Code, so not a primary system right now — but Gemini Notebook (formerly NotebookLM) remains the best research interface, and Gemini Omni’s direct video editing is remarkable. Claude still has no image generator.
Azhar and Petrovic, who visited Moonshot AI and Alibaba in China earlier this year, examine whether Kimi K3 breaks the economic case for Western AI. Some claim it does, since it completes frontier-grade tasks at far lower cost — Microsoft engineers are reportedly testing K3 inside Copilot. Their view: not so fast, and the paid post works through what happens next.
Key data points: open models are now just 4-7 months behind the frontier in cyber capabilities (down from 6-10 months in 2025), and they estimate Chinese labs extract 4-7x more performance per unit of compute than US labs.
Also flagged: Alibaba’s Qwen3.8 (2.4 trillion parameters) is coming soon as an open-weight release — a change from its previous closed launch.
Ben’s daily roundup leads with the story everyone in AI was talking about: the White House accusation that Moonshot AI built Kimi K3 by distilling Anthropic’s Fable — the “caught cheating” framing capturing how the industry read the news.
Lands amid his week’s coverage of K3’s leaderboard win (“Better design than Fable” on Jul 21), underscoring how quickly the narrative flipped from “China caught up” to “did they copy?”
A live-session recap on how non-coders can ship AI-made apps, games, and tools — “without pretending GitHub is obvious.” A friendly on-ramp if you want to build rather than just chat.
From Your AI Feeds
Inoreader AI Folder.
11 articles landed in your “AI” folder in the last 24 hours. The Moonshot-Anthropic distillation story dominated, with the rest split between OpenAI product news and Gemini experiments in cultural heritage.
Amy Castor and David Gerard report that the practice of buying physical books, scanning them, and destroying them for training data — first documented at Anthropic during last year’s copyright litigation — has spread across AI vendors desperate for more high-quality text beyond the Common Crawl. A sharp read for anyone thinking about content provenance and publisher relations.
Kalev Leetaru published two more entries in his series using Gemini Deep Research to identify and contextualize museum artifacts — a Civil War map and a fragile folded 1939 Estonian map — from nothing more than a phone photo. A vivid demonstration of multimodal AI for archives, and an idea worth borrowing for any organization sitting on historical assets. The Estonian map piece is here.
Google’s three-way Gemini split, ChatGPT ads, and more.
See also · Substack Highlights
Discoveries
Workflows & Tool Watch.
MacWhisper 14: redesigned transcript editor, big speed boost, new cloud options
Jordi Bruin shipped MacWhisper 14 with a brand-new Editor View — edit text directly on the main screen, assign speakers, and mark favorite passages without menu-diving. Performance is noticeably faster, and 14.1 added Gladia as a cloud transcription provider plus quick access to your ten most recent dictations from the menu bar. Free upgrade for Pro users. For podcast production, this makes the transcribe-edit-publish loop meaningfully shorter.
RelevantMacWhisper · Red Book podcast · video production
Claude Cowork now handles recurring tasks — and can learn skills by watching your screen
Two Cowork upgrades worth trying this week. First, scheduled recurring tasks: from the Schedule tab (or by typing /schedule), you can have Cowork run daily briefings, weekly competitor scans, or inbox triage automatically — exactly the machinery producing the briefing you’re reading now. Second, per AI Breakfast, Cowork can record your screen while you demonstrate a task and turn the recording into a reusable custom automation skill — teaching by showing rather than writing instructions.
RelevantClaude Cowork · team coordination · media monitoring
Perplexity Comet’s July wave: Custom Skills, Model Council, and Enterprise rollout
Perplexity’s July updates to its Comet browser and “Computer” agent system include Custom Skills for automating repeating tasks, Model Council (running several frontier models in parallel and comparing answers), redesigned tab groups with live agent-activity animations, and Comet for Enterprise — in-page research, summarization, and multi-step tasks like managing email and filling forms. If you want a second agentic browser alongside Claude, this is the one to watch.
Build a newsletter production line with Claude Desktop and MCP servers
A detailed community write-up shows how to chain Claude Desktop with a set of MCP servers (plus Langfuse for tracking) into an end-to-end newsletter workflow — gathering sources, drafting, and formatting on a schedule. The architecture maps directly onto comms use cases: coverage digests, media monitoring summaries, or internal updates assembled the same way.
RelevantClaude · MCP · content creation and distribution
n8n’s community library passes 8,300 AI workflow templates
The n8n template library has grown past 8,300 ready-to-import workflows, with 7,480 in the AI category alone — support triage, invoice extraction, content repurposing, social automation, and RAG chatbots that connect Gmail, Slack, Notion, WordPress, and more. Worth a browse before building anything from scratch: someone has probably already wired up 80 percent of it.
Product Hunt this week: Framer AI Agents, Deck, and CreateOS Sandbox
Notable AI launches on Product Hunt this week: Framer AI Agents (design and publish full websites with AI), Deck (an AI assistant with its own email inbox you can delegate to), nudge2.0 (AI that schedules your week around your goals), and CreateOS Sandbox, which took the top daily spot on July 22. The pattern across launches: specialized agents that execute, rather than general-purpose chatbots.
No new Tencent-specific stories broke in the past 24 hours, but Hunyuan Hy3 remains squarely in the frame as coverage of China’s open-model surge intensifies. Hy3 — 295 billion total parameters, 21 billion activated, 256K context, released under Apache 2.0 with the EU/UK/Korea restrictions lifted — is repeatedly cited alongside Kimi K3, DeepSeek V4, and Qwen3.8 as proof of China’s open-weight momentum. With Washington now accusing Moonshot of distillation and export-control violations, expect provenance questions to reach every Chinese lab: Hy3’s clean Apache 2.0 licensing and documented training story are assets worth having ready in messaging.
If media come calling this week, the strongest proof points remain: 90 percent agent task resolution on the WorkBuddy enterprise platform, integration across WorkBuddy, CodeBuddy, Yuanbao, ima, Marvis, QQ Browser, Tencent News, WeGame, and Sogou Input with nearly 50 more products in the pipeline, and API input pricing cut to 1 yuan per million tokens. Caixin’s coverage of the free AI-agent feature is here.
The agent-specific “Implementation Opinions” became enforceable July 15 — the world’s first dedicated regulatory category for AI agents. Any journalist covering WorkBuddy, Yuanbao’s agent features, or the WeChat AI agent should expect questions about compliance posture; worth aligning with legal on a standing line.