Beijing-based Moonshot AI released Kimi K3, a 2.8-trillion-parameter model that is now the largest openly available AI system in the world. In blind developer testing on Arena’s Frontend Code leaderboard, K3 took the number-one spot ahead of Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 — a first for an open Chinese model. It offers a one-million-token context window (roughly 700,000 words it can consider at once), and Moonshot says new engineering tricks made training about 2.5 times more efficient than its predecessor, a pointed response to US chip export limits. Full model weights are due for public release by July 27. The story dominated this weekend’s newsletters because it suggests the gap between free, open models and the paid US frontier has narrowed to months, not years.
Anthropic is in early discussions to lease AI computing capacity from Meta in a deal reported to be worth as much as $10 billion over two years, paid in monthly increments with early-exit clauses on both sides. Anthropic proposed the arrangement in June, weeks after striking a similar deal to use SpaceX’s Colossus data-center capacity. For Meta, it would validate Mark Zuckerberg’s idea of turning his enormous AI infrastructure spending into a rentable cloud business; for Anthropic, it adds another leg to its compute supply. The talks are preliminary and could still fall apart, but they signal how the biggest AI players are becoming each other’s landlords.
In its second-quarter shareholder letter, Netflix disclosed that around 300 titles in 2026 have used generative AI somewhere in production — spanning conception, pre-visualization, shooting and release, though mostly in post-production. Cited examples include the Indian sports thriller “Glory,” Brazil’s “Brasil 70: A Saga do Tri,” and the docuseries “The American Experiment,” which contains about 17 minutes of AI-enhanced footage. Co-CEO Ted Sarandos framed the tools as supporting creators rather than replacing them, while critics immediately raised questions about a homogenized “Netflix look.” For communicators, this is the clearest signal yet that AI-in-production disclosure is becoming a standard part of earnings narratives — and a reputational flashpoint.
The Apple-versus-OpenAI legal fight kept escalating. Apple’s suit alleges OpenAI encouraged former Apple employees — more than 400 now work there — to bring confidential information about unreleased products, with claims reaching up to OpenAI’s chief hardware officer. OpenAI responded that it is “not aware of any evidence that this complaint has merit,” and legal watchers noted Apple’s filing even mixed up two OpenAI employees’ names. The suit threatens OpenAI’s reported first flagship device: a screenless, moving smart speaker pitched internally as a “humanlike AI companion” for the home. Undeterred, OpenAI released its first physical product last week — the $230 Codex Micro keyboard for commanding fleets of coding agents.
Starting today, July 20, Anthropic is including its top-tier Claude Fable 5 model in Max and Team Premium plans at 50 percent of normal usage limits, and making it available through usage credits for Pro and Team Standard subscribers. The company had extended free Fable 5 access on paid plans through July 19 and kept Claude Code’s weekly rate limits 50 percent higher over the same period. Anthropic also launched Claude for Teachers, giving verified US K-12 educators free access to premium Claude tools with curriculum connections in all 50 states. Taken together, it is a broad push to move Fable 5 from a novelty into the default working model — worth knowing as newsroom and agency contacts increasingly cite it.
Thinking Machines releases Inkling, its first modelMira Murati’s startup shipped a 975-billion-parameter open-weight model trained on text, image, audio and video. It openly admits Inkling is not the strongest model available — the bet is that adaptable AI beats one-size-fits-all.
TechCrunch · Jul 15
02
Google reportedly delays Gemini 3.5 Pro by monthsInternal testing reportedly found the model short of expectations on coding and long-horizon reasoning; it stays in limited enterprise preview while rivals ship.
Tech Startups · Jul 17
03
Google’s Gemini Avatar “digital twin” rolls out widelyPaid Gemini subscribers can now generate video of an AI clone of themselves — face, voice and mannerisms — summonable in chats with “@me.” Every video carries an invisible SynthID watermark; users must be 18+.
IT Voice · Jul 19
04
OpenAI’s Codex Micro: a $230 keyboard for commanding AI agentsCo-designed with Work Louder, it has light-up “Agent Keys” showing agent status, shortcut Command Keys, a workflow joystick and a dial that adjusts how much “reasoning” an agent applies.
OpenAI lays out its case for teen-safe AIAge-appropriate protections, learning tools, parental controls and expert partnerships — a preview of where the youth-safety debate (and regulation) is heading.
The Substack tools were not connected during this run, so today’s newsletter coverage comes from your Substack subscriptions as they arrive in Inoreader. Four AI newsletters published in the past 24 hours.
Google’s Gemini Avatar rollout means paid subscribers can generate video of a photoreal digital clone of themselves — a genuinely new deepfake-adjacent frontier for brand and executive comms.
Covers the reported Meta–Anthropic talks that could see Anthropic lease up to $10 billion of Meta compute over two years, following Anthropic’s SpaceX Colossus arrangement.
Thinking Machines “opens the vault” with Inkling, its 975B-parameter open-weight debut aimed at organizations that want to tune models themselves.
Lead essay “Deep Kimpact” argues Moonshot’s Kimi K3 may be a bigger shock to the US AI establishment than DeepSeek was — open weights, frontier-adjacent performance, trained under chip constraints.
Examines why solar costs are rising even as deployment accelerates, a counterintuitive energy story with direct implications for AI data-center buildout.
Rounds up the evolving AI-and-copyright fight, worth watching for any content-owning company.
Breaks down Netflix’s shareholder-letter disclosure that ~300 titles used generative AI in 2026, and what “every level of production” actually means in practice.
Also flags China’s new national AI plan and MLB’s decision to ban AI tools from dugouts during games.
Takes on Scott Galloway’s argument that AI capital spending is a bubble, asking whether the answer is boom, bust — or both at once.
Useful framing if you are fielding “is this another dot-com?” questions from reporters or leadership.
From Your AI Feeds
Inoreader AI Folder.
Six articles landed in your “AI” folder in the last 24 hours. Kimi K3 and Netflix’s AI disclosure dominated; the GDELT Project published two fresh experiments using Gemini for media and museum analysis.
Kalev Leetaru continues his series testing Google’s Gemini Deep Research as a museum curator’s assistant — this time asking it to weave a collection of Estonian historical artifacts into a coherent national founding narrative. A fascinating look at how AI research agents handle narrative-building, which is essentially what communicators do daily.
GDELT added a new experimental section to its Media Trends reports that asks Gemini to generate contrarian scenarios — deliberately arguing against the prevailing media narrative. For issues managers, this is a ready-made technique: ask an AI to steelman the opposite of your coverage read before you brief leadership.
Claude Cowork moves to the cloud — your scheduled tasks now run even with the Mac asleep
Anthropic is expanding Cowork from the desktop app to web and mobile, with sessions and files shared across devices and — most useful for you — scheduled tasks that execute in the cloud with no device online. It also added Microsoft 365 write tools, meaning Cowork can now draft and send in Outlook, not just read. Beta access starts with Max subscribers, and doubled usage limits run through August 5. Practical upshot: briefings like this one, plus recurring media-monitoring or report tasks, no longer depend on the Mac mini being awake.
RelevantClaude · Cowork · Microsoft 365 · scheduled briefings
MacWhisper can now push transcripts straight into n8n, Zapier or Make — and be driven from the command line
Recent MacWhisper updates add automatic delivery of finished transcriptions to Make.com, n8n, Zapier or any custom webhook, a new Editor display mode (beta) that lets you edit transcripts like a text document, and CLI control so agents and scripts can trigger transcriptions. For podcast production, that means a Red Book episode could be transcribed the moment recording ends and flow automatically into DEVONthink or Notion via an n8n workflow, with show notes drafted by Claude — no manual export step.
RelevantMacWhisper · n8n · Zapier · Red Book podcast · DEVONthink
“Dynamic Workflows in Claude Code” — Hacker News debates how to steer long-running AI sessions
A well-trafficked Hacker News thread digs into techniques for controlling long-running Claude Code sessions: injecting mid-course corrections and nudges instead of restarting, and chaining fresh review passes by new agent sessions. The same pattern applies to any long agent task — think multi-hour media analysis or report generation — where you want to redirect the agent without losing its accumulated context.
OpenAI’s Codex Micro is basically a Stream Deck for AI agents
The $230 Codex Micro keyboard, co-designed with Work Louder, gives physical controls to agent workflows: keys that light up with each agent’s status, programmable shortcut keys, a joystick for launching common workflows and a dial to turn an agent’s “reasoning effort” up or down. As a heavy Stream Deck user, the interesting part is the interaction model — dedicated hardware for supervising multiple AI agents at once is now a product category.
Bloomberg’s hands-on review of Xiaowei, the prototype AI agent inside WeChat, frames it as Tencent’s bid to restore its innovator reputation — an assistant that will eventually run errands for a billion-plus users across the mini-app ecosystem. Expect follow-up enquiries on rollout timing, data handling, and how Xiaowei compares to rival agents from Alibaba and ByteDance.
The piece catalogs Tencent’s 2026 agent lineup — WorkBuddy (8.85 million monthly visits), Zaohua Gongfang, Ardot, Design Miora, WorkRally, DataBuddy and LearnBuddy — arguing Tencent has quietly become one of the most prolific agent producers anywhere. Useful third-party validation to point reporters toward.
Bloomberg frames the divestment as Tencent paring mature internet bets to fund its AI push. Anticipate “what gets sold next” and capital-allocation questions in upcoming earnings conversations.
Coverage of a new Tencent model led by an ex-OpenAI researcher — talent-war framing that ties Tencent directly into the US-China frontier narrative. Worth aligning on messaging about Tencent’s research bench.
Trade coverage of Tencent’s international enterprise AI portfolio, including WorkBuddy’s agentic workspace and Design Miora’s persistent-memory creative studio — a good barometer of how the overseas B2B story is landing.