Sunday edition: Saturday was a quiet publishing day, so this briefing covers the full Friday–Saturday news cycle to make sure nothing slips through the weekend.
Beijing-based Moonshot AI released Kimi K3, a 2.8-trillion-parameter model with a 1-million-token context window and always-on reasoning, making it the largest open-source AI model released to date. It still trails Claude Fable 5 and GPT-5.6 Sol overall, but ranked first on four of eight real-world automation benchmarks, including SpreadsheetBench 2 and BrowseComp. Fortune called it a “new DeepSeek shock” as chip stocks sold off, and Axios framed it as China erasing America’s AI lead. Full weights land July 27, timed just ahead of the World AI Conference in Shanghai — expect this to dominate China-tech coverage this week.
Google has reportedly pushed the broader release of Gemini 3.5 Pro back by several months after internal testing showed it fell short on coding and complex long-horizon reasoning. Alphabet shares fell more than 4% on the report. Coming the same week Moonshot’s open model surged and OpenAI shipped its GPT-5.6 family, the delay sharpens the narrative that Google’s model cadence is slipping just as competition peaks.
Anthropic is reportedly negotiating with Samsung to build a custom AI chip tuned for Claude models, and is said to be preparing an S-1 filing for an IPO as early as October 2026. The company has quietly become the industry’s revenue leader — on track for roughly $47 billion annualized and reportedly profitable this year, driven by Claude Code and enterprise adoption. For comms planning: an Anthropic IPO would be the biggest AI listing ever and will reset every “who’s winning AI” storyline.
Oracle is reportedly cutting up to 30,000 employees to help fund its share of the $500 billion Stargate AI infrastructure partnership with OpenAI and SoftBank. It is the starkest example yet of how the AI capital-spending boom is being paid for — with headcount. Expect “AI pays for AI with your job” framing to spread; it is a storyline every large tech employer, Tencent included, will be asked about.
Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, unveiled Inkling, its first flagship model, and opened the weights. Newsletters across the board treated it as the week’s second-biggest open-model story after Kimi K3 — a notable signal that even elite U.S. labs now see openness as a competitive strategy, not a concession. Ben’s Bites has a good plain-English take in “I’ve got an Inkling.”
Claude Fable 5 free access ends tonight, July 19Anthropic’s free Fable 5 window expires at 11:59 PM PT; users can move to Opus 5, extend, or pay with credits. Worth knowing if your team has been testing it.
Substack’s own tools weren’t available this run, so these summaries come from your newsletter feeds in Inoreader. Saturday was quiet, so this covers Thursday–Friday editions.
A practical head-to-head of the three frontier models for daily work: Fable 5 for deep reasoning and agentic reliability, GPT-5.6 Sol for breadth and speed, Grok 4.5 for price.
Bottom line: pick per task, not per brand — the gaps are now workload-specific rather than absolute.
A subscriber-only walkthrough of loop patterns in Claude Code — having the agent iterate on a task until tests or checks pass rather than one-shotting it.
Directly relevant to your automation experiments: the same “run until verified” idea applies to non-coding workflows like report generation.
The community’s own read on where AI engineering is heading: agents as the default unit of work, evals as the new CI, and context management as the core skill.
Good radar for the vocabulary your technical colleagues will be using next quarter.
ByteDance researchers found newer models “learn on the job” — a possible new scaling law measured by EdgeBench, their real-world environment benchmark.
Azeem’s companion essay “Reading is dying. GPU demand isn’t.” pairs a Brown professor’s AI-cheating suspicions with the relentless compute build-out.
Google is pushing Search further into agent territory — completing tasks rather than returning links — while Moonshot’s Kimi K3 grabs the open-model spotlight.
Worth watching for media strategy: agentic search reshapes how brand content gets discovered and cited.
Matt’s framework for actually getting value from AI: boring, repeatable workflows beat flashy demos — pick one recurring task and automate it end to end.
Leads with Thinking Machines’ Inkling launch as the news hook.
From Your AI Feeds
Inoreader AI Folder.
Only one item landed in the strict last-24-hours window (a quiet Saturday), so here are the notable pieces from your AI folder across Friday–Saturday, deduplicated against the sections above.
David Gerard interviews cybersecurity writer Kim Crawley about her forthcoming book “Recovering After GenAI” — a skeptic’s look at what organizations do when AI projects underdeliver. Available on the Pivot to AI podcast and YouTube; a useful counter-narrative source as bubble talk grows louder.
OpenAI CFO Sarah Friar proposes measuring AI ROI through four lenses: useful work completed, cost per successful task, dependability, and return on compute. Expect these phrases to show up in enterprise AI coverage and analyst notes.
Kalev Leetaru’s GDELT Project added a “Geopolitical Framing” section that uses Gemini to analyze how nations and leaders are framed across global media — genuinely useful for a comms team tracking international narratives. A companion series shows Gemini Deep Research building rich historical backgrounders at scale.
A technical but readable explainer on how Mac apps capture audio from Zoom, Chrome, and other applications — the same plumbing that powers meeting recorders and transcription tools like MacWhisper. Handy background for your podcast and recording stack.
Gerard shows how attackers can hijack reasoning chatbots simply by writing text that mimics the AI’s own style — a plain-English illustration of why “prompt injection” remains unsolved. Relevant to anyone deploying AI agents on public content.
Discoveries
Workflows & Tool Watch.
MacWhisper 14 ships a proper transcript editor
Your transcription workhorse got its biggest update in a while: edit transcripts directly in the main window, add paragraphs faster, reassign speakers with fewer clicks, favorite key passages, and enjoy noticeably better performance. For Red Book episodes, the faster speaker reassignment alone should save real time per edit.
DEVONthink’s MCP server now exposes nearly 60 commands to Claude
The updated “Take Control of DEVONthink 4” covers DEVONthink 4.3.2, which includes a built-in MCP server — meaning Claude can search, read, file, and organize inside your databases directly, with the chat assistant now redacting phone numbers, card numbers, and tokens before anything reaches a model. If you haven’t switched it on yet, this is the bridge between your research archive and Claude.
Perplexity’s “Personal Computer” mode lands on Mac
Perplexity’s desktop app now does local file editing, local browsing via its Comet browser, and voice orchestration, with GPT-5.5 as the default orchestration model plus reusable Workflows and plan approvals. It is the most aggressive “AI runs your computer” push yet from a consumer app — worth a cautious test on non-sensitive tasks.
SystemSculpt’s 2026 roundup sorts the ecosystem into retrieval engines, workflow operators, and agent surfaces. Standouts: Obsilo Agent (55+ tools, hybrid semantic search, MCP connectors, multi-agent tasking inside the vault) and Note Companion (a lighter “AI note operator” for filing and cleanup). A good shortlist if you want AI inside camOS without handing it the keys.
An Obsidian vault that briefs you daily — built from Readwise, n8n, and Claude
A detailed build combining Readwise highlights, n8n automations, and Claude into a vault that summarizes what you read and surfaces connections each morning. A related n8n recipe from GrowwStacks distills your AI chat transcripts into structured Obsidian notes — solving the “great conversation, lost forever” problem. Both map cleanly onto your existing stack.
The MCP servers actually worth connecting to Claude Code
Nimbalyst’s tested ranking says three to six servers is the sweet spot: GitHub first, then project-specific ones, with Context7 (live documentation), Playwright (real browser), and Exa (web search) as the strongest general picks. The same restraint applies to Cowork — more connectors is not better; the right few are.
With CGTN, Bloomberg, Axios, and Fortune all framing Kimi K3 as China’s open-source moment ahead of the World AI Conference in Shanghai, expect reporters to ask how Hunyuan compares. The strongest line: Hunyuan Hy3 shipped first (July 6) under a genuine Apache 2.0 license with no territorial restrictions, and its OpenRouter token share hit 8.7% in June — Tencent is a leader in the open-model wave K3 just made famous.
Pandaily’s “Has Tencent AI Finally Turned the Corner?” is the most quotable of the recent wave, crediting Hy3’s 90% agent-task resolution rate and the sub-six-month rebuild from February’s architecture reset. Caixin and TechNode emphasize the free agent feature, Yuanbao integration, and the 1-yuan-per-million-token API pricing.
International coverage continues tracking Xiaowei’s small-scale testing in Weixin (text and voice, launching mini-programs) and the DeepSeek-powered Dayuan agent rolling out on WeCom, with SCMP noting the share-price jump on launch expectations. No fresh hard news this weekend — but with WAIC opening in Shanghai this coming week, prepare for launch-timing questions.