China’s Moonshot AI dropped Kimi K3 late on July 16 — a roughly 2.8-trillion-parameter model with a 1-million-token context window, instantly the biggest “open” model in history. It ships in two versions (K3 Max for chat and agents, K3 Swarm Max for large-scale parallel work) at $3 per million input tokens and $15 output, with the full model weights promised free by July 27. Early independent estimates place it in the same tier as Claude Opus 4.8 and GPT-5.5, and one widely shared benchmark showed it beating Claude Fable 5 on front-end coding. The timing was surgical: it dropped hours before Google’s biggest launch of the year, guaranteeing every Gemini comparison now includes a Chinese model whose weights will soon be free. For comms teams, this is the sharpest version yet of the question clients and reporters keep asking — what exactly justifies the premium on closed models?
President Xi Jinping delivered the opening keynote of the World AI Conference in Shanghai — his first appearance in the event’s history — urging countries to cooperate on AI and ensure no single nation dominates the technology, a pointed message aimed at Washington. The centrepiece is the World AI Cooperation Organization (WAICO), formally established on July 16 with 29 founding member countries including Brazil, Indonesia, Malaysia, Russia, South Africa and Pakistan, headquartered in Shanghai. Paired with Huawei showing its Atlas 950 SuperPoD domestic computing cluster on the WAIC floor and Kimi K3’s overnight release, Beijing presented a full-stack pitch — chips, models and rules — to the Global South in a single week. Expect this to dominate the China-tech narrative for weeks; worth having language ready on how Tencent’s open-source Hunyuan work fits the broader story.
Google’s Gemini 3.5 Pro was widely reported to launch July 17, but as of this morning the model still is not listed in the public Gemini API, and reporting suggests Google may have missed a third internal deadline after scrapping the original base model entirely and restarting training. Leaked specs — a 2-million-token context window (double anything in the field), a “Deep Think” reasoning mode gated to the $250-a-month Ultra tier, and pricing near $1.25 input per million tokens — remain unconfirmed by Google. The rebuild story cuts both ways: it signals Google refused to ship a flawed flagship, but it also means the launch now lands after GPT-5.6, Grok 4.5 and Kimi K3 have all reset expectations. One thing Google did confirm this week: NotebookLM has been renamed Gemini Notebook, with a built-in cloud computer that runs code inside notebooks.
TSMC — the company that fabricates nearly every advanced AI chip — reported second-quarter net profit of roughly $22 billion, up 77 percent year over year, on revenue of $40.2 billion. It raised 2026 capital spending guidance to $60–64 billion and announced an additional $100 billion for its Arizona operations, lifting total planned US investment to $265 billion with up to four more fabs under consideration. The read-through is simple: while model companies slash prices in an escalating token war, the hardware layer keeps printing records — every price cut requires more silicon to serve more demand, and TSMC collects on all sides. A 77 percent profit surge says the AI buildout is not slowing; it is accelerating, and it is being paid for with purchase orders rather than press releases.
OpenAI discontinued Atlas, its AI web browser, consolidating browsing, work tools, voice and agents into ChatGPT as a single “super app.” The shutdown — weeks ahead of an expected IPO filing — says OpenAI now believes the chat app itself, not the browser, becomes the front door to the internet. Microsoft reportedly reached the same conclusion, opting for deeper ChatGPT integration rather than building a rival browser, even as it simultaneously instructs enterprise sales teams to pitch Copilot as superior to both OpenAI and Anthropic products. A year ago every lab wanted its own browser; now the fight is over which single app you open first each morning.
Huawei demonstrates its Atlas 950 SuperPoD as WAIC opensChina’s most advanced homegrown AI training cluster, built entirely on domestic silicon under US export restrictions, took centre stage in Shanghai during Xi’s keynote week.
The Substack tools were unavailable this run, so these highlights come from the Substack newsletters captured in your Inoreader feeds over roughly the past 36 hours.
Explores why large language models are fundamentally non-deterministic — ask the same question twice and you will get two different answers, a property baked into how they generate text.
Argues this matters enormously for anyone using AI in professional workflows where consistency and auditability are expected, from legal review to corporate communications.
A useful framing for explaining to executives why AI outputs need human review layers rather than being treated like deterministic software.
Leads with Thinking Machines’ release of Inkling, Mira Murati’s first open-weight model — the day’s defining story in the open-versus-closed debate.
Frames the release as a statement move: the executive who helped build the world’s most famous closed models chose open weights for her independent debut.
Rounds up the rest of a busy news day, including the run-up to Google’s expected Gemini 3.5 Pro launch.
A practical decision guide for the current frontier-model field: Claude Fable 5, OpenAI’s GPT-5.6 Sol, and xAI’s Grok 4.5, which owns the value position on price.
The takeaway for non-engineers: model choice is now task-dependent rather than one-size-fits-all — match the model to the job instead of defaulting to a single assistant.
Useful background for anyone fielding “which AI should our team use?” questions at work.
Distills the AI Engineer World’s Fair into five trends shaping how AI products actually get built this year — a good pulse-check on where the practitioner community is heading.
Slightly older than the 24-hour window but included as the week’s most substantive technical read from your paid-attention list.
From Your AI Feeds
Inoreader AI Folder.
Roughly a dozen distinct articles hit your “AI” folder in the last 24 hours (after removing duplicates). Kimi K3 and the open-model surge dominated, with OpenAI product news close behind.
Kimi K3 topped a front-end coding benchmark ahead of Claude Fable 5 — the clearest single data point yet that open Chinese models are competitive at the frontier, not just cheap. The weights go free on July 27, which will make the comparison impossible to ignore.
The Neuron’s read on K3 day: every capable free model erodes the case for paying per token. Also flagged: the EU opened Android to rival app stores, AI inference platform Fireworks AI hit a $17.5 billion valuation, and TSMC surged on earnings.
Beyond the K3 benchmark story, Lore reports OpenAI’s Codex coding tool has passed 9 million users, Thinking Machines launched Inkling, and xAI has open-sourced Grok Build — three more entries in the week’s open-model stampede.
Google is pushing Search’s AI Mode toward doing tasks rather than answering questions, with the rebranded Gemini Notebook syncing across the Gemini app and Search. AI Valley also covers the Kimi K3 release.
OpenAI CFO Sarah Friar published a practical framework for measuring AI return on investment: useful work completed, cost per successful task, dependability, and return on compute. Genuinely useful vocabulary for anyone who has to justify AI spending to a board or finance team.
DoorDash published real numbers on its shopping agent: a 24 percent lift in checkout conversion. The newsletter’s framing — the question now is who can prove their AI worked — pairs neatly with OpenAI’s ROI scorecard above.
OpenAI is introducing wellbeing features that nudge heavy users to take breaks — a notable shift in how AI companies talk about healthy usage, and a likely template for how other labs handle growing scrutiny of AI overuse.
David Gerard interviews cybersecurity author Kim Crawley about her forthcoming book “Recovering After GenAI” — a skeptic’s view worth keeping in the reading mix as a counterweight to launch-week hype.
GDELT added experimental “Geopolitical Framing” and narrative-analysis sections to its daily Media Trends reports, using Gemini to read how nations and leaders are framed across global media. Two companion posts show Gemini Deep Research building rich historical backgrounders for museum collections. For a comms team tracking how China tech is framed internationally, these free reports are worth a look.
Matt Wolfe’s roundup leads with Thinking Machines shipping Inkling and makes the case that boring, reliable AI workflows — not flashy demos — are where real productivity gains come from.
See also · News Roundup
Discoveries
Workflows & Tool Watch.
The Obsidian + Claude “second brain” playbook: 30 copy-paste workflows
A detailed new playbook shows how to pair Obsidian with Claude (including Cowork mode and Claude Code) so the AI reads your vault, processes your inbox folder into properly tagged notes, finds connections between notes, and runs overnight routines. The flagship trick: dump everything into an Inbox folder and tell Claude “Process my inbox” — what took 20–30 minutes of filing happens in under 5. It includes a CLAUDE.md template, recommended plugins and three MCP servers.
RelevantObsidian · Claude Cowork · Claude Code · DEVONthink
Claude meets DEVONthink: an AI assistant for your research library
A Mac Power Users forum thread walks through connecting Claude directly to DEVONthink so it can search, summarize and answer questions against years of accumulated research — turning a static archive into something you can interrogate conversationally. Given the size of your DEVONthink databases, this is one of the most directly applicable discussions of the week, with real-world setup notes from other longtime DEVONthink users.
Perplexity Computer now works inside Word, Excel, PowerPoint, Outlook and Teams
Perplexity added its Computer assistant directly into the Microsoft 365 apps, so it can draft documents, analyze spreadsheets, build presentations and manage email using the context already in your files. It also gained a live task-tracking panel and inline citations for traceable claims. Since Tencent runs on Microsoft Teams and Outlook, this is worth a hands-on test for briefing-document and deck workflows.
RelevantPerplexity · Microsoft Teams · Outlook · PowerPoint
N8N2MCP turns any n8n workflow into an MCP server in three clicks
A new open-source tool on Product Hunt converts existing n8n automations into MCP servers that Claude and other AI assistants can call directly — no coding required. Copy your workflow JSON, map credentials, deploy. Practically, that means any automation you build in n8n (media-monitoring digests, report generation, publishing pipelines) becomes a tool Claude can trigger in conversation. Separately, n8n itself now ships a native MCP server that lets Claude Desktop build and publish workflows inside your n8n instance.
Show HN: AI podcast transcription with auto-generated chapters and show notes
A developer launched a transcription tool on Hacker News claiming 99%+ accuracy using Whisper with optimized chunking for long audio, automatic speaker identification, AI chapter detection based on topic boundaries, and clean exports to Markdown. A free tier offers previews. For Red Book podcast production, the automatic chapters and show notes are the interesting part — comparable to MacWhisper’s output but with structured chapter detection built in.
RelevantPodcast production · MacWhisper · Red Book
Gemini Notebook’s cloud computer: run analysis without leaving your notes
Buried in the NotebookLM rename is a real workflow upgrade: Pro users get a secure cloud computer that executes code inside notebooks. In plain terms, you can upload sources (reports, transcripts, coverage exports), ask questions, and have the notebook run actual data analysis on them — charts, counts, comparisons — without touching a spreadsheet. Worth testing against your media-coverage analysis workflow.
RelevantNotebookLM / Gemini Notebook · Media analysis · DEVONthink
Tencent is reportedly leading a consortium — alongside ZhenFund and HSG — in talks to buy back Meta’s stake in AI-agent startup Manus at roughly the same $2 billion valuation, after China’s National Development and Reform Commission blocked Meta’s acquisition in April on national-security grounds. Reports conflict on whether Tencent would become the largest shareholder or take a majority: Reuters-sourced reporting says Tencent will remain a minority holder and will not seek control, while Digitimes and The Decoder describe a majority stake. Manus is telling staff it expects to remain independent. This is the first use of China’s Foreign Investment Security Review to unwind a completed AI transaction — expect sustained international press interest and questions framing Tencent as Beijing’s instrument in the deal. Coverage: Tom’s Hardware, The Decoder, BigGo Finance.
Tencent’s Hunyuan Hy3 — released July 6 under a genuine Apache 2.0 license with 295 billion total parameters (21 billion active) and a 256K context window — remains part of the WAIC-week narrative about China’s open-model surge alongside Kimi K3 and DeepSeek V4. The free Agent feature in Yuanbao, which generates PowerPoint, Word, Excel, PDF and HTML files from natural conversation, drew fresh coverage from Caixin and TechNode. With reporters comparing every Chinese open model this week, having Hy3’s benchmark position and Apache-license talking points at hand will be useful.
Tencent shares fell 4.63 percent on July 17 amid a broader pullback in Hong Kong tech. Nothing company-specific drove the move in the coverage reviewed; the next scheduled catalyst is Q2 results on August 12, where AI capex — after president Martin Lau’s commitment to double AI investment — will be the dominant analyst question.