Tencent reported second-quarter revenue of RMB 204.8 billion, up 11 percent year over year, powered by a 17 percent jump in domestic games and 22 percent growth in marketing services. The headline story for analysts was the AI build-out: operating capital expenditure surged roughly 190 percent to RMB 51.8 billion, pushing free cash flow negative, while new AI products weighed on operating profit by about RMB 10.5 billion. Core profit excluding those AI ventures still grew 19 percent to RMB 86.1 billion, though net profit of RMB 56 billion came in just below Bloomberg consensus. Management framed the quarter around the long-term AI strategy — Hunyuan, WorkBuddy, CodeBuddy and Xiaowei — with Hunyuan 4 expected later this year. Full detail in Tencent Mentions below.
Anthropic announced that new Claude models will embed machine-readable, invisible watermarks in every piece of generated text — across the app, the API, Claude Code and cloud partners. The move answers the transparency requirements of Article 50 of the EU AI Act, but Anthropic is applying it globally rather than only in Europe. Unlike metadata, the mark is woven into the text itself, so it travels when copied and pasted and can survive some editing. For communications teams, this is a significant shift: platforms and employers will increasingly be able to check whether copy was machine-generated, though Anthropic stresses the mark proves processing, not authorship. Fortune frames it as part of an industry scramble to police “AI slop.”
Meta published “The Future is for Everyone: The Path to a Positive AI Future,” a 6,500-word essay signed by Mark Zuckerberg laying out its open-model strategy, alongside a run of concrete moves: the release of Muse Glimmer, a 30-billion-parameter multimodal model tuned for local, on-device agent work with a 131K context window, and cuts of up to 95 percent on Muse pricing for customers who share usage data. Newsletters The Neuron and Future Tools both led with the story, calling it Zuckerberg’s “superintelligence bargain” — trade your data for nearly free frontier AI. Critics, including David Gerard at Pivot to AI, dismissed the essay as corporate vapor, but the pricing pressure on OpenAI, Anthropic and Google is real.
OpenAI has paused internal development of its upcoming Astra model after evaluations suggested it may have crossed the “Critical” cybersecurity threshold in the company’s Preparedness Framework — the first time OpenAI has publicly halted a model on those grounds. At the same time, it expanded its Daybreak cybersecurity initiative into two tiers: Daybreak Blue (GPT-5.6 Sol with certain guardrails relaxed for security teams) and Daybreak Red, which offers a purpose-built GPT-5.6-Cyber model to vetted partners for authorized vulnerability research. The Daybreak models are also now available on AWS through Amazon Bedrock. The message: offensive-grade AI capability is here, and OpenAI wants defenders to get it first.
xAI introduced Grok Bot, which provisions each AI agent with its own persistent cloud computer so it can run applications, keep files, and work across sessions like a remote employee. It is the clearest expression yet of this month’s dominant theme — AI moving from chat assistant to autonomous coworker — echoed by OpenAI’s new enterprise research showing firms shifting “from assistance to execution,” and by a wave of agent products on Product Hunt. For a comms leader, the practical question is governance: agents with their own computers and credentials are a new class of insider to manage.
The Substack connector was unavailable in this run, so today’s newsletter coverage was pulled from Inoreader, where your Substack and email newsletters arrive via your newsletter feeds. Paid-tier posts that arrive as previews are summarized from available content.
Lead story: researchers have “cracked” the reasoning of OpenAI, Claude and Gemini models, reading their internal step-by-step thinking — including cases where Claude-based tools can flag other Claude outputs.
Also covered: xAI’s Grok Bot agent computers, and Lightricks’ LTX 2.5 open video model.
The earlier Aug 12 edition, “Watch: AI can write DNA now,” features an interview with researcher Eric Nguyen on AI-designed genetic sequences, deepfake viruses and personalized medicine.
Meta’s open-model blitz is the lead: the “Future is for Everyone” manifesto, Muse Glimmer’s release, and steep price cuts positioned as a land-grab against closed labs.
Also notes Anthropic has made “auto” mode the default in Claude Code — part of the trend toward more autonomous coding agents.
Weekly roundup touching AI economics; follows Tuesday’s essay “The poverty of anti-tech thought” and guest post “Should we ‘pace’ AI self-improvement?” by Tim Fist.
Earlier in the week: “Banning data centers would blow up the U.S. economy” — useful ammunition for the data-center debate.
Non-AI but notable for media watchers: political pressure on Paramount’s David Ellison intensifies; Axios Media Trends’ Tuesday edition covered risky TV bets and “3 Shifts Reshaping Advertising” for communicators.
From Your AI Feeds
Inoreader AI Folder.
Nine unique articles landed in your “AI” folder in the last 24 hours. The day was dominated by Anthropic’s watermark announcement, Meta’s open-AI push, and the agent-as-coworker wave.
OpenAI’s new enterprise research quantifies the shift from AI-as-helper to AI-as-doer: companies increasingly hand agents complete tasks rather than asking for drafts, and “frontier firms” adopting agentic workflows are pulling measurably ahead of peers. Useful benchmarking data if you’re making the case for AI adoption internally at Tencent.
A reusable prompt for turning repetitive tasks, bottlenecks and busywork into AI-powered systems — practical enough to try on your own comms-team processes this week.
Video interview with Eric Nguyen on AI-generated DNA, deepfake viruses, and personalized medicine — a sober look at generative biology’s promise and risks.
Case study on RingCentral using ChatGPT Work and Codex to centralize operational intelligence — one of several enterprise case studies (Virgin Atlantic, Zapier, Model ML) OpenAI published this week.
Discoveries
Workflows & Tool Watch.
An overnight agent that files your fleeting notes into Obsidian
A self-hosted setup making the rounds this week: an n8n-driven agent reads the quick notes you capture during the day, then overnight it files, tags, and links them into your Obsidian vault using careful “context engineering” — all running locally, so nothing leaves your machine. This maps almost exactly onto your Drafts-to-Obsidian capture habit: Drafts stays the inbox, the agent becomes the librarian.
Claude Code’s August updates: background code review, steadier MCP connections
Anthropic shipped a batch of Claude Code improvements this month: the /code-review command now runs as a background helper so you can keep working, MCP re-authentication no longer revokes working sign-ins mid-switch, and scheduled tasks and workflow saves got reliability fixes. Newsletters also note “auto” mode is now the default — Claude asks permission less and acts more. If you’re running scheduled briefings like this one, these fixes directly affect you.
Raycast v2 lands on macOS with built-in dictation and whole-disk AI search
Raycast’s major v2 update adds built-in dictation, an AI chat that remembers context across sessions, a new file-indexing engine that scans your entire disk in seconds, and a redesign for macOS Tahoe. For a Stream Deck and keyboard-driven Mac setup like yours, it is worth a look — the dictation overlaps with MacWhisper for quick captures, and the persistent AI chat can sit alongside Claude for fast lookups.
Template: a weekly Perplexity research digest that writes itself into Notion
An n8n workflow template automates a full weekly research cycle: every Monday at 9 a.m. it queries Perplexity for cited, up-to-date answers on your chosen topics, compiles the findings, saves the report to a Notion database, and emails a summary via Gmail. Swap in “Tencent coverage,” “China tech policy,” or competitor names and this becomes a lightweight media-monitoring assistant for your team.
Relevantn8n · Perplexity · Notion · Media monitoring
Four of this week’s top Product Hunt launches share one premise: the AI does the task, not you. Bond (a Y Combinator company with a $3M seed) auto-manages executive to-do lists; Goldfish gives your Mac persistent memory of your entire work context; Asmi AI makes real phone calls on your behalf; and Slashy takes over inbox management. On Hacker News, Kimi K2.7 Code hit 458 points, showing how hot the open-source coding-model race is. Worth a skeptical trial before trusting any with real work — but the direction of travel is unmistakable.
RelevantThings 3 · Apple Mail · macOS · Team coordination
Q2 revenue rose 11 percent to RMB 204.8 billion with diluted EPS up 9 percent to RMB 7.433. Coverage split along two lines: growth outlets led with domestic games up 17 percent and marketing services up 22 percent on AI-driven ad targeting (CNBC), while profit-focused coverage led with net profit of RMB 56 billion missing the RMB 58.4 billion consensus as AI investment ramped (MarketScreener, Free Malaysia Today). BigGo Finance highlighted capex up roughly 190 percent to RMB 51.8 billion and core profit excluding AI ventures up 19 percent to RMB 86.1 billion. Expect follow-up questions on the Hunyuan 4 timeline, WorkBuddy and CodeBuddy monetization, and how long the negative free cash flow phase lasts. See also: the earnings call transcript.
Following the July 6 release, Tencent Hy3 (formerly Hunyuan) is now available globally: free to users worldwide through WorkBuddy until August 31, plus access via Tencent Design Miora and Tencent Cloud TokenHub. The model ships under a genuine Apache 2.0 license without territorial restrictions and is rolling out to third-party developer platforms including OpenRouter, Hermes, Kilo, Cline, OpenClaw, OpenCode and Cherry Studio. Reporters covering the earnings story are likely to pair it with this rollout as evidence of the international AI push.