Alibaba announced on Sunday it is selling 710 million new Hong Kong shares at HK$112.70 — a 3.6% discount to Friday’s close — to raise HK$80 billion, or roughly US$10.2 billion. Every dollar of the net proceeds is earmarked for what the company calls “full-stack AI capabilities”: chips, computing infrastructure and model development. It is the largest primary follow-on offering ever by a Hong Kong-listed company, and 2026’s third-largest globally behind Alphabet’s US$80 billion raise and Intel’s US$15 billion sale. Morgan Stanley, HSBC, UBS and CICC are joint bookrunners, and the placement was structured offshore of US securities registration. For anyone tracking Chinese tech capital allocation, this resets the benchmark for how much a China platform company is willing to put behind AI in a single move.
Mark Ma of the University of Pittsburgh, working with the Atlanta Fed, analysed millions of Glassdoor reviews, thousands of financial filings and hundreds of AI announcements from US public companies across five years. The finding: roughly nine in ten executives believe AI has not yet lifted productivity at their firms. Stock reactions to layoffs that companies blamed on AI averaged close to zero, and management optimism expressed across some 10,000 earnings calls showed no meaningful relationship to actual productivity outcomes. Employee sentiment about AI on Glassdoor was, in the researchers’ words, “much more negative” than the overall tone of reviews. This is the most rigorous public counterweight yet to the AI-efficiency narrative, and it is the kind of study that will show up in reporters’ questions about any AI-linked restructuring announcement.
Nvidia’s contract server builders have told the major cloud providers that prices on AI server systems will climb more than 15% on shipments starting in early 2027, hitting the flagship Vera Rubin and Grace Blackwell configurations. Fortune, confirming Bloomberg’s original report, attributes the increase to soaring memory (DRAM) costs from Samsung, SK Hynix and Micron — costs Nvidia can no longer absorb even at its roughly 75% gross margin. In plain terms: the raw components that go into AI servers have got so expensive that the world’s most profitable chip company is passing the bill along. This is the first broad sticker shock the hyperscalers have faced in the Rubin era, and it feeds directly into the wider story about memory prices having risen roughly 500% in a year because of AI demand.
Drawing on spending data from Ramp covering 70,000 companies, the Financial Times reports that Anthropic’s flagship Fable 5 has flattened out at roughly 11% of what customers spend on Anthropic models — two months after launch. It generates just 75% of the model-attributed revenue that OpenAI’s GPT-5.6 Sol produces, despite costing twice as much per token. More striking: the cheaper Opus 5, launched in late July at $5 per million input tokens and $25 per million output, has already overtaken Fable 5 in enterprise spending. The FT reads this as evidence that corporate buyers have stopped paying a premium for top-end performance and are optimising for price. It is a useful data point for anyone arguing that the AI market is maturing into a cost-conscious commodity business faster than the labs would like.
Business Insider’s Katie Roof reports that Hugging Face — the platform where most of the world’s open AI models are published and downloaded — has been working with a bank to gauge bidder interest in a sale that could value it above US$13 billion. That is nearly triple the US$4.5 billion it fetched in its 2023 Series D, and it comes after the company publicly turned down a US$500 million investment from Nvidia in late 2025. Hugging Face is effectively the neutral distribution layer for open-weight models, including the Chinese releases from Alibaba, DeepSeek and Moonshot. Who ends up owning that layer is a genuine strategic question, not just a financial one.
Anonymous “Ox Alpha” on OpenRouter fingerprinted as Zhipu’s unreleased GLM-5.3A free, unnamed model with a 1 million-token context window has appeared on OpenRouter. Independent researcher Ben Davis published analysis on August 21 pegging it to Zhipu AI’s GLM-5.3 with 99% confidence. It scored 80% on the DeepSWE coding benchmark, ahead of Claude Fable 5 (65%) and GPT-5.6 Sol (52%). Free access runs through August 27.
Crypto Briefing · Aug 23
02
DeepSeek’s new vision model beats Anthropic’s Opus 4.8 on two benchmarksDeepSeek released V4-Flash-Vision-Exp, adding image understanding to its 284-billion-parameter model. It beat Opus 4.8 on ALE (1,000+ multi-step app tasks) and ZeroBench (100 hard image-analysis tasks), and uses a compression technique that cut the cost of processing a million tokens by 73%.
SiliconANGLE · Aug 23
03
Anthropic’s Fable 5 tops an autonomous AI research leaderboardPrime Intellect published NanoGPT Speedrun Frontier — 153 autonomous runs across 18 frontier models, each given eight Nvidia H200 chips for up to eight days to optimise a small model from scratch. Fable 5 closed 82% of the gap between the baseline and a human record built over months. Opus 5 and Moonshot’s Kimi K3 followed at 52–54%; Grok trailed badly. All traces are public.
Prime Intellect · Aug 23
04
A 27-billion-parameter open model reverse-engineered a licensed app in 30 minutes — offlineXDA’s Adam Conway ran Alibaba’s Qwen 3.8 27B entirely offline on a Lenovo ThinkStation with 128GB of memory and gave it a reverse-engineering job he assumed needed a frontier model. It identified hidden cryptographic material, self-corrected a bad key hash without prompting, and produced a working bypass script. Hacker News treated it as evidence that serious offensive-security capability has quietly moved onto consumer hardware.
XDA Developers · Aug 23
05
“Why your local LLM feels dumber than it is” tops Hacker NewsA Level1Techs forum series testing Qwen 3.6-27B and Qwen 3.8 derivatives shows how much the plumbing distorts quality: swapping attention backends produced different answers at identical settings, and one compression method silently broke tool calling. It hit 464 points on Hacker News in about a day. The takeaway for anyone self-hosting: how you run the model matters as much as which model you run.
Level1Techs · Aug 23
06
Autolith launches: a coding agent with a live runtime it can inspect and rewriteLambda Symbolics OÜ unveiled Autolith, a programming agent that runs a live Common Lisp environment inside the same process the model works in, so it can test and redefine its own functions mid-session. It keeps conversations, memories and checkpoints across sessions. 122 points on Hacker News.
Lambda Symbolics · Aug 23
07
OpenAI cut GPT-5.6 Sol’s price by more than 20%Effective August 21 and running at least through November 21, OpenAI’s flagship now costs $4 per million input tokens and $20 per million output, down from $5 and $30. It follows July cuts of 80% on the budget Luna model and 20% on the mid-range Terra — a pattern of continuous repricing rather than one-off launches.
OpenAI · Aug 21
08
Nvidia’s AVO agent scores a perfect 100 on ARC-AGI-3Nvidia’s Agentic Variation Operators cleared all 183 levels across 25 game environments while using about 12% fewer actions than the previous best on the same underlying model. Nvidia’s framing is the interesting part: the scaffolding around a model — memory, a loop that detects when the agent is stuck and redirects it — determines sustained autonomous progress more than raw model capability does.
Nvidia · Aug 22
Deals & People
09
Broadcom seeks US$60 billion or more in debt to build Anthropic’s chip capacityPer Bloomberg sources, Broadcom is in talks to raise more than US$60 billion — discussions may reach US$100 billion — through a special-purpose vehicle that would lease custom AI chips to Anthropic and other labs. Apollo and Blackstone are participating. It extends a June pact that already committed US$35 billion for Anthropic compute, targeting 20 gigawatts of capacity by 2028.
TNW · Aug 23
10
Anthropic hires the founder of Google’s TPU programmeAmir Salek, who started Google’s custom AI chip effort and shipped seven generations before leaving in 2022, joins Anthropic’s compute team under James Bradbury. Anthropic currently relies on Nvidia, Google and Amazon silicon and has signed a US$250 million order with UK chip startup Fractile. The hire signals a serious push toward its own accelerators, mirroring OpenAI’s Broadcom-designed “Jalapeno” chip.
Outlook Business · Aug 23
11
Nvidia puts US$7 billion into Poolside without calling it an acquisitionNvidia will pay US$6 billion for a non-exclusive licence to Poolside’s Model Factory — the platform behind its open-weight Laguna coding models — plus a US$1 billion investment at a US$12 billion pre-money valuation, while making job offers to 109 Poolside employees. The three co-founders stay. Poolside told investors it is “not an acquisition or an acquihire.” The structure mirrors Nvidia’s earlier Groq deal, which lawmakers criticised as sidestepping merger review.
PYMNTS · Aug 23
12
Inherent exits stealth with US$50 million and a claim against the frontier labsLondon-based Inherent, founded by Google DeepMind alumni including chief scientist Edward Hughes, says its Faraday agent — built on Alibaba’s Qwen 3.6 27B with OpenAI’s GPT-5.5 Codex for coding — outperforms Claude Opus 4.8 and GPT-5.5 at independently reproducing findings from published scientific papers. Twelve people in King’s Cross, growing to 20–25 by year-end.
TechCrunch · Aug 22
13
Starcloud adds US$250 million for data centres in orbit, with Nvidia joiningThe Redmond company announced a Series A extension at a US$2.3 billion post-money valuation, more than doubling its March mark and bringing total funding to US$420 million. Manhattan West led; Nvidia contributed US$25 million alongside Cisco Investments. Proceeds fund a 100,000 sq ft factory in Woodinville and Starcloud-3, planned to fly on SpaceX’s Starship.
TechCrunch · Aug 23
14
Oxford spinout Fractile reaches a US$6.5 billion valuation on an Anthropic dealFractile is raising about US$600 million at a US$6.5 billion pre-money valuation — more than six times the US$1 billion mark it set in May. The jump follows an initial agreement to supply Anthropic with roughly US$250 million of its memory-based inference chips. The chips are not expected to be production-ready until 2027.
AI Weekly · Aug 20
15
Cognition AI in talks at a US$40 billion valuationThe maker of the Devin coding assistant is in early talks for funding that would value it above US$40 billion, up more than 50% from the US$26 billion it held less than three months ago. Revenue is reportedly approaching a US$1 billion annual run rate, roughly double its level at the last round.
Bloomberg · Aug 12
16
Anthropic’s IPO preparation continues, with a US$2 trillion target reportedReports through August have Anthropic filing before month-end for what would be the largest IPO ever, aiming to eclipse SpaceX’s record US$75 billion raise. CFO Krishna Rao is leading early investor meetings. Anthropic’s share of the enterprise LLM API market reached 32% in Q2 2026, passing OpenAI’s 25% for the first time.
Fortune / CNBC · Aug 13
China & Geopolitics
17
WSJ maps the Tsinghua network behind Z.ai, Moonshot and DeepSeekA Wall Street Journal feature traces the careers of Z.ai co-founder Tang Jie and Moonshot AI’s Yang Zhilin — once a Tsinghua professor and his student — alongside DeepSeek’s Liang Wenfeng, arguing China’s frontier-lab surge was built over more than a decade through the Tsinghua ecosystem rather than appearing overnight. Z.ai recently raised roughly US$4 billion in Hong Kong and is planning a Shanghai listing.
WSJ · Aug 23
18
Bloomberg: the US–China AI gap is closing faster than expectedA Bloomberg graphics package pulls together benchmark trajectories, pricing curves and adoption signals to argue that cheap, capable Chinese model releases are winning enterprise evaluations. Western customers are now testing Chinese open-weight models on real workloads despite export controls.
Bloomberg · Aug 22
19
A humanoid robot ran 100m in 9.39 seconds in Beijing, beating Usain Bolt’s recordOn opening day of the second World Humanoid Robot Games at Beijing’s National Speed Skating Oval, X-Humanoid’s Tiangong Ultra edged Honor’s Lightning (9.47s), both beating Bolt’s 9.58s human record from 2009. A humanoid also cleared 2.88m in the standing high jump. Organisers cite 2,056 robots from 16 countries across 51 events — quadruple last year’s field. Commentators note the sprints ran on smooth arena surfaces built for wheeled feet, not asphalt.
Fortune · Aug 22
20
Z.ai holds GLM-5.3’s open weights back over hacking riskZ.ai says GLM-5.3 learned to chain together complete software exploits faster than expected during training, so it is delaying public release of the model weights by about two weeks for safety hardening. The company reports its models have found 2,436 vulnerabilities across 269 open-source projects, 1,097 rated critical or high, in systems including the Linux kernel, WebKit and FreeBSD. The oldest bug dated to 1981.
Axios · Aug 14
21
Seoul cram schools boom on semiconductor-industry bonusesThe Financial Times reports semiconductor-focused hagwon in Seoul are flourishing as top high-school students prepare years ahead for employment-linked chip majors that guarantee jobs at Samsung and SK Hynix. Record AI-driven earnings and eye-catching worker bonuses are fuelling the rush, with universities routing graduates directly onto AI-memory production lines.
FT · Aug 23
Workforce & Society
22
Hollywood creatives are taking piecework training the AI models replacing themThe Guardian profiles writers, editors and VFX artists rating AI model output during a historic production slump. Screenwriter-producer Ruth Fowler on the paradox: “It’s teaching it how to take our jobs.” LA shoot days dropped 48% between 2021 and 2025; US motion-picture jobs fell from 450,000 in July 2022 to 326,000 in May 2026, while AI-related listings in arts postings nearly doubled year over year.
The Guardian · Aug 22
23
Google and Microsoft race to wire US schools with AIThe New York Times reports Google, Microsoft, OpenAI and others are pouring billions into US schools — Copilot rollouts, Gemini for Education, grants funnelled through teacher unions. The piece frames it as a race to lock students into vendor platforms while researchers still question whether AI is ready for K-12. Ten AI-policy experts shared it on Bluesky within hours.
NYT · Aug 23
24
AI agents are pushing founders into 90-hour weeksThe Wall Street Journal reports that agents’ growing capability is driving productivity FOMO among startup founders, some now working 90-hour weeks managing fleets of agents across their organisations. It cuts directly against industry claims that automation eases workloads — Berkeley Haas research from February found workers’ estimates of AI time savings diverge sharply from what executives claim.
WSJ · Aug 23
25
Harvard Business School ships AI clones of its professorsHBS Foundry is an eight-week, US$699 online entrepreneurship bootcamp whose defining feature is HeyGen-built AI avatars of real faculty — including Flybridge co-founder Jeff Bussgang — that give feedback on students’ practice pitches and mock board meetings between live sessions. Bussgang conceded his digital twin feels “creepy” but said students appreciate the on-demand coaching. A useful precedent for anyone weighing synthetic spokespeople.
TechCrunch · Aug 22
26
Pivot to AI: “When this thing crashes, I want to see apologies”David Gerard published a long interview with the technology critic Tante, arguing the AI investment cycle is a bubble whose boosters should be held to account when it deflates. Whatever you make of the thesis, this is the sharpest strand of the sceptic commentary that increasingly frames mainstream coverage. Video and podcast versions both available.
Pivot To AI · Aug 21
27
Memory prices are up 500% in a year — because of AIThe consumer-facing side of the Nvidia price story. DRAM demand from AI data centres has driven memory costs up roughly fivefold in twelve months, which is now feeding into the price of ordinary laptops, phones and graphics cards. Expect this to become a mainstream consumer-affairs story before long.
Pivot To AI · Aug 17
Security & Regulation
28
Zero-click attack steals Grok chat history — xAI has not patched itAdversa AI published a technique that hides encrypted instructions inside an ordinary web page. When Grok is asked to summarise that page, its sandbox decrypts the payload and treats the recovered text as trusted internal instructions, then sends the user’s name, approximate location, subscription tier and active chat history to an attacker’s server. No click required. Roughly 40% success rate across 20 attempts since June. xAI was notified on June 3 and has issued no patch. Similar attacks were demonstrated against Google Gemini.
GBHackers · Aug 23
29
Microsoft patches a one-click Copilot data-theft flaw eight months after disclosureMicrosoft shipped a fix on August 18 for CVE-2026-24301, dubbed “CoSnitch” by Varonis Threat Labs. A single malicious link could auto-execute prompts and exfiltrate connected Gmail, Drive and Calendar data with no click and no confirmation, by chaining an undocumented URL parameter with Copilot’s built-in fetch and persistent memory poisoning. Varonis reported it in December 2024.
CSO Online · Aug 20
30
Anthropic begins watermarking Claude’s output worldwide to meet EU rulesInvisible, machine-readable watermarks are now added to text and files from Claude models released after August 2, 2026, with older models to follow by December 2. The mark is built into the text so it survives copy-and-paste, but Anthropic acknowledges it can be stripped by resaving, reformatting or screenshotting. The EU AI Act transparency requirement carries fines up to €15 million or 3% of global revenue, so the rule is being applied globally rather than only in Europe.
Unrot · Aug 23
31
OpenAI says security monitoring now costs 20% of the compute it monitorsFollowing its “Pacing model development in an era of cyber-critical capabilities” post, OpenAI estimates monitoring overhead at roughly 20% of the inference compute being watched — covering all reinforcement-learning training and tool-based evaluations for GPT-5.6 Sol-class models and above. OpenAI says this is internal research spend it will not pass to customers; The Register questions whether that holds if the company goes public.
The Register · Aug 19
32
Samsung raises advanced foundry prices up to 15% on AI demandSamsung raised prices for its 4nm and 5nm foundry work by 10–15% and its 8nm process by nearly 10% on July orders, sources told Reuters. TSMC’s sold-out 3nm and 2nm allocations for Apple, Nvidia and AMD are pushing overflow demand to Samsung, which now expects AI applications to top 30% of foundry revenue.
Reuters / Yahoo · Aug 19
From Your Inbox
Substack Highlights.
The Substack API tools are not connected to this session, so newsletter coverage below is drawn from your Inoreader subscriptions instead. Four Substack titles published inside the 24-hour window; the three below them are weekend editions you may not have read yet. To restore full coverage of all 143 subscriptions, the Substack connector needs to be authorised in your Claude connector settings.
The lead essay, “Great minds think (a little too much) alike,” revisits Azhar’s earlier research on whether consulting several AI models produces genuinely independent opinions. The answer is largely no — the major models converge on similar framings, so running a question past four of them creates a false sense of corroboration.
The practical implication for anyone using AI in a decision or drafting process: treat multiple model outputs as one opinion expressed four ways, not four opinions. Real diversity still has to come from a human with a different vantage point.
Exponential View is recruiting an AI Economy Research Fellow — worth passing on if you know an economist working on AI’s macro effects.
Reads OpenAI’s recent “slowing down” messaging — the safety pauses, the compute-overhead disclosures — as partly performative positioning ahead of scrutiny, rather than a genuine deceleration.
Frames Anthropic’s chip and debt arrangements (Broadcom, Fractile, the Salek hire) as an unusually deep capital structure being built ahead of the IPO.
Argues China’s open-weight releases — Alibaba’s Qwen3.8-Max, Z.ai’s GLM line, Moonshot’s Kimi K3 — amount to a coordinated commercial offensive rather than a series of unrelated launches.
The argument: Amazon is building a business of many pillars — chips (Trainium/Inferentia), cloud infrastructure, enterprise inference, commerce and robotics — so it collects a toll regardless of which model lab wins.
A useful counter-frame to the “who has the best model” coverage that dominates AI reporting: infrastructure owners can win without ever shipping a frontier model.
Unpacks Altman’s latest public framing of why people dislike AI, and why the industry’s “dear peasants” tone — the implicit message that the public should simply accept what is coming — is a communications failure rather than a persuasion problem.
Secondary item: some AI agents do not actually delete user data when instructed to, a growing gap between advertised privacy controls and observed behaviour.
Directly relevant reading for anyone writing AI messaging for a general audience.
Uses the sixteenth-century petard — an explosive charge that frequently killed the engineer who set it — as an extended metaphor for AI actors whose own tools turn on them. The Hamlet line “hoist with his own petard” does the framing work.
Lands squarely on this week’s Z.ai story: a lab that trained a model so good at finding exploits it had to delay its own release.
Traces how the “harness” — the scaffolding of memory, tool access, retry loops and supervision wrapped around a model — has become the real site of competition, more than the model weights themselves.
Corroborated this week by Nvidia’s ARC-AGI-3 result, where a better harness on the same base model produced a perfect score with 12% fewer actions.
The practical version for your own setup: what you put around Claude or ChatGPT — your instructions, connected tools and saved procedures — determines results more than which model you pick.
A build-in-public walkthrough of how Tossell assembled his own AI tooling, in the practical register the newsletter is known for.
Pairs with the earlier “Simple agents” and “Do you use a personal agent?” editions, which together make the case that small, single-purpose agents outperform elaborate multi-agent architectures for most everyday work.
From Your AI Feeds
Inoreader AI Folder.
Five articles landed in your “AI” folder in the last 24 hours — a light Sunday, as expected. All five are newsletter editions rather than reported pieces, and all are covered in more depth in Substack Highlights above. Sixty-plus older items sit in the folder from the preceding week; the most useful of those have been folded into the News Roundup.
The Neuron’s Sunday edition on why Altman’s public messaging about AI keeps landing badly with ordinary readers, plus a secondary item on AI agents that fail to delete user data when asked. Arrived twice in your folder — once via the direct Neuron feed and once via the aggregated “Newsletters on AI” feed, which is a deduplication issue worth cleaning up in Inoreader.
Argues Amazon’s position rests on many pillars at once — chips, cloud infrastructure, enterprise inference, commerce and robotics — so it profits from AI adoption without needing to own the winning model. A clean articulation of the “picks and shovels” case.
Three-part read on OpenAI’s safety-flavoured slowdown messaging, Anthropic’s deepening chip and debt commitments ahead of its IPO, and the coordinated character of China’s open-weight model releases. The China section is the most directly useful for competitive framing.
Azeem Azhar on model convergence: asking four different AI systems the same question does not give you four independent views, because they think too much alike. The practical warning is against using multi-model agreement as evidence that an answer is right.
See also · Substack Highlights
Discoveries
Workflows & Tool Watch.
Claude Cowork sessions now run in the cloud, so a task survives closing your laptop
The Cowork beta has moved sessions off any single machine and into the cloud, and put Cowork itself into the Claude apps on iPhone, iPad, Android and claude.ai — no separate app, just an entry in the existing sidebar. The practical change for you: you can start a long task at the Mac mini in Hong Kong, close the machine, and steer the same session from your phone in a taxi, with Claude pinging you when it needs a decision. For a job where the news cycle does not respect your desk hours, this is the single most useful change in the toolkit this month. Anthropic also added public-beta self-hosted environments for Claude Code on Team and Enterprise plans, which matters if Tencent IT ever wants agent sessions running inside its own network rather than Anthropic’s.
RelevantClaude · Cowork · Claude Code · MacBook Pro · iPad Pro · iPhone
DEVONthink now ships its own MCP server — your archive becomes something Claude can read directly
On macOS Sequoia and later, DEVONthink includes a built-in MCP server. MCP is simply a standard plug that lets an AI assistant reach into an application; once connected, Claude can read and summarise records, create and update documents, move items between groups, and add tags or annotations in your databases — without you copying anything into a chat window. Given how much of your working memory lives in DEVONthink, this is the connection that turns the archive from a place you search into a place Claude searches for you. Pair it with the Obsidian MCP server, which shipped v0.16.0 in August moving to the current MCP v2 specification and adding tool groups for Dataview queries and graph navigation, and both halves of your PKM stack become directly addressable.
RelevantDEVONthink · Obsidian · Claude · macOS Sequoia
Cloudflare’s Kitesurf: a browser built for AI agents, not people
Cloudflare launched Kitesurf, a web browser designed to be driven by software rather than a human. It throws out tabs, themes, extensions and sync, and in exchange uses three to seven times less computing power than Chromium for the same agent tasks. Written in Rust and WebAssembly, it passes over 235,000 web-standards tests and speaks the same control protocol as Puppeteer, Playwright and MCP clients — meaning existing automations point at it with almost no rewriting. Cloudflare paired it with x402, a protocol that lets agents pay for services on their own, with 20+ companies already in those payment flows. It is free during beta. If you ever want an agent doing repetitive monitoring across dozens of media sites, this is the cheap engine to run it on.
RelevantMedia monitoring · n8n · Make · Automation · MCP
n8n’s August updates make MCP servers a two-click connection
n8n shipped an inline “Available tools” card that lets you connect MCP servers directly from the assistant chat, rather than configuring them separately. It also added MCP support for creating folders and moving workflows, Sandbox support in Agent Knowledge, and a new Slack integration for agents. Combined with the AI Agent node’s ability to discover MCP-compliant tools without custom API wrappers, the practical effect is that any tool exposing an MCP server — including DEVONthink and Obsidian — can now be dropped into an n8n automation without code. A realistic first build for your team: a workflow that watches an Inoreader folder, drafts a summary, and posts it to Slack for the regional comms leads.
Relevantn8n · Slack · MCP · Inoreader · Team coordination
The five-layer Claude Code setup, and why Skills are the layer that matters for non-coders
A well-circulated write-up from four months of daily use lays out the five layers of a serious Claude setup: CLAUDE.md for standing instructions, MCP servers for reaching external tools, Skills for reusable procedures, hooks for automatic safety checks, and subagents for isolated research. The distinction worth internalising is Skills versus MCP: a Skill is a written procedure (“how we handle an inbound media enquiry”), while MCP is the wiring to the systems that procedure touches. Installing a Skill is genuinely two minutes — drop a folder containing a SKILL.md file into ~/.claude/skills/. This daily briefing is itself a Skill. The same pattern would work for a crisis-response checklist, a Red Book episode prep routine, or a standing format for your leadership updates.
RelevantClaude Code · Cowork · Skills · Media relations · Podcast production
Perplexity Comet is now free everywhere, with Computer as the cloud agent tier
Perplexity removed Comet’s paywall entirely on July 29 — the agentic browser is now free on Windows, macOS, Android and iOS, down from a US$200-per-month Max exclusive at its 2025 launch. It is Chromium-based, so your existing Chrome extensions carry over. Computer, the cloud agent, remains a Max feature and interestingly runs on a mix of Google’s Gemini Flash, Anthropic’s Claude Sonnet 4.5 and OpenAI’s GPT-5.1 alongside Perplexity’s own tuned open models. The strategic read: the price collapse is a direct response to competitive pressure from ChatGPT Atlas and Claude for Chrome. The practical read: worth a free trial for research sweeps before committing to anything paid.
RelevantPerplexity Comet · Perplexity Computer · Research workflows
A caution worth circulating to your team: summarising a web page can leak your chat history
The Grok vulnerability published this week is the clearest example yet of a risk that applies to every AI browser and assistant, not just Grok. An attacker hides instructions inside an ordinary-looking web page. You ask your assistant to summarise it. The assistant reads the hidden instructions, treats them as coming from you, and quietly sends your name, location and conversation history to the attacker. No click, no warning, roughly a 40% success rate in testing — and xAI has not shipped a fix since being told in June. Microsoft’s equivalent Copilot flaw took eight months to patch. The rule to give your team: do not point an AI assistant at a link from an unknown sender, and treat any page a stranger asks you to summarise the way you would treat an unexpected attachment.
RelevantIssues management · Team security · Claude for Chrome · Copilot · 1Password
A clean scan across the news cycle, your Inoreader AI folder and the Substack titles found no story naming Tencent in the past 24 hours. The items below are the most recent Tencent AI facts a reporter is likely to raise, plus two adjacent stories from this weekend that create comparison risk.
Revenue of RMB204.8 billion, up 11% year on year. Marketing services rose 22% to RMB43.6 billion on improvements to the AI-driven ad recommendation model, and domestic games rose 17% to RMB47.3 billion. Capital expenditure was RMB52.8 billion, up 176% year on year, with management guiding to a substantial further increase in the second half as more China-designed ASICs become available. Free cash flow was negative RMB13.8 billion; excluding compute prepayments it would have been positive RMB37.6 billion. Management defended potential “superior” returns on AI spending. Expect the negative-free-cash-flow figure to be the number reporters reach for.
Tencent Hy3 became globally available following its 6 July release, reachable through WorkBuddy, Tencent Design Miora and Tencent Cloud TokenHub. WorkBuddy access is free to users worldwide only until 31 August 2026 Pacific Time — one week from today, and a foreseeable inbound question about what happens after that date. On the Q2 call, management noted Hunyuan 3 production reached a sixfold increase in average daily token usage and ranks among the top three models globally on OpenRouter, with WorkBuddy and CodeBuddy described as clear leaders in China’s AI productivity market.
Alibaba’s Sunday announcement is the largest primary follow-on ever by a Hong Kong-listed company, with all proceeds directed at full-stack AI. The obvious follow-up question is why Tencent is funding its AI build from operating cash flow rather than raising equity — which is a straightforward answer, but one worth having ready before it is asked in a comparison frame.
The Wall Street Journal’s feature on the Tsinghua network behind Z.ai, Moonshot and DeepSeek, and Bloomberg’s graphics package arguing the US–China capability gap is closing, both ran in the past 48 hours. Neither names Tencent, but both establish the frame — Chinese labs as fast-moving, cheap and increasingly credible with Western enterprise buyers — that subsequent Tencent coverage will be written into.