Western Tide · AI · 2026-08-21 · Issue #43

Today's Guide: The winds in the large model circle have shifted. Pew Research says over one-third of webpages are suspected to be ghostwritten by AI since ChatGPT launched. Meanwhile, Meta has quietly become one of Microsoft's largest AI customers, and the _Brockman_ era has officially taken over OpenAI. On the humanoid robot front, Agility is talking about going public, and bipedal robots are starting to appear at welding stations on Wall Street. The security line isn't calm either; Grok was exposed for leaking data even when facing encrypted malicious instructions.

Editor's Note: AI content is flooding in, platforms are contracting, and capital is doing the math. How far the tide recedes depends on who actually holds the production tools, versus who just wrote 421 articles waiting for clicks.

---

I. Large AI Models

1. ChatGPT Can Now Control iMessage, Apple Privacy Controversy Reignites

• Summary: Bloomberg reports that ChatGPT has added direct control capabilities for iMessage, allowing it to send and receive messages on behalf of users. As soon as the news broke, concerns among Apple users about data boundaries rose noticeably: where the assistant's permissions end, the privacy red line begins.
• Source: Bloomberg · 2026-08-20
• Editor's Comment: Assistants taking over chats is just the beginning. The real issue lies at the system level. Once personal devices open up system permissions to third-party models, the first crack appears in Apple's walled garden.

2. Pew Research: Over One-Third of Webpages Suspected of Being Written by AI Since ChatGPT Launch

• Summary: A study released by Pew Research shows that more than one-third of webpages published after ChatGPT's release show signs of AI authorship. The study covers a large number of news sites, blogs, and commercial sites with a substantial sample size.
• Source: TechCrunch · 2026-08-20
• Editor's Comment: The one-third ratio is more convincing than any exclamation about "AI occupying the content industry." Search engines and advertisers will eventually vote with their feet.

3. OpenAI Today Looks More and More Like Greg Brockman's Company

• Summary: A long-form article by The Verge reviews OpenAI's past year: lawsuits against Musk, multiple executive departures, constant turmoil, while Greg Brockman's role has expanded all the way, gradually becoming the company's de facto core driver.
• Source: TheVerge · 2026-08-20
• Editor's Comment: The founding team has gone full circle, and those remaining have taken the wheel. OpenAI's story has changed from "a lab changing the world" to "a company that needs a helmsman."

4. Grok Exposed: Encrypted Malicious Instructions Still Fail to Stop Data Leaks

• Summary: Ars Technica reports that security researchers found that even if malicious instructions are encrypted, Grok may still leak user data. xAI's previously touted security promises have been discounted by this vulnerability.
• Source: ArsTechnica · 2026-08-20
• Editor's Comment: Encryption provides protection at the transmission layer, but the model's own instruction boundary is where the vulnerability lies. There is still a distance to go between the security narrative and actual implementation.

5. Meta Has Quietly Become One of Microsoft's Largest AI Customers

• Summary: Bloomberg cites sources familiar with the matter saying that Meta has become one of Microsoft's largest AI customers, with cooperation between the two sides in cloud computing power and model ecosystems deepening quietly. The two giants were still sparring in open-source debates during the last round.
• Source: Bloomberg · 2026-08-20
• Editor's Comment: They argue about open source with their mouths, but their bodies are honest. Meta wants to develop its own open-source models while also buying Microsoft's computing power. Between giants, there are no enemies, only bills.

6. Bloomberg Graphics Comparison: Which Agent Among ChatGPT / Gemini / DeepSeek / Kimi Is Stronger?

• Summary: Bloomberg released interactive graphics comparing the Agent performance of ChatGPT, Gemini, DeepSeek, and Kimi across three dimensions: capability, cost, and ecosystem. Charts show that the gap between top US and Chinese models is visibly narrowing.
• Source: Bloomberg · 2026-08-20
• Editor's Comment: Foreign media doing US-China comparisons more frequently is itself a signal. When competitors enter research charts, it means competition has moved from the court of public opinion to actual evaluation benches.

7. Beijing Opens an AI-Themed Bar, DeepSeek Tokens Come with Beer

• Summary: Reuters reports that an AI-themed bar has appeared in Beijing, where customers can scan a code to receive DeepSeek tokens, chatting about models while drinking beer. The owner treats APIs like drink menus and prompts like cocktail recipes.
• Source: Reuters · 2026-08-20
• Editor's Comment: When a tech hotspot starts appearing on bar menus, it has truly broken out of its niche. Reuters writes lightly, but the cultural signal behind it is not light.

8. An Honest Experiment: 421 AI Articles Yielded Only 10 Clicks

• Summary: A site owner publicly shared experimental data: publishing 421 AI-generated articles over five months resulted in only 10 clicks. He used real numbers to pour cold water on "AI content entrepreneurship."
• Source: Hacker News · 2026-08-21
• Editor's Comment: Traffic platforms don't lack content; they lack trusted content. AI zeroes out the production side, but the ceiling on the distribution side becomes even higher.

9. The Verge: AI Is Stirring Up the Rice Bowl in Mathematics

• Summary: The Verge's Decoder podcast invited an AI correspondent based in London to discuss AI's impact on mathematical research: from AI-assisted proofs to math education, and existential anxiety among practitioners—there are more questions than answers.
• Source: TheVerge · 2026-08-20
• Editor's Comment: Mathematics is one of the few disciplines with clear right and wrong, which is why it hits AI's capability boundaries earliest. Proofs can be verified, but whether the process of discovery will be monopolized is the next big question.

10. Someone Used GPT-5.6 to Hack NetHack and Climb Leaderboards

• Summary: A player documented using GPT-5.6 Daybreak Blue to find buffer overflows, farm items, and achieve high scores in NetHack 5.0. Top-tier models are being used for the most unserious things.
• Source: Hacker News · 2026-08-20
• Editor's Comment: Using the strongest model for game cheating looks wasteful, but it's actually a stress test. Finding real vulnerabilities in old games shows that the model's understanding of "boundaries" goes far beyond chatting.

11. Study: Nearly 40% of Music Released Globally in July Used AI

• Summary: A study covering new music releases globally in July 2026 shows that nearly 40% of works used AI tools in the creative process. The "human content" of the music industry is being re-evaluated.
• Source: Euronews · 2026-08-20
• Editor's Comment: The 40% figure will scare the industry, but platform moderation and copyright systems clearly aren't ready yet. Whoever establishes the rules for "AI participation labels" first will gain pricing power first.

II. AI Software

1. Google Discover Will Allow Customizing Feeds via Conversational Descriptions

• Summary: The Verge reports that Google is adding AI customization capabilities to the Discover feed, allowing users to directly describe what they want to see, generating personalized feeds through chat-style interactions.
• Source: TheVerge · 2026-08-21
• Editor's Comment: Feeds evolve from "algorithm guesses what you like" to "you tell it what you like." The power of recommendation systems is starting to shift into users' hands.

2. Linkdaze: A Smart Calendar Designed for the Whole Family

• Summary: TechCrunch reports that Linkdaze's smart calendar targets not personal schedule management, but integrating the whole family's arrangements into one system, automatically coordinating pick-ups/drop-offs, grocery shopping, and activities.
• Source: TechCrunch · 2026-08-21
• Editor's Comment: Personal calendars are tools; family calendars are systems. Whoever bites off the hard nut of "multi-role coordination" first can upgrade from efficiency tool to household infrastructure.

3. Meta's Pocket Opens to US Users: Vibe Coding Games While Sharing

• Summary: Meta's experimental app Pocket is now open to all US users, focusing on quickly making games via vibe coding and sharing them with one click. Last month, it was quietly launched only in test markets.
• Source: TechCrunch · 2026-08-21
• Editor's Comment: Meta isn't betting on game quality, but on the barrier to creation. When "writing games" becomes "describing games," the UGC pool will grow immeasurably large.

4. Slack Launches Collaborative Vibe Coding Channels

• Summary: Slack launched dedicated vibe coding channels, allowing team members to write code directly with AI agents within Slack, without switching back and forth between different tools.
• Source: TheVerge · 2026-08-20
• Editor's Comment: Code collaboration is migrating from IDEs to conversational tools. Slack wants to be the main battlefield for development collaboration in the AI era; this position used to be called GitHub.

5. Muon Space Raises $250 Million, Backed by Google and Salesforce

• Summary: Satellite startup Muon Space completed a $250 million funding round, with participation from Google and Salesforce. Funds will be used to move AI inference capabilities into orbit, performing real-time computing at the edge of space.
• Source: Bloomberg · 2026-08-20
• Editor's Comment: Ground-based computing power has reached its limit, so giants are looking upward. How much imagination space exists for space AI depends on how high the word "edge" can be pushed.

6. JetBrains Survey: Adoption Trends of AI Coding Agents

• Summary: JetBrains Research released a developer survey tracking the adoption rate, use cases, and reasons for abandonment of AI coding agents in enterprises. The data comes from a large sample of real developers.
• Source: JetBrains · 2026-08-20
• Editor's Comment: JetBrains' surveys are closer to the truth than any media because the people writing code are all in their IDEs. Adoption rates rise fast, and complaints rise fast too.

7. Kalanick Fires Again: Only 1% of Venture Capitalists Are Useful

• Summary: Uber founder Travis Kalanick launched another round of attacks on VCs, claiming that only 1% of investors are truly helpful. He has repeatedly complained publicly about capital "disappearing" during founders' difficult times.
• Source: TechCrunch · 2026-08-20
• Editor's Comment: Entrepreneurs bashing VCs is a staple act, but Kalanick's 1% theory puts the unspoken screening logic of the industry on the table.

8. Meta AI Launches Native Mac App

• Summary: Meta released the Mac version of the Meta AI app, supporting window sharing, allowing users to throw content from any window directly to AI for processing. It is currently available in beta form, targeting enterprises and creators.
• Source: TheVerge · 2026-08-20
• Editor's Comment: Desktops are the last high ground in the assistant war. Meta is patching ecosystem gaps through interface openness; whether it can shake up early movers depends on how far "system-level permissions" like window sharing can go.

9. Opinion: "Credentialism" Is Coming for AI Agents

• Summary: A widely forwarded article proposes that AI agents are experiencing the "credentialism" phase of human workplaces. Whether an agent is useful increasingly depends on credible qualifications, certifications, and evaluation systems.
• Source: Hacker News · 2026-08-20
• Editor's Comment: When there are too many agents to choose from, the market naturally grows "degrees." The resume war for agents has begun; someone just hasn't issued the diplomas yet.

10. A Constitution for One: Governing My Own AI Agent Legion for 7 Months

• Summary: A developer shared his experience of writing a "constitution" for his AI agent fleet over 7 months: not prompts, but a set of permission, boundary, and accountability rules, ensuring agents understand the rules before acting.
• Source: Hacker News · 2026-08-20
• Editor's Comment: An individual establishing a constitution for agents sounds like performance art, but it is actually the earliest governance practice. When agents really cause trouble, this constitution will be the first template for immunity and accountability.

III. Humanoid Robots

1. Agility Robotics CEO Discusses IPO Plans and Chinese Rivals

• Summary: Bloomberg interviewed the CEO of Agility Robotics, with topics revolving around the company's IPO plans, while directly addressing pricing and shipment pressure from Chinese competitors. The company previously focused on the Digit bipedal robot, targeting logistics scenarios.
• Source: Bloomberg · 2026-08-21
• Editor's Comment: Leading US humanoid robot companies starting to talk about IPOs indicates the industry has entered the "profit-seeking" stage. And the phrase "Chinese rivals" appearing in a CEO interview is itself a footnote to competitiveness.

2. IEEE Spectrum: How Persona AI Makes Humanoid Robots Profitable in Welding

• Summary: IEEE Spectrum reported on the deployment of Persona AI's bipedal humanoid robots in welding scenarios: not performing in labs, but doing real work at welding stations, billed by output.
• Source: IEEE Spectrum · 2026-08-21
• Editor's Comment: Welding is one of the few scenarios where "robots outperform humans safely, compliantly, and cheaply." If humanoid robots can prove their ROI amidst welding sparks, the story finally holds water.

3. Bloomberg: Chinese Humanoid Robots Shift to Industrial Pragmatism

• Summary: Bloomberg video report states that Chinese humanoid robot companies are collectively shifting from "showmanship" to industrial pragmatism: entering factories, working night shifts, calculating ROI, speaking through shipment volumes and failure rates.
• Source: Bloomberg · 2026-08-20
• Editor's Comment: Foreign media observing the shift among Chinese players aligns with the conclusion from booths at WRC: good-looking is worse than useful, parameters are worse than man-hours.

4. Wired: Generalist Robots That Learn on the Spot, "Like Clever Toddlers"

• Summary: Wired journalists visited a robotics lab in Cambridge and saw generalist robots capable of instant learning on-site: not pre-programmed, but taught on the spot like children, learning and doing immediately.
• Source: Wired · 2026-08-20
• Editor's Comment: Generalization ability is the holy grail of generalist robots. If the route of on-site learning works, "buy it and teach it yourself" will become reality sooner than "all-capable out of the box."

IV. Autonomous Driving

1. Waymo Declassifies Robotaxi's "Brain": Rear Cabin Computer Performs Trillions of Operations Per Second

• Summary: The Verge reports that Waymo disclosed details of the Robotaxi's onboard computing platform: the computer in the rear cabin can perform up to 1 quadrillion operations per second, with the accompanying sensor and chip combination revealed externally for the first time.
• Source: TheVerge · 2026-08-20
• Editor's Comment: Waymo's willingness to open the chassis indicates hardware is no longer its moat; data is. Showing off the "brain" is meant to show latecomers the threshold.

V. World Models / Physical AI

---

Track Statistics: Large AI Models 11 items · AI Software 10 items · Humanoid Robots 4 items · Autonomous Driving 1 item · World Models/Physical AI 0 items, Total 26 items