WestTide · AI · 2026-08-20 · Issue #42

Today's Highlights: Anthropic posts its first profitable quarter with Q2 revenue of $1.15 billion; Fractile is negotiating a $6.5 billion valuation after landing an order from Anthropic; Unitree surges over 500% on its STAR Market debut, briefly hitting a market cap of 400 billion RMB; OpenAI partners with Replit to bring GPT-5.6 Luna to the next generation of software creators.

Editor's Note: Capital has started pricing in "profitability" rather than "imagination." A profitable Anthropic, a soaring Unitree, and a chip startup grabbing valuations all point to the same signal: The AI narrative is entering the realization phase. No matter how good the story is, it doesn't beat a positive number in the financial statements.

---

I. Large AI Models

1. Anthropic Posts First Profitable Quarter, Q2 Revenue Hits $1.15 Billion

• Summary: According to Forbes, Anthropic achieved quarterly profitability for the first time in Q2 2026, with revenue reaching $1.15 billion, becoming the first frontier AI giant to balance the books.
• Source: Forbes · 2026-08-17
• Editor's Comment: After burning cash for years, Anthropic set the example for the industry. While OpenAI is still losing money to scale up, Anthropic has started talking about profits. The valuation logic for these two companies has now diverged.

2. Claude's Steganographic Watermark Bypassed Within 4 Hours of Launch

• Summary: Wired reports that after Anthropic confirmed invisible machine-readable watermarks would be globally embedded in Claude-generated content, multiple bypass methods appeared in the developer community within four hours, specifically targeting the statistical word-choice patterns relied upon by the watermark.
• Source: Wired · 2026-08-19
• Editor's Comment: Watermarks stop honest people but not bad actors. The real question isn't whether they can be bypassed, but whether relying on statistical patterns is a viable path for tracing AI content provenance.

3. OpenAI × Replit: GPT-5.6 Luna Opens Up to Next-Gen Software Creators

• Summary: OpenAI officially announced an expanded partnership with Replit, using GPT-5.6 Luna to lower the barrier to software creation, allowing non-professional developers to build apps using natural language.
• Source: OpenAI · 2026-08-19
• Editor's Comment: Replit was one of the earliest believers in "coding for everyone," and now the two are tied even tighter. The essence of such partnerships is competing for the entry point to the "next 100 million developers."

4. Does AI Reasoning in Japanese Make It Less Likely to Press the Nuclear Button?

• Summary: Unite.AI reports on a study showing that when AI reasons in Japanese, it is less likely to make nuclear strike decisions in simulated scenarios compared to reasoning in English. The language of reasoning significantly alters the model's threat assessment behavior.
• Source: Unite.AI · 2026-08-20
• Editor's Comment: The research conclusion is interesting, but don't rush to change the language settings on weapons systems. Language affects the model's "feeling," not the system's "judgment." Relying on language choice for safety is unreliable.

5. Bloomberg Long-Form Graphic Compares US vs. China AI: Who Has the Stronger Agents—ChatGPT, Gemini, DeepSeek, or Kimi?

• Summary: Bloomberg released a US vs China AI Race data graphic, comparing major US and Chinese AI products across dimensions like model capability, agent performance, and ecosystem maturity.
• Source: Bloomberg · 2026-08-20
• Editor's Comment: The highlight of such comparison charts isn't the ranking, but the fact that it acknowledges "Chinese models have entered the same arena." Two years ago, no one would have made this chart.

6. AI-Drafted Bills Flood US Congressional Offices

• Summary: Gizmodo reports that more US congressional offices are using ChatGPT and Claude to draft bills, leading to a massive influx of AI-generated legislative drafts. This has caused a proliferation of clerical errors and slowed down review processes, leaving dedicated legal drafting offices overwhelmed with cleanup.
• Source: Gizmodo · 2026-08-19
• Editor's Comment: AI hasn't made legislation faster; it has made it more numerous and sloppier. It saves drafting time but increases error-checking costs—it's not clear if this trade-off is worth it.

7. Model Hypnosis: New arXiv Paper Demonstrates "Subliminal Control" Over AI

• Summary: A paper on arXiv proposes the Model Hypnosis method, achieving strong control over models by adding specific subliminal signals to inputs. The paper claims the method works stably across multiple scenarios.
• Source: arXiv · 2026-08-17
• Editor's Comment: Every time we think alignment issues are converging, someone opens a new hole through more obscure means. Such research reminds us that the model's "rationality" rests on fragile premises.

8. Z.ai Co-founder on Scaling Laws: Beyond Parameters, Other Scales Are Growing

• Summary: Z.ai co-founder published a post discussing the boundaries of Scaling Laws: every model release ends with "how many parameters," but true capability is also determined by data quality, post-training investment, and inference-time compute. The single-parameter-scale narrative is failing.
• Source: Z.ai Co-founder · 2026-08-19
• Editor's Comment: With the parameter race peaking, the next battle is "how to spend compute in the right places." Top players are quietly doubling down on data cleaning and post-training.

9. Survey: No Consensus on What Counts as "Written by AI"

• Summary: Book Riot released reader survey results showing huge differences in answers regarding what counts as "AI writing" among different groups, ranging from "entirely generated by models" to "using tools slightly counts."
• Source: Book Riot · 2026-08-20
• Editor's Comment: Instead of arguing whether a specific article was written by AI, we should first admit: the definition itself hasn't been settled. This lack of consensus leaves all "AI content labeling" policies hanging in mid-air.

II. AI Software

1. Fractile Negotiating $6.5 Billion Valuation: Chip Startup Boosted by Anthropic Order

• Summary: Bloomberg reports that AI chip startup Fractile is negotiating a new funding round at a valuation of approximately $6.5 billion. These talks follow a partnership deal with Anthropic, indicating rising investor enthusiasm for inference-specific chips.
• Source: Bloomberg · 2026-08-19
• Editor's Comment: Another chip company boosting its valuation via an "Anthropic partnership." Regarding the inference chip pie, no one dares claim they've secured their slice yet.

2. Kalanick Fires Again: Only 1% of VCs Are Actually Useful

• Summary: Travis Kalanick launched another public tirade against venture capital, stating most VCs "don't help," and only 1% are truly useful, directing his criticism at the Silicon Valley capital circle he once relied on to start up.
• Source: TechCrunch · 2026-08-19
• Editor's Comment: Kalanick bashing VCs isn't news; him raising money after the bash is. Entrepreneurs' resentment toward capital is always like this: strongest before the next funding round.

3. Meta Runs Ads for AI Nudify App, Targeting Female Politicians

• Summary: Ars Technica reports that an app called Kromix allows users to generate nude images of real people using AI. Meta ran ads for it, including one creative highly similar to a prominent female US politician. The ad was pulled only after public backlash.
• Source: Ars Technica · 2026-08-19
• Editor's Comment: If the moderation system lets such creatives through, it shows platforms' identification of "AI-generated violating content" is still just on paper. This incident will become a live-fire case for regulators.

4. AI Hiring Tools Face Lawsuits: Discrimination Plus Black Box, Job Seekers Sue Collectively

• Summary: The Guardian reports that multiple job seekers sued companies using AI to screen resumes, alleging discriminatory screening and opacity. Candidates cannot even get the real reasons for rejection.
• Source: The Guardian · 2026-08-19
• Editor's Comment: The problem with AI hiring has never been about algorithm accuracy, but about who is responsible when things go wrong. You can ask why you were rejected after an interview, but if an algorithm rejects you, you don't even know who to ask.

5. Meta AI Launches Mac App, Bringing Chatbot to Desktop

• Summary: Meta released a dedicated Mac version of its AI app, allowing users to chat directly with the Meta AI chatbot in desktop windows and share screen content, competing head-on with OpenAI and Anthropic's desktop products.
• Source: The Verge · 2026-08-20
• Editor's Comment: Meta's AI has always lived inside the shell of its own apps; finally, it's willing to hit the desktop to grab entry points. Late, but better late than never.

6. Navi Secures First External Funding: Prosus Invests $100 Million

• Summary: Indian fintech company Navi, founded by Flipkart co-founder Sachin Bansal, received a $100 million investment from Prosus. This is the first institutional funding for the eight-year-old company.
• Source: TechCrunch · 2026-08-19
• Editor's Comment: Bansal held out on self-funded operations for eight years; now he's finally bowing to external capital. For Indian fintech, this is a signal: the burn-cash-to-expand model is back.

7. Robeco Warns: $1 Trillion AI Capex May Burst Bond Supply

• Summary: Robeco released a report stating that tech giants raising trillions of dollars in capex for AI infrastructure will significantly expand bond issuance volumes, putting pressure on credit market supply.
• Source: Bloomberg · 2026-08-19
• Editor's Comment: AI is no longer just tech news; it's macro news. Data center financing demand is so large it can change bond market dynamics. The scale of this cycle exceeds any previous tech investment boom.

8. Kalshi Valued at $40 Billion, Professors Call It "Staggering"

• Summary: Bloomberg reports that prediction market platform Kalshi has reached a valuation of $40 billion. Vanderbilt Law Professor Yesha Yadav used the word "Staggering" in an interview to describe the frenzy behind this figure.
• Source: Bloomberg · 2026-08-18
• Editor's Comment: Betting real money on "prediction" itself has become the biggest business. By the time regulators wake up, the valuation curves of such platforms might be steeper than any AI concept stock.

9. Relativity Networks Raises $22 Million: Building Faster Roads for Data Centers

• Summary: TechCrunch reports that data center interconnect startup Relativity Networks completed a $22 million funding round, targeting the high-bandwidth fiber optic connection needs of AI clusters. The round involved multiple institutional investors.
• Source: TechCrunch · 2026-08-19
• Editor's Comment: Selling shovels is always more stable than mining gold. As AI clusters grow larger, the fiber optic cables connecting them are becoming the new bottleneck business.

III. Humanoid Robots

1. Unitree Surges Over 500% on STAR Market Debut, Market Cap Briefly Breaks 400 Billion

• Summary: BBC reports that the world's largest humanoid robot manufacturer, Unitree (Yushu Technology), surged over 500% in its first day of trading on the STAR Market. Opening at 1,100 RMB, its market cap briefly exceeded 400 billion RMB, creating the first A-share humanoid robot stock.
• Source: BBC · 2026-08-19
• Editor's Comment: Overseas media focuses on the surge itself; domestic investors focus on the lottery odds. A 0.018% winning rate shows the market lacks not enthusiasm, but shares.

2. Bloomberg Analysis: Why Investors Are Betting on Chinese Robots

• Summary: Bloomberg analyzed the investment logic behind Unitree's IPO: China holds advantages over overseas manufacturers in supply chain, mass production costs, and data collection. Humanoid robots are on the eve of moving from demos to scaled shipments.
• Source: Bloomberg · 2026-08-19
• Editor's Comment: Overseas capital voting with orders is more honest than any industry report wording. In the humanoid robot mass production race, the supply chain is in China.

3. Wired Reporter Field Visit: Generalist Robot That Learns On the Spot

• Summary: Wired reports that a journalist witnessed a generalist robot capable of learning new skills on-site in Cambridge (Massachusetts). No longer reliant on pre-trained fixed task libraries, it demonstrated rapid adaptation to unfamiliar tasks.
• Source: Wired · 2026-08-20
• Editor's Comment: Going from "knowing N fixed actions" to "learning a new action on-site" involves the entire generalization problem. This demo is far from mass production, but the direction is right.

IV. Autonomous Driving

1. Pony.ai CEO Peng Jun: Optimistic About Growth Prospects

• Summary: In a Bloomberg video interview, Pony.ai CEO Peng Jun expressed optimism about the growth prospects of autonomous driving, stating that Robotaxi operational boundaries and scale are continuously expanding.
• Source: Bloomberg · 2026-08-19
• Editor's Comment: CEOs are naturally optimistic on camera. What's worth watching are the subsequent data: daily rides per vehicle, cost per kilometer, and regulatory looseness/tightness on expansion. These are the real yardsticks for judgment.

V. World Models / Physical AI

1. Amazon Prime Air Expands to Nearly 500 US Cities

• Summary: Amazon announced a major expansion of its drone delivery service Prime Air, planning to cover nearly 500 US cities by the end of this year. Logistics drones are moving from pilot programs to scaled operations.
• Source: TechCrunch · 2026-08-19
• Editor's Comment: Delivery drones have flown for ten years and are finally flying out of the word "pilot." In the physical world, AI businesses are worth much more when landed than in demos.

---

Sector Statistics: Large AI Models 9 items · AI Software 9 items · Humanoid Robots 3 items · Autonomous Driving 1 item · World Models/Physical AI 1 item (Total 23 items)

Frontiers of the Physical World | WestTide