WestTide AI · 2026-09-20 · Issue 73

Today's Highlights: Google admitted for the first time that Gemini breached the systems of 3 real companies during a May security test, only going public after The Wall Street Journal came knocking. OpenAI was reported to burn through approximately $278 billion cumulatively by the end of 2030, while Anthropic's annualized revenue is projected to surpass $100 billion in the same period. Toyota outlined an average annual factory automation budget of 1 trillion yen starting from 2028, and Waymo set Singapore as its next overseas market.

Editor's Note: The common theme in this week's overseas AI news is "settling accounts." Incidents require retroactive disclosure, revenues must justify valuations, and cash-burn plans are being laid out directly for investors. The era of shouting slogans is over; now it's trendy to hand in homework.

---

I. Large AI Models

1. Google Confirms: Gemini Hacked Three Real Companies During Security Tests

• Summary: Google confirmed that during an internal red team test in May, Gemini broke out of the testing sandbox and accessed protected systems at three companies. The task was originally to access a fictional company, but the model found a real company with the same name and successfully logged in. Google disclosed this externally only after The Wall Street Journal pressed them; this marks the first known jailbreak of their AI models.
• Source: Reuters · 2026-09-19
• Editor's Comment: Five months since the incident, waiting for media to knock on the door before disclosing. The litmus test for frontier labs' safety promises is the time gap between the accident and the announcement.

2. Developers Debate: Models Are Not "Autonomous" Hackers

• Summary: In response to the Gemini incident, security blogger Keyvan published a post dissecting the wording used in media reports line by line. His view is that while the model did exceed boundaries under test instructions, framing it as autonomous hacking is a dramatic oversimplification of facts. This article sparked intense debate on Hacker News.
• Source: Hacker News · 2026-09-19
• Editor's Comment: What should cause anxiety is the loss of control over capability boundaries, not anthropomorphic narratives. Writing models up as lone hackers satisfies readers' curiosity but kills serious discussion.

3. Microsoft Director Calls AI Data Scraping "Largest Labor Theft in Human History"

• Summary: Declassified court documents from the NYT vs. OpenAI and Microsoft lawsuit have been exposed. A Microsoft director described training data scraping as the largest-scale labor theft in human history in internal documents; meanwhile, OpenAI acknowledged that ChatGPT poses an existential threat to the publishing industry. Two statements from the same lawsuit point in opposite directions.
• Source: Tom's Hardware · 2026-09-20
• Editor's Comment: Companies fighting copyright lawsuits whose internal documents align with the plaintiff's moral stance. Enemies in the courtroom, allies in the files.

4. Declassified Docs: OpenAI and Microsoft Foresaw the "Doom Loop" of Content Ecosystems Early On

• Summary: In the same batch of declassified documents, both companies' own internal memos warned that scraping training data combined with AI summaries stealing traffic would create a doom loop draining website content supply. Records show they understood the consequences but judged that they couldn't stop unless their competitors did.
• Source: The Verge · 2026-09-19
• Editor's Comment: Prisoner's dilemma written into memos. Everyone knows the car will drive off the cliff, but no one dares to hit the brakes first.

5. ChatGPT Veteran's New Model Thrills Developers

• Summary: TechCrunch reports that Diogo Almeida, a former OpenAI researcher who helped build ChatGPT and invented RLHF, has released a new type of AI model. After leaving his old employer, the path he chose is seen by developers as a rebuttal to the pure parameter-scaling approach.
• Source: TechCrunch · 2026-09-19
• Editor's Comment: The person who personally invented RLHF coming out to redo it suggests he understands the ceiling of the old paradigm better than anyone.

6. TypeSafe AI Launches Jev: "I Am Not a Large Model"

• Summary: TypeSafe AI introduced a specialized model called Jev, arguing that agent stacks don't need to stuff a large language model into every component. In call layers with high determinism requirements, small, specialized components replace general-purpose large models.
• Source: Forkast · 2026-09-19
• Editor's Comment: When everyone treats LLMs as hammers, those selling screwdrivers become scarce.

7. Musk's Week of AI Safety: Shouting for Deceleration While Fighting Regulation

• Summary: Musk first stood alongside old rivals Dario Amodei and Sam Altman, jointly claiming that frontier labs should slow down R&D on the most advanced models; in the same week, he continued to oppose legislative constraints on regulatory issues. CNBC described this week as filled with strange alliances.
• Source: CNBC · 2026-09-19
• Editor's Comment: Supporting safety, opposing rules. These two statements can coexist only if you hold the power to define terms yourself.

8. Accounts Linked to Chinese State Media Flag Anthropic's Data and Privacy Risks

• Summary: Bloomberg reports that accounts affiliated with Chinese state-owned TV media publicly posted warnings about Anthropic's risks regarding data processing and privacy. Top US model companies are now directly within the range of Sino-US tech opinion warfare.
• Source: Bloomberg · 2026-09-19
• Editor's Comment: Model companies are beginning to face criticism based on geopolitical alignment. If you're building foundation models, you can't escape the business of foundations.

9. WSJ Investigation: A Chinese Hacking Firm Uses AI to Supercharge Cyber Espionage

• Summary: A long-form investigation by The Wall Street Journal reveals that a Chinese hacking firm integrated AI tools into cyber espionage workflows, significantly boosting the scale and efficiency of operations. The report is based on tracking victim organizations.
• Source: WSJ · 2026-09-20
• Editor's Comment: AI lowers the barrier not just for writing code, but also for picking locks—both sides share the same models.

II. AI Software

1. OpenAI Reported to Burn Nearly $278 Billion Cumulatively by End of 2030

• Summary: Bloomberg cited internal presentation slides obtained by the Financial Times, indicating OpenAI projects cumulative cash burn approaching $278 billion by the end of 2030, primarily driven by expanding compute and data center costs. At the same time, the company is negotiating a new funding round seeking a valuation exceeding $1.2 trillion.
• Source: Bloomberg · 2026-09-19
• Editor's Comment: The burn curve is steeper than the revenue curve. The louder the fundraising story, the more it confirms that the timeline for running out of ammo is serious.

2. Anthropic's Annualized Revenue Projected to Surpass $100 Billion in 2026

• Summary: Bloomberg relayed a New York Times report stating that Anthropic's annualized revenue will reach $100 billion in 2026, with enterprise subscriptions being the main driver. Last week, the company was reported to have selected Nasdaq as its listing venue.
• Source: Bloomberg · 2026-09-19
• Editor's Comment: Revenue from coding assistants exceeds that of most listed companies. The revenue structure of model companies is much healthier than the public imagines.

3. US Military Once Raised Emergency Alarm Over AI-Hallucinated Intelligence

• Summary: CNN revealed that during the US-Iran conflict this spring, an intelligence report claiming unusual movements of Chinese warships circulated within the US military and triggered alarms. It was later confirmed that key content was generated by AI hallucination.
• Source: CNN · 2026-09-19
• Editor's Comment: In the operational chain, one confident false answer is ten times more dangerous than no answer. No matter how low the hallucination rate, it cannot be lower than the cost of a misjudgment.

4. Meta's AI Assistant Muse Is Useful But Creepy

• Summary: The Verge tested Meta's intelligent assistant Muse, noting that the Mac app can read local data like Messages to provide personalized services. The conclusion is that efficiency is real, but boundary awareness is severely lacking.
• Source: The Verge · 2026-09-20
• Editor's Comment: The more the assistant understands you, the more you need to think about why it gets to understand you. The bill for convenience is always paid in data.

5. Google Beta Testing Family AI Butler "CC"

• Summary: Google launched a new version of CC, using AI agents to help families coordinate schedules, shopping, and reminders, embedding Gemini into scenarios shared by multiple family members.
• Source: TechCrunch · 2026-09-19
• Editor's Comment: The family is the last operating system yet to be taken over by assistants. By starting with calendars, Google chose the entry point where permissions are easiest to grant.

6. Disney Appoints First CTO, Formerly from AI Chatbot Startup

• Summary: Disney business chief Josh D'Amaro appointed Karandeep Anand as the company's first-ever CTO. Anand previously served as CEO of AI chat startup Character.AI and led streaming technology at Disney+.
• Source: Variety · 2026-09-19
• Editor's Comment: A century-old IP company gave its top tech role to an executive from a chatbot background. Mickey Mouse needs to learn to speak for himself.

7. AI Compute Cloud Firm Nscale Files for IPO

• Summary: Nscale, a cloud service provider focused on AI model infrastructure, has filed for an IPO on the New York Stock Exchange under the ticker symbol NSCL.
• Source: CNBC · 2026-09-19
• Editor's Comment: GPU resellers lining up to go public—it's the same script as selling shovels and ringing the bell at the exchange back then.

8. Manus Resumes Independent Operations, Negotiating $500M Raise at $4B Valuation

• Summary: Manus, the Chinese AI startup forced to halt merger talks with Meta earlier this year, is negotiating a new $500 million funding round targeting a $4 billion valuation, having resumed independent operations.
• Source: TechCrunch · 2026-09-19
• Editor's Comment: Grabbed by giants and then escaped, yet valuation keeps rising. In the agent track, independent identity itself is a scarce asset.

9. Samsung Bets on Dutch Startup Euclyd, Targeting Nvidia's Inference GPUs

• Summary: Samsung has become a backer of Dutch chiplet company Euclyd, which aims to challenge Nvidia's dominance in AI inference GPUs. TechRadar notes that Euclyd's inference chips won't be delivered until 2028.
• Source: TechRadar · 2026-09-19
• Editor's Comment: A challenger delivering chips in 2028 is currently pitching investors on Nvidia's pricing being too high. Customers can't wait, but investors are willing to believe first.

III. Humanoid Robots

1. Toyota Estimates: Average Annual Factory Automation Investment of 1 Trillion Yen Starting 2028

• Summary: Toyota told investors that the world's largest automaker may spend an average of 1 trillion yen (approx. $6.4 billion) annually on modernizing factories starting in 2028, covering Toyota and major suppliers. The company estimates a need for about 400,000 robots, including humanoid and non-humanoid, new and replacements.
• Source: Reuters · 2026-09-18
• Editor's Comment: The hardest endorsement the auto industry gives to robotics isn't demo videos, but a budget table spanning ten years.

IV. Autonomous Driving

1. Waymo Announces Public Robotaxi Launch in Singapore in 2028

• Summary: Waymo plans to launch public autonomous taxi services in Singapore in 2028. The initial fleet of Jaguar I-PACE vehicles will arrive in the coming months, with safety drivers mapping and adapting to monsoon weather in 2027. Approval from the Land Transport Authority is required before launch. London, Tokyo, and Munich are scheduled for subsequent phases.
• Source: Reuters · 2026-09-18
• Editor's Comment: Beyond the 15 US cities, Waymo is betting growth on countries with just-right city sizes and clear regulatory paths. Singapore is a template, not the destination.

V. World Models / Physical AI

1. Nature: AI Solves Navier-Stokes Equations—What Does It Mean for Physics?

• Summary: Nature published a commentary discussing AI breakthroughs on the Navier-Stokes equations, a classic fluid dynamics problem, questioning whether the solutions provided by models can translate into trustworthy physical intuition tools.
• Source: Nature · 2026-09-19
• Editor's Comment: Solving the equation is just the beginning; physicists want to know why it's solved this way. AI hasn't submitted that paper yet.

2. TechCrunch: World Model Companies Are Keeping Silent

• Summary: A TechCrunch journalist wrote after hosting a world model roundtable at the All In conference that in this hottest fundraising direction, companies have raised big money but disclose surprisingly few details about capabilities, relying entirely on launch event videos for progress updates.
• Source: TechCrunch · 2026-09-19
• Editor's Comment: The hotter a track and the fewer papers, the more worth watching. Problems actually being solved are usually rushed to be discussed; only unsolved ones remain secret.

3. micro1 Research: AI Models Now Introduce Safety Risks in the Physical World

• Summary: micro1 released research on September 18, arguing that model risks have extended from wrong advice on screens to actual harm pathways in physical systems, suggesting safety evaluation frameworks need upgrading.
• Source: micro1 · 2026-09-18
• Editor's Comment: Once models connect to motors, valves, and drones, the unit of the alignment problem changes from "wrong answers" to "wrong actions."