WestTide · AI · 2026-09-22 · Issue 75

Today's Highlights: British Columbia, Canada, sues OpenAI, alleging it failed to issue warnings prior to a shooting incident. Amazon blocks Meta's shopping agent Muse from its site. Boston Dynamics opens a new application center in Georgia, where Atlas begins practicing at production line pace.

Editor's Note: Today, money and rules are pushing in two directions. The model side continues to compete for talent and compute, while the application side starts setting barriers against each other. Whoever secures the entry point first gets to set the prices.

---

I. AI Large Models (LLM / Foundation Model)

1. British Columbia Sues OpenAI, Alleging Failure to Warn About Shooting

• Summary: Bloomberg reported on September 21 that the government of British Columbia, Canada, has filed a lawsuit against OpenAI, citing the company's failure to issue warnings before a mass shooting incident. It is rare for local governments to sue leading model companies over model warning responsibilities.
• Source: Bloomberg · 2026-09-21
• Editor's Comment: Once warning responsibility is written into case law, compliance costs for model companies will extend far beyond just legal budgets.

2. Meta's Muse Launches Faster on Mobile Than ChatGPT Did Back Then

• Summary: TechCrunch cited estimates released by market intelligence firms on September 21, stating that the initial growth curve of Meta's new AI app Muse outperformed ChatGPT's mobile performance during its early launch period.
• Source: TechCrunch · 2026-09-21
• Editor's Comment: Social distribution bypassed the cold start hurdle. ChatGPT's curve back then was built up step by step.

3. Alibaba Replaces Qwen Leadership Ahead of Apsara Conference Opening

• Summary: On September 21, Alibaba appointed senior AI researcher Liu Dayiheng as the head of the Qwen large language model project. In the guest list on the official Apsara Conference website, his position ranks after Joe Tsai and Eddie Wu.
• Source: The Information · 2026-09-21
• Editor's Comment: The ranking on the list is the public version of internal weight. Releasing this the day before the conference indicates that Qwen's roadmap needs to be realigned.

4. US District Court Ruling Favorable to AI Makes Open Source Developers Uneasy

• Summary: DevOps.com analyzed the impact of this ruling on September 21. Open source developers worry that the court's lenient stance on training data will further weaken the actual binding force of open source licenses regarding data usage.
• Source: DevOps.com · 2026-09-21
• Editor's Comment: Licenses are the dividing line in the open source world. If the court's stance loosens, this line can only be held together by self-discipline.

5. Solving LLM Pruning as a Physics Problem

• Summary: HuggingFace blog introduced Multiverse Computing's approach on September 21, formulating the removal of entire structural blocks in Transformers as an Ising optimization problem, using annealing concepts to determine which layers to cut.
• Source: HuggingFace · 2026-09-21
• Editor's Comment: Pruning shifts from an artisan's intuition to a reproducible optimization problem. The benefit is verifiability; the cost is that experience becomes less valuable.

6. Running Frontier Models on Own Hardware

• Summary: Tim Dettmers published content for DLab Open Source Week on September 21, discussing how to run frontier models on own hardware, along with the costs and engineering details of this route.
• Source: timdettmers.com · 2026-09-21
• Editor's Comment: The compute narrative has been told for three years. Finally, someone is seriously calculating the bill for building clusters themselves, rather than just reporting purchase orders.

7. FT Article: AI Is Already a Powerful but Troublesome Collaborator in Mathematics

• Summary: A Financial Times article on September 21 discussed AI's actual role in mathematical research, noting both genuine progress and issues with verifying results.
• Source: Financial Times · 2026-09-21
• Editor's Comment: The math community's stance is the most divided. They use it but don't fully trust it; this attitude will likely spread to other disciplines.

---

II. AI Software (AI Application / SaaS)

1. Amazon Blocks Meta's Muse Agent, Preventing It From Placing Orders on Its Site

• Summary: Amazon has blocked access for Meta's personal AI agent Muse to shop on Amazon's website on behalf of users. Previously, Amazon had asked Meta to proactively exclude Amazon e-commerce sites from Muse, but negotiations failed. Bloomberg and The Register reported on this matter on the same day.
• Source: The Register · 2026-09-21
• Editor's Comment: When agents place orders for people, platforms' first reaction is to shut the gate. The battle for entry points has moved from browsers to agents.

2. Kairos Power Secures Up to $100 Million from Samsung Group to Build Reactors for Google

• Summary: Kairos Power selected engineering firm Samsung C&T to construct its 50 MW demonstration reactor for Google, with Samsung contributing up to $100 million.
• Source: TechCrunch · 2026-09-21
• Editor's Comment: Data center power moves from signing contracts to driving piles; this deal is what truly lands it.

3. Ex-xAI Employee Raises $50 Million for Deepfake Detection Firm

• Summary: Bloomberg reported on September 21 that a former xAI employee is raising $50 million for their anti-deepfake technology company.
• Source: Bloomberg · 2026-09-21
• Editor's Comment: Funding on the generation side has long been piled high. The detection side is only now starting to catch up.

4. Oura's $2.2 Billion IPO Is Mostly a Payday for Existing Shareholders

• Summary: Smart ring maker Oura plans to raise up to $2.2 billion in this IPO. Disclosures show that most of this money flows to existing shareholders.
• Source: TechCrunch · 2026-09-21
• Editor's Comment: Hardware companies going public sell existing shares. Don't price rings using software valuation models.

5. FintechOS Completes $28 Million Equity and Debt Financing

• Summary: Romanian-origin FintechOS provides AI and software services to banks and insurance companies. This round of $28 million consists of both equity and debt.
• Source: Tech.eu · 2026-09-21
• Editor's Comment: Startups serving traditional financial institutions are seeing their financing structures become more like those of financial institutions themselves.

6. Corridor Lands $25 Million Seed Round for SMB Health Benefits Brokerage

• Summary: Founder Jackson Wagner left his role as Product Lead at Scale AI in July 2022 and is now focusing on health benefits brokerage for small and medium businesses, securing a $25 million seed round.
• Source: TechCrunch · 2026-09-21
• Editor's Comment: This round of money mainly buys people. An AI background itself is part of the valuation.

7. Unit1 Studio Raises £20 Million for Avatar Concerts

• Summary: UK-based Unit1 Studio, primarily focused on mixed reality avatar concerts, raised £20 million in this round to expand its technology and performance lineup.
• Source: Tech.eu · 2026-09-21
• Editor's Comment: The cost structure of performances is being rewritten, yet ticket prices are still set based on live human standards.

8. Fidelity's Durance Avoids Tech Bonds Feeding the AI Boom

• Summary: Bloomberg reported on September 21 that Fidelity's Durance is avoiding bonds from tech companies financing AI expansion.
• Source: Bloomberg · 2026-09-21
• Editor's Comment: Buyers are starting to scrutinize bond quality. Stratification is appearing on the financing side of this capital expenditure cycle.

9. Factlabel Puts a Nutrition Label on AI-Written Content

• Summary: Open source project Factlabel compares AI-generated conclusions item by item with raw data, blocking unsubstantiated claims. It hit Hacker News on September 22.
• Source: GitHub · 2026-09-22
• Editor's Comment: Moving verification from prompts to toolchains is more reliable than repeatedly telling the model not to hallucinate.

10. Casbin Gateway Adds a Gateway for Local Coding Agents

• Summary: Apache project Casbin Gateway provides a unified security gateway for various coding agents on machines, restricting the system resources they can touch via policies.
• Source: GitHub · 2026-09-21
• Editor's Comment: Tightening agent permissions is a lesson that must be learned before entering production environments.

---

III. Humanoid Robots

1. Boston Dynamics Opens Metaplant Application Center Specifically for Training Atlas

• Summary: A robotics application center located within Hyundai's Metaplant America facility in Georgia is now operational, used to train the Atlas humanoid robot, as reported by The Robot Report on September 18.
• Source: The Robot Report · 2026-09-18
• Editor's Comment: Training grounds follow the production lines. The capabilities of the next version of Atlas will likely be determined by factory data.

2. SoftBank Agrees to Acquire Robotics and AI Institute

• Summary: This transaction, rumored for two months, has been confirmed and is currently under review by the Committee on Foreign Investment in the United States (CFIUS). The Robot Report reported on September 17.
• Source: The Robot Report · 2026-09-17
• Editor's Comment: Top-tier robotics research teams are returning to industrial capital. The survival space for pure research institutions is being compressed.

---

IV. Autonomous Driving

1. Flooded Road Traps Autonomous Car in a Loop

• Summary: A video released by The Weather Network on September 21 shows an autonomous car repeatedly probing a flooded road but failing to pass through.
• Source: The Weather Network · 2026-09-21
• Editor's Comment: In long-tail scenarios, the hardest step is often deciding to give up.

2. CarPlay Now Directly Accessible to Major Chatbots

• Summary: Engadget reviewed the availability of various chatbots on CarPlay on September 21, noting that ChatGPT, Claude, etc., can already be invoked via voice in driving scenarios.
• Source: Engadget · 2026-09-21
• Editor's Comment: Cockpits have become AI entry points first. Automakers' roles risk being reduced to hardware suppliers.

---

V. World Models / Physical AI

1. Companies Building World Models Are Unwilling to Discuss Many Things

• Summary: An observation article by TechCrunch on September 20 noted that world model companies are secretive about data, technical routes, and evaluation standards.
• Source: TechCrunch · 2026-09-20
• Editor's Comment: Lack of transparency hinders industry-wide benchmarking and trust-building.