1. Video Introduction: Topic Overview

• The video was recorded on a Saturday afternoon, March 29, 2025.
• The presenter shares content from a presentation given at an event on March 22.
• Topic: "How can startup founders and AI engineers survive amid the attacks of frontier models?"
• This topic is presented as a thought experiment aimed at tech entrepreneurs and those interested in AI business, and the presenter notes it may be somewhat subjective.

"I prepared these materials for tech entrepreneurs who want to seize opportunities in the AI era. I'd appreciate if you view it from this perspective."

2. Two Axes for Making Money in AI Business

• The presenter emphasizes that there are only two areas where money can be made in AI business:
  1. Infrastructure: Companies that provide chips like Nvidia or offer orchestration layers like cloud services.
  2. Vertical integration: Companies like Tesla that leverage AI through vertical integration in specific industries (e.g., automotive).

"There are only about two areas in the AI world where you can make money. Beyond that, most don't make money."

• The key is either securing "proprietary data" that only you can access or running to a new territory.

3. Three Layers of AI Technology

• The presenter divides AI technology into three layers:
  1. Computing/Engineering layer: Currently the most profitable area.
  2. Algorithm layer: Currently the most fiercely competitive area where creating capital value is difficult.
  3. Service layer: The area of providing AI-powered services.

"The algorithm layer seems to be the area losing the most in terms of capital value creation right now. Unless you get hired by big tech or become a university professor."

4. Big Tech Strategies

• The presenter analyzes the strategies of major big tech companies: Nvidia, OpenAI, Google, Meta, Tesla.
  • Nvidia: Started with chips, expanding into middleware and services.
  • OpenAI: Started with services, expanding into chip development.
  • Google: Covering all areas across the board.
  • Meta: Starting from the top service layer, expanding downward.
  • Tesla: Leveraging AI through vertical integration in the specific industry of automotive.

"Tesla leverages open-domain technology well, vertically integrating from top to bottom."

5. Two Options for Startups

• Two options available to startups:
  1. Build AI services on top of frontier models.
  2. Develop vertically integrated AI services in a specific vertical domain.

"From a startup's perspective, you either build a service startup for AGI or, like Tesla, grab a specific vertical and build something through vertical integration."

• The presenter mentions he hasn't seen alternatives beyond these two options.

6. Data Flywheel and Virtuous Cycle

• Using Tesla's case, the presenter explains the importance of the data flywheel.
  • Data leads to better services, which attract more users, generating more data, which improves services further -- a virtuous cycle.

"Tesla created a virtuous cycle where data makes Autopilot more accurate, which sells more cars, which brings in more data."

7. AI Models Surpassing Humans by 2027

• The presenter predicts that AI models surpassing humans in most domains will emerge by 2027.
• He emphasizes that current models already outperform humans in many areas.

"By 2027, I think models that surpass humans in every area will emerge. In fact, they're already outperforming humans in many areas right now."

8. Verifiable Reward Functions and AI's Limits

• The presenter says big tech will dominate areas where verifiable reward functions can be created.
• However, non-verifiable domains still present opportunities.

"Big tech will take all the verifiable domains. But non-verifiable domains still have opportunities."

• For example, areas requiring people's tastes or subjective judgment are parts that AI cannot easily replace.

9. AI Services Using Non-Verifiable Data

• The presenter says systems that convert non-verifiable data into verifiable data are important.
• This requires simulator environments or domain-specific AI services.

"Closed-loop systems that convert non-verifiable data into verifiable data are important. These could be simulators or AI services."

10. Conclusion: Find Somewhere to Run

• The presenter emphasizes that to survive in AI business, you must find domains where big tech can't easily follow.
• In particular, leveraging domain-specific data and environments is important.

"Those of you who've spent a lot of time thinking about where to run will relate. You don't want to go into an area that OpenAI will finish off next year."

11. Closing and Practical Suggestion

• The presenter suggests viewers look for examples of closed-loop systems using non-verifiable data in their own domains.
• Using AI models like GPT or Claude to explore this is also a good approach.

"Look for examples of closed-loop systems that convert non-verifiable data to verifiable data in your domain. AI models will help you much better."

12. Notable Quotes

• "There are only about two areas in the AI world where you can make money."
• "By 2027, models that surpass humans in every area will emerge."
• "Systems that convert non-verifiable data to verifiable data are important."
• "Those of you who've spent time thinking about where to run will relate."

13. Key Keywords

• Frontier models
• AI business
• Data flywheel
• Verifiable reward functions
• Non-verifiable data
• Closed-loop systems
• Vertical integration
• Vertical domains

This video provides many insights for those contemplating the present and future of AI business. The emphasis on differentiation strategies from big tech and opportunities in non-verifiable domains was particularly impressive. Think about what opportunities exist in your own domain