1. Intro & Framing

• Recorded: Morning of June 7, 2025
• Topic: How AI will reshape the future of companies — discussion on recent AI developments and structural change within organizations
• Target audience: Enthusiasts with a solid understanding of recent AI trends and context

"The range of understanding among people participating in the AI world is incredibly wide. So I'm always wrestling with which level to pitch the conversation at."

2. AGI and the Current State of AI

• Emphasizes that AGI is better understood as a transition period, not a single moment in time
• The transition to AGI has already begun, marked by early 2025 as a turning point
• Models are advancing rapidly through events like AI Super Week, entering a virtuous cycle loop (model → service → data → better model)

"AGI is happening right now." "Frontier models are already, in certain domains, what you could call super intelligence — far beyond human-level performance."

• Domains where AI excels: mathematics, coding, biology, and more
• Hallucination: occurs when the model lacks direct knowledge or context → Still a limitation in domains where proprietary data is critical

3. The Core of AI Progress: Evaluation Functions and Reinforcement Learning

• Fields with a verifiable reward function (math, coding, biology, etc.) see rapid AI advancement
• The key to AI development is a closed-loop feedback system that converts non-verifiable into verifiable

"Once we realized that just by pouring in electricity and scaling test-time compute, we could surpass human-level performance in a given domain — that was the breakthrough."

• Tools like AlphaEvolve even leverage hallucination as a driver of innovation, using it to explore the space of possibilities through diverse attempts

4. The Evolution of AI Coding Tools and Market Shifts

• Progress across AI coding tools: Aider, Claude Code, Devin, Cursor, and more
• Cursor's trajectory:
  • Started as a simple code assistant → progressively strengthened agentic workflows
  • Developed its own model (Cursor-small) from user data; uses a hybrid approach (frontier models for complex tasks, own model for simple ones)
  • Execution speed and market share reinforced each other in a virtuous cycle, cementing its top position

"Is Cursor number one? ... There's no answer other than maintaining execution speed."

• How AI is transforming the software engineering market:
  • Senior engineers: leverage AI to strengthen full-stack capabilities — biggest beneficiaries
  • Junior engineers: reduced training incentives, declining hiring demand

"What matters now is less about the effort that goes into actual implementation, and far more about setting direction, defining strategy, and the ability to evaluate and verify what AI produces..."

5. The Future of Junior Engineers and Entrepreneurial Opportunities

• Junior engineers:
  • Opportunities for internal training and hiring will shrink for a while, but new opportunities like entrepreneurship will open up
  • As AI natives, they have the potential to build one-person companies or create innovative business models

"Junior engineers don't need to be sad about not getting hired right now... What they really need is just an entrepreneurial mindset."

• Established companies: those slow to adopt AI see little change; those moving fast are already adjusting headcount and restructuring

6. AI Service Architecture and the Importance of Data

• Cursor for X:
  • A keyword VCs are watching closely
  • Build a meaningful service on top of a frontier model, accumulate user traffic and proprietary data, and capitalize on first-mover effects
• DPO (Direct Preference Optimization):
  • Enables preference learning more simply than RLHF (Reinforcement Learning from Human Feedback)

"At first, our company also tried to build our own model — but we quickly realized the data volume just wasn't there..."

• Agent services:
  • Agent frameworks (Google ADK, Pydantic AI, OpenAI Agent SDK, etc.)
  • Managing hallucination prevention and trade-offs (expensive models = smart but slow; cheap models = fast but dumb) is critical

7. New Organizational Structures for AI Companies

• AI-era companies need organic integration across all layers — from model research at the bottom to UX at the top
• Building AI services is fundamentally different from building mobile apps; it's much closer to building a search service
  • The core loop is data collection, evaluation, and iterative improvement

"Building an AI product is very different from building a mobile app. ... It's much more similar to building a search product."

• Vertical AI (domain-specialized AI) is proliferating
  • AI services tailored to specific domains will continue to emerge

8. The 3 Elements of an AI-Native Company

• Three core elements:
  1. Leader: the person who connects AI with the domain
  2. AI engineer: rapidly applies the latest AI trends and technologies
  3. Internal customer: actual domain experts (non-engineers)

"Even getting the internal customers and the AI engineering group aligned takes about two years."

• Division of labor between Forward Research Engineers and Product Engineering
• Feedback from internal customers combined with AI engineers' rapid experimentation (MVP) → gradual alignment

"I'm now hearing more and more internal customers saying, 'We're not a cosmetics company — we're an AI company.'"

9. Future Stages of the Company

• AI-assisted company → AI-driven company → Autonomous company → Self-evolving company
• The human role:
  • Gradually shifting toward evaluator and strategist
  • Prompting and evaluation become the primary human contributions

"The core internal and external processes run on AI, and humans set direction and strategy via prompting — with prompting and evaluation being the only roles left for humans. That feels like the ideal end state."

• Maximizing efficiency is the logic of capital and markets
• Companies slow to adopt AI need to start by building their data loop first

10. Ethics, Human Fulfillment, and Reality

• AI automation can replace human jobs
  • Replacement may prove more efficient than augmentation
• New roles:
  • Talent who can support AI transformation (AX)
  • Prompting and evaluation skills become critical

"When that era arrives, if you can't clearly position yourself as either someone who prompts AI well or someone who evaluates AI output well — I'd say that's a dangerous place to be."

11. Replication and Domain Innovation

• Replicating the Cursor-for-X architecture into each domain is the methodology for innovation
• Designing AI services suited to specific domains — such as the beauty industry — is essential

"I need to bring this methodology into my own domain. Every domain is different. In my case, it's beauty..."

• Alignment of the core three elements (leader, AI engineer, internal customer) is critical
  • This process takes time — typically two to three years

12. Closing & Conclusion

• The AI-transformed future of companies has already begun:
  • Domain-specific AI services — as illustrated by the Cursor case — are key
  • Collaboration between AI engineers and domain experts (internal customers), bridged by a connecting leader, is what matters
• Prompting and evaluation will remain the enduring human roles
• Efficiency maximization is the market imperative; companies that fail to adapt to AI transformation risk being left behind

"This change is happening with something like 99.8% probability — the right answer is to start fast." "A company is, ultimately, a closed-loop feedback system that converts non-verifiable into verifiable."

Key Keyword Summary

• AGI transition
• Super intelligence
• Hallucination
• Verifiable reward function
• Closed-loop feedback
• Agentic workflow
• Cursor for X
• Proprietary data
• DPO, RLHF
• Vertical AI
• AI-native company
• Prompting & Evaluation
• Autonomous company
• Alignment
• Replication
• Efficiency maximization

🎯 Core Message of This Episode

• AI is already changing how companies are structured and how work gets done.
• Building AI services and data loops tailored to your domain is the competitive advantage of the future.
• Prompting and evaluation, along with the alignment of domain expertise and technical capability, will define the new human role.
• Adapting quickly to change and pursuing domain-specific innovation is the key to survival and growth!

"I've shared the reality without sugarcoating it. ... A company is, in the end, a closed-loop feedback system that converts non-verifiable into verifiable."

Thank you for reading! 🚀 Feel free to leave questions or thoughts in the comments anytime!