-->

※ You can watch the full video (≈ 24 minutes) and check the talk directly.

Key Messages at a Glance

  • AI is still underestimated. The public views AI only through the lens of ChatGPT, but rapid developments in reinforcement learning, planning, and strategy are driving even greater changes.
  • Computing power and data are the biggest challenges. The world may need the equivalent of 90 nuclear power plants to secure the electricity needed for AI training.
  • Making AI invent “completely new concepts” requires computations far beyond human capacity.
  • Autonomous agent safety issues must be solved with “guardrails” and “emergency shutoffs.” They emphasize designing safety mechanisms instead of stopping research.
  • US-China competition is heading toward a winner-takes-all structure. They warn that lagging countries might feel pressured enough to want to strike opponent data centers.
  • The double-edged nature of open-source models: They accelerate innovation but also increase risks of misuse for biological and cyber threats.
  • AI’s positive potential in medicine, science, education, and healthcare will fundamentally change human life.
  • AI adoption is a marathon. Individuals and organizations that don’t continuously learn and apply will likely fall behind.

1. The Era of Creative AI Heralded by AlphaGo’s Divine Move

  • The moment in 2016 when AlphaGo created “a move no human would ever make” was a watershed for AI creativity.
  • It was a completely new strategy that emerged in a game humans had explored for 2,500 years, opening the paradigm of ‘AI = creative partner.’

2. Why Is “AI Still Underestimated”?

  • There are perceptual limitations in defining AI only through language models.
  • Reinforcement learning-based planning and reasoning systems are rapidly acquiring strategic capabilities that surpass human experts.
  • When these trends are applied to actual industrial, military, and scientific domains, the ripple effects will grow exponentially.

3. Two Walls to Overcome: Energy and Data

Challenge Description Implications
Power The US alone may need an additional 90 GW. Green energy generation and efficient AI chip development are essential.
Data Web data is approaching its limits. Synthetic data, simulations, and diverse sensing data become important.

4. How to Design Safety Systems for Autonomous Agents?

  • “Guardrails” rather than “research suspension”: We need step-by-step permission restrictions, human oversight loops, and emergency shutoff switches.
  • International cooperation and transparent monitoring systems are essential.

5. The Tectonic Shift from US-China Competition

  • First-mover advantage: The country that first secures superintelligence could virtually monopolize global hegemony.
  • “Strike the data centers” scenario: The temptation to physically prevent the birth of opponent superintelligence could become reality.

6. The Open Source Dilemma

  • We need to balance between accelerating research and innovation ↔ increasing misuse risks.
  • When releasing models, we must design risk assessment, usage policy enforcement, and misuse detection systems together.

7. The Positive Revolution AI Will Bring

  1. Drug development — We can comprehensively identify all ‘drug-able’ sites and dramatically reduce clinical costs.
  2. Basic science — Can contribute to solving ultra-difficult problems like dark matter and energy interpretation.
  3. Personalized education — Can optimize learning paths like a personal tutor and reduce global educational gaps.
  4. Expanding healthcare access — Can provide AI-powered medical consultation and diagnostic services even in underdeveloped regions.

8. Action Guidelines for Us

AI adoption is not a sprint but a 42 km marathon. We must start small projects right now and continuously iterate and improve to survive in the future.

Conclusion

Eric Schmidt asserts that “the most significant event for humanity in the next 500-1000 years will happen right now, within our lifetime.”

While directly facing both the positive changes and potential risks of AI, it’s most important to continuously learn and experiment while establishing solid safety measures.

One sentence not to miss If you don’t run with AI, you’ll inevitably fall behind.

참고