This talk features Anthropic co-founder Jared Kaplan explaining AI scaling laws and the journey toward human-level AI (AGI). He discusses how AI has gradually advanced, what role scaling has played, and what changes lie ahead. He also shares advice on AI-human collaboration, the future of work, and how to effectively leverage AI.

1. Jared Kaplan's Background and Transition to AI

Jared Kaplan was originally a theoretical physicist who spent many years in academia. From a young age, he was fascinated by science -- influenced by his mother, a science fiction writer -- captivated by fundamental questions like "How does the universe work?" and "Do we have free will?"

"The reason I started physics is that my mom was a sci-fi writer, and I wanted to find out if we could build a faster-than-light drive."

However, as progress in physics research slowed, friends encouraged him to explore AI. Initially skeptical, thinking AI had made little progress in 50 years, he gradually realized AI was advancing rapidly and made the transition.

2. Training Modern AI Models and Scaling Laws

Kaplan explains that modern AI models (e.g., Claude, ChatGPT) are trained in two stages:

1. Pre-training: AI learns by imitating vast amounts of human-written text, picking up correlations between words.

   "Pre-training is the process where the model learns to predict what word comes next from large-scale text."

2. Reinforcement Learning (RL): AI is reinforced through reward signals to perform genuinely useful tasks.

   "We reinforce good behavior -- helpful, honest, harmless behavior -- and suppress bad behavior."

Kaplan emphasizes that scaling laws apply to both stages. That is, as data and computing resources increase, AI performance improves in a predictably consistent manner.

"When we experimented in 2019, scaling up resources -- compute, dataset size, neural network size -- we found that performance improved in a very precisely predictable way. As precise as laws in physics."

3. How Scaling Has Expanded AI Capabilities

Kaplan describes AI capabilities along two axes:

• Flexibility (ability to perform diverse tasks): AI has evolved from narrow systems specialized in specific games (like the old AlphaGo) to general-purpose AI handling text, images, speech, and more.

• Expansion of task duration (performing complex tasks over longer periods): The temporal range of tasks AI can handle keeps growing.

  "The length of tasks AI can handle doubles roughly every 7 months."

Kaplan projects that AI will soon handle not just minutes or hours of work, but complex projects spanning days, months, or even years.

"Eventually, millions of AIs could collaborate to replace the work of entire human organizations. What the theoretical physics community achieved in 50 years, AI might accomplish in days."

4. Remaining Challenges for Human-Level AI

Kaplan acknowledges that while scaling alone can significantly advance AI, achieving human-level AI requires several additional elements:

• Organizational Knowledge: AI needs to understand context and work like someone who has been at an organization for years, not starting from a blank slate.

• Memory: For long-term tasks, AI must be able to remember its progress and store and utilize necessary information.

  "Memory is knowledge. To work on long-horizon tasks, AI needs to remember its progress and make use of it."

• Sophisticated Oversight and Reward Signals: Beyond problems with clear answers (coding, math), AI must learn to handle ambiguous, subjective tasks like humor, poetry, and research taste more precisely.

• Scaling to Complex Tasks: AI's application scope must expand from text to multimodal, robotics, and more complex domains.

5. Preparing for the AI Era

Kaplan offers several pieces of advice on how to prepare as AI advances rapidly:

1. Build things that aren't perfect yet

   "AI models are improving quickly, so even if things are a bit lacking now, better models will come soon and complete the product. Experiment at the frontier."

2. Use AI to integrate AI The pace of AI development is so fast that actually integrating it into products and organizations can't keep up, so AI itself should be used to accelerate this process.

3. Identify fields where AI can be adopted quickly Think about which fields, beyond software engineering, will see the next wave of rapid AI adoption.

"The frontier of what AI can do is moving very fast. Experimenting at that frontier is what matters."

6. Claude 4 and Changes in the Latest AI Models

During Q&A, questions about the recently released Claude 4's features and changes came up:

• Improved capabilities as a coding agent

  "Claude 3.7 had a tendency to rush through tests in coding, but Claude 4 has better oversight and code quality."

• Enhanced memory features

  "Claude 4 can store memories in files or logs while working on complex tasks, retrieving them across multiple work sessions."

• Gradual, steady progress

  "What scaling laws show is a curve of incremental improvement. Each version gets a little better. We expect to eventually reach human-level AI."

7. AI-Human Collaboration and the Future of Work

As AI advances, the way humans and AI collaborate is also changing.

• AI mistakes and the human role

  "AI is sometimes astonishingly smart, yet makes absurd mistakes. Humans will become increasingly important as managers who verify AI output and do sanity checks."

• AI replacing entire workflows Recently, AI has been replacing entire work flows rather than simply serving as a "copilot."

• The future of collaboration

  "The most advanced work will still need humans, but more and more tasks will be fully automated."

• AI's breadth and depth of knowledge AI can synthesize vast knowledge that no single human could hold, providing new insights especially across biology, psychology, history, and other fields.

"AI absorbs all of human civilization's knowledge. It can combine knowledge from diverse fields that no single expert could know, offering new insights."

8. AI Research and the Physics Mindset

Kaplan explains how his experience as a physicist has helped in AI research.

• Focusing on macro trends

  "What I learned from physics was to find the biggest picture, the macro trends, and make them as precise as possible."

• Discovering scaling laws

  "The really important thing about scaling laws is improving the slope. The more compute you invest, the greater the advantage."

• AI interpretability research AI interpretability research is actually closer to biology or neuroscience, he explains.

  "Unlike the brain, with AI we can measure every neuron's activity, providing far more data for interpretability research."

9. Limits and Future of Scaling Laws

Discussion continues on how long scaling laws will remain valid and what signals would indicate their limits.

"If scaling laws break down, it's most likely because we made a mistake in training. In the past 5 years, whenever scaling appeared to break, it was usually because we were doing something wrong."

There's also discussion of how to increase efficiency when computing resources become scarce, including lower-precision arithmetic and various other methods.

"AI will keep getting cheaper, and eventually models that can handle very complex tasks end-to-end will hold the greatest value."

10. Skills Individuals Need for the AI Era

Finally, advice follows on how individuals should prepare in an era of advancing AI.

• Understanding how AI models work

  "Understanding how these models work, and the ability to efficiently use and integrate them, is what matters."

• Experiment and take on challenges at the frontier

  "Building and experimenting directly at the most advanced frontier will be extremely valuable."

11. Q&A: Expanding AI Task Duration, RL Data, the Human Role, and More

During Q&A, discussions cover how AI can perform longer tasks, the limits of data labeling in RL, and methods for AI to generate its own tasks.

"For AI to work on longer tasks, it needs the ability to recognize and correct its own mistakes. Even small improvements can double the task duration limit."

"To train AI on more complex tasks, you can create increasingly complex tasks and train with RL. In the worst case, you'd have to create and train on many tasks one by one, but AI can learn more efficiently by self-supervising and providing its own feedback."

Conclusion

This talk emphasizes that AI's advancement is not just a technological leap but is progressing predictably according to clear trends called scaling laws. AI will increasingly handle more complex and longer tasks, and the way humans and AI collaborate will change significantly. The most important thing in this era of change is understanding AI's principles and having the attitude to experiment and take on challenges at the rapidly shifting frontier.

"The frontier of AI is moving fast. Experimenting and taking on challenges at that frontier is the best way to prepare for the future."