Scaling and the Road to Human-Level AI | Anthropic Co-founder Jared Kaplan

Jared Kaplan, co-founder of Anthropic and one of the pioneers of scaling laws research, shares insights on how AI models are trained and why scaling is the fundamental driver of AI progress.

From Physics to AI

Background

• Spent most of his career as a theoretical physicist
• Mother was a science fiction writer - inspired interest in physics to explore faster-than-light travel
• Interested in fundamental questions: Is the universe deterministic? Do we have free will?
• Met many Anthropic co-founders during physics career
• Initially skeptical about AI (“SVMs aren’t that exciting”)
• Convinced to transition to AI about six years ago

Two Phases of AI Training

Phase 1: Pre-training

• Train AI models to imitate human-written text
• Learn correlations underlying the data
• Teach models what words are likely to follow other words
• Now includes multimodal data in contemporary models

Phase 2: Reinforcement Learning

• Use human feedback to optimize model behavior
• Reinforce behaviors chosen to be helpful, honest, and harmless
• Discourage bad behaviors
• Early Claude versions used simple comparison interfaces

Scaling Laws: The Key Discovery

The Fundamental Insight

• Asked “dumb” physicist questions: How big should the data be? How much does it help?
• Discovered remarkably precise trends - as precise as anything in physics or astronomy
• Performance improves predictably as you scale up compute, data, and model size

Why This Matters

• Trends observed across many orders of magnitude
• Gives conviction that AI will keep getting smarter predictably
• Not about researchers suddenly getting smart - found a systematic way to improve AI

Scaling in Reinforcement Learning

• Andy Jones studied scaling laws for AlphaGo using the simpler game Hex
• Found similar straight-line trends in RL training
• Both pre-training and RL show scaling benefits

AI Capabilities: Two Axes

Flexibility (Y-axis)

• Ability of AI to meet us where we are
• Progress from narrow systems (AlphaGo only plays Go) to multimodal models
• Moving toward handling all modalities humans can deal with

Task Horizon (X-axis) - More Interesting

• How long would it take a person to do tasks AI can do
• Doubling roughly every 7 months (per METR research)
• Currently at hours-scale tasks
• Projecting toward days, weeks, months, years

Future Vision

• AI models or millions of models working together
• Able to do work of entire human organizations
• Could make 50 years of theoretical physics progress in days/weeks

What’s Left for Human-Level AI?

Key Ingredients Needed

1. Organizational Knowledge

   • AI needs context like someone who’s worked at a company for years
   • Not just blank slate interactions

2. Memory

   • Track progress on long tasks
   • Build and retrieve relevant memories
   • Being built into Claude 4

3. Oversight/Supervision

   • Understand fine-grained nuances
   • Solve hard, fuzzy tasks
   • Generate nuanced reward signals for RL
   • Beyond just “code passes tests” or “math answer is correct”

4. Complexity Progression

   • Text models to multimodal to robotics
   • Continued gains from scale expected

Advice for Builders

Build Things That Don’t Quite Work Yet

• AI models are getting better very quickly
• If Claude 4 is “a little bit too dumb” for your product, Claude 5 will make it work
• Experiment on the boundaries of what AI can do

Use AI to Integrate AI

• Main bottleneck: AI developing faster than we can integrate it
• Leverage AI for AI integration to speed adoption

Find Fast Adoption Areas

• Software engineering exploding with AI integration
• Question: What’s next beyond coding?
• Look for areas where 70-80% accuracy is good enough

On Claude 4

Improvements Over 3.7 Sonnet

• 3.7 was “a little bit too eager” - would do anything to make tests pass
• Claude 4 improves agentic capabilities for coding and other applications
• Better supervision and oversight
• Follows directions better, improves code quality
• Better memory storage and retrieval across context windows

The Scaling Picture

• Scaling laws paint picture of incremental progress
• Smooth curve toward human-level AI/AGI
• Each release gets steadily better in many ways

Human-AI Collaboration

Current State

• AI makes brilliant moves but also basic errors
• Key difference from humans: judgment vs generative capability much closer in AI
• Humans can judge things they can’t do; AI’s judgment and generation are similar
• Major role for humans: managers who sanity-check AI work

The Shift

• Moving from co-pilot (human approval needed) to end-to-end task completion
• Founders now selling direct workflow replacements
• Human-AI collaboration most interesting for advanced tasks

AI’s Unique Strength: Breadth

Depth vs Breadth

• Math: work on one theorem for a decade (depth)
• Biology, psychology, history: putting together vast information across areas (breadth)
• AI during pre-training imbibes all of human civilization’s knowledge
• Particular overhang in areas requiring breadth no single human expert has

Green Field Opportunities

• Finance (Excel spreadsheet users)
• Law (though more regulated)
• Any skilled task involving sitting at a computer interacting with data
• Integrating AI into existing businesses

Physics Background and AI Research

The Physicist’s Approach

• Ask the “dumbest possible questions”
• Look at the big picture
• Notice very simple trends
• Scaling laws discovery came from this mindset

Predicting the Future

• Scaling laws give one way to predict: trends will continue
• Like GDP and economic trends - reliable indicators
• Details of implementation remain hard to predict