AI summary by TwinMind
AI Won’t Plateau — if We Give It Time To Think | Noam Brown | TED
TED
12m 9s
AI Progress Through Scale (2017-2024)
- AI advancement past 5 years = one word: scale
- Frontier models still use transformer architecture from 2017
- Training methods similar to 2019 models
- Main difference: scale of data and compute
- GPT-2 (2019): $5,000 training cost
- Current frontier models: hundreds of millions of dollars
- Concerns about scaling plateau at hundreds of billions/trillions cost
PhD Research: Poker AI Development (2012-2017)
- Started PhD 2012, worked on poker-playing AIs
- Poker = chess with cards (strategy + luck)
- Research paradigm: scale larger models for longer on more data
- Annual improvements through scaling approach
First Competition (2015)
- Bot challenged 4 world's top poker players
- 80,000-hand competition
- $120,000 prize money
- Bot lost by wide margin, outmatched from day one
Key Observation: System 1 vs System 2 Thinking
- Bot: played ~1 trillion hands over 3 months training
- Made decisions in 10 milliseconds instantly
- Human experts: ~10 million lifetime hands
- Took time to think on difficult decisions
- Easy decisions: couple seconds
- Hard decisions: longer thinking time
- Reference: Daniel Kahneman's "Thinking Fast and Slow"
- System 1: fast, intuitive (face recognition, joke comprehension)
- System 2: slow, methodical (vacation planning, essay writing, math problems)
Breakthrough Discovery
- Experiment: bot thinking 20 seconds per hand
- Result: same performance boost as 100,000x model scaling + 100,000x training time
- Initially thought result was bug
- First 3 PhD years: achieved 100x scaling through multiple papers
- Realized System 2 thinking would overshadow all previous scaling work
Second Competition (2017)
- Redesigned poker AI focusing on System 2 thinking
- 120,000-hand competition vs 4 top pros
- $200,000 prize money
- Result: beat all players by huge margin
- Surprised poker community, AI community, and researchers
- Betting odds progression:
- Start: 4:1 against AI
- After 3 days: 50:50
- Day 8: could only bet on which human loses least
System 2 Thinking in Other Games
Chess: Deep Blue (1997)
- IBM's Deep Blue beat world champion Garry Kasparov
- Took couple minutes per move (not instant)
Go: AlphaGo (2016)
- DeepMind's AlphaGo beat world champion Lee Sedol
- Took couple minutes per move
- Later paper findings:
- With thinking time: beat any human by huge margin
- Without thinking time: performed worse than top humans
Scientific Measurement (2021)
- Paper quantified thinking time impact
- 10x thinking time scaling ≈ 10x model size/training scaling
- Clear relationship between System 2 thinking time and System 1 training
Economic Implications
- Current frontier models: hundreds of millions training cost
- Query cost: fractions of penny
- Two scaling paths:
- Continue System 1 scaling: hundreds of millions → billions of dollars
- Scale System 2 thinking: penny per query → 10 cents per query
- Trade-off becomes worthwhile at certain point
OpenAI o1: System 2 for Language
- Released o1 series: language models that think before responding
- Easy questions: few seconds thinking
- Difficult questions: few minutes thinking
- Opens new scaling dimension beyond System 1 training
- Largely untapped scaling direction
Cost-Benefit Analysis
- Current frontier models: <1 penny per query
- Common objection: people won't wait minutes or pay dollars
- Audience experiment: willingness to pay >$1 for:
- Cancer treatment: nearly everyone
- $1,000: many hands still up
- Efficient solar panels: many willing
- Riemann hypothesis proof: some willing
Current Impact and Future
- AI conception: chatbots → broader applications
- Revolution happening now (not 10 years or 2 years away)
- o1 preview already released
- Users report saving days of work
- Top university researchers using successfully
- Just the preview version
Conclusion
- AI progress history: System 1 scaling only
- New paradigm: System 1 + System 2 scaling
- Very beginning of System 2 scaling direction
- Challenge to skeptics: "want to bet?"
