AI summary by TwinMind
Everyone Was Wrong About Intelligence – Dario Amodei (Anthropic CEO)
Dwarkesh Patel
8m 18s
Predictability and Limits of Scaling Laws
- Scaling laws for AI models: predictable so far
- Uncertainty about commercial explosion, forms, AI replacing or pairing with humans
- Track record of predictions: self-admitted poor, estimates ~10% accuracy
- No one has a notably good prediction record
- Intelligence spectrum: village idiot to Einstein diagram referenced
- Human intelligence range broad, not neatly covered by AI at same time or tasks
Human Range and Task Performance Variability
- Tasks like writing a sonnet in the style of Cormac McCarthy: high-level human skill, models improving
- Constrained writing (e.g., no letter 'e'): models possibly superhuman or close
- Mathematical theorem proving: models only beginning, make simple mistakes, lack broad error correction or extended reasoning
- Intelligence not a linear spectrum; involves multiple domains, skills (memory, etc.)
- Theories of intelligence often dissolve into a continuum
- Less thinking in terms of "intelligence," more in terms of observed abilities
Reflections on Progress and Surprises
- In 2018, would have considered current models (2023) as AGI-level based on benchmarks (e.g., writing theorems in Shakespeare’s style, open-ended standardized test performance)
- Now, despite impressive benchmarks, models clearly not human-level intelligence
- Discrepancy between benchmark performance and general intelligence noted
Model Architecture, Scaling, and Biological Analogies
- Initial belief (2020): Models had grasped language essence; scaling vs. RL (reinforcement learning) tradeoff considered
- Ongoing scaling: models kept improving, contrary to expectations that language plateau was near
- AI models: 2-3 orders of magnitude fewer parameters/synapses than human brain
- Trained on 3-4+ orders of magnitude more data than a human child to age 18
- Biological analogy seems weak; models much smaller, require more data, but do many human tasks
- Scaling remains key; measuring ability vs. humans more important than analogies
Creativity, Discovery, and Knowledge Connections
- Models have "memorized" vast human knowledge, but no genuinely new discoveries
- Human with same knowledge base might make novel connections/discoveries (e.g., medical cures)
- Definitions of creativity/discovery become fuzzy in large compute context
- Models show "ordinary creativity" (e.g., writing sonnets in unique styles)
- Models can draw new connections similar to ordinary person; no major scientific discoveries yet
- Attributed to current skill level, expected to change with further scaling
- Question: Should models’ vast knowledge already confer an advantage in making discoveries?
- Models are ahead in memorization/fact-connection
- Biology: Models potentially close to making new discoveries due to knowledge density, but not quite at skill threshold
- Physics: Requires more abstract thinking and formula creation
Conclusions and Open Questions
- AI models: On the cusp of integrating knowledge to make new discoveries, especially in complex fields like biology
- Progress surprises: Abilities emerge unevenly across tasks/domains
- Theories and analogies (esp. biological) less useful than empirical measurement and scaling
