Interview
Everyone Was Wrong About Intelligence – Dario Amodei (Anthropic CEO)
- Predictive accuracy regarding the timeline and form of commercial AI explosion remains poor across the industry, with even experts averaging success rates of only ~10%.
- Intelligence in current models is not a linear spectrum but a wide distribution of domain-specific skills that develop at disparate rates.
- Models demonstrate superhuman or near-superhuman capability in constrained creative writing tasks (e.g., writing a sonnet in a specific author's style or omitting specific letters).
- Models remain in early developmental stages regarding rigorous mathematical proof and exhibit high error rates in extended, multi-step reasoning tasks without self-correction.
- Pre-training scaling has proven unexpectedly efficient compared to alternative approaches like Reinforcement Learning (RL), contradicting earlier hypotheses that RL would be the primary driver of future capability.
- There is a significant discrepancy in resource efficiency between artificial and biological intelligence:
- Current models possess 2–3 orders of magnitude fewer synapses than the human brain.
- Conversely, models are trained on 3–4 orders of magnitude more data than a human encounters by age 18.
- Skepticism regarding biological analogies has increased as empirical data shows models can replicate human-level tasks despite these massive architectural and data efficiency gaps.
- Current-generation models, despite memorizing the entire corpus of human knowledge, have not yet generated new scientific discoveries or novel connections leading to major breakthroughs.
- The speaker attributes this gap to current model skill levels being insufficient to synthesize known facts into new insights, rather than a lack of inherent creativity.
- Biology is identified as a near-term domain where this synthesis may occur, as success relies heavily on breadth of knowledge ("knowing A and B") rather than the derivation of new physical laws.
- Future improvements are expected to bridge the gap between memorization and discovery, with the models currently described as being "on the cusp" of integrating vast knowledge bases to produce novel connections.