Interview
Everyone Was Wrong About Intelligence – Dario Amodei (Anthropic CEO)
- Scaling laws are considered predictable, though the timing, form, and human-model integration of a commercial explosion remain uncertain.
- Current models may achieve superhuman or near-superhuman performance on constrained writing tasks but will likely continue making errors in proving mathematical theorems and executing extended tasks for an unspecified duration.
- Cognitive abilities are expected to develop asynchronously across models, creating disparities where specific skills, such as coding, are mastered while others, like proving the prime number theorem, remain out of reach.
- Theoretical understanding of intelligence may shift toward viewing it as a continuum rather than a fixed threshold, bridging the gap between high benchmark scores and human-level intelligence.
- Current and future model generations are not expected to represent human-level intelligence despite demonstrating impressive capabilities on standardized tests and creative writing.
- Strategic focus may shift from reinforcement learning to identifying alternative objectives or efficiency gains that could explain the discrepancy between model size and data volume.
- Continued scaling is anticipated as the primary trajectory, with the relevant metric being the remaining distance to human performance, which is estimated to be relatively small.
- Significant scientific discoveries are predicted to emerge as model capabilities reach the threshold necessary to synthesize vast amounts of memorized knowledge, particularly in complex biological fields.