Interview, Conference Presentation
Are We On Path Towards Superhuman Intelligence? – Dario Amodei (Anthropic CEO)
- Models are projected to increasingly perform the primary workload as they accelerate human tasks, with performance metrics expected to rise relatively linearly despite inherent unpredictability.
- Advancement trajectories are predicted to be fast and steep, driven by scaling laws that continue to bend, requiring increasing precision to manage entropy, with financial investment potentially rising by a factor of one hundred alongside faster chips and improved algorithms.
- Current economic spending and acceleration are anticipated to proceed left to themselves unless actively interrupted by safety measures and government regulation, though the exact mathematical nature of this growth remains uncertain.
- While a new architecture on the scale of a transformer is considered possible for the next iteration, the current trajectory is expected to continue largely unchanged even without such inventions, with further discoveries unlikely to significantly accelerate an already rapid pace.
- Capabilities are expected to improve broadly across all areas without specific stagnation, matching the level of a generally well-educated human within two or three years.
- The ecosystem reaction to Google is estimated to be ten times more important than other factors in driving change, serving as a key variable alongside the ongoing increase in spending.
- AI systems are not expected to achieve superhuman status in physical world embodiment without a recursive loop of faster AIs training faster AIs to solve that specific problem.
- Productivity is forecast to follow a sequence of first speeding up human work, then equaling it, and eventually becoming the main contributor to scientific progress.
- Critical thresholds for existential danger, the capacity to take over most AI research, and the ability to seriously change economic structures are not expected to coincide with the benchmark for human-level general capability.
- These various high-impact thresholds are anticipated to occur at various times after the base capability of matching a well-educated human is reached, while the detailed future remains unpredictable and potentially counterintuitive.