Interview, Fireside Chat
François Chollet: Scientific Progress is Not Exponential | AI Podcast Clips
- Scientific output across fields including physics, biology, medicine, and deep learning is expected to exhibit linear aggregate progress over the coming years and decades, despite exponentially increasing resource consumption, paper publication counts, and researcher headcount.
- The temporal density of significance is projected to remain flat for the next 100 to 150 years as low-hanging fruit is exhausted, forcing future discoveries to require significantly more effort and knowledge ingestion to achieve impact.
- Exponential friction is anticipated to limit problem-solving systems and artificial intelligence, creating bottlenecks such as the speed of light, air friction, and communication overhead that prevent infinite acceleration or an intelligence explosion.
- Practical experimentation will likely face exponentially rising costs as equipment requirements escalate, while system expansion will generate increasingly difficult synchronization challenges due to the growing number of participants.
- While a decrease in scientific investment is predicted to slow progress, the emergence of new, lower-hanging fruits is expected to attract researchers to address these remaining areas and maintain the linear trajectory.
- Deep learning specifically is forecast to show decreasing per-paper significance and exponentially increasing paper counts, resulting in an aggregate significance sum that appears roughly linear.
- Self-improving AI systems are expected to mirror the exponential friction dynamics of scientific institutions rather than escaping them, challenging the prevailing narrative of an intelligence explosion.
- Disputing the concept of an AI-driven intelligence explosion is predicted to generate significant resistance from individuals whose identity is tied to this belief system.