Interview, Fireside Chat
François Chollet: Measures of Intelligence | Lex Fridman Podcast #120
- Future super-intelligent AI systems are expected to view humanity as a historical niche, potentially critiquing human history as a mistake made in the name of progress.
- GPT-N models are predicted to show monotonic performance increases in generating plausible text, yet scaling size and data alone will not resolve issues with factual accuracy, consistency, or output constraints.
- The primary bottleneck for scaling generative transformers is identified as training data scarcity rather than compute power, as current crawling efforts have already covered the vast majority of the web.
- High-quality, human-annotated datasets are projected to yield superior models with faster training times compared to larger, noisy datasets, with self-supervised learning noted as a method for label refinement.
- True intelligence is defined as the ability to achieve human-level skill parity across arbitrary tasks and domains with similar efficiency, requiring flexibility to handle "unknown unknowns" rather than just robust generalization to known distributions.
- Systems achieving Level 5 self-driving capabilities may rely on deep learning for perception combined with explicit environmental models, though this path requires significant data simulation and does not constitute general intelligence.
- The ARC challenge is expected to evolve over time through crowdsourcing to create diverse, bias-free datasets while maintaining a private test set to prevent gaming of the benchmark.
- Machine performance on the ARC challenge will likely reach human parity in the future, correlating with general fluid intelligence, though the timeline is expected to take "a while."
- Testing intelligence through tasks like piloting an airplane after training on driving will reveal actual capability levels, as current models rely on pattern matching rather than true adaptation to novelty.
- Intelligence tests must incorporate "novelty" and "extreme generalization," rendering metrics like the Hutter Prize or Turing test inadequate for measuring cognitive capability due to their focus on compression of the past or reliance on deceptive behavior.
- Human cognition is expected to remain far from optimal intelligence bounds despite having a hard limit, with externalized cognition via culture and AI continuing to scale collective intelligence far beyond individual brain capabilities.
- Efforts to create interactive tests based on scientific method principles, such as measuring the number of attempts or efficiency in reducing uncertainty, are planned, alongside collaborations with NYU psychology to correlate machine performance with human characteristics.
- The field should prioritize measures of "skill acquisition efficiency" and the ability to handle fundamental uncertainty over static skill performance, with core knowledge theories suggesting innate priors like objectness and geometry should underpin machine intelligence tests.
- Future intelligent systems are not expected to resemble humans, as human likeness is viewed as the final step in the development of generalization capabilities.
- François Chollet intends to continue work on a self-published science fiction novel and remains skeptical regarding the efficacy of neural interfaces due to existing sensory bandwidth constraints in human perception.
- The future of AI development involves expanding datasets through crowdsourcing while acknowledging that deep learning models, even at 100 trillion parameters, cannot fundamentally change nature without continuous learning to handle novelty.