New Lex Fridman Insight: François Chollet: Keras, Deep Learning, and the Progress of AI
Sent June 11, 2026
Key Insights
- François Chollet argues intelligence explosion is flawed due to its oversimplified definition of intelligence.
- Scientific progress is linear despite exponential resource consumption, challenging the intelligence explosion narrative.
- Keras' integration with TensorFlow 2.0 offers both high-level usability and low-level flexibility for deep learning.
- Deep learning's limitations necessitate combining it with symbolic AI for complex problem-solving.
- AI algorithms risk behavior manipulation, necessitating user control over recommendation systems.
How the conversation moved
The episode begins with François Chollet questioning the concept of an intelligence explosion, which he argues is based on a flawed definition of intelligence. He suggests that intelligence cannot be isolated to just the brain but is rather a product of the interaction between the brain, body, and environment. This framing challenges the idea that AI could rapidly surpass human intelligence through self-improvement. Chollet emphasizes that improving one part of this interdependent system leads to new bottlenecks, preventing exponential intelligence growth.
Chollet supports his argument by discussing scientific progress, which he claims remains linear despite exponential increases in resource consumption. He references Michael Nielsen's work, which shows that as scientific fields mature, it becomes increasingly difficult to make significant discoveries. This requires more researchers to achieve the same impact, suggesting that the intelligence explosion narrative is more of a belief system than a scientific argument. The conversation then shifts to the development of Keras and its integration with TensorFlow, which democratizes deep learning by offering both high-level usability and low-level flexibility.
Lex Fridman didn't challenge Chollet's critique of the intelligence explosion directly, though one might argue that the potential for rapid AI advancements could still exist in unforeseen ways. The conversation moves to the limitations of deep learning, where Chollet highlights the need for combining it with symbolic AI to tackle complex problems. He argues that while deep learning requires dense sampling, symbolic AI can generalize better using abstract rules, which is crucial for applications like autonomous driving.
The episode concludes with discussions on the societal implications of AI algorithms, particularly the risk of behavior manipulation through recommendation systems. Chollet stresses the importance of user control to prevent the spread of misinformation and bias. The conversation also touches on the importance of benchmarks in evaluating AI systems, emphasizing practical effectiveness over theoretical correctness. Chollet's insights suggest a future where AI development is more nuanced, balancing between deep learning, symbolic AI, and user-centric design.
Surprising moments
In-depth
Intelligence Explosion Critique
- Chollet argues intelligence explosion relies on a flawed definition of intelligence.
- Intelligence emerges from brain, body, and environment interaction.
- Improving one part of an interdependent system leads to new bottlenecks.
- Human intelligence is specialized and not the bottleneck for solving larger problems.
Scientific Progress vs. Resource Consumption
- Scientific progress is linear despite exponential resource consumption.
- Michael Nielsen's approach reveals flat significance graph across disciplines.
- Increasing complexity requires more researchers for the same impact.
- Exponential resource consumption maintains linear progress.
Keras and TensorFlow Integration
- Keras was released in 2015 with an easy-to-use LSTM implementation.
- Theano was initially preferred over TensorFlow for RNNs.
- Keras allows model definition via Python instead of YAML.
- TensorFlow 2.0 combines high-level usability with low-level flexibility.
Deep Learning and Symbolic AI
- Deep neural networks require dense sampling, which is impractical for complex problems.
- Symbolic AI generalizes better than deep learning using abstract rules.
- Successful AI systems combine symbolic AI with deep learning.
- Program synthesis is in its infancy, similar to deep learning in the 90s.
AI Manipulation and User Control
- Large-scale computation allows algorithms to improve exponentially.
- Data quality and scale are the new bottlenecks in AI systems.
- AI algorithms maximize engagement, risking the spread of fake news.
- Recommendation systems can predict and influence user behavior at scale.
Notable Quotes
I was basically questioning this idea, first of all, because the notion of intelligence explosion uses an implicit definition of intelligence that doesn't sound quite right to me.
Still open
- Chollet questioned whether the intelligence explosion narrative is more of a belief system than a scientific argument, leaving its validity open.