François Chollet: Keras, Deep Learning, and the Progress of AI

In the conversation with Lex Fridman, François Chollet, the creator of Keras and AI researcher at Google, delves into a wide-ranging discussion on the development of Keras, the state of deep learning, and his views on the progress and future of AI. Chollet begins by reflecting on the controversial ideas he's shared online, particularly his skepticism of the 'intelligence explosion' hypothesis, which suggests that an AI could recursively self-improve to the point of superhuman intelligence. Chollet argues that intelligence is not a single, scalar property but emerges from the complex interactions between a brain, a body, and an environment. He also emphasizes that intelligence is inherently specialized and that the idea of an 'intelligence explosion' is flawed because it overlooks the bottlenecks and interdependencies within complex systems.

Chollet then recounts the history of Keras, its beginnings as a response to the lack of user-friendly deep learning libraries, and its eventual integration with TensorFlow. He highlights the design decisions made during Keras' development, such as the choice to define models using Python code rather than static configuration files, which was against the mainstream at the time. He also discusses his work on integrating the Keras API into TensorFlow and the development of TensorFlow 2.0, expressing excitement about the higher-level APIs and the potential of automated machine learning.

Throughout the episode, Chollet touches on various topics, including the limitations of deep learning, the need for benchmarks that measure the generalization power of AI, and the risks associated with AI, such as mass surveillance and the manipulation of human behavior through recommendation algorithms. He expresses concern about the overhyping of AI capabilities and the potential for backlash if promises about technologies like autonomous vehicles are not fulfilled. Chollet concludes with the idea that AI's usefulness, rather than its theoretical correctness, is what ultimately matters in assessing its value.