How François Chollet Is Building A New Path To AGI
Y Combinator Startup Podcast Podcast Recap
Published:
Duration: 57 min
Guests: François Chollet
Summary
François Chollet discusses his innovative approach to achieving artificial general intelligence (AGI) through symbolic models and program synthesis, challenging the reliance on deep learning. His work with ndia aims to create more efficient AI systems by focusing on concise, symbolic models.
What Happened
François Chollet is pioneering a new direction in AI research with his lab, ndea, where he focuses on program synthesis as an alternative to deep learning. He introduced the ARC Prize, a global competition aimed at solving the ARC AGI benchmark, which evaluates reasoning and skill acquisition efficiency.
Chollet's approach relies on symbolic models, which he claims are more concise and efficient than traditional parametric models like deep learning. He utilizes a technique called symbolic descent, which requires less data and computational power, making it potentially more sustainable in the long term.
He expresses skepticism about the future of current AI systems, such as large language models (LLMs), predicting they may not be the foundation of AI in 50 years. Chollet estimates a 10-15% chance that his approach will succeed but believes that the potential benefits make it worth pursuing.
Chollet explains his definition of general intelligence as the human-level efficiency in learning various tasks, emphasizing that current technology can automate domains with verifiable rewards, such as coding. He foresees a similar revolution in mathematics due to the presence of verifiable rewards.
The ARC AGI benchmark has evolved through versions, with ARC AGI V3 focusing on agentic intelligence, measuring exploration efficiency and goal setting in environments without prior instructions. A video game studio was established to create over 250 unique games for ARC 3, designed to test fluid intelligence by evaluating the efficiency of solving new problems.
He argues that AGI could be achieved with a compact codebase, potentially less than 10,000 lines, suggesting that even 1980s computational resources could suffice. This contrasts with projects like Cyc, which lacked learning and relied on handcrafted symbolic knowledge.
Chollet's vision for AI involves building from first principles, emphasizing symbolic compression and program synthesis, as seen in his development of Keras, which focused on usability and community building. His predictions include the possibility of achieving AGI by the early 2030s, while also acknowledging the potential for other AI approaches such as genetic algorithms.
Chollet advises leveraging AI for personal empowerment and encourages learning about AI and its applications. He emphasizes the importance of creating AI systems that are efficient and adaptable, rather than merely following current trends in deep learning.
Key Insights
- François Chollet's ndea lab focuses on program synthesis, aiming to develop AI systems that are more efficient than deep learning by using symbolic models that require less data and computational resources.
- The ARC AGI benchmark, created by Chollet, evaluates AI's reasoning and skill acquisition efficiency, with ARC AGI V3 specifically measuring agentic intelligence and exploration efficiency without prior instructions.
- Chollet believes that current AI technologies like large language models may not remain foundational in the next 50 years, advocating for a shift towards methods that emphasize human-level learning efficiency.
- AGI, according to Chollet, could be achieved with a minimal codebase, potentially less than 10,000 lines, highlighting the importance of symbolic compression and program synthesis in future AI development.