François Chollet: Why Scaling Alone Isn’t Enough for AGI
François Chollet, creator of Keras and founder of NDIA (a new AGI research lab), argues that scaling alone is insufficient for achieving AGI and proposes an alternative approach based on symbolic program synthesis rather than deep learning. He discusses why current LLM-based systems excel at formally verifiable domains like coding but struggle with open-ended reasoning, and introduces ARC AGI v3, a benchmark measuring agentic intelligence through interactive environments rather than pattern recognition.
Key takeaways
- • Scaling deep learning has hit diminishing returns for reasoning tasks; LLMs score consistently below 10% on ARC v1 regardless of model size, indicating a fundamental architectural limitation rather than a data problem.
- • Symbolic program synthesis seeks to replace parametric neural networks with minimal, interpretable symbolic models that generalize better with less data—analogous to how scientific laws compress observations into elegant equations rather than fitting curves.
- • Domains with formally verifiable reward signals (code, mathematics) enable rapid automation through reinforcement learning, while subjective domains (essay writing) remain bottlenecked by costly human annotation and slow progress.
- • ARC AGI v3 measures agentic intelligence by dropping AI agents into novel game environments with no instructions, requiring them to match human-level exploration efficiency and goal discovery—a fundamentally different challenge than static pattern recognition.
- • Current reasoning model breakthroughs represent improved training and knowledge, not higher fluid intelligence; models like OpenAI o3 succeed through trial-and-error in verification-rich environments, not because they became "smarter."
- • AGI is likely achievable by the early 2030s through either scaling approaches hitting a breakthrough or alternative methods like NDIA's symbolic approach, suggesting the path forward remains uncertain but the timeline is compressed.
Mentioned (19)
More from these creators
Inside The Startup Reinventing America’s Trillion Dollar Chemical Industry
India’s Fastest Growing AI Startup
The Future Of Brain-Computer Interfaces
Common Mistakes With Vibe Coded Websites
The Powerful Alternative To Fine-Tuning
The AI Agent Economy Is Here