Chollet: How We Get to AGI

Notes from François Chollet’s talk: intelligence as process, the ARC suite, and why we need to combine type-1 (continuous) and type-2 (discrete) abstractions.

Watch Chollet’s talk on YouTube.

François Chollet, creator of Keras and the ARC suite of AGI benchmarks, gave a very thoughtful talk on how we can actually get to AGI [1]:

• Intelligence is a process, not skills
  • Intelligence is an efficient conversion from past experience to potential operating area
  • AGI is not about automation, it’s about invention
• Three characteristics of intelligence
  • Fluid intelligence
  • Broad operational area
  • Information-efficient program acquisition and synthesis
• The evolution of ARC benchmarks
  • ARC-1: Prior-free reasoning tasks
  • ARC-2: Compositional reasoning tasks
  • ARC-3: Reasoning with interactive agency
• Intelligence implementation requires two components
  • Abstraction acquisition
  • Abstraction recombination (application)
• Two types of abstractions
  • Type-1: prototype/value-centric, continuous — transformer is good at this
  • Type-2: program-centric abstraction/graphs, discrete/search — ???
  • Type-2 is data efficient but test-time inefficient due to combinatorial explosion.
• Intelligent systems need to combine Type-1 and 2 abstractions
  • Type-2 can be guided by type-1
  • Need global library of abstract subroutines

---

References

[1] Chollet, François. “How We Get To AGI.” YouTube. https://youtu.be/5QcCeSsNRks

Originally posted on LinkedIn.