Conference Presentation, Keynote
François Chollet: How We Get To AGI
- Compute costs are projected to decline by two orders of magnitude every decade with no indication of the trend stopping.
- General intelligence is predicted to emerge spontaneously from scaling data into larger models, though a community pivot toward test-time adaptation began in 2024.
- There is a possibility among some observers that AGI already exists, yet current systems like the O3 model are considered below human level as they struggle with tasks easy for humans but difficult for AI.
- Intelligence is defined by the efficiency of operationalizing past information for the future, with fluid intelligence requiring test and adaptation rather than pre-training scaling alone.
- Future development will target fluid intelligence, specifically the ability to adapt and invent, moving beyond the automation of known tasks.
- The research community is directed toward solving bottlenecks identified by the ARC benchmark, which serves as a directional guide rather than the final destination for AGI.
- The ARC-1 benchmark is currently saturating, necessitating more sensitive evaluation tools like ARC AGI 2 for granular assessment.
- ARC AGI 3, designed to assess agency, exploration, interactive learning, and autonomous goal achievement, is scheduled to launch in early 2026 following a developer preview in July.
- Solving ARC levels is viewed as a progressive milestone; solving ARC-2 remains very difficult currently, ARC-3 is further away, and ARC-4 is required to achieve AGI.
- The "kaleidoscope hypothesis" posits that novelty arises from the recombination of a small number of unique meaning atoms rather than true invention.
- Achieving intelligence requires the acquisition of abstractions and their on-the-fly recombination, specifically merging continuous value-centric and discrete program-centric types.
- AI systems are expected to evolve into entities resembling programmers that synthesize task-specific programs or models using discrete program search guided by deep learning intuition.
- A global library of reusable abstraction building blocks is planned to be constantly evolved, allowing the system to upload new blocks similar to a GitHub repository.
- The ultimate objective is an AI capable of facing novel situations and quickly assembling working models using its library, mimicking a human software engineer.
- The Endia research lab system is expected to continuously improve itself over time by expanding its abstraction library and refining its intuition about program space.
- The first milestone for the Endia system involves solving ARC-GIS using a system initialized with zero prior knowledge of that domain.
- The system is intended to be leveraged for scientific research to empower human researchers and accelerate the timeline of scientific discovery.