Keep pulling the thread on François Chollet.
François Chollet predicts that Artificial General Intelligence (AGI) will most likely be achieved in the early 2030s.
Endia is developing a new machine learning substrate based on program synthesis as an alternative to the parametric learning used in deep learning.
François Chollet predicts that in 50 years, AI will not be built on the current deep learning and LLM stack due to its inefficiency.
Any problem domain with formally verifiable solutions and a trusted reward signal can be fully automated with current LLM-based technology.
The performance of base LLMs on ARC-AGI V1 remained low despite a 50,000x increase in model scale, indicating that scaling pre-training alone was insufficient for fluid intelligence.
The ARC-AGI V3 benchmark is designed to measure "agentic intelligence" by placing an AI in a novel, game-like environment where it must discover goals and controls through trial and error.
François Chollet predicts that in retrospect, AGI will be discovered to be achievable with a codebase of less than 10,000 lines and could have been built in the 1980s with the computing resources available then.
François Chollet predicts AGI will be achieved around the time the ARC 6 or ARC 7 benchmarks are released.
Endia is a research lab founded by François Chollet that is focused on program synthesis.
Endia is developing a method called "symbolic descent," designed as the symbolic space equivalent of gradient descent.
Endia's symbolic models are expected to require less training data and be more efficient at inference time compared to deep learning models.
François Chollet estimates that Endia's ambitious vision has a 10% to 15% chance of success.