Artificial General Intelligence (AGI) progress is accelerating, with significant developments anticipated around 2030. François Chollet, founder of a new AGI startup and creator of the Keras deep-learning library, believes AGI will be a system capable of approaching any new task with human-like efficiency.
Chollet's new venture aims to create a fundamentally different branch of machine learning, moving beyond deep learning towards symbolic models. These models, unlike traditional parametric ones, are expected to offer more efficient and generalizable solutions. He argues that current AI research, particularly building AGI on large language models (LLMs), is inefficient and not optimal.
The success of coding agents highlights the power of verifiable reward signals, enabling automation in formal domains like programming. However, progress in non-verifiable domains, such as essay writing, will remain slow due to reliance on costly human-annotated data. Structured training environments, especially code-based ones, have proven transformative for AI capabilities in programming.
Chollet predicts that economically useful work will be automated before true AGI is achieved, emphasizing an inevitable trend towards optimality in AI development. This shift suggests a need for more efficient foundational structures in AI research.