Keep pulling the thread on Francois Chollet.
The most significant recent advance in AI is the development of models that can adapt at test-time to novel problems, a capability demonstrated by models like OpenAI's O3 and top entries in the Arc Rust competition.
My estimated timeline for the arrival of AGI has shortened from approximately 10 years to 5 years.
A future AGI system will achieve continual learning through a global, shared database of reusable abstractions, programs, and templates.
An AGI's superhuman learning ability will stem from millions of instances learning in parallel and sharing knowledge synergistically, rather than from any single instance being more efficient than a human.
I have become more bearish on the near-term usefulness of AI models due to their inability to learn on the job.
The narrative that simply scaling up the GPT-4 architecture would lead to AGI has not materialized.
The introduction of AGI will not lead to a 'singularity' of continuous exponential change because real-world systems, which are mostly composed of humans, have inherent bottlenecks that will limit the pace of progress.
A technological singularity should be defined by its economic impact, such as achieving over 30% annual economic growth, rather than by the raw intelligence of an individual AI model.
Despite dramatically more resources being dedicated to science compared to 100 years ago, the pace of major scientific progress is not accelerating.
True intelligence is defined by computational and data efficiency—the ability to achieve more with fewer resources—not by brute-force search.
Current memory features in commercial LLMs work well for coding because the codebase provides an external, language-based scaffold, a structure that is absent in most other types of work.
A system that requires enormous amounts of compute at test time to solve simple puzzles, like those in the ARC benchmark, is demonstrating brute-force search, not genuine intelligence.