François Chollet says OpenAI set AGI progress back by five to 10 years

ARC Prize co-founder François Chollet said OpenAI had set progress toward artificial general intelligence (AGI) back by five to 10 years, according to TechRadar’s account of his comments on the Dwarkesh Podcast.

Chollet argued that large language models (LLMs), which are based on neural networks, are a dead end for achieving AGI. He said LLMs could still become more powerful, autonomous and useful to businesses, but that such improvements would not amount to AGI.

In Chollet’s view, funding LLM development at the expense of other approaches means the path to AGI is being neglected. TechRadar also reported his assessment that progress had been set back by up to a decade.

The ARC-AGI benchmark, co-founded by Chollet, is intended to measure progress toward AGI. TechRadar reported that LLMs are performing better on the benchmark, but none has come close to meeting its mark. The article describes AGI as human-like intelligence that can reason and learn new capabilities outside a model’s training data.

Chollet’s comments distinguish improvements in current models from achieving AGI; they do not establish when, or whether, AGI will be achieved.