In a recent statement, François Chollet, co‑founder of the ARC Prize, claimed that OpenAI’s focus on large language models (LLMs) has pushed the goal of artificial general intelligence (AGI) back by five to ten years. The remark was made during a conference in San Francisco on 3 October 2026, drawing attention from researchers and industry observers worldwide.
Chollet argues that while LLMs demonstrate impressive language abilities, they rely on massive data sets and computational resources rather than novel architectural breakthroughs. He believes that the industry’s emphasis on scaling current models diverts talent and funding from exploring fundamentally different approaches that could accelerate AGI. According to him, this shift in priority has created a bottleneck, delaying progress that might otherwise have been achieved sooner.
The debate centers on whether progress in AI should be measured by the size of models or by advances in underlying design. Proponents of LLMs say that larger models improve performance across tasks, reducing the need for task‑specific engineering. Critics, including Chollet, contend that without new architectures, these models will plateau, offering diminishing returns. They point to research on modular, symbolic, and neuro‑symbolic systems as promising alternatives that could break current limitations. The industry’s current investment patterns reflect a preference for incremental scaling, raising concerns about long‑term innovation.
Chollet’s estimate of a five‑to‑ten‑year setback is based on comparative timelines of past AI milestones. He cites the rapid rise of LLMs since 2020, yet notes that breakthroughs in AGI remain elusive. Critics argue that the delay is speculative and that LLMs may still indirectly contribute to AGI by uncovering new training techniques. Others suggest that the gap between LLM capability and true general intelligence is narrower than Chollet implies. Thus, the claim remains controversial, with evidence both supporting and challenging the projected timeline.
The broader implications of this viewpoint are significant. If the industry continues to prioritize scaling over exploration, funding agencies may face a misallocation of resources. Academic institutions might shift curricula toward data‑driven methods, potentially neglecting foundational research. Meanwhile, policymakers could be misled into assuming that AGI is nearer than it truly is, affecting regulation and investment decisions.
What evidence supports Chollet’s claim of a delay? Chollet references the lack of AGI breakthroughs despite large model deployments and argues that scaling alone has not produced new architectural insights, suggesting a lag in progress.
How might the industry respond to this criticism? Some firms are already investing in hybrid models that combine symbolic Could LLMs still play a role in achieving AGI? Yes, many researchers believe that LLMs can serve as foundational tools, providing data and training signals that inform more advanced architectures, even if they are not AGI themselves.