AI Pioneer Chollet Declares Scaling Dead, Pursues Inventive Human-Like AGI

Beyond bigger models: François Chollet redefines AI's future, prioritizing adaptive reasoning and novel problem-solving for true general intelligence.

July 4, 2025

AI Pioneer Chollet Declares Scaling Dead, Pursues Inventive Human-Like AGI
A prominent voice in artificial intelligence research is challenging the industry's prevailing dogma, arguing that the relentless pursuit of bigger models is a path of diminishing returns. François Chollet, a Google AI researcher and the creator of the popular Keras deep-learning library, contends that the era of achieving greater intelligence simply by scaling up current architectures has reached its end.[1][2][3] Instead, he posits that the future of AI lies in developing systems that can adapt, reason, and solve novel problems with the ingenuity of a human programmer, a stark departure from the brute-force memorization capabilities of today's large language models (LLMs).[1][4] This perspective signals a potential paradigm shift in the quest for artificial general intelligence (AGI), moving the focus from sheer size to the quality and efficiency of intelligence itself.[5][6]
Chollet's critique of the scaling-first approach is rooted in the fundamental limitations he observes in current LLMs.[7] While these models demonstrate impressive fluency with language and can perform a vast number of tasks, he argues this is a form of "crystallized skill" rather than true intelligence.[8] Their performance is based on recognizing patterns within massive training datasets, effectively a high-tech form of memorization.[9] When faced with genuinely new problems that fall outside their training distribution, their performance plummets.[10][3] Chollet highlights that many current AI benchmarks are flawed because they can be "gamed" by ensuring similar problems are included in the training data, thus measuring task-specific skill rather than genuine problem-solving ability.[7][9] He points to the fact that despite exponential increases in model size and computing power, fundamental issues like logical reasoning failures and an inability to generalize from a few examples persist.[8][3] This suggests that simply making models bigger will not spontaneously give rise to the flexible, adaptive reasoning that characterizes human intelligence. A majority of AI experts appear to share this skepticism, with one survey finding that 76% believe scaling current approaches is unlikely to lead to AGI.[7]
To steer research toward what he defines as "fluid intelligence"—the ability to efficiently acquire new skills and adapt to unfamiliar situations—Chollet introduced the Abstraction and Reasoning Corpus (ARC) in 2019.[11][12][10] ARC is a benchmark composed of a series of visual reasoning puzzles. Each task provides a few "before and after" examples of a grid transformation, and the AI must deduce the underlying rule to solve a new, unseen input grid.[11][13] The tasks are designed to be easy for humans, who average around 84-85% accuracy, but have proven exceptionally difficult for even the most advanced AI systems.[11][14] Unlike many other benchmarks, ARC is resistant to memorization and brute-force data training, as the evaluation tasks are novel.[9][15] It specifically tests for the ability to form abstractions and perform reasoning from a small number of examples, probing the core cognitive abilities that Chollet believes are essential for true intelligence.[10][5] For years, even the largest models scored near zero on ARC, a result Chollet points to as decisive evidence that fluid intelligence does not simply emerge from scaling up pre-training.[4]
In response to recent progress and the evolving landscape of AI, Chollet has introduced newer iterations of the benchmark, including ARC-2 and a previewed ARC-3.[4] While the first version of ARC (ARC-1) was a binary test for the presence of fluid intelligence, the newer versions are designed to be more sensitive and probe more sophisticated abilities like compositional reasoning.[4] The recent shift in the research community toward "test-time adaptation"—models that can modify their own behavior and learn during inference—has led to significant progress on ARC.[4] However, Chollet emphasizes that even with these advances, the performance of top AI models still lags far behind that of humans.[4][2] ARC-3 is expected to further raise the bar by introducing interactive agency, requiring systems to not just solve static puzzles but to interact with an environment to achieve goals.[4] Through the ARC Prize, a global competition with a substantial prize pool, Chollet is incentivizing the research community to tackle these harder problems and explore new avenues beyond the mainstream LLM paradigm.[2][16][9]
Chollet’s path to AGI diverges sharply from the current industry focus. He is a strong advocate for approaches that combine the pattern-recognition strengths of deep learning with the logical capabilities of symbolic reasoning, often termed neuro-symbolic AI.[2][17][18] His vision for AGI is not a super-charged chatbot but a system capable of autonomous invention and scientific discovery.[4][18] He argues the goal should be to create systems that can generate new solutions and skills on the fly, a process he likens to program synthesis, where an AI writes a new program to solve a novel problem.[7][8][19] This requires a shift away from static, pre-trained models toward dynamic systems that can learn efficiently from very little experience.[8][6] To pursue this vision, Chollet recently left Google to launch a new AI research lab called Ndea, which will focus on developing AGI through these alternative methods.[4][20]
In conclusion, François Chollet presents a compelling counter-narrative to the dominant "scaling is all you need" philosophy in AI. He argues that the industry's focus on ever-larger models has led to systems that are skilled but not intelligent, adept at regurgitating information but incapable of genuine reasoning or adaptation.[1][8] Through his work on the ARC benchmark, he is pushing the field to redefine its metrics for success, prioritizing skill-acquisition efficiency and the ability to solve truly novel problems.[11][5] His proposed path forward, centered on principles like program synthesis and the fusion of neural and symbolic methods, outlines a vision for an AGI that functions less like a vast database and more like an inventive human scientist.[4][2] As the AI community grapples with the limitations of current approaches, Chollet's ideas and the challenges posed by ARC may prove instrumental in navigating the complex and uncertain road to creating truly general artificial intelligence.

Sources
Share this article