François Chollet States that the Purpose of the AI Industry Should Be to Create Tools in Human Hands that Enhance Human Prosperity and Well-being
François Chollet stated that the purpose of the AI industry should be to create tools in human hands that enhance human prosperity and well-being, rather than to create a "successor species" to humanity; entertaining this idea effectively becomes an enemy to all current and future humans.
Chollet is the author of the Keras deep learning library, which he released in 2015 before joining Google, where he worked until November 2024 as a Senior Staff Engineer and contributed to TensorFlow. In 2019, he released the ARC-AGI benchmark to measure a system's generalization capabilities on novel reasoning tasks, rather than skills that can be pre-trained.
In 2024, he co-founded the ARC Prize with Zapier co-founder Mike Knoop, with an initial prize of $1 million aimed at achieving 85% accuracy on a private evaluation set; this later expanded into the ARC Prize Foundation, a non-profit organization, with total prizes exceeding $2 million by 2026, and introduced an interactive ARC-AGI-3 environment where human scores are perfect and cutting-edge agents score below 1%.
After leaving Google in November 2024, he co-founded Ndea with Knoop, focusing on general intelligence in program synthesis rather than merely scaling language models. He has long publicly opposed extrapolating current technology as a risk of extinction or autonomous takeover, believing that existing systems lack autonomy and that the strongest coding agents still require human control.
He also emphasized that intelligence is largely externalized in civilization and culture, with individual beings approaching primate levels; he pointed out that if AI causes long-term harm, it is more likely to resemble the tragedy of the commons seen with cars or fast fashion, rather than a supervillain takeover. In recent years, he has acknowledged that synthetic pandemics and other biological risks could be amplified by AI, which are real risks that need to be taken seriously but are not the current priority narrative of extinction.
In market mechanisms, the narratives of superintelligence and "successor species" are used for financing, talent acquisition, and regulatory lobbying, directing capital towards large-scale training clusters and closed laboratories; the tool theory and benchmark competitions redirect resources towards verifiable generalization tests and open evaluations. Beneficiaries are laboratories that can convert AGI narratives into valuations and computing power quotas, while those under pressure are research paths that insist on human tool positioning and struggle to raise equivalent funds along the same narrative line.
Source: Public Information
ABAB AI Insight
Chollet's path over the past decade has been to first create interfaces like Keras that put deep learning in the hands of developers, then embed that interface into TensorFlow within Google, and finally use ARC to separate "brushing skills" from "novel task generalization." Leaving Google to found Ndea represents advancing the same logic from libraries and benchmarks to program synthesis laboratories, rather than joining the then-expanding end-to-end scaling camp.
After 2023, capital and attention have clearly shifted towards closed laboratories that can tell the story of "next-generation species/control superintelligence," concentrating computing power, talent, and policy windows. Chollet's side mobilizes prizes, public-private sets, non-profit foundations, and joint ventures with Knoop, keeping evaluation rights within reproducible grids and interactive environments, rather than handing narrative rights to non-auditable internal demonstrations.
Similar positions have appeared multiple times: early engineering opposition to "strong AI equals extinction" directed resources towards toolchains and benchmarks, while safety narrative advocates directed resources towards governance frameworks and large pre-training. The current phase is one of route contention during an expansion period—one side exchanges scale for narrative, while the other exchanges non-preparable tasks for definition rights.
The structural judgment pertains to the transfer of pricing power: whoever defines "intelligence is close enough to a successor" can set pricing for computing power, regulation, and valuation. The mechanism is that once an evaluation set can be brushed, and a demonstration cannot be reproduced, pricing power shifts from verifiable tool performance to unfalsifiable species narratives.
ABAB News · Cognitive Laws
When narratives can raise funds, evaluations will become political.
When tools are in human hands, species narratives lose pricing power.
The successor story sells fear, while the tool story sells control.