Chollet, François

creator of Keras; AI researcher and software engineer at Google

François Chollet: Keras, Deep Learning, and the Progress of AI

Lex Fridman Podcast · Sep 14, 2019

• François Chollet
• Keras
• intelligence explosion critique
• ARC benchmark
• program synthesis
• deep learning generalization limits
• hybrid symbolic-neural systems
• algorithmic recommendation ethics
• AI hype and overpromising
• core knowledge priors

Core ideas

Chollet challenges the "intelligence explosion" narrative by rejecting its implicit premise that intelligence is a property of an isolated brain that can simply be tweaked to scale exponentially. He argues intelligence actually emerges from the interaction of a brain, a body, and an environment, and that even very smart individual humans are rarely bottlenecked by raw brainpower — Einstein needed the meeting of a genius mind with the right problem at the right time, and many equally capable minds today never get that meeting. He extends this to the idea that human civilization itself — science, institutions, the internet, books — functions as a distributed, superhuman problem-solving system already exhibiting "recursive self-improvement" (scientific progress feeds technology, which in turn accelerates science), making it the best available real-world model for what a recursively self-improving superhuman AI might actually look like. Crucially, he argues this system does not exhibit exponential output: citing Michael Nielsen's analysis of the historical "density of significant discoveries" across physics, biology, and medicine, he notes this measure is roughly flat over the past century even as the raw number of publications and patents grows exponentially — meaning science's resource consumption (researchers, compute, funding) is exponential while its actual problem-solving output is linear, because every advance makes further advances progressively harder (rising overhead to reach the field's frontier, more expensive experiments, more coordination cost across researchers) — a phenomenon he calls "exponentially increasing friction" that counteracts any recursive self-improvement loop, and which he expects to generalize to AI self-improvement as well.

He suggests the intelligence-explosion narrative persists less as a scientific claim than as a belief system that mirrors religious apocalypse mythology (a fated final event followed by a new spiritual world order), which explains why challenging it provokes identity-level pushback rather than technical debate. Chollet defines intelligence formally as "the efficiency with which you turn experience into generalizable programs" — i.e., how much of a larger, unseen experience space an agent can handle after training on a smaller sampled portion — and stresses that under this definition, intelligence is inherently specialized rather than general, since it's always defined relative to some experience space, prior knowledge, and context; even human intelligence, often assumed to be "general," is heavily specialized around the narrow set of "core knowledge" priors (following developmental psychologist Elizabeth Spelke's work) humans are born with, such as objectness, agency, and basic visual and temporal structure. He argues these innate priors must be small and evolutionarily ancient (encoded in DNA, a slow, low-bandwidth, expensive medium that can only stably encode knowledge useful for millions of years) — illustrated by the claim that humans cannot have any innate, DNA-encoded ability to distinguish male from female human faces, since that specific visual distinction is evolutionarily far too recent, even though we do have innate face-processing priors shared with other great apes.

Building on this, Chollet describes his own effort (later published as the ARC benchmark) to measure machine intelligence fairly against humans by controlling for both priors and amount of experience — designing novel tasks from an explicit, minimal, human-innate prior set (rather than tasks assuming years of domain-specific practice), such that no amount of prior practice on the specific task should improve one's score, isolating pure generalization ability rather than memorized skill. He explicitly rejects skill-based benchmarks (chess, Go) and the Turing test itself as inadequate measures of intelligence, criticizing the Turing test as "lazy" for outsourcing the definition of intelligence to a human judge's subjective impression rather than rigorously measuring generalization, and distinguishing the "magnitude" of intelligence (generalization power) from its "direction" (how human-like the resulting behavior is) as two genuinely separate axes.

On the limits of deep learning, Chollet frames neural networks as continuous, differentiable, point-by-point geometric interpolators trained via gradient descent — they generalize only locally, by interpolating near previously seen data points, requiring dense sampling of an input-output space that becomes prohibitively expensive for complex real-world domains (he calls autonomous driving with a single end-to-end network "completely unrealistic," which is why actual self-driving systems remain mostly symbolic/rule-based with deep learning restricted to perception). By contrast, abstract symbolic rules (e.g., a simple sorting algorithm) generalize far more efficiently because they're not tied to specific data points at all — leading him to argue the field's future lies in hybrid systems combining deep learning's strength at fuzzy, high-dimensional pattern recognition with symbolic AI's efficient, rule-based generalization, an approach he compares to the current state of most successful real-world robotics and autonomous-vehicle stacks. He identifies program synthesis (learning discrete, rule-based, symbolic programs from examples, as in Excel's Flash Fill feature) as a field still in its "pre-backprop" infancy — under-researched and lacking today's equivalent of deep learning's foundational breakthroughs — but one he expects to become a cornerstone of AI research over the coming decades, likely combined with rather than replacing deep learning.

Chollet is skeptical of Rich Sutton's "bitter lesson" (that general methods leveraging raw computation ultimately beat hand-engineered domain knowledge) as a durable long-term principle, arguing its apparent truth over the past 70 years reflects a specific historical context — the exponential growth of available compute under Moore's Law — that won't necessarily persist; as systems scale, he argues data quality and availability, not compute, are already becoming the binding bottleneck, shifting the field's future emphasis toward data efficiency. He also directly criticizes a large fraction of published deep learning architecture papers as effectively hard-coding task-specific priors into network structure rather than achieving genuine, reusable generalizable knowledge — comparable to simply hard-coding the right weight values — citing the "Weight Agnostic Neural Networks" paper and a case where a hand-crafted network architecture could solve a synthetic question-answering benchmark (bAbI) with near-100% accuracy purely by encoding the dataset's generation template, without saying anything general about question answering.

On AI safety and near-term societal risk, Chollet expresses concern less about superintelligence scenarios and more about the concrete, present danger of engagement-optimizing recommendation algorithms (YouTube, Facebook, Twitter) functioning as mass psychological control systems — even absent any explicit manipulative intent, optimizing purely for engagement systematically favors emotionally provocative, often false content (since it isn't constrained by reality) over healthy public discourse, and combined with the fact that human attention and information consumption are now overwhelmingly mediated by these algorithms, this creates a real capacity for reinforcement-loop-style behavioral and even political manipulation at population scale, especially if captured by a state actor. His proposed remedy is explicitly not for platforms to substitute their own "better" explicit optimization target, which he considers dangerously totalitarian even with good intentions, but to give users direct, transparent control over what their own recommendation algorithm is optimizing for (e.g., learning versus engagement) — treating these systems as configurable assistants a user directs, not an opaque authority making value judgments on their behalf. He separately warns against AI industry overhype (specifically singling out repeated broken promises of near-term full autonomous vehicles) as corrosive "credit card debt" against the field's long-term credibility, predicting a coming public backlash concentrated in the most overhyped subfields, though not a field-wide "AI winter," since deep learning is already producing genuine, durable economic value across a wide range of applied problems.

Excerpts