Creator of Keras. Now at Anthropic. Author of "Deep Learning with Python". Most thoughtful voice on AGI benchmarking via the ARC challenge.
François Chollet is the creator of Keras, author of 'Deep Learning with Python,' and a leading critic of LLM scaling hype, now at Anthropic pushing symbolic program synthesis for AGI. He defines intelligence as efficient skill acquisition in novel situations, benchmarked via his ARC challenge where LLMs score near-zero while humans excel. Chollet advocates hybrid AI combining deep learning's perception with discrete reasoning for true generalization, rejecting exponential 'intelligence explosion' narratives.
François Chollet created Keras, a high-level deep learning library [8][28][40], authored Deep Learning with Python with companion notebooks [28], and founded the ARC-AGI benchmark to measure fluid intelligence [1][32][33]. Now at Anthropic, he critiques DL myopia [3] and leads efforts in symbolic learning [22].
Chollet champions Keras with JAX for optimal AI development [8][12][13][14], introducing Keras Kinetic for cloud TPU/GPU execution [13][14]. He provides extensive code resources: transfer learning tutorials [55][56], data augmentation CNNs [57], functional API designs [58], and model repos [40][46]. Early Gists cover RNN optimizations [52], Nelder-Mead [44], and backend integrations [49][50][51][59].
DL excels at interpolative perception but fails discrete reasoning and generalization [1][10][29][30][34][36][37][43][45]. LLMs are 'vector program databases' for memorization, lacking fluid intelligence [5][6][29][36][37][42]; they score <1% on ARC vs. human 97-98% [26][29][33][36]. Curve-fitting limits complex systems [9][10]; symbolic methods outperform [4][11].
AGI is human-level learning efficiency on novel tasks, not benchmark scores [25][47]. ARC-AGI series [1][2][23][24][26][27][32][33][35][38] tests core knowledge priors and program synthesis; ARC-3 adds interactivity [26][27]. Roadmap: annual unsaturated releases [24]; ARC Prize incentivizes progress [33][38].
Chollet pushes 'program synthesis' over scaling [1][4][22][30][34][36][38][45], using DL for intuition + search for reasoning [30][36]. Symbolic compression enables extreme generalization [11]; NDIA lab targets hybrid systems [22][30]. Active inference aids LLMs on ARC [39].
Intelligence is skill acquisition efficiency, not skill [20][25][47]; bounded optimality ratio [20]. Humans near-optimal with tools [19][20]; AGI shifts class divides to cognitive agency [21]. Rejects intelligence explosion [48][60]; progress linear despite resources [60].
AI favors established firms [16][17]; enterprise AI needs security [18]. Critiques web for 'collective stupidity' [61][62]; calls for creativity-focused platforms [61]. Scientific progress linear [60][11].
LLMs memorize patterns but lack reasoning, generalization, and fluid intelligence; ARC exposes this gap.
Core tool for measuring AGI via novel task efficiency; annual evolution to stay unsaturated.
Path to AGI: hybrid DL perception + discrete search for optimal generalization.
True intelligence is rapid skill acquisition in novelty, not memorized performance.
Advocates Keras/JAX/Kinetic for efficient ML; extensive tutorials and tools.
Exponential resources yield linear gains due to rising discovery difficulty.
François Chollet
Sam Altman
François Chollet
Sam Altman
Every entry that fed the multi-agent compile above. Inline citation markers in the wiki text (like [1], [2]) are not yet individually linked to specific sources — this is the full set of sources the compile considered.