François Chollet created Keras, introduced the ARC-AGI challenge and is a co-founder of the Ndea research lab. His writing and interviews distinguish performance on familiar tasks from the ability to acquire skills for genuinely new ones. — Deep Learning with Python — About the Book.

Infographic for "Lessons from Francois Chollet".

On the Nature of Intelligence

  1. Skill Acquisition, Not Stored Skill: In his 2019 framework, intelligence concerns how efficiently a system acquires new skills across a stated range of tasks, accounting for its priors, experience and the difficulty of generalization. — On the Measure of Intelligence.
  2. Human Generalization: In a 2017 essay, Chollet contrasted humans’ ability to adapt to radically unfamiliar situations with the more local generalization of the deep-learning systems he described at the time. — The Limitations of Deep Learning.
  3. Handle Novel Situations: Chollet treats adaptation to genuinely new situations as a central test of intelligence: the learner must form a useful model for the situation rather than retrieve a rehearsed answer. — Machine Learning Street Talk Interview.
  4. Separate Intelligence From Goals: Chollet distinguishes intelligence—the capacity to acquire useful skills—from an agent’s goals or autonomy. Giving a system a goal does not by itself establish that it can generalize. — Mindscape #280 Transcript.
  5. Generate New Solutions: His account separates an intelligent process that can develop a new task-solving program from the finished program that performs a learned task; high task skill alone does not show broad generalization. — On the Measure of Intelligence.
  6. Consciousness as a Gradual Process: In a speculative discussion, Chollet proposed that consciousness develops gradually in early childhood as a child builds models of self and world. He called this his theory, not an established developmental result. — Machine Learning Street Talk Interview.
  7. Degrees of Consciousness: Chollet said he believes consciousness may exist in degrees and that at least some animals may experience it. This is his tentative philosophical view, not a measured scale. — Machine Learning Street Talk Interview.

On Large Language Models (LLMs) and AGI

  1. Useful on Familiar Tasks: In 2024 Chollet argued that large language models can be extremely useful on tasks resembling their training experience, even while that usefulness does not settle how they handle unfamiliar problems. — Mindscape #280 Transcript.
  2. Do Not Equate LLMs With AGI: In a 2024 interview, Chollet rejected claims that then-current language models already constituted AGI or had human-like general intelligence. — Mindscape #280 Transcript.
  3. Scaling Is Not the Whole Path: Chollet argued in 2024 that merely scaling the available approaches would not, in his view, yield AGI; achieving broader intelligence would require a new idea. — Mindscape #280 Transcript.
  4. Test Outside Familiar Distributions: Chollet’s criticism of language models focuses on tasks that differ materially from what a system has encountered: strong familiar-task performance need not imply comparable performance on novel tasks. — Mindscape #280 Transcript.
  5. A Database Is Not a Program Generator: In the 2024 Mindscape interview, Chollet distinguished models that can retrieve or combine learned patterns from a system that can synthesize a new solution program for a novel task. — Mindscape #280 Transcript.
  6. Stored Programs Do Not Suffice: Chollet argued that adding more memorized task programs is different from giving a system the ability to construct a new program when it meets an unfamiliar problem. — Mindscape #280 Transcript.
  7. A Scientist-Like AGI: In a 2024 interview, Chollet described a possible AGI as a highly capable scientist—a vision of what general intelligence might do, not a capability he claimed existing systems had achieved. — TIME100 AI — François Chollet.
  8. Define What AI Measures: Chollet argued that AI research needs an actionable definition and measure of intelligence; task-specific scores alone can mask the effects of prior knowledge and training experience. — On the Measure of Intelligence.
  9. Question the Intelligence-Explosion Story: In a signed essay, Chollet challenged simple intelligence-explosion narratives, arguing that recursive improvement is constrained by the nature of intelligence and problem solving. — The Impossibility of Intelligence Explosion.

On Machine Learning and Deep Learning

  1. Learn Useful Representations: The authors compare deep learning to unfolding a crumpled paper ball: successive transformations can make intertwined data classes easier to separate. — Deep Learning with Python — Chapter 2.
  2. Layered Representation Learning: The authors explain deep learning as learning successive layers of useful representations from data, with each layer transforming the input for a task. — Deep Learning with Python — Chapter 1.
  3. Check Whether Inputs Can Predict Targets: Before fitting a model, ask whether the available inputs contain information about the intended target; examples of X and Y cannot solve a problem if X is not predictive of Y. — Deep Learning with Python — Chapter 6.
  4. Data Work Matters: In his 2017 outlook, Chollet described data preparation and architecture or hyperparameter tuning as substantial parts of deep-learning practice at that time. — The Future of Deep Learning.
  5. Initial Hyperparameters Need Testing: Even experienced practitioners’ first architecture and hyperparameter choices are usually not optimal; empirical validation is needed to improve them. — Deep Learning with Python — Chapter 18.
  6. Automate the Search, Keep Judgment: Systematic hyperparameter search can replace repetitive manual tweaking, but practitioners still need to design a sensible search space and guard against overfitting the validation set. — Deep Learning with Python — Chapter 18.
  7. Combine Formal and Geometric Methods: In a 2017 essay, Chollet proposed a research direction combining formal reasoning and abstraction modules with learned pattern-recognition modules; he presented it as a forecast, not an inevitable result. — The Future of Deep Learning.

On Evaluation and Benchmarking

  1. Use Novel Tasks to Test Intelligence: Chollet argues that an intelligence benchmark should require unfamiliar tasks whose solutions cannot simply have been memorized in advance. — On the Measure of Intelligence.
  2. Design ARC to Resist Memorization: ARC was designed around explicitly stated human-like priors and unfamiliar tasks so success would test adaptation rather than only rehearsed task skill; leakage or training on test items remains a separate concern. — On the Measure of Intelligence.
  3. Test Research Claims Empirically: In a direct interview, Chollet argued that strong ML research needs meaningful baselines, ablation studies, rigorous validation and significance testing rather than inflated results. — PyImageSearch Interview.
  4. Iterate with a reliable evaluation loop: Chollet argues that successful Kaggle teams improve by testing and refining their approaches repeatedly, guided by reliable evaluation. He describes an ideation–experiment–analysis loop; he does not report a measured correlation between experiment counts and leaderboard rank. — François Chollet — The Loop of Progress.
  5. Past Prices May Not Predict Future Returns: The authors caution that a model trained on historical data assumes the future will behave like the past; market-price history is a case where that assumption may fail. — Deep Learning with Python — Chapter 6.

On the Future of AI and its Impact

  1. Research Should Create Knowledge: Chollet criticized PR-driven projects and publication incentives that neglect what can actually be learned; he described creating knowledge as the purpose of research. — PyImageSearch Interview.
  2. Give Users More Agency: In a personal essay, Chollet urged developers to use AI to expand users’ agency rather than manipulate their behavior. He framed this as a concern and design preference, not a demonstrated outcome. — What Worries Me About AI.
  3. Build Ethics Into Design: Chollet urged software makers to consider their values during design, so ethical choices are part of what they build rather than an afterthought. — Notes to Myself on Software Engineering.
  4. Locate the Intelligence in the Builder: For a fixed hand-coded program, Chollet located the adaptive intelligence in the human programmer who designed it, rather than in the program’s already specified task behavior. — Mindscape #280 Transcript.
  5. Hybrid AI Was a Research Forecast: Chollet’s 2017 outlook envisioned learned systems that combine formal algorithmic operations with geometric pattern learning. This was a research proposal, not a prediction of a guaranteed future architecture. — The Future of Deep Learning.
  6. AGI Timing Is Uncertain: Chollet said in 2024 that if broader intelligence requires a new invention, its arrival cannot be reliably forecast from scaling curves alone. — Mindscape #280 Transcript.
  7. Care About API Users: In his API design essay, Chollet argued that good developer tools begin with attention to users’ needs; practical design rules follow from that principle. — User Experience Design for APIs.

Learn more:

  1. On the Measure of Intelligence
  2. ARC Prize — What Is ARC-AGI?
  3. Machine Learning Street Talk Interview
  4. François Chollet — Personal Site
  5. The Limitations of Deep Learning
  6. What Worries Me About AI
  7. The Impossibility of Intelligence Explosion
  8. Mindscape #280 Transcript
  9. TIME100 AI — François Chollet
  10. Deep Learning with Python — Chapter 1
  11. Deep Learning with Python — Chapter 2
  12. Deep Learning with Python — Chapter 6
  13. PyImageSearch Interview
  14. Notes to Myself on Software Engineering
  15. The Future of Deep Learning
  16. User Experience Design for APIs
  17. Deep Learning with Python — Chapter 18
  18. Deep Learning with Python — About the Book
  19. Deep Learning with Python — Chapter 5