DOSSIER · №75 OF 81 · ACADEMIC · ch09
← →
← Back to The 81
75 · A FC ACADEMIC

"The Abstraction Benchmark"

François Chollet

CHAPTER
ch09 · No Way Know-How
TIER
Academic
STATUS
Living · Active
ACTIVE
1987 – present
AFFILIATION
Google DeepMind · ex-Keras
JURISDICTION
AI Architecture / Reasoning Benchmarks / Abstraction Theory
COLD OPEN
"Current AI systems acquire skills through exposure. That is not intelligence — it is sophisticated retrieval. Intelligence is the ability to efficiently acquire new skills from minimal data. The ARC benchmark tests that. Almost every modern model fails it."
OVERVIEW

Created Keras, making deep learning accessible to millions, then turned around and built the test that proves deep learning cannot reason. His ARC benchmark — the Abstraction and Reasoning Corpus — presents simple visual puzzles that any five-year-old can solve but that state-of-the-art AI consistently fails, because they require genuine abstraction, not pattern matching.

WHY THIS VOICE MATTERS TO KNOWWARE
Chollet proves empirically that current AI is a two-body system: training data maps to trained pattern. The third body — the ability to abstract across contexts and reason from minimal examples — is missing. His benchmark is the clearest existing test of whether coordination intelligence has been achieved, and so far the answer is no.
OUTSTANDING NOTES
  • ◦Google Distinguished Engineer
  • ◦Keras Lifetime Achievement
CLASSIFICATION
FC
ACADEMIC · TIER A
ch09
METHODS & FRAMEWORKS
5 entries
ARC benchmarkAbstraction and reasoning corpusCritique of deep learning generalizationKeras creatorSkill vs. intelligence distinction
KEY WORKS
  1. 01 Deep Learning with Python (2017)
  2. 02 On the Measure of Intelligence (2019)
  3. 03 ARC-AGI Benchmark (2019)
  4. 04 Keras framework
← All 81
← 074 076 →