Dwarkesh PodcastFrancois Chollet — Why the biggest AI models can't solve simple puzzles
CHAPTERS
- 0:00 – 2:25
ARC as an anti-memorization IQ test for AI systems
Chollet introduces the Abstraction and Reasoning Corpus (ARC) as an “IQ test” aimed at measuring generalization to novel problems rather than recall. He explains how ARC tries to minimize reliance on world knowledge and instead target core cognitive primitives that even young children possess.
- •ARC is designed to resist memorization and dataset overlap
- •LLMs scale by cramming patterns/knowledge; ARC minimizes need for that
- •ARC focuses on 'core knowledge' (objects, counting, simple physics, symmetry)
- •Novelty per puzzle is the main difficulty: you must infer a new rule each time
- 2:25 – 6:54
What would change Chollet’s mind about LLMs and AGI?
Dwarkesh presses on what evidence would convince Chollet that LLM-style scaling is on track to AGI. Chollet’s bar is repeated demonstration of rapid adaptation to genuinely novel tasks, not just higher benchmark scores achieved via training on similar data.
- •A high ARC score is only meaningful if achieved without training on ARC-like task distributions
- •Key criterion: on-the-fly adaptation to novelty (sample-efficient skill acquisition)
- •ARC may have flaws and could be 'hacked' by brute-force task generation
- •The method of achieving performance matters as much as the score
- 6:54 – 8:43
What ARC puzzles look like (and why they’re easy for humans)
Chollet describes the ARC task format: a few input-output grid examples, then a new input requiring the inferred transformation. He emphasizes that the puzzles are simple and obvious to humans but demand reasoning from sparse demonstrations rather than retrieval of known solutions.
- •Each task: small 2D symbol grids with a few demonstrations plus a test input
- •Goal: infer a transformation rule from examples and apply it once
- •Requires only core knowledge, not specialized domain knowledge
- •Humans can often solve quickly; memorization alone shouldn’t help
- 8:43 – 16:57
Why LLMs struggle: not parsing—program synthesis under novelty
Dwarkesh suggests multimodal models may fix ARC by improving spatial reasoning; Chollet argues the representation isn’t the bottleneck. The core issue is that each task requires synthesizing a new solution program, whereas LLMs tend to reuse and retrieve familiar templates.
- •ARC grids are small; flattening to sequences is not the main challenge
- •LLMs can learn ARC-like tasks when trained on variants—evidence they can encode solutions
- •Failure mode is unfamiliarity: each ARC task is structurally novel
- •ARC demands on-the-fly program synthesis, not program fetching
- 16:57 – 22:01
Skill vs intelligence: scaling benchmarks vs real generalization
Chollet draws a sharp distinction between accumulating skills through training data and demonstrating general intelligence. He argues many popular benchmarks are solvable by memorizing a finite set of reasoning templates, so scaling improves benchmark performance without increasing true adaptability.
- •Scaling laws track benchmark performance—but benchmarks often reward memorization
- •Many 'reasoning' tests reduce to recalling program templates and plugging in values
- •LLMs as “interpolative memories”: bigger models = larger pattern/program banks
- •Skill (breadth of templates) is not intelligence (fast adaptation to new tasks)
- 22:01 – 44:03
Debate over ‘memorization’: humans, school learning, and novelty in life
Dwarkesh challenges whether calling LLM success “memorization” is fair, since humans also learn by drilling and internalizing procedures. Chollet agrees humans use memorization heavily, but insists humans are not limited to it—everyday life requires handling genuine novelty that static retrieval can’t cover.
- •Humans mix pattern-matching and reasoning; education is largely memorization of procedures
- •Critical difference: humans can still adapt when templates don’t fit
- •Examples: Caesar cipher works for common shifts but fails for uncommon ones—suggesting template recall
- •Real environments change; you can’t pretrain on ‘everything you’ll see’
- 44:03 – 46:27
Can automation happen without AGI? Static distributions and job impact
The conversation turns to whether large-scale automation requires true intelligence. Chollet argues that many tasks can be automated through memorization-like systems if the environment is stable enough, but intelligence becomes necessary when dealing with change, uncertainty, and open-ended novelty.
- •Automation ≠ intelligence; value can be generated without AGI
- •If tasks are stable and specifiable, memorization-based systems can perform extremely well
- •Intelligence is needed when distributions shift and novelty is unavoidable
- •Chollet predicts programming will still involve continual novelty and remain hard to fully automate
- 46:27 – 49:11
Grokking, compression, and why LLM generalization is still ‘local’
Dwarkesh cites grokking and meta-learning as evidence neural nets can move from memorization to genuine generalization. Chollet reframes grokking via minimum description length and argues that, while LLMs do generalize through compression, the form of generalization remains limited by the parametric-curve substrate.
- •Grokking explained as compression/regularization yielding shorter programs that generalize
- •LLMs must compress, so they learn reusable program fragments and some generalization
- •Chollet’s claim: deep nets primarily support local generalization within distribution
- •To get extreme generalization, he expects a different modeling paradigm is required
- 49:11 – 1:01:04
Future path: hybridizing deep learning with discrete program search
Chollet lays out his preferred direction: combining deep learning’s pattern-based intuition with discrete program synthesis and search. He argues program search is data-efficient but compute-expensive; deep learning can guide search to reduce combinatorial explosion, approximating a System 1 + System 2 architecture.
- •Deep learning: compute-efficient training but data-hungry and distribution-bound
- •Program synthesis/search: data-efficient (few-shot) but compute-inefficient (combinatorial)
- •Humans blend intuition (System 1) with explicit search/planning (System 2)
- •Hybrid approach: discrete search outer loop guided by learned models for next-step suggestions and evaluation
- 1:01:04 – 1:06:10
Mike Knoop’s ARC ‘nerd-snipe’ story and why ARC lacked traction
Mike Knoop explains how ARC re-captured his attention after chain-of-thought results and why slow progress on ARC felt important. Chollet suggests ARC didn’t gain the usual benchmark momentum because it’s not easily tractable with existing methods—success likely requires genuinely new ideas.
- •Knoop expected ARC to saturate like MMLU/GSM8K but saw minimal progress
- •Few ML researchers knew about ARC despite its conceptual importance
- •Benchmarks spread when early wins make them tractable; ARC resists that dynamic
- •ARC is intended as inspiration to confront limitations of current techniques
- 1:06:10 – 1:09:30
Open vs closed frontier research and ‘oxygen’ being sucked into LLMs
Knoop and Chollet lament reduced openness in frontier research publishing and argue it slows ecosystem-wide progress. Chollet claims the LLM boom concentrated attention and resources on scaling a narrow paradigm, potentially delaying breakthroughs needed for more general intelligence.
- •Frontier labs publish fewer technical details, reducing collective iteration
- •Chollet argues OpenAI catalyzed both secrecy and LLM-centric hype
- •Research diversity felt higher circa 2015–2016 despite fewer practitioners
- •Negative results (e.g., big labs trying ARC) often go unpublished
- 1:09:30 – 1:12:30
ARC Prize details: scoring targets, timelines, and open-source requirements
They announce the ARC Prize structure: a headline reward for reaching human-level performance and smaller progress/paper awards to build community momentum. A key requirement is that prize-winning solutions be released publicly to ensure the benchmark accelerates shared progress.
- •~$1M prize pool; $500k for first team to reach 85% (human average lower bound)
- •Annual contest cycle with progress prizes ($50k top score; $50k best paper)
- •Submissions must be made public domain to claim prize money
- •Contest designed to re-baseline community knowledge each year via sharing
- 1:12:30 – 1:27:00
Resisting benchmark saturation: compute limits, private test sets, and ARC 2.0
Dwarkesh probes whether ARC might follow the typical pattern of sudden saturation once key tricks are found. Chollet explains why Jack Cole’s method is different (test-time adaptation) and discusses evaluation design choices: constrained compute, private test sets, and plans to evolve ARC to reduce leakage and redundancy.
- •Jack Cole approach: test-time fine-tuning as a form of active inference/program synthesis
- •Two paradigms: huge learned DSL + shallow search (LLM) vs small DSL + deep search (classic ARC solvers)
- •Kaggle constraints (no internet/API; limited GPU/runtime) force efficiency but allow per-task compute
- •Plans: private H100 ‘VM track’, gated evaluation API, and ARC 2.0 to reduce redundancy/leakage
- 1:27:00 – 1:34:39
What counts as ‘cheating’ vs legitimate progress—and why solving ARC matters
They discuss what solutions would satisfy ARC’s intent: systems that synthesize generalizable programs from a handful of examples without brute-forcing overlap with the hidden test distribution. Chollet argues a real ARC breakthrough would represent a new programming paradigm—specifying behavior by examples rather than hand-coding algorithms.
- •Legitimate: combining LLMs with code execution/program synthesis and principled search
- •Cheating: brute-force generating massive ARC-like task corpora to ensure overlap with test set
- •Core knowledge fine-tuning is acceptable; tens of millions of near-ARC tasks is not the goal
- •A strong ARC solver implies a powerful new method for programming-by-demonstration
