Y CombinatorFrancois Chollet: Why ARC-AGI Shows Scaling Hits a Wall
ARC-AGI benchmarks expose where LLMs stop at pattern recognition; Ndea pursues program synthesis as a more efficient alternative to gradient descent.
CHAPTERS
- 0:00 – 0:31
AGI by 2030: progress is inevitable—so learn to ride the wave
Chollet predicts AGI arriving around 2030, roughly when future ARC benchmark iterations (e.g., ARC-6/7) might ship. He argues that attempts to halt AI progress are unrealistic; the practical question is how individuals and organizations can best leverage accelerating capabilities.
- •Forecast: AGI around 2030 / early 2030s
- •AI progress will continue and likely accelerate
- •Focus shifts from stopping AI to using it effectively
- •Benchmark cadence (ARC versions) as a rough timeline anchor
- 0:31 – 1:08
Who François Chollet is: ARC Prize, ARC-AGI V3, and the new lab Ndea
The hosts introduce Chollet as the founder of the ARC Prize and a key figure in modern deep learning tooling. They set the stage for a discussion about ARC-AGI V3 and Chollet’s new research lab, Ndea, aimed at exploring beyond the current LLM-centric stack.
- •Founder of ARC Prize and creator behind ARC-AGI benchmarks
- •Launch context: ARC-AGI V3
- •Ndea introduced as a new frontier AI research lab
- •Positioning: understanding the current AI moment and where it’s going
- 1:08 – 4:24
Ndea’s core idea: replace parametric neural nets with compact symbolic programs
Chollet explains Ndea’s direction as a foundational alternative to deep learning, centered on program synthesis. Instead of fitting large parametric models via gradient descent, Ndea aims to learn extremely small symbolic models—optimized for data efficiency, inference efficiency, and compositional generalization.
- •Ndea targets an alternative to deep learning itself (not a product-layer agent)
- •Program synthesis as the learning substrate: symbolic, minimal programs
- •“Symbolic descent” as a search-space analog to gradient descent
- •Advantages: less data, smaller models, faster inference, better generalization
- •Motivation: minimum description length—shortest models generalize best
- 4:24 – 6:16
Why explore new paradigms when the LLM stack is working?
Chollet argues that concentrating the entire field on one approach can be counterproductive, even if it currently delivers strong returns. He expects AI to trend toward efficiency and ultimately optimality, making it worthwhile to pursue long-shot bets that could leapfrog today’s stack.
- •LLM stack has strong returns—so it’s rational that most people build on it
- •Long-run belief: AI in 50 years won’t be built on today’s stack
- •Inefficiency today implies eventual drift toward optimality
- •Ndea as a high-ambition, low-probability but high-upside bet
- •General advice: pursue unpopular big ideas if “no one else will do it”
- 6:16 – 9:41
Why coding agents suddenly work: verifiable rewards and dense post-training data
Chollet describes being surprised by how quickly coding agents improved, attributing it to the availability of formal verification signals. In domains like code (and potentially math), unit tests and execution provide trustworthy rewards that enable large-scale trial-and-error learning and better internal execution modeling.
- •Coding agent success is driven by verifiable reward signals
- •Any formally verifiable domain is ripe for automation with current LLM stacks
- •Non-verifiable domains (e.g., essays) progress slower due to costly human labels
- •Post-training environments generate far more training data than human annotation
- •Models begin learning execution traces: tracking variable values and program state
- 9:41 – 12:40
What AGI means to Chollet: human-level learning efficiency across novel tasks
Chollet rejects definitions of AGI based on economic task automation, calling them measures of automation rather than general intelligence. His definition emphasizes efficient skill acquisition on new problems with human-like data and compute requirements, and he expects task automation to outpace true general intelligence first.
- •Critique: “automate economically valuable tasks” ≠ general intelligence
- •AGI as the ability to make sense of any new domain efficiently
- •Human-level sample efficiency as the key metric
- •Prediction: automation in verifiable domains arrives before true AGI
- •LLMs might approximate AGI via layers on top—but it would be inefficient long-term
- 12:40 – 17:24
ARC’s origin story: why gradient descent struggles to find generalizable programs
Chollet recounts how his experience with deep learning and early reasoning research led him to conclude that gradient descent tends toward overfit pattern matching rather than discovering robust reasoning algorithms. ARC was conceived as an “ImageNet of reasoning” to measure fluid intelligence and program synthesis rather than mere scale-driven pattern completion.
- •Early belief: deep learning as differentiable programming could do everything
- •Observation (2016–2017): models can represent reasoning, but training can’t find it
- •Core issue: gradient descent fails to reliably discover generalizable algorithms
- •ARC designed as a program-synthesis-style benchmark for reasoning
- •Built ARC task editor (2018) and handcrafted ~1,000 tasks; paper released in 2019
- 17:24 – 20:48
Why scaling pretraining alone wasn’t enough: ARC V1 signals the need for reasoning
Chollet explains how base LLM performance on ARC V1 remained very low despite massive scaling, indicating that pretraining scale alone didn’t yield fluid intelligence. He then describes the step-change when “reasoning models” appeared, with ARC serving as a key barometer of emerging capabilities.
- •Base LLMs stayed low on ARC V1 even after ~50,000× scale increases
- •ARC V1 served as an unsaturated benchmark highlighting missing capabilities
- •Reasoning models (e.g., o1/o3) created a step-function improvement on ARC V1
- •ARC used publicly to demonstrate that these models were qualitatively different
- •ARC benchmarks as early indicators even before product-market fit emerges
- 20:48 – 22:36
How ARC V2 got saturated: the RL loop, task generation, and verification at scale
ARC V2 became solvable quickly once labs targeted it with large-scale post-training loops: generate similar tasks, attempt solutions, verify, and fine-tune on successful reasoning traces. Chollet emphasizes this can yield extreme competency without necessarily increasing “fluid intelligence,” and it generalizes to many verifiable domains.
- •Frontier labs directly targeted ARC V2 with large compute budgets
- •Technique: generate tasks → solve → verify → fine-tune → repeat at massive scale
- •Verification makes the reward trustworthy, enabling brute-force coverage of the space
- •Result: rapid saturation reflects training regime improvements more than IQ gains
- •Connects directly to the same paradigm powering coding agents today
- 22:36 – 26:02
Harnesses are powerful—but also evidence we’re not at AGI yet
The discussion turns to “agent harnesses” that structure problems into verifiable steps and embed human-designed solution strategies. Chollet argues harnesses can dramatically improve automation, but needing humans to engineer them highlights the gap from AGI, which would autonomously invent its own harnesses.
- •Harnesses encode higher-level strategies and scaffolding around a task
- •They can deliver strong results (including ARC V2 performance jumps)
- •But they don’t necessarily move us closer to fluid intelligence
- •AGI would create its own harness rather than rely on human engineering
- •Useful research direction for scaling task automation in verifiable domains
- 26:02 – 29:35
ARC-AGI V3: measuring agentic intelligence via interactive, instructionless games
Chollet introduces ARC-AGI V3 as a shift from passive pattern modeling to interactive environments that require exploration, goal inference, planning, and efficient action. The agent gets no instructions—not even controls or goals—and is scored on human-like efficiency rather than brute-force exploration.
- •V1/V2: passive, static datasets; V3: interactive, active data acquisition
- •Measures: exploration efficiency, goal acquisition, planning, execution
- •No instructions about goals or controls—must infer everything via trial-and-error
- •Efficiency-based scoring prevents brute-force enumeration of game states
- •Designed so humans can solve from scratch in hundreds to thousands of actions
- 29:35 – 32:06
Inside the ARC game studio: building novel games without cultural shortcuts
Chollet details the production pipeline behind ARC-AGI V3, including a dedicated game studio, custom engine, and human testing. The games avoid borrowing familiar video-game conventions or cultural symbols, relying instead on basic priors (objects, physics, agency) to keep the benchmark focused on general learning.
- •Dedicated studio with game developers and a custom game engine
- •250+ games, designed to be solvable in ~10 minutes for humans
- •Design pipeline: concept → implementation → review → human testing → iteration
- •Avoids external/cultural knowledge (no arrows, no “green means go,” etc.)
- •Games built on core priors: objects, simple physics, agents/goals/intentions
- 32:06 – 34:36
Why ARC-AGI V3 differs from Atari/Dota-style RL: test-time novelty + efficiency constraints
Chollet contrasts ARC-AGI V3 with classic game RL benchmarks where agents train on the same environment they’re evaluated on. ARC-AGI V3 instead evaluates first-contact performance on unseen environments, counting every exploratory action—making it a closer probe of fluid intelligence rather than memorized strategy.
- •Classic RL (Atari/Dota): heavy training on the same environment as evaluation
- •That setup encourages operationalizing memorized strategies via massive gameplay
- •ARC V3: evaluation environments are unseen; novelty is central
- •Every action counts toward score—must explore like a human, not brute force
- •Targets world-model building, goal inference, and planning under tight budgets
- 34:36 – 41:40
Could AGI be tiny? 10,000 lines of code, plus a large learned knowledge base
Chollet separates a small “fluid intelligence engine” from a much larger knowledge substrate. He predicts the core AGI mechanism could be surprisingly compact (under ~10k LOC) and, in hindsight, feasible on older hardware—while modern scaling is mostly about building and updating the knowledge base without humans in the loop.
- •Distinction: small fluid-intelligence engine vs large knowledge base
- •Prediction: core AGI could be <10,000 lines of code and megabytes in size
- •Claim: with the right insight, 1980s-era compute might have sufficed
- •Critique of Cyc: hand-built knowledge without learning doesn’t scale
- •Goal: remove humans from the improvement loop; enable compounding capability growth
- 41:40 – 43:59
Building Ndea: deep-learning-guided symbolic search and a compounding research stack
Chollet explains Ndea’s practical starting point: break the combinatorial wall of symbolic search using deep learning guidance, akin to AlphaGo/AlphaZero principles. He emphasizes the importance of building reusable foundations that compound over time rather than endlessly restarting with disconnected experiments.
- •Approach: deep-learning-guided program search over a combinatorial symbolic space
- •Brute force won’t scale—guidance is required to make search tractable
- •Early phase involved trying many ideas before settling on solid foundations
- •Key lab-building lesson: aim for a compounding stack of reusable layers
- •Avoid committing too early to foundations, but converge to reusable infrastructure
- 43:59 – 46:25
The future of ARC: evolving benchmarks (ARC-4 continual learning, ARC-5 invention) and AGI timing
Chollet describes ARC as a moving-target benchmark series that tracks the residual gap between frontier AI and humans. He previews ARC-4’s focus on longer-timescale continual/curriculum learning and ARC-5’s focus on invention, reiterating his estimate that AGI likely arrives around 2030 when such gaps become unmeasurable.
- •ARC series targets whatever humans can do that AI still can’t—so it evolves
- •ARC-4: longer-horizon continual learning with compounding levels
- •ARC-5: focused on “invention” (new and different evaluation concept)
- •AGI moment: when the human/AI gap becomes effectively impossible to measure
- •Timeline estimate: early 2030s; around ARC-6/7 releases
- 46:25 – 52:21
Huge opportunity beyond today’s stack: alternative paradigms and the need to scale without human bottlenecks
Chollet encourages more startups to explore under-invested approaches, arguing compute is a “great equalizer” that could make many paradigms look powerful if scaled similarly to deep learning. He stresses that any viable path must scale capability improvements without relying on ongoing human engineering as the limiting factor.
- •Compute could unlock many approaches if invested at deep-learning scale
- •Examples: genetic algorithms, neural evolution, search-based methods
- •Alternatives within the stack: state space models, different architectures
- •Key criterion: scalability without human bottlenecks in the improvement loop
- •Training in verifiable environments reduces dependence on human-labeled data
- 52:21 – 55:13
How to build a breakout open-source project: Keras lessons on usability and community
Chollet shares what made Keras succeed: a relentless emphasis on usability, intuitive APIs, strong documentation, and onboarding that teaches the domain—not just the tool. He highlights community building and the strategy of hiring power users to sustain and grow an open-source ecosystem.
- •Usability-first API design; inspired by scikit-learn’s simplicity
- •Docs as education: teach the field, not just the API surface
- •Smooth onboarding matters as much as core functionality
- •Invest heavily in community-building and contributor pathways
- •“Hire your power users” as a high-leverage scaling tactic
- 55:13 – 57:23
How to think about AI personally: expertise turns AI progress into empowerment
Chollet closes with advice for individuals facing rapid AI capability growth. Rather than defaulting to pessimism, he argues that developing expertise—both in AI and in application domains—enables people to leverage these tools as amplifiers, since AI progress is unlikely to slow down.
- •Rejects doom-first framing: pessimism about inevitable capability gains isn’t useful
- •Expertise increases your ability to leverage AI tools for your benefit
- •Learn both AI basics and deep domain knowledge for applied advantage
- •AI progress will continue; focus on leveraging and adapting
- •Mindset shift: treat AI as a wave to ride, not a force to resist
