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 timeline: why Chollet expects “AGI around 2030”
Chollet opens with a concrete forecast: AGI in the early 2030s, roughly when ARC-AGI would be at versions 6–7. He frames the key question as not whether progress can be stopped, but how people and companies can best ride and leverage accelerating AI capability gains.
- •Predicts AGI around 2030/early 2030s
- •AI progress is accelerating and likely unstoppable
- •Shifts focus from fear to leverage/agency: how to use the wave
- •Links AGI timing to ARC benchmark roadmap cadence
- 0:31 – 1:30
Introducing Ndea: rebuilding the ML stack beyond deep learning
Ndea is presented as an AGI research lab pursuing an alternative learning substrate rather than incremental improvements on the current LLM stack. Chollet argues the long-run trajectory of AI should move toward efficiency and ultimately optimality, motivating exploration outside today’s dominant paradigm.
- •Ndea’s mission: a new branch of ML closer to optimal than deep learning
- •Not a “layer on top” of LLMs—aims to rebuild foundations
- •Industry focus on LLMs is rational short-term, risky long-term monoculture
- •Belief that AI stacks will change significantly over decades toward optimality
- 1:30 – 3:04
From neural nets to compact symbolic programs: program synthesis + “symbolic descent”
Chollet explains Ndea’s core technical idea: replace parametric function fitting (gradient descent on neural nets) with learning the smallest symbolic program consistent with data. Because gradients don’t apply in symbolic space, Ndea explores an analogue he calls “symbolic descent,” aiming for concise models that generalize and compose better.
- •Replaces parametric curves with minimal symbolic models (MDL principle)
- •Symbolic models promise greater data efficiency and faster inference
- •Conciseness as a driver of generalization and compositionality
- •Introduces “symbolic descent” as a search/process analogue to gradient descent
- 3:04 – 5:20
Why Ndea isn’t competing with coding agents (it targets the substrate)
He clarifies a common misconception: Ndea isn’t building an alternative code-generation agent. Instead it targets a lower layer of the stack—learning mechanisms themselves—while coding agents are characterized as a high-level application layer built atop today’s LLM substrate.
- •Program synthesis here ≠ codegen product/agent
- •Coding agents are “last-layer” applications; Ndea rebuilds foundations
- •Goal: alternative to deep learning itself, not an agent wrapper
- •Focus on general learning optimality, not tool automation features
- 5:20 – 7:22
Why “scaling LLMs” may mislead: efficiency, optimality, and the gradient descent wall
Chollet argues scaling is powerful but may be an inefficient route to human-like learning efficiency. He traces his shift from early deep learning optimism to the view that gradient descent often finds overfit pattern-matching rather than generalizable reasoning programs, suggesting deep learning may hit a wall on certain forms of abstraction.
- •Scaling compute can approximate many things, but may remain inefficient
- •Deep learning can represent algorithms; training often fails to find them
- •Gradient descent tends toward pattern matching vs discovering programs
- •Long-term AI should trend toward efficiency/optimality, not just capability
- 7:22 – 8:50
Why coding agents suddenly work: verifiable rewards + RL post-training loops
The conversation attributes the rapid success of coding agents to formal verification signals (tests, compilers, execution) that enable reliable reinforcement learning. This produces dense synthetic training data, teaches execution-trace modeling, and drives large gains in usefulness without necessarily increasing “fluid intelligence.”
- •Code is tractable because rewards are verifiable (unit tests, compilation)
- •RL-style loops generate massive training data via trial/verify/fine-tune
- •Models learn execution-trace reasoning (variable tracking, mental execution)
- •Capability gains can come from better training/harnesses, not higher IQ
- 8:50 – 14:00
The hard boundary: non-verifiable domains and slow progress (essays, law, etc.)
Chollet contrasts fast automation in verifiable domains with stalled progress in fuzzy domains where reward is subjective. Without trusted verification, systems depend on scarce, costly human annotation and struggle to improve via self-generated data loops.
- •Non-verifiable tasks lack reliable reward signals
- •Human labeling is expensive; progress becomes slower and may stall
- •Verifiable environments enable compounding data generation
- •Highlights why “reasoning” gains don’t transfer uniformly across domains
- 14:00 – 27:03
ARC’s origin story: from Keras and Google Brain to the ‘ImageNet of reasoning’
Chollet recounts how work on reasoning/theorem proving at Google Brain revealed limitations of gradient descent in learning generalizable algorithms. Seeking a benchmark analogous to ImageNet for reasoning, he iterated through ideas and ultimately built ARC tasks by hand, culminating in the 2019 ARC paper and dataset.
- •Keras background and early belief in deep learning’s generality
- •2016–2017: theorem proving experiments expose optimization limits
- •Goal: “ImageNet of reasoning” to measure fluid intelligence
- •2018: built task editor; created ~1,000 tasks; published ARC in 2019
- 27:03 – 31:14
ARC-AGI V1 → V2 → V3: what each version revealed about the field
He explains how ARC versions served as barometers: V1 resisted pretraining scale and only jumped with reasoning models; V2 then saturated once labs applied large-scale targeted RL loops and task harnessing. V3 is designed to be harder to “target” and aims to measure interactive, agentic learning efficiency under novelty.
- •V1: base LLMs stayed low despite massive scaling; reasoning models caused step-change
- •V2: saturation driven by targeted generation/verification/fine-tuning loops
- •Harness engineering boosts performance but indicates lack of AGI (humans in loop)
- •V3: interactive settings and private sets intended to resist overfitting/targeting
- 31:14 – 35:31
Inside ARC-AGI V3: measuring agentic intelligence with novel mini-games
ARC-AGI V3 shifts from passive pattern inference to active exploration: agents must infer controls, goals, and environment dynamics without instructions. Scoring emphasizes action efficiency, rewarding systems that learn and plan like humans within hundreds to thousands of actions rather than brute-forcing state spaces.
- •Agents dropped into unknown games: no instructions, unknown goals/controls
- •Measures exploration efficiency, goal inference/setting, planning and execution
- •Human solvability validated via user testing; humans learn quickly from scratch
- •Efficiency-based scoring prevents brute-force exploration from scoring well
- 35:31 – 46:46
Could AGI be tiny? 10,000 lines of code, knowledge vs. fluid intelligence, and compounding stacks
Chollet separates a small ‘fluid intelligence engine’ from a large knowledge base, predicting the core AGI algorithm may be compact (possibly <10k LOC) and retrospectively obvious. He critiques hand-built knowledge bases like Cyc for lacking learning, and argues scalable systems must remove humans from the improvement loop while enabling compounding capability gains.
- •Distinguishes small reasoning/learning engine from large knowledge substrate
- •Speculates AGI core could be <10k LOC; could have been built decades ago (in principle)
- •Cyc critique: hand-crafted knowledge without learning doesn’t scale
- •Ndea approach: deep-learning-guided program search to break combinatorial search walls
- •Emphasis on compounding research stacks and minimizing human bottlenecks
- 46:46
Future ARC roadmap + advice: evolving benchmarks, alternate paradigms, and building open source
He positions ARC as a moving target aimed at the residual gap between humans and frontier AI, foreshadowing ARC4 (continual/curriculum learning) and ARC5 (invention). He encourages exploring neglected paradigms (e.g., genetic algorithms, alternative architectures, search vs. gradients), offers principles for approaches that scale, and closes with practical open-source lessons from Keras plus career advice: build expertise to leverage AI rather than fear it.
- •ARC will continue evolving: ARC4 (continual/curriculum), ARC5 (invention)
- •Benchmarks should track the residual gap, not declare a single ‘AGI test’
- •Calls for more AI paradigms; suggests revisiting older ideas (70s/80s)
- •Key criterion: approaches must scale without heavy ongoing human engineering
- •Open source lessons: prioritize usability/docs/onboarding; build community; hire power users
- •Personal guidance: AI progress won’t stop—learn domains deeply to leverage it
What AGI means here: human-like skill acquisition efficiency (not task automation)
Chollet rejects popular economic definitions of AGI as “automating valuable work,” arguing that measures automation rather than general intelligence. He defines AGI as the ability to learn new tasks/domains with human-like sample and compute efficiency across the breadth of human-learnable tasks.
- •Critiques “economically valuable tasks” definition as automation-centric
- •Defines AGI as human-level learning efficiency across novel tasks
- •Humans are extremely sample-efficient; this is the benchmark
- •Predicts automation in verifiable domains will precede true AGI
Building a ‘game studio’ for benchmarks: pipeline, engine, and concept design constraints
Chollet details the production approach behind ARC3: a dedicated studio with professional game developers, a custom engine, and iterative human testing to produce hundreds of short games. The games avoid cultural/language symbols and instead rely on core priors (objects, physics, agency) to reduce external knowledge leakage.
- •Dedicated studio + custom engine; 250+ games produced
- •Pipeline: design → implementation → review → human testing → iteration
- •Avoids cultural conventions (arrows, red/green meanings); no language reliance
- •Uses core priors (objects, physics, intentions) to test general learning
