How We Get To AGI: key ideas from François Chollet’s talk
In How We Get To AGI, François Chollet argues that general intelligence will require more than larger models: he defines intelligence as efficiency in dealing with unfamiliar situations, shows how ARC-AGI resisted scaling, and outlines an approach that combines learned intuition with discrete search to write a program for each new task. He delivered the talk on June 16, 2025, at Y Combinator’s AI Startup School in San Francisco; Y Combinator publishes the video.
This summary follows the order of the talk and draws on the video and the transcript published by The Singju Post. The transcript is automated and has not been reviewed by the speaker; the explanations below paraphrase his arguments rather than quote him. For a full section-by-section account, including subsequent developments, see our reference page on the talk.
The starting point: computing gets cheaper
The talk begins with a material trend. According to François Chollet, the cost of computing has fallen by two orders of magnitude per decade since 1940, with no sign of slowing. For a long time, compute and data were the limiting factors in AI.
Around 2010, affordable graphics processors and large datasets allowed deep learning to start delivering results in both vision and language. Self-supervised training on text then became the dominant approach: researchers increased model size and training data while keeping the same architecture and method, and benchmark scores improved predictably. These relationships are known as scaling laws.
This led to a widely held belief around 2023: keep scaling, and general intelligence would emerge. Chollet recalls that this was the dominant view two years earlier, but says that by 2025 almost no one still defended it.
Why scaling alone fell short
In his account, the problem lay in what benchmarks measured. He distinguishes two things that are often treated as interchangeable:
- Memorized skills, which are static and specific to a task.
- Fluid intelligence, the ability to make sense of something unfamiliar as you encounter it.
To make that distinction measurable, he released ARC, the Abstraction and Reasoning Corpus, in 2019. Between 2019 and the time of GPT-4.5, he estimates that base models were scaled up by a factor of roughly 50,000, while their ARC-AGI-1 scores rose from 0% to around 10%. A person in the audience, he says, would score well above 95%.
His conclusion is that large-scale pretraining does not cause fluid intelligence to emerge. A system can acquire an extensive collection of skills without becoming correspondingly better at handling situations outside that collection.
The 2024 shift: test-time adaptation
In 2024, research shifted toward another approach: test-time adaptation. Instead of simply retrieving an answer from what it has learned, a model changes its own state during inference to adapt to the task at hand. Chollet includes test-time training, program synthesis, and chain-of-thought synthesis in this category.
ARC results then changed substantially. In December 2024, OpenAI presented a version of o3 specifically fine-tuned on ARC, reaching human-level performance on ARC-AGI-1 for the first time. According to Chollet, every approach performing well on ARC in 2025 used one of these techniques.
That leaves him with three questions: why scaling was insufficient, whether test-time adaptation will be enough, and what comes next.
What intelligence means to François Chollet
To answer those questions, he returns to the definition of intelligence. He contrasts two traditions:
| Minsky-style view | McCarthy-style view | |
|---|---|---|
| Goal of AI | Perform tasks normally done by humans | Handle problems the machine has not been prepared for |
| Current interpretation | A model capable of most economically useful tasks | A system that adapts to unfamiliar situations |
| What it produces | Automation | Invention |
He sides with McCarthy. In his view, intelligence is a process, and skill is its product. He illustrates this with roads: a road network connects predetermined points A and B, while a construction company can build new roads as needs change. Attributing intelligence to a fixed collection of programmed skills confuses the roads with the organization that builds them.
He formalizes this as an efficiency ratio: what a system can do when faced with future, unfamiliar, uncertain situations, relative to the information available to it. That information includes both its experience and the knowledge supplied by its designers. This follows the definition in his 2019 essay, On the Measure of Intelligence, explored in our article on François Chollet’s approach to measuring intelligence.
Three considerations follow when measuring intelligence:
- Static skill versus fluid intelligence: whether a system has ready-made programs or creates new ones as needed, with a continuum between those possibilities.
- Scope of application: whether it can drive only in a designated area or in any city.
- Information efficiency: how much data and practice it needed to acquire the skill.
This explains his criticism of human exams as AI evaluations. Such exams assume that the test-taker has not memorized the questions beforehand, an assumption that does not hold for a model trained on the web.
He also describes the tendency to take shortcuts: optimizing a single measure can hit the target while missing the underlying goal. His examples include the Netflix Prize, whose winning system was never deployed, and chess, where Deep Blue’s victory taught us almost nothing about intelligence.
ARC-AGI: a direction, not a destination
ARC-AGI-1 contains 1,000 different tasks, designed to prevent preparation through memorization. Each task requires only core knowledge that a 4-year-old child possesses: objects, basic physics, geometry, and counting. Chollet stresses that passing ARC does not establish that a system is an AGI. Its purpose is to direct research toward unresolved bottlenecks.
Our ARC-AGI reference page and ARC-AGI explainer cover the versions and scores in more detail.
He then introduces the successive versions:
- ARC-AGI-1 behaves like a binary test: without fluid intelligence, a system scores close to 0%; with a little, its score rises sharply. The benchmark therefore saturates well below the human level of intelligence.
- ARC-AGI-2, released in March 2025, targets compositional generalization. Every task requires deliberate thought but remains solvable. The team tested it in person with around 400 people in San Diego; at least two people solved every task, and a group of 10 people using majority voting would score 100%. Base models score 0%, while reasoning systems using a single chain score 1–2%.
- ARC-AGI-3, still in development at the time, evaluates agency. A system enters a game without knowing its controls, rules, or objective. It is evaluated on efficiency, with an action budget limited to the number of actions observed in human play.
His practical criterion is straightforward: as long as it remains easy to invent tasks that humans find simple but AI cannot solve, regardless of the computing power available, AGI has not been achieved.
What is still missing: abstraction and recombination
The second half of the talk examines the remaining gaps. Its starting point is the kaleidoscope hypothesis. The world seems infinitely varied, yet it can be described using a small set of building blocks of meaning that are continually recombined. One tree resembles another tree, but also a neuron.
In this account, intelligence consists of extracting those building blocks, or abstractions, from experience, then recombining them as needed to understand a new situation.
Chollet identifies two shortcomings in large models:
- No recombination at test time: they learned extensively during training, then remained fixed. Test-time adaptation addresses this limitation, which he considers an important step toward AGI.
- Very low efficiency: gradient descent requires three to four orders of magnitude more data than a human needs to learn a simple abstraction. Even the best techniques in 2025 still required thousands of dollars in compute to reach human-level performance on ARC-AGI-1.
He then distinguishes two types of abstraction:
| Type 1: value-centric | Type 2: program-centric | |
|---|---|---|
| Domain | Continuous | Discrete, involving graphs |
| Comparison method | Distance | Exact structural matching |
| What it enables | Perception, intuition, pattern recognition | Reasoning, planning, code refactoring |
| A strength of Transformers | Yes | No; even sorting a list remains difficult to learn |
For type 2, he argues that discrete program search is necessary and enables invention. He points to design through genetic algorithms in the 1990s, AlphaGo’s move 37, and DeepMind’s AlphaEvolve system.
Program search is highly data-efficient: two or three examples can be enough. Its difficulty is combinatorial explosion, as the number of possible programs becomes too large to explore. Deep learning has the opposite strengths and weaknesses.
The proposed approach: a meta-learner that writes programs
Chollet’s proposed solution combines the two approaches. A chess player considers only a few moves suggested by intuition. Similarly, the system would use quick, approximate judgments learned through deep learning to guide program search and keep combinatorial explosion under control.
In practical terms, he describes a system that writes a program tailored to each new task. That program would combine deep learning modules for perception with algorithmic modules for reasoning.
It would be assembled from a library of reusable abstractions, which grows with every problem solved. Chollet compares this to a developer publishing a useful library that others can reuse: solving one problem creates building blocks for solving later ones.
This is the approach pursued by Ndea, the lab he co-founded in 2025 to accelerate scientific discovery. The first milestone announced in the talk was to solve ARC-AGI with a system that starts with no knowledge of ARC.
What has happened since
The talk’s main announced release took place: ARC-AGI-3 launched in March 2026. According to its technical report, humans could solve all the environments at that point, while leading systems scored below 1%.
By October 2026, verified results published by ARC Prize showed a striking difference depending on the harness—the software framework surrounding the model. Gemini 4 Argon reached 97.5% with the harness supplied by its developer, but 49.8% with the standard ARC Prize harness. Our benchmark comparison tool tracks these figures.
The questions raised in June 2025 therefore remain relevant: how much of a score comes from the model, how much comes from the surrounding tools, and at what cost? These are precisely the questions of efficiency that the talk placed at its center.
Key takeaways
- For François Chollet, intelligence is the efficiency with which a system acquires new skills, rather than the number of skills it already possesses.
- ARC-AGI showed that scaling pretraining did not produce fluid intelligence; test-time adaptation changed the picture in 2024.
- In his view, compositional generalization and efficiency remain unresolved. His proposed approach combines learned intuition with discrete program search.
- For further reading, the AGI and research section brings together reference pages on these concepts.
Frequently asked questions
What is the main argument of How We Get To AGI?
Intelligence is not the sum of a system’s skills, but how efficiently it acquires new skills when facing unfamiliar situations. Making models and their training datasets larger has produced skills without delivering that ability to adapt. François Chollet proposes combining the intuition learned by neural networks with discrete program search.
Where can I watch the talk and read the transcript?
The video is available on Y Combinator’s YouTube channel. The Singju Post provides a full automated transcript that has not been reviewed by the speaker. If the wording is unclear, the video is the authoritative source.
Did the predictions in the talk come true?
ARC-AGI-3 was released in March 2026, in line with the announcement of an early 2026 launch, and leading systems scored below 1% at release. By October 2026, some models exceeded 95% with an agent harness supplied by their developer, but remained around 50% with the standard ARC Prize harness. The question of efficiency remains open.
What is the kaleidoscope hypothesis?
It is the idea that the world appears infinitely varied but can be described through a small set of building blocks of meaning, continually recombined. Intelligence involves extracting those building blocks, or abstractions, from experience and quickly recombining them to understand a new situation.