FRENES

Deep dive

AGI: understanding artificial general intelligence

AGI is an AI able to learn efficiently to solve problems it has never seen, in any domain. It does not exist yet: the best model solves 95% of ARC-AGI-2, but at a cost and with a learning efficiency far from human. This 12-page deep dive explains how to measure it and what is missing.

Colourful grid puzzles printed on a table, next to a pencil and a laptop
12pages in the deep dive
95 %ARC-AGI-2 record
73models scored on ARC-AGI-2

AGI: what are we talking about?

Artificial general intelligence (AGI) refers to a system able to learn and efficiently solve new problems across varied domains, without being prepared for each one. The definition is debated: some measure it by the number of human tasks automated, others by the ability to acquire a new skill from very few examples. This deep dive follows the second approach, centred on generalisation, in detail.

The pages below explain the key notions (generalisation, test-time adaptation, program synthesis), the benchmarks that try to measure progress, the history of the concept, forecasts and risks. They are published under the responsibility of François Chollet, creator of the ARC-AGI benchmark.

The “How We Get To AGI” talk

François Chollet : How We Get To AGI (AI Startup School, Y Combinator, 2025). The video only loads from YouTube when you click.

The structured, annotated summary of the talk is in How We Get To AGI.

The deep dive, in reading order

  • ARC-AGI Beginner
    ARC-AGI tests adaptation to unfamiliar problems. Learn how its three versions work, what published scores show, and where comparisons need greater care.
  • ARC Prize Beginner
    ARC Prize rewards open methods for solving ARC-AGI tasks. Learn how its rules work, what past editions achieved, and what those scores can actually tell us.
  • How We Get To AGI Intermediate
    A detailed guide to François Chollet’s AI Startup School 2025 talk: key ideas, cited figures, what has held up since, and the questions still open to debate.
  • Generalization Intermediate
    Generalization means solving new problems with what you have learned. Explore definitions, examples, and ways to measure intelligence in Chollet's framework.
  • Program synthesis Advanced
    Program synthesis builds code from constraints or examples. Explore its methods, history, practical limits, and role in research on artificial general intelligence.
  • Reasoning models Intermediate
    Reasoning models improve some tasks by using more computation. Learn how they are trained, what published results show, and where their limitations emerge.
  • Test-time adaptation Advanced
    Test-time adaptation adjusts a model to a new task. Learn how it works, what published results show, and how to assess the full cost of inference-time compute.
  • Measuring progress Intermediate
    Benchmarks measure specific capabilities, not the distance to AGI. Understand their limits, saturation, and contamination to interpret scores more clearly.
  • History of AGI Beginner
    From Turing to today's labs, the history of AGI explains its competing definitions, major research approaches, and disagreements over how to measure intelligence.
  • AGI risks Intermediate
    AGI risks span documented harms and possible future scenarios. Understand scientific disagreements, the available evidence, and practical safeguards for AI.
  • AGI forecasts Intermediate
    No AGI timeline has won broad agreement. Compare published forecasts, their definitions, methods, and the uncertainties behind their widely differing estimates.
  • Limits of deep learning Advanced
    Deep learning struggles with novelty, causality, and long tasks. Explore researchers' findings and learn how to assess these limits in practical applications.

Where do models stand on ARC-AGI?

ARC-AGI measures the ability to solve never-seen visual puzzles that are easy for humans. As of October 25, 2026, 73 models have a published ARC-AGI-2 score; the best is GPT-6 Astra (95%). On the first version, ARC-AGI-1, the record is 98.5% (Claude Opus 5.5). A high score is not enough to conclude AGI: cost per task and learning efficiency matter too.

GPT-6 AstraGPT-6 Astra : 95 %95 %GPT-6.1 SolGPT-6.1 Sol : 94.2 %94.2 %Claude Opus 5.5Claude Opus 5.5 : 93.3 %93.3 %GPT-5.6 SolGPT-5.6 Sol : 92.5 %92.5 %Gemini 4 ArgonGemini 4 Argon : 91.3 %91.3 %Claude Opus 5Claude Opus 5 : 90.4 %90.4 %Claude Fable 5.1Claude Fable 5.1 : 90 %90 %GPT-6 SolGPT-6 Sol : 89.6 %89.6 %Claude Fable 5Claude Fable 5 : 89.2 %89.2 %Gemini 3.8 FlashGemini 3.8 Flash : 89.2 %89.2 %
ARC-AGI-2, share of tasks solved. Source: Epoch AI (CC BY 4.0), checked October 25, 2026.

Recent articles on AGI and research

All articles in this section →

Learn more

Frequently asked questions

Does AGI already exist?

No, by the generalisation-based definition: the best systems remain far from human learning efficiency on truly new problems, even when their scores rise quickly. Other, more economic definitions give different answers; the deep dive compares them.

What is the ARC-AGI benchmark?

A set of grid-based visual puzzles, created in 2019, where you must infer a rule from two or three examples. Each task is new, which prevents success through mere memorisation.

When will AGI arrive?

Nobody knows. Public forecasts range from a few years to several decades and depend mostly on the definition used. The forecasts page summarises the main estimates and their assumptions.

Related articles

Neutrality. No affiliate links. No vendor pays to be reviewed, rated or ranked. Reviews are directed by François Chollet, neutral and factual, and every data point is dated and sourced. How we score