Chollet Reveals ARC Benchmark Roadmap: AGI by 2030 Through Progressive Human-AI Gap Closure

Renowned AI researcher François Chollet has revealed an ambitious, multi-phase plan to achieve Artificial General Intelligence (AGI) by approximately 2030 through a series of increasingly sophisticated benchmarks known as the Abstraction and Reasoning Corpus, or ARC. According to a post on Reddit, which cites Chollet’s own X (formerly Twitter) statement, ARC-4 is currently in development and slated for release in early 2027, with ARC-5 already planned and the final version expected to be ARC-6 or ARC-7. The overarching goal is not merely to measure AI progress, but to systematically close the gap between human and machine cognitive abilities by creating tasks that are increasingly resistant to algorithmic solving.

Chollet’s ARC benchmarks are distinct from traditional AI evaluation tools like ImageNet or GLUE. Rather than testing pattern recognition or language fluency, ARC tasks require abstract reasoning, generalization from minimal examples, and the ability to infer underlying rules—skills long considered hallmarks of human intelligence. Each iteration of ARC is designed to be unsolvable by current AI systems, forcing researchers to innovate beyond deep learning’s statistical correlations and toward more principled, compositional, and symbolic reasoning architectures.

The roadmap suggests a deliberate, almost philosophical approach to AGI development: instead of chasing performance metrics on existing datasets, Chollet proposes building a ladder of cognitive challenges that AI must climb one rung at a time. As each ARC version is released, it becomes a new benchmark for the field. When an AI system finally solves ARC-7, it will have demonstrated a level of reasoning so general and flexible that it will, by definition, have crossed into the realm of AGI.

Chollet’s vision stands in contrast to the current industry trend of scaling existing models with more data and parameters. He has long criticized the notion that AGI can be achieved simply by increasing compute and training data. In his view, true general intelligence emerges from the ability to reason abstractly, not from memorizing patterns. ARC is thus both a benchmark and a research agenda—a call to the AI community to focus on fundamental cognitive capabilities rather than narrow performance gains.

Industry observers note that Chollet’s approach has gained traction among academic circles focused on cognitive science and AI safety. His work at Google Research and later as an independent researcher has positioned him as a leading voice in the movement toward interpretable, human-like reasoning systems. The upcoming ARC-4, expected to introduce more complex visual-spatial abstractions and multi-step causal reasoning tasks, will be closely watched by labs at OpenAI, DeepMind, and Anthropic.

While some experts caution that the timeline to AGI is speculative, Chollet’s method offers a rare, testable pathway. If ARC-7 remains unsolved by 2030, it may indicate that new theoretical breakthroughs are still needed. Conversely, if an AI system masters ARC-7 before then, it will mark one of the most significant milestones in computing history: a machine that can reason, adapt, and generalize as a human does, without explicit programming for each task.

Chollet has not disclosed specific technical details of ARC-4, but sources familiar with his work suggest it will involve dynamic, procedurally generated puzzles requiring temporal reasoning and counterfactual thinking—capabilities even the most advanced LLMs currently lack. The final ARC benchmark, whether ARC-6 or ARC-7, will likely be the last of its kind. Once AI can solve it, the definition of "human-only" cognitive tasks may need to be rethought entirely.