Public transcript

AGI: Francois Chollet + Sam Altman

By: ARC Prize
Length: 29 min
Language: English
Text from: the video's captions
Transcribed: 12 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

This is a fireside conversation between François Chollet and Sam Altman, moderated by a Menlo Ventures partner, covering parenting in an AI world, ARC-AGI 3 and other benchmarks, differing approaches to AGI (symbolic learning vs. scaling deep learning), OpenAI's priorities, and a rapid-fire round. Chollet estimates at least a year until ARC-AGI 3 saturates and rates progress toward AGI at 5 out of 10, while Altman says AGI feels very close, missing mainly continuous learning or long-term memory. Altman also signals OpenAI is concentrating compute on accelerating science and the economy.

Key points

François Chollet worries about human agency and whether people will outsource their cognition to machines, while Sam Altman is optimistic that kids will grow up with more agency in a 'Star Trek world'.
Chollet predicts ARC-AGI 3 will take at least a year to saturate, and longer if frontier labs don't explicitly target it.
Altman says benchmarks that are interesting, real, and not saturated matter to OpenAI, and there are not many left.
Chollet critiques the METER benchmark as noisy, domain-specific, and heavily indexed on software engineering, while Altman argues time-on-task is an imperfect proxy for discovering new science.
Chollet's company is building 'symbolic learning', replacing parametric curves with minimal symbolic programs to approach optimal machine learning.
Altman says OpenAI is concentrating its compute on accelerating science and the economy, comparing a million GPUs for cancer research versus making videos.
Altman says his goals of automated AI research interns by September 2026 and fully automated AI researchers by March 2028 remain on trajectory.
In rapid fire, Chollet rates AGI progress at 5/10 and Altman says it feels very close, missing only continuous learning or long-term memory.

Questions it answers

00:00Parenting in an AI world

How do Chollet and Altman think about raising kids amid rapid AI change?

Chollet worries about human agency and whether people will outsource their cognition to machines, while Altman is optimistic that kids will grow up in a Star Trek world with more agency and astonishing achievements.

Chollet: intelligence is the ability to adapt, and kids are the most adaptable.
Altman: his kids will never be smarter than a computer and will be fine with that.

04:50Recent anecdotes

What have they been doing in the last few weeks?

Chollet has been live-coding a Minecraft clone with Codex based on his son's feature requests, writing no code himself. Altman says the world is finally taking seriously that the social contract must evolve.

Chollet used Codex to build a Minecraft clone without writing code.
Altman sees real cross-party willingness to discuss the evolving social contract.

07:46ARC-AGI 3 predictions

When will ARC-AGI 3 break 85%?

Chollet says it depends on how much frontier labs target it: at least a year in practice, several years if ignored. Altman says real, unsaturated benchmarks matter to OpenAI but notes goalposts move and AGI is hard to define.

ARC-AGI 3 is deliberately out of distribution and much harder than ARC-AGI 2.
Altman expects an ARC-AGI 4, 5, and eventually ASI benchmarks.

10:44The METER benchmark

Is the METER time-on-task benchmark a valid measure of AGI?

Chollet calls it good but noisy signal with small sample sizes and heavy software-engineering focus, while Altman argues time-on-task is an imperfect proxy for what matters: discovering new science and economic value.

Chollet: current models can't solve tasks humans do in 5-10 minutes, so 12-hour autonomy claims are overstated.
Altman: some tasks take geniuses minutes that others can't solve in years.

14:53Symbolic learning vs. scaling

How do their approaches to AGI differ?

Chollet is rebuilding machine learning on a symbolic substrate, seeking minimal programs for optimal learning and generalization. Altman says OpenAI pushes the paradigm it understands, aiming at an aligned AI researcher, and welcomes diverse approaches.

Chollet expects future large-scale symbolic foundation models.
Altman: OpenAI's automated AI researcher goals for 2026 and 2028 are on trajectory.

18:28Underexplored research

What research directions deserve more effort?

Chollet names evolutionary algorithms and program synthesis as compute-starved but promising. Altman highlights using frontier models to accelerate physics, math, and bio, and is glad academics are again trying very different ideas.

Chollet: compute is a great equalizer for untried approaches.
Altman: non-AI research fields feel ready to be burst open.

21:48OpenAI's priorities

Why is OpenAI concentrating its efforts?

Altman says the impact on accelerating science and the economy feels about to be massive, so OpenAI is shutting down other projects to concentrate compute — he'd pick a million GPUs for curing cancer over making videos.

OpenAI is focusing on coding and enterprise work.
Altman compares the moment to the original GPT days of concentrating firepower.

23:24Rapid fire

Where do they stand on AGI, favorites, and influences?

Chollet rates AGI progress 5/10; Altman says it feels very close, missing only continuous learning or long-term memory. They name Codex and magnetic tiles, Asimov books, and Pierre Boudier and Altman's mom as most respected.

Chollet is bullish on hardware, robots, and drones.
Altman muses on national debt in a possible deflationary world.

Notes

Parenting and the Future of Humanity

François Chollet is concerned about human agency: whether people will outsource their minds and cognition to machines, and what it will mean to be human in 50 years.
Sam Altman is optimistic: kids will grow up in a 'Star Trek world', have more agency than us, and achieve things that astonish us.
Chollet defines intelligence as the ability to adapt, and trusts kids to adapt to rapid change.
Chollet shares an anecdote of live-coding a Minecraft clone with Codex at his son's direction, without writing a single line of code himself.
Altman notes the world is finally taking seriously that the social contract will have to evolve, with real willingness across political lines to work on it.

Benchmarks: ARC-AGI 3 and METER

ARC-AGI 3 currently sits at 0.37; Chollet predicts at least a year until saturation, longer if labs don't explicitly target it.
Altman says real, unsaturated benchmarks matter to OpenAI, but notes goalposts move and AGI is hard to define — he expects an ARC-AGI 4, 5, and eventually ASI benchmarks.
On METER (12-hour autonomous tasks by Opus), Chollet calls it good but noisy signal: small sample sizes, weak human baselines, and heavy indexing on software engineering.
Altman argues time-on-task is an imperfect proxy — what matters is discovering new science and new economic value, not duration.

Approaches to AGI

Chollet's company pursues 'symbolic learning': replacing parametric curves with minimal symbolic programs, aiming for optimal learning, better generalization, and compute efficiency.
Altman says it's a great thing to try; OpenAI bets with conviction on its own paradigm and aims to build an aligned AI researcher.
Altman says his goals — automated AI research interns by September 2026 and fully automated AI researchers by March 2028 — are on trajectory.

Research Directions and OpenAI Priorities

Chollet points to evolutionary algorithms and program synthesis as underexplored, compute-limited directions.
Altman recalls a researcher's sadness that academics only wanted to reproduce GPT-4, and highlights using frontier models to accelerate physics, math, and bio research.
Altman says OpenAI is concentrating compute on accelerating science and the economy — he'd pick a million GPUs for curing cancer over making videos.

Rapid Fire

AGI distance: Chollet says 5/10; Altman says it feels very close, missing only continuous learning or long-term memory.
Most impressive product: Chollet names Codex; Altman names magnetic tiles for kids.
Bullish: Chollet on hardware, robots, and drones; Altman muses on debt in a deflationary world.
Influential books: Asimov's robot and Foundation stories for both.
Most respected person: Chollet names French AI curiosity researcher Pierre Boudier; Altman says his mom.

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.