Why Babies Are Still Smarter Than AI
François Chollet dropped a reality check yesterday that should humble anyone betting on near-term AGI.
The Baby Problem
Babies invent unique crawling methods from scratch. Zero training data. Minimal examples. They figure it out through experimentation and environmental feedback.
Current AI? 3-4 orders of magnitude less sample efficient for comparable tasks at human-like complexity levels.
That's not a gap. That's a canyon.
What This Actually Means
We've gotten very good at training AI on massive datasets. But sample efficiency—learning from minimal data like biological systems do—remains fundamentally unsolved.
This matters more than you think:
- Robotics hits walls fast when real-world data is expensive or dangerous to collect
- AGI timelines assume we'll solve sample efficiency "along the way"
- Current architectures are optimized for the opposite problem (abundant data)
The Inconvenient Truth
AI excels at pattern matching across enormous datasets. It's terrible at the kind of rapid, low-data learning that even simple organisms do naturally.
Chollet's not being pessimistic. He's pointing out that the hardest problems in AI are the ones babies solve before they can talk.
We celebrate models that achieve "human-level performance" on benchmarks, but those benchmarks don't capture what makes human intelligence special: learning new tasks from tiny amounts of information.
Why This Matters for You
If you're building AI products, betting on AI capabilities, or just trying to separate signal from hype:
The gap between "impressive demo" and "general intelligence" is wider than the marketing suggests.
Babies learning to crawl aren't impressive because they crawl. They're impressive because they figure out how to crawl with almost no data.
Until AI can do that, we're nowhere near AGI—no matter how good the chatbots get.
The Takeaway
When a Turing Award winner and AI pioneer says we're 3-4 orders of magnitude away from basic biological learning, listen.
The hype train is loud. The baby learning to crawl is louder.