An AI Startup Raised $180M on a Claim That May Be Nonsense

A new essay on the AI forum LessWrong takes aim at one of the tech world's favorite claims: that human brains are dramatically more efficient learners than AI models. The author, writing as axel_sdq, points out that people compare a chatbot to a three-year-old as if it were an apples-to-apples contest. It isn't. A toddler has seen, heard, felt and touched the world for roughly 12,000 waking hours. Run that through a rough calculation and you get an estimated 27 terabytes of video alone — before you count sound, balance, or the fact that children learn from consequences, not web pages.

So when someone says a small AI model matches a preschooler's language skills using only 40 gigabytes of text, that's not proof machines are wasteful. It's proof the two things were never doing the same job. GPT-2 learned words. A preschooler learns faces, voices, walking and which behaviors get snacks. The author also notes the awkward fact that for most of human history, people never read anything at all — so 'text tokens per human' isn't really a meaningful measure of anything.

Why does this matter beyond an online argument? Because a startup called Flapping Airplanes raised $180 million in seed funding and reportedly reached a $5 billion valuation on precisely this thesis. Its pitch: humans are 100,000 to 1,000,000 times more sample-efficient than AI, so a smarter approach could build powerful models without swallowing the internet. That's a huge, very expensive bet. The essay isn't saying the startup will fail — it's saying the headline statistic underneath it is shaky.

Even OpenAI's Dan Selsam has echoed the idea that today's algorithms need far too much data. The author's response is blunt: of course a researcher watching text-based gains flatten would say that. The real takeaway for the rest of us is simpler. When you hear a stunning AI comparison, ask what's actually being measured — and who profits from the answer.