Many entrepreneurs have lived this moment: a new hire crushes their first few weeks, a vendor’s first shipment comes back flawless, or a piece of software sails through every test you throw at it. You relax. You hand over more responsibility. You stop watching as closely.
Cognitive scientist Gary Marcus, professor emeritus at NYU and an entrepreneur in his own right, has spent decades studying why that instinct gets smart people in trouble. In his conversation at the World Science Festival, which you can watch below, he shares a story that captures the problem.
In the early days of self-driving car testing, Google let its own engineers try the experimental vehicles on one condition: stay alert at all times because the technology was still unreliable. These were sharp, technically sophisticated people who understood exactly how unfinished the system was. A dashboard camera caught what happened anyway. After about twenty minutes of smooth, confident driving, engineer after engineer reached into the back seat for their briefcase and stopped watching the road.
Marcus calls this a small sample error. The engineers had watched the car handle a limited set of conditions well, and they extended that performance to every condition, the same way you’d trust a driver after watching them merge cleanly a few times. The car was matching patterns from its training data, and those patterns broke down exactly where the road got unfamiliar. The same early systems that drove smoothly for twenty minutes also ran into stopped emergency vehicles, mistakes no attentive human driver would make.
This applies well beyond cars. Evaluating a new hire, a manufacturing partner, a marketing channel, or a piece of software all involve the same trap. A handful of strong weeks, one excellent shipment, or a clean demo shows you how something performs under the conditions you happened to test. It says little about what happens when the order triples, the client asks for something unusual, or the market shifts underneath you.
The useful habit is noticing the exact moment you’re tempted to reach for the briefcase, and using that moment to ask what you haven’t tested yet. Push toward the situation that looks least like what you’ve already seen: the unusually large order, the request slightly outside scope, the month conditions change. Real confidence comes from how something performs at those edges, more than from how smoothly the first few weeks went.
If you’re an entrepreneur making fast calls in a shifting economy, this is worth checking before you scale anything, whether that’s a hire, a partnership, or a new tool. The early version of the story is real. It’s just rarely the complete one.
Watch Marcus’s full conversation below to hear more about why AI systems hallucinate, what’s actually driving recent progress in the field, and where he thinks artificial intelligence goes from here.
