Generalist AI Showed a Robot Counting Cash. The Boring Part Is the Hardest.

Generalist AI released Gen-1 on Saturday. The demo video making the rounds shows a robot picking up a stack of bills, counting them into groups of ten, and handing one group to a cashier. There is nothing flashy about it. Counting currency is exactly the kind of small physical task a robot has never been good at.

For the last decade, robot companies have been showing off robots that walk, run, backflip, and occasionally climb stairs. Those demos are impressive and not very useful. The actual economic value of robotics is in whether you can do the fiddly, repetitive tasks that humans do in warehouses, restaurants, hospitals, and small shops. Counting bills is closer to the center of that market than a backflip will ever be.

Gen-1 is described as a physical intelligence model. It is trained to map visual input and proprioceptive feedback to motor commands in a way that generalizes across tasks. The robot does not have to be reprogrammed for every new task. You show it a new task, and it can attempt a version of it based on the patterns it has already learned.

That is the trick nobody has nailed until now. Robot policies have historically been good at one task and bad at everything else. A robot that packs boxes is terrible at folding towels. A robot that folds towels is terrible at packing boxes. Generalist AI is claiming that one model can do both, given a minimum of task-specific examples.

Why aren't we talking about this like a Waymo moment? Because the Waymo story is about revenue, and Gen-1 is still a research demo. The bills get counted correctly. They are counted slowly. The environment is controlled. The robot does not yet handle a surprising case like a bill stuck to another bill. None of that will stop the research from improving, and the pace of improvement in this class of model is fast.

The real question for an operator is which physical workflows in your business are waiting for a capability like this. If you run a retail operation, a restaurant, a clinic, or a small warehouse, you already have a list of tasks that are hard to hire for, hard to train, and hard to keep staffed. Those are the places where a robot that can handle small dexterous tasks would actually move a number on your P&L.

You are not going to install a Gen-1 robot next quarter. That is fine. What you should do now is make a list of those tasks so that when the robots are good enough and cheap enough, you already know where they go first. The companies that deploy physical AI fastest over the next three years will be the ones that knew the day before the robot arrived exactly which counter it was going behind.

There is also a broader signal here about AI's trajectory. The first wave was in the browser. The second wave is moving into the physical world. The second wave is harder in every way. More failure modes. Higher cost per unit. Slower iteration. Bigger safety questions. It is also where a lot of real human labor actually lives, which is what makes it the wave that matters most.

If you lead an operations team, the shift in mental model is this. For the past few years the AI conversation has been about knowledge work. The next few years it will increasingly be about physical work. The tools are different, the constraints are different, and the bottleneck is different. Start paying attention to the companies releasing physical intelligence models now, because they will be the vendors your operations team evaluates in three years.