Andy Jordan
← Field Notes

Field note

Shake

By Andy Jordan · October 8, 2026

The paw

Banjo, shaking. Sound on for the full effect.

Banjo shakes hands. Watch it twice. I do.

The paw comes up. It lands in my palm. For one second, a dog and a human have agreed on something.

We taught him that. No manual, no hack. Treats, repetition, and a lot of patience on both ends. He's not a genius. He's a good dog who figured out that this move, offered to this hand, gets good things back.

Everything I want to say about AI is in that clip.

Who's holding the treat

What we did with Banjo has a name in AI: reinforcement learning. Reward the behavior you want. Repeat until it sticks. The frontier labs use a version of it to shape the models you talk to every day. Human raters, reward signals, millions of tiny treats.

The method isn't the problem. The method works. Banjo is proof.

The question is who's holding the treat, and what they're paying out for.

When I reward Banjo, the incentive is clean. I want a good companion. He wants a treat and my approval. Our interests point the same way. Nobody is paying me to make him shake more often so they can sell ads against his paw.

A handshake is a promise

Nobody really knows where the handshake came from. The story we tell is that it started as proof: an open hand, no weapon. True or not, that's what it means now. I'm not going to hurt you.

Banjo offers his paw because he trusts the hand on the other side. That's the whole trick.

Now look at the incentives behind the AI agents we're being asked to shake hands with. Products tuned to keep you engaged. Labs racing each other to ship before they're sure. Business models that pay out for attention, and benchmarks that pay out for looking smart over being right.

Not every lab. Not every model. But enough that people feel it. And people can always tell when the hand might be holding something.

We've done this before

Edison opened the Pearl Street power station in Manhattan in 1882. Thirty-five years later, in 1917, only about a quarter of American homes had electricity. By 1930 it was nearly 70 percent. On farms, it was about one in ten well into the 1930s.

The technology wasn't the bottleneck. It worked. It just set houses on fire, cost a fortune to wire, and a lot of people saw no reason to pay for something they'd lived fine without.

So the industry did something interesting in 1920. Western Electric made a silent film called The Go-Getter. A farm kid comes home from the war, his old-school father won't modernize, and the plot turns on showing the family what electric light and appliances actually do for them. A story, built to earn trust.

The gap closed because we built the trust around the technology, not because the generators got better:

  • Underwriters Laboratories (1894): someone had to certify that the thing in your wall wouldn't burn your house down.
  • Electrical codes and regulated utilities: rules everyone could see, and someone accountable when they broke.
  • The Rural Electrification Administration (1935): wiring the places the market decided weren't worth the trip.

Safety, accountability, and access. That's what a handshake looks like at national scale.

Where AI actually is

Here's the number that doesn't make the keynotes. As of May 2026, about one in five U.S. businesses, 19.8 percent, told the Census Bureau they'd used AI in any business function in the past two weeks. Among the biggest firms, 250-plus employees, it was 37 percent. Among the smallest, under 20 percent, and it barely moved in six months.

That's 1917 electricity. A quarter of the houses lit, the rest watching from the road.

And this is where the jobs story gets honest. Economist Paul David showed that factories took decades to reorganize around the electric motor before the productivity showed up. The new work came after the adoption: electricians, appliance makers, radio, the whole modern economy. Not before it.

The AI jobs everyone keeps promising work the same way. Agent builders, AI operators, the roles we don't have names for yet. They only exist when people actually let AI into the work. And people don't let in what they don't trust.

We keep building better generators. We've barely started on the Underwriters Laboratories.

Worth trusting

Back to Banjo.

He doesn't shake because he's smart. He shakes because, over and over, the hand on the other side has been worth it. Fair rewards. No tricks. Every time.

The AI industry keeps trying to solve trust with capability. Bigger models, better benchmarks, smarter dogs. But a smarter dog doesn't trust you more. It just figures out faster whether it should.

Trust is the last frontier. And you don't get it by making the dog smarter.

You get it by being worth trusting.

Sources