What Is Left

The models are basically done. They're not perfect, but for most work the bottleneck isn't how smart the model is anymore. They can write the code, read the contract and plan the project well enough that the hard part has moved somewhere else. And everyone has access to them. Every startup is running the same models in the same kind of harness, so everyone ends up building technically pretty similar things.

Recently I talked to a founder who clearly had a gold vein. The startup was what I call a Palantir lite: put an FDE inside a business, map how the business works, slap AI on top. There are a ton of these now and a lot of them are great businesses.

Afterwards I kept thinking: what's stopping me from just taking their startup? What do they actually have? The problem was obvious. The solution was so obvious an AI could figure it out. What they had was the doing. They understood the customer's business, the customer let them in, and they were really good at it.

I see this everywhere now. The tech moat is gone. The idea moat is gone too, because everyone has interesting ideas and everyone has the same tools. Most startups are basically AI henchmen now. We are the henchmen of AI.

Being the henchman isn't trivial though. It sounds simple, and kind of is. But actually doing it is hard. You can have the best idea ever and the most powerful tool in history in your pocket, and most people will still go "nothing ever happens" and ngmi. Rolling up your sleeves and executing is worth a lot right now.

So what is left? If the intelligence is there and everyone has it, what is still between AI and the real world?

I think three things: hardware, ontology and permission.

Everyone is already doing this

Look at what AI startups actually do. Almost nobody is building new intelligence. Everyone takes the same model and tries to get it somewhere it can do real work. Some do it with software. A lot do it with people: engineers who sit with a customer, figure out how the business works, and wire the model in. Forward deployed engineers. Palantir came up with the role, and now OpenAI, Anthropic and Microsoft are all hiring armies of them.

If the hard part was the model, nobody would pay people to sit next to the customer. They pay because the hard part is everything around it.

1. Hardware

The first thing is the physical world. A model can tell you how to check a pipe for cracks. It can't actually check it. No body, no sensors, no hands.

Plugging a device into a model is the easy part. Building the device is not. Hardware is basically the last place where the tech moat still exists. You can build a whole app in a weekend with a good harness. You can't do that with a robot arm, a supply chain or a chip. Building, testing and shipping physical stuff takes months no matter how good your model is.

That's why I think most of the purely technical startups that are left are turning into hardware startups. It's the last place where being better at tech is still hard to copy. Bonus: every sensor you put out there sees stuff no model has ever read about, so hardware also gives the models new data.

2. Ontology

The second thing is understanding a specific place. The model knows a lot in general. It doesn't know how this company works. What does "an order" mean here? Which system is the real one when two disagree? Who is allowed to approve what?

Palantir calls this the ontology: a map of a company's data, the actions you can take on it, and the rules around them. Today humans build that map. FDEs sit with a customer for weeks or months and turn the mess into something a model can work with.

The question is whether AI can build that map by itself. Some companies are trying. One reads a company's code to figure out its business rules, since that's where the rules actually live. Others have AI draft the ontology and humans fix it. Nobody has shown AI doing the whole thing alone yet. And the skeptics have a point: a definition is only worth something if someone owns it.

If AI learns to build the ontology itself, the FDE layer goes away. If it can't, the people building these maps stay valuable for a long time.

3. Permission

The third thing is the one I understand the least.

Say the model is smart enough and knows the company perfectly. It still has to be let in. Someone has to give it the data. Someone has to trust it to change a system. And when it screws up, someone has to answer for it.

Today that someone is always a human, or a company full of humans. That's a big part of what you pay for when you buy from a startup instead of just building it yourself with AI. You can rebuild most products in a few weeks now. You can't rebuild the customer trusting you, or there being someone to call when it breaks.

So what would it take for AI to make an important decision with nobody behind it? Is it a tech problem, a legal one, or just human nature? I don't know. AI already makes small calls alone, like blocking a card payment or filtering spam. It just hasn't moved to anything that really matters yet.

Where this leaves us

I don't think the next big companies will be the ones with the best model. Everyone has the best model. It'll be the ones that close these three gaps: getting AI into the physical world, getting it to understand specific places, and getting it trusted enough to act.

Right now people close all three. That's why so many AI startups look more like consulting companies than software companies. The value is people going in and doing the work, with AI behind them. Henchmen.

My guess is hardware gets plugged in first, then ontology, then permission. But I'm writing this partly to find out where I'm wrong.

And I'm not sure how long henchmanmaxxing lasts. If AI learns to build the ontology itself, a big part of the henchman work is gone. If it gets trusted to act on its own, the rest goes too. Being a good henchman is worth a lot today. How long AI will still need one, I genuinely don't know.