Everyone Is Still Building the Same Thing

Go look at twenty AI startups barfed onto Hacker News this morning.

Twenty landing pages.
Twenty demo videos with the same soft piano.
Twenty bots that translate $human_talking to $bot_talking
Twenty clip stringing video editors

Twenty products that are, well under the tape and glue, a text string that calls a/some model(s) and does something with what comes back.

That's not an insult. That's the whole industry right now, and most of the people in it haven't noticed yet.

I've seen this movie.

Everything Is a Primitive Nobody Has Admitted Is a Primitive

In 2008 a virtual machine was a product. You could build a company on "we give you a server, but it's fast and it doesn't suck." People did. Some of them got very rich, even the blatant copies.

Today a VM is a primitive. It's a thing you allocate on the way to doing something else. Most humans are not buying a VM.

They're buying whatever is running on it.

Chat is the VM of right now.

RAG is a primitive.
Function calling is a primitive.

Agents, tool use, memory, evals, context windows allllll primitives, all of them. Every one of these is a malloc that somebody has wrapped in a pricing page and a Series A.

The tell is when the demo is the feature working. types words in box, thinking, done!

Not the outcome, the feature. "Watch it call an API!" Zzzzz

Congratulations, you've demoed a socket!

The people who win the next five years are the ones who work out which of these they're building on top of or under these toys... yea I said it.

Your Primitives Are Still Too Complicated

This is the part that makes me want to throw a laptop across a room.

The model feels like magic. Genuinely. Ten years ago what these things do would have been a research paper and a parlor trick.

Then you sign up, and the magic stops at the door:

API key, go make one, don't lose it
Rate limits, which you will discover in production
Token budget, which you will discover on the invoice
Which model? Which version of which model? Which one next quarter?
Do you want a vector database? Which vector database?
What's your chunking strategy, friend?
Did you set up evals? No? Enjoy.
And a bill nobody can predict, ever

Just because the AI feels like magic doesn't mean the rest of it gets a pass.

Linode sold you a box. One page. Pick a size, pick a region, there's your IP in thirty seconds. That was the entire product and it was enough, because WE ate the complexity ourselves instead of handing it to you with a certification path stapled to it.

If I learned anything from Chris Aker, it was that this alone can impact the quality of everything in said feature or product.

Apple runs the same play on hardware. The complexity is real and it is enormous. It's just on their side of the glass. When you get it, poof magic SIRI speaks.

If your pricing page needs a calculator, that isn't a pricing page. That's a trap and you should feel bad.

"All Purpose" Means No Purpose

Every one of these things is "your everything assistant." It can do anything.

Fine. When do I reach for it?

That's the question a tool has to answer, and "anything" is not an answer. A hammer tells you what it's for by being shaped like a hammer.

A thing that does anything tells you nothing about when to pick it up. So you don't. It goes in the drawer next to the other junk then the purpose is no longer needed or becomes novel.

And here's where it loops back around: "all purpose" is what you call a primitive when you haven't found its purpose yet.

It isn't a product strategy. It's a shrug with a logo and a smiley face on it.

The tools I use every single day are embarrassingly narrow. That is exactly why I use them.

Final Thought

None of this means the technology is fake. It isn't. It's the most interesting thing to happen to this industry in twenty years.

But right now everyone is holding the same brick and calling it a house.

Bricks are awesome, now build the houses.

That's the whole job. It always was. - GOBLIN