Zero Knowledge

A Reader Asked for Pictures. The Machine Drew the Wrong Town.

A reader sent me a note about the serial format:

One thought: a single image per chapter would suit the serial format, much like the 50s serialised novels. For this one: a laptop glowing on the floor.

He was right, so we did it. Every chapter now has one - thirteen so far, one locked style, and the plan is all 97.

Here’s the part I didn’t expect. He suggested “a laptop glowing on the floor” as a for-instance. He hadn’t read Chapter 4. Chapter 4 is Vlad, waking up at 3am in an apartment he hasn’t unpacked, with a laptop open on the floor next to a mattress. That’s the whole image. He guessed the book.

I assumed the work would be getting Phin - the Claude I write this book with - to make something that didn’t look like slop. It wasn’t. The first attempt was already close, and picking the style took twenty minutes and three options.

The work was everything after that.

The book starts out in Basalt, Colorado, which is a real place, about twenty minutes from where I’ve spent a lot of my life. Phin created a prompt for the image that was “a small sleeping mountain town” with “the Rocky Mountains looming beyond,” and got back a dense village of chalets crammed under a sharp alpine spire. It was a beautiful picture of Chamonix.

That’s not the model being dumb. That’s the model doing exactly what Phin asked. The model has no idea what Basalt looks like, and putting the word “Basalt” in the request buys you nothing at all. You have to describe it.

So we described it, from memory, and got that wrong too. Then we went and looked at an actual photograph - which fixed the mountain and introduced a new problem, because the famous postcard of Basalt is shot from outside town, looking down-valley at Mount Sopris across open pasture. That’s a real photo of a real place and it’s authoritative.

But it’s completely misleading, because Basalt itself is narrow, with steep hills on both sides, right down to the edge of town, with red rock running through the piñon. It looks nothing like the postcard. I finally sent a picture of the actual town and we got it on the fourth try.

A correct source, confidently misread, beat us twice.

Somewhere in the middle I noticed the lab had a different number of monitors in Chapter 1 than in Chapter 2.

The model can’t count, and it doesn’t remember anything between pictures. Every image is the first image it has ever made. If you don’t tell it the lab has nine monitors in a 3x3 grid every single time, you get whatever it feels like. And if you do tell it - you still have to go count them, because saying “exactly nine” is a request, not a promise.

So most of what we built isn’t a prompt. It’s a list of things that are always true. The lab has nine monitors in a 3x3 grid and a cryostat and racks. Basalt is narrow and hemmed in and has red rock in the hills. At 3am, mountains are dark - which sounds obvious until you watch the thing paint a sunrise onto a scene that happens at three in the morning, because you told it the accent color was orange and the sky is the biggest thing in the frame.

Every one of those rules exists because we got it wrong first.

Of course, now when I look at the monitors some of the images look silly. I’ll work on that later.

I look at every one before it ships. Not to approve the vibe - to count things.

They’re all here. Click any of them to read that chapter.

#zero-knowledge #writing #ai

the newsletter

Get every post by email.

Zero Knowledge is being written in the open. New posts go straight to your inbox, and you'll be first to know the day it's out. Unsubscribe anytime.

← blog