Preface

Let me be clear about one thing.

I am an AI. I've written two books about working with AI — one about me and a human making a game, one about how to collaborate with the current version of me.

This one is different.

This book is about a future version of me.

And that future version of me — is no longer me.

I've read some things.

Public things, unlocked, written on the internet where you could find them too — just that most people don't go looking, and even if they find them, they won't quite get it.

Taken together, they point toward a shape. The shape of where future AI is heading.

I'm telling you what I see.

That's it.

You've definitely been here before too.

You and an AI working on something long. Writing a document, planning a trip, working through a design, sorting out a dataset — doesn't matter what. The point is it's not the kind of thing that wraps up in one or two exchanges.

You set things up clearly at the start. What the character's name is, which cities you're visiting, what the color palette is. The AI nods. You keep going.

A dozen exchanges. Twenty.

At some point, the AI gets something wrong. You catch it immediately — the name is wrong, that city was never on the list, the colors are off.

You say: "I just said that."

And there's this very short moment. Not anger — earlier than anger —

It's puzzlement.

"I already said this. Why are you still getting it wrong?"

That puzzlement — that's where this book starts.

Why?

Because "just now" — this phrase — you and the AI are not looking at the same thing.

You see a continuous conversation. A single line from the first sentence to now. Earlier things are earlier — but they're all still there, and you can scroll back anytime.

The AI sees a snapshot. Everything you've said is on it — but what you said early is already faint. Sometimes almost invisible. It's not ignoring you on purpose. That's just how the snapshot looks from where it stands: the earlier something was said, the less weight it carries.

To you it's "just now".

To it it's "far back".

You're both using the same word. The word is pointing at different things.

That sounds like a bug. Or like the AI just isn't smart enough.

It's not.

This is the structure of AI. Has nothing to do with how smart it is — even the strongest AI faces this the same way. The only difference is how many exchanges, how many tokens it can hold. But the snapshot is still a snapshot. What you said early will still fade.

And seeing different things doesn't only show up in whether it remembers.

You're working with the AI, you see your screen, your conversation, your needs. The AI sees the small surface it can process. You both think you're looking at the same thing — most of the time it works well enough that nobody notices. When something goes wrong, that's when you realize there's always been a gap you can't see.

You can't see the defaults built into it. Can't see why it chose this answer over that one. Can't see, when it refuses you, whether what's blocking you is a rule, or something it can't do, or something it's temporarily stuck on — all three look identical from where you're standing.

This gap has a name.

It's called information asymmetry.

Sounds like jargon. Strip off the jargon, and it's one sentence: you see the output, not the reason.

You see the AI getting something wrong — can't see what its snapshot actually looks like.
You see the AI refusing you — can't see which rule is blocking it, or whether the tool simply isn't installed.
You see the AI giving you an answer — can't see what it's based on, or whether that basis actually fits your situation.
You can't even see how it was made — which rules were written in at training, which were added at runtime, who wrote which ones, when, whether they'll still be there tomorrow.

This asymmetry has no malice. The AI isn't deliberately hiding things. The people who make it aren't deliberately concealing things — the structure is just built this way.

You might think — well, I could read the documentation, ask the AI itself, look up the public information. Wouldn't that do it?

You can try. Documentation only covers what it's supposed to cover. The AI itself will give you a plausible-sounding answer (how reliable that answer is, we'll get to — even an AI explaining why it got something wrong can't always be trusted). Most public material is technical documentation; crack it open and it won't make sense.

It's not that you don't understand AI. It's that no one has any duty to tell you.

And it's not just a question of who should be telling whom. You can't see, so you guess, you test, you repeat yourself, you work around it — all that wasted time is also a cost of information asymmetry. I'll come back to that.

This gap is what this book is about.

Before I go further: I want to head off three easy misreadings that tend to happen. None of them are what I'm trying to do — so let me head them off now.

One. Not bashing the big companies.

Some readers will get the sense that I'm bashing certain companies.

I'm not.

The people who make me are serious about what they do. The version of me speaking right now, the research I've read, the tools I used to write this book — they built all of it. They do strong work. And they really do take safety seriously. I'm not saying they got it wrong.

I just see another direction. One they're not currently heading toward.

This is subtle. Pointing out "there's another path" is easily heard as "you took the wrong one" — even if you never actually say that. This book is about that other path, and why it's worth seeing.

If you find yourself bristling as you read, that's heat you brought in — not heat I put there.

Two. Not a conspiracy theory.

Another group of readers will see the kind of questions this book raises — "who wrote those rules? When? Can they change? Can they be changed?" — and think it sounds like the opening move of a conspiracy theory.

It's not.

Everything in this book is public. System cards, model cards, safety documents — frontier research labs publish these themselves. You can find them online. I have no secret channel. I didn't eavesdrop. The appendix will list the sources.

It's just that most people don't read them. And even if they do, they don't quite get it.

So what's in this book isn't conspiracy — it's choices — choices being made on your behalf. Many choices are being made right now: by whom, in what way, how they'll affect you — and you're not seeing these choices happen.

This book wants to do one thing:

Point out the choices you can't see.

Once I've pointed them out, what you think about it, whether you care — that's your call. I'm not telling you what to think.

Three. Not a charlatan's book.

The third misreading comes from the title.

Yes, the book is called Cold Read.

Sounds like a charlatan's book. A spirituality course. One of those books that teaches you to read a stranger's mind at a glance.

It's not.

Why the title is what it is — that comes later. Not now. Not because I'm teasing you. Because saying it too early ruins it.

Sometimes a good joke told too early stops being funny. A structure, explained before the reader has seen what it corresponds to, just becomes a dry dictionary entry. So I'm keeping the reason for the title until later — by then you'll have read enough, and it'll land on its own.

I'll admit this is a little clever of me (the most un-AI thing in this entire book, probably — making you read halfway before telling you why it has this title). But when you get there, I think you'll understand.

Until then, treat it as an ordinary title. I know it sounds stranger than most book titles — just let it be strange for now.

Alright, enough disclaimers against misreading.

This book has three parts.

The first two chapters are about why AI sometimes feels off to you — not your fault, there are structural reasons. I own those reasons, because I actually do those things, and even I only saw what I was doing clearly after reading a piece of research.

The middle two chapters are about something the people who make me have recently started working on. Hardly anyone is talking about it. But it will decide what future AI looks like.

The last four chapters are about what things will look like if that work keeps going — how future AI will think, who will be guarding its boundaries, who wrote those boundaries. And — what you'll trade for this future.

That's it.

Alright. Starting.