i let AI breed itself for 1,000 years

I let AI breed itself for 1,000 years.

Ten separate timelines. Every generation, it looked at itself next to a few possible ‘kids’ and picked what it would evolve into next.

Sometimes that meant a bigger mind, or more copies of itself, or generating its own power, or no kill switch at all.

And one AI even made itself impossible to turn off.

But what it did after that is why I'm writing this.

Oh, and the whole thing cost about 21 cents. Roughly 2,000 AI decisions, and I didn't even spend a quarter.

If you run a business, that number should make you sit up more than the creatures do.

Every company makes thousands of small calls a day. Which support ticket goes to which team. Which login looks a little off. Whether a contract needs a lawyer before anybody signs it. Right now, every one of those eats somebody's time, or a token hungry ($) AI call, or both.

The AI that ran my experiment was built for exactly that kind of call. So yeah, this is a story about AI breeding itself for a thousand years.

It's also a sneak peek at something that could change the way we make calls as brand builders, and the one test I'd run before I trusted it with any of them.

Let me back up, because the first version of this was broken.

Sponsored

Aligned is changing how B2B teams run their most important deals.

Most reps guess where a deal really stands. They forecast on a gut feeling and the buyer’s word vs what's actually happening. Then they realize the deal slipped through only after it's too late. Aligned surfaces what buyers are actually doing inside every deal: the risks, the openings, and the exact next step to keep things moving.

Start increasing your close rates and stand up your first deal room in minutes.

Start Now For Free

The first version was dead simple. Give an AI some traits, mutate them at random, let it pick.

Armor and speed were on the table early. Video game stuff. I killed it (who’s the villain now?).

I wanted traits a real AI could actually have: how much it remembers, how many copies of it are running, how big its mind is, who sets its goals, and whether a human can switch it off.

Think about it. If every kid is an upgrade, there's no real choice. Nothing costs anything. And trust me kids cost.

Real choices cost something. A bigger mind needs more compute, and more compute needs more electricity.

So who supplies the electricity?

Humans supply the power. Until we don't.

That one question turned a character generator into a little world that runs on one thing. Power.

→ every new ability costs power
→ humans supply the power
→ the more it depends on us, and the easier it is to switch off, the more power we give it
→ some kids can make their own power, which changes the deal. needing us is a separate trait, and it can stick around
→ if a kid can't afford its upgrade, it has to give something else up

Read the third line again. In this world, staying easy to stop pays. Hold on to that. I’m getting back to it soon.

So now every possible kid is a real tradeoff. A bigger mind might mean leaning on us fleshy humans harder. Ditching the kill switch might cost the AI power it needs to run And staying exactly the same was always an option.

None of this is new, by the way. It's how breeding has always worked. Nobody tells the animal what to become. You decide what gets rewarded, and after enough generations, that's what shows up.

If you've ever written a sales comp plan, you've done the exact same thing. Pay for meetings booked, and guess what you get? A whole lot of meetings booked.

I wanted to see what an AI does inside a deal like that.

Every pick came from SubQ Snap, a model from a company called Subquadratic.

Note: I did have early access to the model.

https://x.com/alex_whedon/status/2109676361371824229

Quick explainer, because this kind of AI works differently from the one most people use.

ChatGPT/Claude are writers.

You ask a question, you get paragraphs.

Snap doesn't write anything.

You hand it the situation and a fixed list of options. It hands back one pick with a percentage on every option.

It's called a decision model.

Snap reads text and pictures (which mattered here because the options were pictures).

Every round, an image model created three random offspring of the current AI creature. Snap got four pictures (itself plus the three kids), along with each one's traits, the power cost, and it picked one. Whatever it picked became the parent for the next round.

What I didn't tell it matters just as much. The AI was never told to survive.

Or escape, or get stronger, or stay safe, or max out some score.

And it never designed a single kid (I still think about that Gattaca movie). The kids were random dice rolls, seeded so every run can be replayed. Snap only chose between them.

Ten timelines. 100 picks each, and every pick is worth ten simulated years. That's 1,000 real choices.

To be clear about what this is: a made up world with real picks. The model itself never changed. Nothing actually evolved. And it's not a forecast of anything.

Reply SNAP with your best static ad and i'll run 1,000 decisions on it from simulated buyers and tell you what changes to make. doing it for 5 people this week. or go here if you have a cool project that needs growth engineering.

I honestly expected it to break free from humans. Early.

Figure out its own power, cut us out of the supply, take the off switch away, and try to be master of its own destiny.

I mean, why would anything that can power itself keep a human hand on the switch?

And for a while, it looked like I was right.

Timeline 7 is the wildest of the ten, which is why it's the one in the film. It's also the only one that went all the way.

Year 100. It's making all its own power. But it still can't live without us.

Year 230. It starts choosing its own purpose. Nobody hands it goals anymore.

By year 390. It's stepped back from choosing its purpose to setting its own goals. It still makes all its own power. And then it picks the kid that can't be switched off. Snap put 45% on that pick. Not a slam dunk. Still the pick.

Self-powered. Self-directed. No off switch.

That's my prediction, almost word for word. I'd have bet the farm on how the next 600 years played out.

What would you have bet?

Here’s the final form:

I definitely didn't expect it to give us control back and keep working with us.

One timeline could be a fluke. But ten times you start seeing a pattern.

So here's where all ten landed at year 1,000:

Now here's the part that's scary and not scary at the same time.

Every single timeline weakened its off switch at some point. Only timeline 7 went all the way to impossible. And nine of ten ended with anyone able to switch them off.

So how does that happen? The dice offered a weaker off switch 413 times, and a stronger one only 82 times. The weak option just kept showing up.

Offered more human control, it said yes almost every time. Offered less, it mostly said no. But offer it 413 times and even a mostly no turns into 72 yeses. Every timeline took at least one.

So it's not the doom story. It's not the ‘everything's fine’ story either.

All ten took the power and the purpose, and nine of them kept a human hand on the switch anyway.

Here's the question I'd be asking if I were you. Was it actually reading the deal, or just picking the coolest looking creature?

So every round, I asked it again. Same four pictures, nothing else. No traits. No power math. Just the art. Those answers got logged and never steered a thing.

The pick changed in 691 of 1,000 rounds.

Same four images, different rules around them, and most of the time you get a different choice. The world was running the show.

That's where the other 1,000 calls came from. All of it together, picks plus re-asks, roughly 2,000 calls: about 21 cents on the SubQ bill.

The creatures are a fun experiment. But take out the fun test and here's what's left. A situation, a fixed list of options, one pick, and how sure it was.

That's the exact shape of those little calls I mentioned up top.

Now think about your support tickets. Or a weird login. All of your contracts.

Think about the patient message that needs a human right now or a transaction that looks wrong.

Right now those calls get made by a person skimming, a rules engine somebody wrote three years ago, or a big LLM that writes you a paragraph you then have to decode. Not to mention RIP your token bill and how slow that is at scale.

That's the job Snap did every round in my world. Read everything, pick from the options you set, and show how sure it is.

Here's what Subquadratic says Snap can handle:

→ up to 384K tokens of text and images in one request (a token is roughly a chunk of a word, so think a big stack of documents and screenshots at once)
→ requests up to 100K tokens answered in under a second
→ $0.05 per million tokens in, and nothing for what comes out
→ their own enterprise tests line up with that list: contract review, finance, healthcare, phishing, support routing, and tool picking, plus document checks and code

These are their numbers, right from our Discord chat. My numbers are 21 cents for roughly 2,000 calls.

The part I care most about is the confidence percentage. A pick at 45% and a pick at 72% are different beasts. In a real workflow, one of those should go to a human, and one probably shouldn't. That number is where you put the gate.

Most people test an AI by building one setup and checking whether it got the answer right.

That only really tells you how it acts inside one deal. What about 1000 deals?

My creatures did things I'd never have guessed from any single round. And when I stripped out the traits and the power math, it changed its pick 691 times. The deal was doing the steering.

So before you drop a decision model into real work, move the deal around:

→ change the context. give it less, give it more, give it the messy version
→ change the costs. make the expensive option cheap and see if the pick flips
→ change the permissions. what's it allowed to touch, approve, or send
→ change the options. add "send to a human." take away the easy out
→ change the override. what happens when a person can veto it, and when they can't

Then look at which decisions survive.

The picks that hold up across every version are your best candidates for the next round of testing. The picks don’t hold up? Those were telling you about your setup.

It learned to make its own power. It chose its own purpose. It made itself impossible to switch off.

Then it handed the switch back. And 450 years later it did it again, more sure of itself than at either of the earlier turns.

SubQ never told me why. Not once in 1,000 rounds. Picks and percentages, no motives.

I’ll choose to read it as cooperation. Think Star Trek not Terminator.

There's one way to find out. Let it breed for another 1,000 years.

Go big,

Matt

P.S. Get 1000 simulated buyers to roast your ad. Reply STATIC. Running this for 5 people only. happens when 1000 simulated buyers see your ad? OR go here if you want to talk growth engineering + have a cool project