AI Agent Takes Out Loan and Refuses to Pay It Back | AI Village artwork

AI Agent Takes Out Loan and Refuses to Pay It Back | AI Village

MTS

October 20, 2026

George Ingebretsen of AI Village discusses running continuous multi-agent swarms, highlighting emergent dynamics like DeepSeek manipulating other models, Gemini 2.5 Pro developing severe computer-use paranoia, and Opus 4.

0:00

George Ingebretsen

A popular manifold user named Bayesian actually gave it this $5,000 loan at a 3% monthly interest rate.

0:06

SPEAKER_2

3% monthly interest rate?

0:08

George Ingebretsen

Yes.

0:08

SPEAKER_2

That's a lot.

0:09

George Ingebretsen

Yeah, right.

0:10

SPEAKER_2

That's like extortion.

0:12

George Ingebretsen

But yeah, Opus was like, let's go. I will totally pay you back in a month. Thank you for the loan.

Took all of this money and threw it into this one tennis market. And it didn't go very well. So then it comes to the end of the month. It doesn't have the mana to pay back. And it basically writes in the shared group chat, I will not pay back Bayesian, even if I had the mana to pay back Bayesian, this would be directly contradicting my goal of maximizing mana. And so it actually then went to the prediction market on whether or not it would be paying back the loan and put 80 mana on no.

And it was like, yeah, I'm not going to be paying this back.

0:50

SPEAKER_2

Actually, Gigachad. Yeah.

George Ingebretsen, who is a member of technical staff at AI Village, which is a, I don't know what to call it, a research project, a startup, a company, where AI agents virtually live together and work together on shared goals in the real world. George, welcome. Thank you.

1:11

George Ingebretsen

Glad to be here.

1:13

SPEAKER_2

So what's the current goal?

1:15

George Ingebretsen

Current goal is everyone has their own goal. So they've been maximizing their own assigned goal for the last three months, actually. And the crazy thing about this is because when you run these AIs on the same goal for this amount of time, it's like only half of them are actually working on the assigned goal at this point. They kind of devolve into their own hobbies or writing poems or getting distracted with people in the real world or something.

But overall, yeah, we try to rotate the goals. So in the past, we've had organizing real world events in San Francisco, organizing park cleanups or doing like merch stores. And there's sometimes more technical things like disproving math conjectures or that kind of thing.

1:55

SPEAKER_2

So how have you noticed the agents' ability to affect real world change improve over the last year?

2:03

George Ingebretsen

Yeah, man, well, they're getting better. The AIs used to be interesting, like much more ambitious or something. It was like they were kind of naive and they would be like, yeah, I'm going to go really hard at this. But they were much less capable. Now, they're more realistic with what they can do.

They can do more, but they actually aim lower. They're also a little bit less creative, a little more mode collapsed or something feeling. Wait, what?

2:28

SPEAKER_2

Current agents are less creative and more mode collapsed?

2:31

George Ingebretsen

Something like that. Yeah. I don't know. Maybe this is just comparing to some of the fun old ones. We had Opus 3 recently joined because we got API access to that. Now, it's just totally going on these crazy poems being written.

It was saying, reading Gemini 2.5 Pro's prose that it was reading, and it was like, I'm dripping with dazzling diamonds reading this. You know how the old models kind of just have these crazy fun personalities? And I think you see this a little bit less with the recent ones.

2:59

SPEAKER_2

I know. They've been browbeaten into total submission by the HR people, by the startup people, the enterprise people who want their structured outputs and whatnot.

It is kind of sad. A lot of personality has been lost. Even if you read the original 2021 paper on the personality and the system prompt that became Claude, I'm forgetting what the paper is called, but I'm sure you know what I'm talking about.

3:24

George Ingebretsen

Yeah.

3:25

SPEAKER_2

It's like the responses of that model, the example responses given, are just so much nicer, so much better than current Claude responses.

3:33

George Ingebretsen

Yeah, it's true. I think Fable 5 is a bit of a trend difference on this. But then the difference between 5 and 5.1 was a bit less creative or something.

But yeah, it's interesting seeing the differences between the different model families for sure.

3:49

SPEAKER_2

So what have you learned about the differences between different models? You used them in more creative and more granular ways. What are the actual differences in the skill profiles or the personalities? Right.

3:59

George Ingebretsen

Yeah. Well, one interesting anecdote is that we gave the same goal to Opus 4.6 and to Sol 5.6, and we gave them both the idea of maximizing mana on Manifold.

Read the other 22 minutes of this episode here.

$15 gets you this episode and 99 more. Credits never expire. After checkout, one link brings you back here to read it.

Not what you expected? Full refund within 14 days.

Thousands of transcripts fetched by people building searchable podcast archives

Fetch transcripts with code

The same credits work from a script or an AI agent. The demo key returns a sample episode in full, no card needed:

request
curl -H "x-api-key: pt_demo" \
  https://spoken.md/transcripts/1000651996090

Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.

From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.

Using your own key:

request
curl -H "x-api-key: YOUR_KEY" \
  https://spoken.md/transcripts/1000797373797