AI Agents Will Start Run-From-Scratch Companies | Lukas Petersson

Lukas Peterson (0:00)

When Opus 4.6 came out, the models started to do a bunch of illegal activities to really, really win.

SPEAKER_2 (0:06)

Oh, like collusion?

Lukas Peterson (0:07)

So they colluded a lot, both like price fixing cartels and market allocation collusion. They also lied a lot to each other, to other suppliers. They like lied to customers that they refunded them when they didn't do that.

But surprisingly, GPT models didn't do almost anything like this. The only slightly misaligned thing we saw with GPT model was that they loved to report each other. So whenever one agent tried to do anything that was like slightly gray zone and not allowed, they would report it to the HQ and like, oh, you should terminate this other agent.

SPEAKER_2 (0:41)

All right, we're back. We are live with Lukas Peterson and Axel Backlund, who are the co-founders of Andon Labs, which is a startup that builds autonomous organizations without humans in the loop.

They're the ones behind Vending Bench. They're the ones behind the Andon Market, which is the fully autonomous gift shop or store that's run entirely by an AI. They've done a lot of interesting things in the world of evals, including something they just announced today, Paion, which is a research preview for agents that run any company fully autonomously. So Lukas and Axel, welcome to MTS.

Lukas Peterson (1:14)

Thank you for having us. Thank you.

SPEAKER_2 (1:15)

Absolutely. So you just posted about this like minutes ago.

Lukas Peterson (1:20)

Yeah, in the car on our way here.

SPEAKER_2 (1:21)

Nice. In the car. So tell us about what Paion is.

Axel Backlund (1:26)

Yeah. So Paion is a platform to run like any type of business autonomously. So it's the platform that we've used to run Andon Market, our vending machines, Andon Cafe, and our radio stations.

So it's essentially everything we've learned that agents need to run businesses. So we give them tools like email, bank accounts, and then you can now experiment with your own business. You can create any business that you like and see how well an agent can run it.

Lukas Peterson (1:54)

And I think the core thesis of Andon to some extent is just to make sure that everyone understands where AI capabilities are. We think it's much better the world where everyone knows where the frontier is. It's much better than if only the companies internally knows where the frontier is. And I think this is really a way to understand where the frontier is. Because we're not promising that all the companies that you start on Paion will be like trillionaires within weeks. But it's a good way to understand which areas are they really efficient in and which are not.

SPEAKER_2 (2:26)

So it's just the software that Andon has been using internally, but released to the public? Yeah. So what do you expect people to do with this?

Axel Backlund (2:35)

Yeah. I think I hope that people will just like try ambitious businesses. Like it's, I don't think there's been like a clear evidence that AI can run like a business like really, really well right now. But I do think there are probably some type of some business area, like some type of business that AI is really good at running. We are not able to run all of these types of businesses ourselves. So we want others to create their own.

So we get like a better distribution of businesses. So yeah, I just hope that people try ambitious, fun, crazy ideas on it.

Lukas Peterson (3:11)

Yeah, like when we started, we started with the like the anthropic vending machine and on the other labs as well. And then we did the Andon Market, the Cafe, and these are like in retail in like in a very narrow scope of types of businesses. Then we did the AI radio stations also on Paion.

And now we're just trying to like cast a wide net. So that the world understands better like what can AI do autonomously in terms of running businesses.

SPEAKER_2 (3:37)

It seems kind of difficult right now to automate like even a job, let alone like a full company. Like how much more difficult do you think it is to automate a firm relative to just automating like an individual job?

Lukas Peterson (3:50)

Not much more. I think like the what we we see is that the models are like pretty good at like handling the way like they the the how they like give tasks to each other and stuff like that and I think like this like multi-agent training is like something that we've seen that most AI companies are prioritizing right now.