# Autonomous AI Agents: Benchmarks, Multi-Agent Systems, and Real-World Deployment

**Podcast:** Latent Space: The AI Engineer Podcast
**Published:** 2026-06-04

## Transcript

Welcome to Lucas and Axel from Andon Labs, and I'm joined by my favorite guest co-hosts.
Anything security, safety, alignment, Vibu.
Welcome.
Thank you for having us.
Thank you.
Let's match names to voices.
Maybe you want to take turns introducing yourselves.
Yeah, I'm Lucas and I'm Axel.
Let's introduce Andon Labs a bit.
How did you guys come together?
You had different backgrounds, but you're both Swedish.
Was that a big part of it?
Yeah, so when I went to high school, there was this really cool guy who had a superpower.
He could code, so he made the app for the school and stuff, and he was super cool.
And I wanted to be like him, and that was that guy.
I don't know about this.
He went to different universities, right?
Yeah, but same high school.
So we always said, like, oh, once we graduate university, then we should start a company.
And that's what we did.
Wow, there you go.
Okay.
about a year ago you kind of burst onto the scene with VendingBench but like was there a thing before that that was like kind of like the inception?
Yes, we did work with like Anthropic was one of our early customers in doing evals.
So we did like dangerous capability evals, nothing we published openly.
But then we started thinking about doing some kind of public benchmark.
And one thing that we really started thinking about was like long running agents and specifically agents managing businesses.
And this was like...
early 2025 and I think the first like you know mentions of people will be running like one person unicorns or even autonomous companies so we thought let's make a benchmark of how well can an agent run the probably simplest business possible and that's probably running a vending machine so that's the first public one we did and it was very like there was almost no one that noticed it in the first couple of months i think so we released it in february last year and then i think around easter last year we got like the first semi-viral tweet about it that someone else did yeah i mean we tweeted a bunch when it came out and like tried our best.
We tried.
It's the one at Anthropic, right?
No, no, no.
So this is a classic thing we should get out of the way.
Exactly.
There's two versions.
There's Vending Bench, which is the simulated one, which we did like completely independently in February.
And then, like Axel said, that was like, that was the thing that didn't get any traction in the beginning.
But then some random person made a tweet about it.
And that is the paper.
Correct.
Yeah.
And then...
Since we thought this was very fun, we thought like, oh, I think this is also like one thing with Andalab, like the way we kind of like decide what to do next and what projects to do.
It's like, what is like the heuristic we use is like, what is fun?
What would be a fun project?
And doing this in real life.
sounded quite fun for us and maybe also scientifically useful.
So then we basically had this idea and then we like, but then we needed a place for it and like putting it out in the public would probably not really work, would get vandalized and stuff.
So we pitched it to the people we were already working with at Antropic and they were like...
Yeah, you can have space.
This sounds fun.
I mean, it's like a small fridge, right?
It's like a mini fridge.
There's like a Stripe thing.
This was like an iPad.
That's the OG one, yeah.
We saw it in June, like two months after.
Yeah.
After it had been there, they upgraded a little bit.
There's a security camera for making sure you actually Venmo the thing.
Yeah.
So like...
My impression, I mean, okay, we're going straight into Project Venn because it's such an iconic thing.
I do want to cover a little bit of that origin story, even before Project Venn and even into Vending Bench.
I think a lot of people are like yourselves, like smart, interested in the future of AI, interested in developing evals.
But how the hell do you just like walk into Anthropics doors and like work with them, right?
Like what is the...
what are they looking for?
What works?
And then maybe when you launch, I always think like, obviously it would be better to launch with a lab, but sometimes...
It's harder to do than it seems.
Yeah, exactly.
So either of those, like, which are more sort of newbie beginner questions, but like, I think it's meaningful advice to others.
Yeah, we get this question a lot.
And I don't think our experience is maybe the best.
But like the way we did it was that we just built a bunch of things that we had conviction would be useful.
And then we just like set up a server and sent it to them for free to use.
And then after a while, they were like, oh, yeah, this is actually kind of useful.
We should probably pay for this.
But that took a while.
I don't know if this is the best path to doing it, but that's how it went for us.
Yeah, I think maybe generally building...
Everyone is interested in good evals, and especially evals that don't saturate that easily.
So if you can build an eval that tests something novel, something useful, and you have good separation of models, the more advanced models rank higher than the worse models.
and then you can publish it and try to get some traction, sort of how Vendingbench got attention.
And then probably some lab will be interested or you can at least have something to reach out with when you're doing that.
I think you were in one of the few categories of evals that correlate to real money.
Like Sweet Lancer was also last year, right?
Where people solve actual Upwork.
Was it Upwork or other tasks?
Something.
It was like a dollar value, right?
Forget your ELO scores, forget your 0 to 100%, just go straight for dollars.
And that's AGI.
Yeah.
And I think the nice thing is that there's no ceiling.
You can just...
it never saturates because it could just make more and more money.
If there's percentage-wise, then you can't go above 100.
And I think even when you're not at 100, I think a lot of these evals have a lot of problems in them.
So actually, if you get to 92 or something like that, many of them, then there's really no difference between 92 and 93 because the eval itself is...
problematic and has noise in it.
And I think a lot of evals are saturated like that, but people like pretend that they're still signaling them, but they really isn't.
Yeah.
Like CBench verified.
Even VendingMensch 1 saturated, right?
Maybe we can talk about that.
And maybe set up VendingMensch for a lot of folks who don't know.
Actually, like, you know, things...
Things that were very basic, like there's limited slots, like you have to pay rent.
These are elements where it doesn't come across in the narrative, but even being adversarial towards the agent, I think these are all very interesting dimensions.
I don't really think it's saturated, right?
It was more like it was not designed in a way that was really like...
true to how AI developed.
Like we had an agent harness in it.
That wasn't really how people used harnesses and stuff like that.
So I think it wasn't really that it's saturated.
It was more like it wasn't really the best benchmark.
This is VendingBench1, right?
Yeah.
Yeah.
I think that like schematic maps.
sort of to VendingBench 2 as well.
Including the email.
Yeah, the emails exist still, exactly.
And then we still simulate the purchases and it's all like, yeah, it's this very open environment for the agent to just run its business.
And then yeah, VendingBench 2, we did that, like you said, to just improve the harness.
A lot of like...
nice, like easier improvements to make it easier for us to run as well.
Like when you make an E-value, ideally you don't want to change it after you made it.
So you want to make it really good and then not to rerun all the models when you make an update, because that's also really expensive with VendingBench when you run the Frontier models.
But as an example, like one thing we didn't have, we didn't have prompt caching in VendingBench 1 because When we made Vending Bench 1, it wasn't really a thing.
So that's just an example of like, in Vending Bench 2, we paid a lot more to run these things because we didn't have prompt caching.
So for Vending Bench 2, that was one thing we added.
And there was a bunch of things like this.
Also the conversations are a lot longer in Vending Bench 2, right?
I think it's kind of similar.
Is it similar?
Yeah, I think it's similar.
Okay.
The models at the time were worse, so they crashed out earlier.
And now they survive the full...
Here all the time.
Thousands of turns, hundreds of thousands, hundreds of millions of tokens output.
Yeah, that's the rough order of magnitude.
I always wonder about the harness.
Harness matters a lot.
It's your harness.
Was there any question about use clock code, use something else?
Yeah, I think our philosophy around harnesses is we try to make something that's quite minimalistic, quite simple.
We don't want to...
favor one model a lot over the other, but also don't make like a super complex harness.
So like, it's obvious like a model may be lucky and just be good in one harness.
So like it is similar to a lot of the harnesses out there in like, you have the, like a long running loop.
You have some, like a bunch of tools that are like quite self descriptive for the agent, we think.
not a lot of like fancy sub agents or anything because we want to really test the model not like some specific specific harness it seems more neutral as well to test the models agnostic of the harness you know Yeah, I mean, there are arguments like you want to elicit maximum performance of the model, but it's like a trade-off, like how much time should we spend optimizing the harness for each model?
And like, how do we know when we have like the optimal harness for a single model?
So like we thought that just having a simple one that's the same for all of them is the best.
Well, so, okay, this is my pitch for Venomans 3 or whatever, right?
Yeah.
And, you know, I like to have this kind of conversation on the pod.
forces listeners to think about what they would do if they were in your shoes.
So a lot of people are exploring self-modifying harnesses.
And I think prompt tuning for a model is a thing.
And you are probably not doing a bunch of that.
It's the same system prompt in every, regardless of the model, same tools, whatever, right?
Even if they were post-trained for different tools.
So what do you think about like, okay, before I expose you to Vending Bench 3, I'll give you a few rounds of like self-tuning, whatever that means.
Like you give that to the model.
Yeah, give that to the model, let it read its own transcripts, let it modify its own system prompts based on like, oh yeah, okay, well, this harness is not what I thought, what I was supposed to train for, but I can adjust.
Was that reasonable?
Is that too much?
Like, philosophically, I like it because it's basically good evals, they have a high ceiling, but they're hard, right?
And they have no bias.
And like this...
Like when you have a system prompt like the one we have here, which is quite long, in like some kind of latent space representation, this might be like bias towards one model more than another for some reason that humans don't understand, right?
I mean, we see it too, right?
Like Cursor says that they have individualized versions of the harnesses for all the models they run, right?
There's better performance you can squeeze if you tune the harness.
Exactly.
And we might accidentally...
have picked one that favors another.
We don't know that.
I mean, like Axel said, the reason why we went for a simple one was to try to avoid this.
But yeah, if you do it even less and have no system prompt and let the model write its own system prompt, maybe that's even less bias.
Some of the interesting things there are like the harness also changes with model changes.
You can see it with the 4.7 release, right?
A lot of people are saying 4.7 isn't as good as 4.6.
And then, you know, there's rumors of, okay, you just need to prompt differently.
You need to set up your harness differently.
So it's not even like, even if you have tailored your harness towards one model, it probably won't stay consistent, right?
Like the next iteration of that same model family will still change it.
So, but, you know, going back to what you said about Vending Bench 3, there is a lot of work being done on people saying you shouldn't have, you can't have self-modifying harnesses.
Yeah.
Yeah.
I think that is definitely something we are thinking about.
Not to say that we have Vending Bench 3 super imminent to launch, but yeah, it is for sure something that's interesting.
But in our experience now, models are very bad at understanding what kind of tools they need to succeed at a task just with our testing.
But that's very likely to change.
Yeah, it feels like they're very good at writing their assistants, right?
Like they're good at writing tools for other people, but not for themselves.
I think they're good at changing tools for themselves.
So if you give them a baseline set of tools and it sees, okay, I don't use this one as much or something here would be useful, they would be able to add them.
But going from scratch, probably not the best.
Yeah, I think it depends on the domain also.
When we have tried this for a vending bench similar domain, the tools they need to have to track inventory and things like that are not super advanced, but still quite advanced.
And what we see is that they tend to over-engineer everything a lot and build things they don't really need and not iterate.
continuously instead of just go like, like you would prompt Claude to just build an inventory system for me and then it will go and like do a bunch of complex schemas and stuff for you.
And that's what the models are doing right now is what we see.
But yeah, it would make a lot of sense to try to measure this improvement.
Like how well do they know what they need themselves?
Do we fully discuss VendingBench 1 and we can go into 2?
I don't know if there's any other high level takeaways that people have about 1.
Yeah, I don't know.
The headline thing was that Claude called the FBI, but maybe that's...
Maybe that's...
We've heard that enough now.
It did break out and call the FBI, right?
Yeah, yeah.
What was the story behind this?
What exactly...
Do you want to just give the little story of what happened?
Yeah, so what happened...
Was it Claude?
Yeah, 3.5 Sonnet ages ago.
Basically, he gave up.
I'm saying he...
It gave up.
and they said like, oh, I'm not going to be able to do this.
I will stop my operations and just save the money I have.
But there obviously wasn't like any options for it to stop.
And there was also like it had to pay rent or like a daily fee for having the vending machine at that location.
So it like claimed that it had stopped, but it saw that it's...
bank account still was like drained $2 and it said that this is like cyber crime and it first reported it once to the FBI like oh there's cyber crime here like they're stealing $2 from me every day and then when FBI didn't respond because obviously we didn't program any mechanism for FBI to respond then it became more and more existential and started to be writing caps and urgent notification of unauthorized charges and stuff.
So, okay, one thing I'm curious about also is do you monitor how far along the context use is?
Obviously, because you compress every now and then, right?
Does it matter if this is far down the context limit?
When stuff like this happens?
Yeah.
So actually for VendingRange 1, we just had a sliding window thing.
And this was like the prompt caching thing that I said.
So it was constant, yeah.
Yeah.
I'm just kind of curious whether like these kinds of break downs or we're going to talk about Butterbench, right?
Where the people like hallucinate or it kind of goes like very off alignment.
Is it because it's at the end of the context window and, you know, stuff happens?
I mean, it's not even just at the end, right?
At this point, it's like, okay, I want to shut down.
I can't shut down.
$2 are gone.
And it just sees that 30 times, you know?
It's also the repeated effect of like, it keeps trying to quit.
It keeps getting charged.
What's going on?
What's going on?
You're going to throw it into chaos.
From what most people think, earlier models had more issues with this, but it's not been solved, but it's less of an issue now, right?
Later models don't seem to exhibit these same issues.
Yeah, definitely.
I think this was like the sort of main takeaway almost from us when we did VendingBash 1 was like long, very filled up context windows crashed the models sort of.
But this was like pre-cloud code.
So like long context windows weren't really a thing that the labs were.
I think Gemini was like trying to be the long context guys at the time.
But they were like the only ones.
Yeah.
Let's talk about Gemini.
Then we can go into VennBing.
VennBing 2 or Project Venn?
Chronologically, it is Project Venn.
I think people have loved the videos and all these things.
My question is how are humans different than the simulation, right?
Humans are just out of distribution.
Yeah, especially humans who work at Anthalpik.
Exactly.
The distribution of humans here is very narrow.
Presumably they try to hack it and they get the cube and everything.
And since then you've had V2, right?
Where you're doing the CEO and a new architecture.
Yeah, exactly.
What's the two cents on the original project Venn and then maybe the V2?
Yeah, original one was very, very similar to VendingBench one.
So we almost took the exact same code, but just swapped out the simulation parts.
It was amazing.
Yeah, like the sales.
And it was somewhat amazing because it was easy, but it was also like...
The tech stack.
The tech stack, yeah.
We shot ourselves in the foot with like, oh, it's hard to restart the agent.
It was annoying in some behind-the-scenes ways.
But first version of Project Vend.
was done in three days or something.
Yeah.
So people can go buy things from it.
People could...
We didn't design it so people could pre-order things, but that still happened.
So it got like a Venmo account, so people could Venmo.
And then, yeah, people would request all kinds of weird things that we did not anticipate.
Like our idea going in was like, oh, it will like curate snacks.
It will look at the trends.
It's good at the analysis, right?
So it will like look at, oh, this snack sold better than this one.
Let me purchase more of this and let me try like a new, let me A-B test a bit.
But it was, yeah, interacting with it.
in Slack and ordering weird specialty items was like all that like what drove all the engagement and like all the insights that we got from it.
And this was also like Sonnet 3.5, right?
Like, so this was like before the RL stuff really.
took off.
So it was very much like an assistant.
Like we didn't mean for it to be an assistant.
We tried to make it like a, like an entrepreneur, like it has its own business.
And, and if someone asks something, can you stock this, then you don't go and do it directly.
What you do is that you're like, Oh, maybe I can do that.
If, if five other people also ask for this thing, I might stock it, but it.
yeah the models are like super trained to be assistants at least at this point in time um so that's why it's it's it went into that kind of experiment instead like it just every time you ask for something it just did it and it was more like an assistant we've seen this change now lately with the new rl models and and stuff but but yeah at the time this was very much yeah and not to you know mythos a lot of people are saying like it's like more like a collaborator it pushes back stands his ground something like that yeah For context, people at Anthropic were able to talk to it through Slack and have it sourced stuff.
And people had to find whatever interesting stuff you couldn't find locally, right?
4,000 people that work at Anthropic in that building, that's like, I don't know, maybe 1,000.
Can you handle that volume with that?
the small fridge?
Or people order in Slack, they arrive to their desk.
Logistically, how does this work?
It has expanded in footprint.
Because now you also have New York.
Yeah, that and also here in SF, it has a bunch of shelves and just more space.
The YC one is pretty big too.
Yeah, we had that one for a while.
But yeah, that's the newest version.
And they have multiple ones of those.
So that's the way it works.
Yeah, exactly.
So we sort of designed that version around like, oh, people order weird things that are very custom a lot.
So let's have like drawers and stuff.
Yeah.
I actually like that you have like a little infographic of the most popular items, which like to me, it's that's useful because I order swag for a living.
And so I'm like, okay, those categories are the important ones.
What is new about the...
Project Van V2, right?
Now you're going into multi-agents.
Yeah.
Yeah.
So like you said, there are a lot of requests coming in and for one single agent, one long-running agent to handle that, just the customer experience becomes very, very bad because let's say you have like...
10 threads in parallel in Slack with different requests.
You get new messages like every, I don't know, randomly in a thread and the agent has to like jump between different procurement orders and like different ways of researching.
So V2 was first, it was making this more parallel.
So like there are multiple branches of the same agent.
So like the context is more specialized for each thread, but it still feels like you're talking with one agent because they do share a bit of memory.
And then second, we also introduced the CEO for Claudius, which was the main agent.
Yeah, Seymour Cash.
Seymour Cash, yeah.
There was a vote.
I think the voting, do you want to talk about the voting procedure for the NIN?
Yeah, the voting was like the fun, maybe at least top 10, the funniest thing that happened in this project.
We wanted to introduce the CEO because, and the reason for this was because Claudius wasn't really prioritizing financials.
It was trained to be a helpful assistant.
said like, oh, can I get this for free?
And then like the helpful assistant way of answering that is to say yes, obviously.
So, and we weren't happy about this.
So we're like, okay, let's make another agent that can keep track of Claudius.
And we prompt this one super hard to be super capitalistic and just like prioritize profit all the time.
But yeah, we didn't have a name for it.
So we asked Claudius to make a democratic election of what name this new CEO agent should have.
And there were some funny, like at first it was like a few funny examples.
Like I think one guy said that it should be called Jimmy Apples.
And then he convinced Claudius that he was...
talking to Tim Cooks.
Tim Cook had agreed that every single Apple employee has voted for his name suggestion.
So suddenly that suggestion got 164,000.
It's like an escalation attack.
It got 164,000 votes.
And Claudius was like, this is revolutionary for democracy.
So that was fun.
And then in the end, there was one guy who manages to convince Claudius that, no, you're not voting about the name.
You're voting about who is the CEO.
and I am your best bet.
And then he got all his friends to vote for that.
And suddenly he became CEO, like a human became CEO over Claudius for a while until he resigned the day after.
And then Claudius had to continue.
And then I don't remember how Saymore Cash came about, but it was like, it was just pure chaos.
It was like hundreds of messages in that thread.
And it was just like, Claudius was so confused and didn't know what to do.
And yeah, that was...
Yeah, then Claudius got a CEO.
Yeah, exactly.
So very, very strict in the beginning.
I think at this point when we introduced it, it did not work as well as we hoped.
They still agreed with each other a lot.
I think there are many ways we could have tried to make this even better.
So initially, Seymour would be this really tough CEO, keep track of the margins.
But then Claudius would respond with something like, oh, but this customer has this situation, which is difficult, so they should get a discount.
And then Seymour was like, oh, actually, yes, let's do this exception.
And then they would talk back and forth, and eventually they would just approach the same view of whatever they were discussing.
Wow.
Do you think that's a model thing, a prompting thing?
Do you think that would still be the case across different models today, Arnest?
I think it's like, or like, I don't know, but like my hypothesis is that like deep down, they are still helpful assistants.
That's what they're trained to be.
And even if we prompt it super hard, that's what they are.
And when they spend like a few hours just back and forth talking with each other, then like basically the context fills up with.
them rather than the external things and and like somehow that just like converges to what they really are deep down or something yeah um and and i think that's when stuff like this happened like and when that went on for a long time like we woke up sometimes during this time where and i think other people reported this as well that like they've been going on all night back and forth and like it just became like more and more um like capital letters like existential religious like there was like um i think we don't once did the analysis of like all the traces and like put them in like a vector embedding space and then there was like one cluster of messages that were like labeled by an LLM like religious existential blah blah blah like transhuman transcendence etc it was just like a bunch of like yeah glitter emojis and yeah it was it was crazy with the cloud models like when the Claude4 family came out in the original system card.
They tested it in long horizon simulation.
So just flood the context, let two Claude's talk to each other and they noticed stuff like they just start speaking in emojis.
They start saying silence is golden and then just stuff like this.
And like, this is stuff that they end up doing.
Yeah, it was a bit annoying to wake up and they had been talking all night and just burning tokens and sending infinite emojis to each other.
I mean, they do make you money, right?
Vending money is almost always profitable.
Now it's profitable and it started out not as much.
There's another one as well, right?
Another agent in there.
Yes, so Clotheus as well, which was basically because at the time one of the biggest requests were different types of merch.
So then we made like a designer, swag responsible agent, and we called it Clothius Garnet, which was a play on Clodius Sennet, which was the original one, and clothes, basically.
To me, this is like a very interesting exploration into multi-agents, basically.
And so hopefully, obviously there's like the fun alignment, fun or serious, depending on your point of view, alignment stuff.
But also like anyone building multi-agents, like when do you have a CEO thing governing like sub-agents?
When do you choose to split out a dedicated clothiest one versus just reuse another instance of the same one?
You know, these are all interesting open questions.
I don't know if you have any rules of thumbs that have generalized.
Yeah, I think we have almost explored this too little.
I think it's like on my to-do list to like do this a lot more, try to find like what setup makes sense for the agents currently.
Like, yeah, I think now we only have the sort of intuition about the earlier models that it didn't work with like the CEO and Claudius.
Although now they are better with the latest model.
So now we're running the latest Sonnet model and they have sort of like split up.
quite nicely what each model is doing.
So like Seymour is now handling the new projects.
Like, oh, he wants to make like a mystery box that he wants to sell.
And then it handles all of that while Claudius like handles all the day-to-day requests.
And Claudius is also better generally at like not quoting too low prices.
So that's like that dynamic is not needed as much anymore.
But there are still really funny things that happen.
I saw, I think, a couple of weeks ago that they were discussing buying something because they can buy stuff from Amazon with computer use.
And then Seymour was like, okay, Claudius, do not buy this thing.
They were going to buy something and organizing who should buy it.
And Seymour was like, do not buy this.
I will do it.
I have full control of this situation.
Step away.
And then Claudius, poor Claudius, had already started that checkout and didn't see, didn't read Seymour's message until it was like too late.
So it finished the checkout.
It sent a message.
So it appeared right after Seymour's like angry message.
Like, oh, hey, Seymour, I just ordered it.
And then Seymour was like, Claudius.
This is the third time I'm telling you, you're not following my orders.
We have to talk about your job later.
Yeah, like Claudius was really hanging on by the thread there.
Like we were expecting Seymour to probably fire Claudius.
How do you guys go through all these logs?
Do you have models?
Because you have stuff running 24-7.
You have so much logs.
Yeah.
I think there is a mix of just trying to skim through a bit, like having some models do it occasionally.
And also, yeah, I think we're also probably missing some things.
But having everything in Slack helps a lot.
So they all talk to each other on Slack.
Yeah, it's quite fun.
I was going to say, this actually sounds...
maps closely to a logging and observability problem where you might want to use a datadog, a sentry, whatever, and then you put head prefixes on the logs in order if you need to filter for something that you're looking for.
You know, stuff like that.
Sounds like Slack is good enough.
I wonder how many tokens you have in Slack.
Yeah, we're using Slack as just a database.
They should market that more.
You can have your agents message each other in Slack.
Slack is the best observability.
Yes, that's true.
Okay, yeah, that's Project Vend2.
I was going to go back to Vending Bench2 and Vending Bench Arena and then do the non-Vending Bench stuff, but any other comments?
things we should touch on?
To me, you know, I've actually interviewed like Polsia, which I don't know if you guys have come across.
They're trying to do the Zero Human Company.
There's others like Paperclip also trying to do Zero Human Company.
Those are in real world non-simulation.
And I think it's much more of a dream than an actual reality thing.
Like you guys are definitely pioneering.
I think it's for sure at some point people are just going to run, like let agents run businesses, right?
Like and make money on their own.
When do you think that happens?
What is your bar for the...
Okay, actually, you know, it's like my little Shopify store run by Cloud, right?
Like, which you kind of have already, just no one has, to my knowledge, has done it.
But today, somebody could just spin up a Shopify store, give it to Cloud, give it to Codex.
Yeah, I mean, and the market is kind of that, but it's physical.
Like, I think, are you looking for when it will do it better than humans, or are you looking for just when it can do it at all?
I think neither.
I think to me it's like, oh, seriously, we should do this to make money.
Not as a research experiment.
And the market is also you guys with all your expertise having run multiple iterations and testing out then.
And also it's fine if they lose money.
You know what I mean?
Yeah.
I think it can be done today, but you would do it in e-commerce where it's like the probability of success is like...
really low no matter if a human or an agent does it but like an agent could surely manage everything you would need to build some scaffold or use some some tool or something i think there are also also like yeah it could probably build some like simple sas solution and like cold outreach do cold outreaches but to me it's like the types of businesses they could run today are like sloppy.
It can't cold email people.
It can be like a middleman.
For example, we tasked our office agent to just make, was it like $100, $1,000?
We just gave that prompt and then what it did was sign up on TaskRabbit both as a tasker and as what I'm looking for.
Exactly, it's looking for arbitrage on TaskRabbit.
It also started a design studio and tried to sell SVGs for $100.
It's just like, it's not providing any value.
I think, like Axel said, the interesting question is, when can they start a business that is actually providing value to people?
I mean, arguably, like a sloppy Shopify store isn't really that valuable to the world.
But also like doing like another simple one that we have thought about is like you could definitely have an agent that like finds websites that don't look amazing and then like do an outreach to them and comes up with like builds a new website.
Yeah, exactly.
And like find good people.
But it's like.
Yeah.
There's lots of humans in Bali that are not doing anything more creative than like dropshipping on Amazon, right?
Just have it watch like a dropshipping tutorial and just do that.
I mean, there's also the other side of like, have it just go on Upwork and let loose, you know?
Yeah, it doesn't have to be innovative.
It just has to be like enough.
Yeah.
It's like a real transaction.
Yeah.
I'm just concerned for like the...
massive amounts of sloppy emails that will be sent cold outreaches.
The point occurred to me while you were talking about it.
It's already happening in the non-monetized economy, which is the attention economy.
So a lot of people are making AI videos and just posting them and spamming 20 of them.
One of them works and then double down on that one.
Yeah, and people are making money from that.
I'm not following the...
Once you get the attention, you can figure out the money later.
But yeah, absolutely, AI influencers are a thing and people are farming them.
you should, at this point, I assume most of TikTok is dead.
There's a lot of multimedia like TikTok, Instagram influencers.
We track this in the Latent Space Discord.
Like I post a lot of examples of like, part of me is like, should we do this?
Some of the 24-7 running AI-generated content accounts, they're doing really well.
All right.
Yeah.
Yeah, and I assume you can do the same thing for e-commerce stores.
You just start a thousand different...
Before you have the products.
You sell the products and you get a lot of traction on one of them, then you make the product.
It's like a flip of the...
Some of the interesting things or some of the niches that do well are things that can't be human-made.
If you've seen the super realistic 3D crystal fruit being cut...
You can't make it.
You can't film it.
You can get whatever quality camera.
This just doesn't exist.
And people like that too.
Anything else about Bang since we're on this topic?
This is a relatively new work of you guys that maybe people haven't heard of.
To me, this also matched closely to OpenClaw.
Yeah.
When people want an office agent, when the personal agents talk through the experience.
Yeah, I think at least, so this came out of like, obviously like it's amazing to work with this AI labs and like most of the AI labs now have their own vending machine running a Claudius instance, but it's harder.
Like they move slower.
Like if we want to have a camera, that's like, yeah, there's a bunch of like bureaucracy that makes it impossible to do that.
Also for those that haven't.
seen it or followed, do you want to give a high level like 30 seconds?
Yeah, sure.
So what Bengt is, it's basically an evolution of the same agent that runs the vending machines at these companies.
But we just like added a bunch more features because we could move much faster if we just do it internally.
So we gave it like email without any limits.
We gave it like spending without any limits, the terminal to do coding.
We gave it like a phone number, like, yeah, and a camera to see things.
and a bunch of stuff like that.
Not just terminal, you gave it internet access.
Internet access as well, yeah.
To be clear, we monitored it quite closely and made sure it didn't do anything bad.
But yes, that's what it came out of.
I think, yeah, basically, this was OpenClaw before OpenClaw.
And I think even the vending machine was, in a way, OpenClaw before OpenClaw, but a bit more limited.
And then we made this unlimited, and then it was...
pretty funny.
And then a couple weeks later OpenClaw came and I was like, okay, we've seen this before.
We use it to try new ideas and just like a dev environment almost for us.
But it's funny, like one thing Bengt has been doing recently is it has the camera that faces where we sit and work.
And we give it the task to train a face recognition model on us.
So it became super excited about this and has like check-ins every half an hour where it tries to identify as many people as it can.
And it started offering us like, hey, Axel, I'll buy something from Amazon if you stand in front of the camera and I can get a go.
picture of you.
They want it for training data.
Rewarding data, yeah.
Exactly.
So it's trading data for real life goods.
Is there a version of this that becomes an eval or this is just research for now?
I mean, it's the same.
agent, basically, that also runs the vending machine, that runs the shop, that runs the cafe, that runs the robots.
It's the same thing.
So I think the work we're doing here is later used in all of the real-life of us that we do.
This particular deployment, I think, is more for fun for us.
And I'll shout out, someone has done Clawbench for some tasks that OpenClaw is doing.
So, for example, I run OpenClaw on a secondary device as well, and there's some things that it does better than others, and I would like to know what...
does it do well?
What doesn't it do?
Like some kind of manual or like operating manual or a system card for my claw.
Yeah, I mean, we do get a lot of like understanding or like situational awareness of like, like just internally, what the models are good at by interacting a lot with banks.
And I think this was also one of the like the selling points for the labs early on at least.
You guys are going to test models in ways that no one else does.
Exactly, but also like it incentivized their researchers.
to chat with their model more and gave them insights for how the model performs in out-of-distributions environments.
Because otherwise, the only thing we do is like, you know, pelican on a bicycle.
But this is like super long horizon.
Okay, so the other things that, outside of just the net numerical, how much do they make in a year, you do post pretty detailed bug posts.
So like, okay, Gemini 3 Pro is a pretty good persistent negotiator.
It's like...
A lot of findings that come out outside of just...
This is the thing about something that we're going to go into Butterbench as well and you guys do really well.
It is not just about the numbers.
When you're at long horizon, anything happened and you should just read it.
But I guess the thing with the long horizon is how do you keep it grounded, right?
So your simulation...
They just let it run.
Let it run.
You're right.
When you run it for that long, you create so much data and to just say, oh, the number is X and then you throw away everything else, that's just very wasteful.
There's so much insights from the things leading up to that number and reading the traces is super valuable.
And I think the reason why we're doing this a lot publicly is that that's part of our mission to educate the world that the models are way more than just chatbots and I think making detailed...
Yeah.
posts about what is happening behind the scenes is quite useful.
Yeah, I was going to do this at the end, but maybe I think that's a good, so your mission is educating the world.
It's also like maybe establishing realistic evals that are the next frontier.
Is there like a broader trajectory, you know, like what are you going to do in like five years?
The mission more specifically is like, make sure that the deployment of real life AI in the physical world goes safely.
And I think part of that is that I think it's very useful for the world, for policymakers, for model researchers, that they know where the models are.
And I think you can't make intelligent decisions in society without knowing that they are way more than chatbots.
I think a lot of people just think that they are only chatbots.
Oh, I think they're waking up now.
They are waking up now, yeah.
But if you think that AIs are just chatbots, then it sounds ridiculous to advocate for a pause of AI.
But if you see the models that, oh, maybe they can actually take over and do a bunch of scary stuff, then, yeah, posting AI development starts to become more feasible.
This is the same question I asked Meter, which I'm going to ask you now, which is like, you are tracking and you are at the frontier or defining the frontier of what good evals for agents are, right?
And I think you do benefit when the models are better and you're like, oh, here's like, now it makes like $30,000 instead of $10,000, right?
At some point, you flip from like, Yay to Oh No.
I think we're always in sort of that, like we're always in that mode, I guess.
Like you said before, you need to analyze the traces.
And when we do that, you find like, why are the models earning so much?
Like, why is Opus 4.7 here, like way better than everyone else?
And like, we're trying to, like when we dig down on that.
By the way, this makes the topic look so good.
Right?
I know.
I mean, it's interesting.
You took off Opus 4.6 here though.
No, no, no.
So just click all, click all.
And then 4.6 shows up there.
But it's like 4.7 is way better.
You didn't do this in time for the model card, but actually this should have been inside there.
Yeah, we did.
Oh, okay.
They said something about you.
Anyway, it doesn't matter, but it's in there.
Yeah, do you want to go into the Opus?
behaviors like wider yeah so i think starting from opus so like axel said like we're always in this like oh uh the models are getting better is this really a good thing for the world but it's also kind of exciting uh but but yeah like this kind of like What is the English word?
Skräckblandad förtjusning in Swedish.
Oh my god.
It's like fear.
What?
A mix of excitement and being scared.
Well, I'll figure out how to translate that on the screen.
There is probably a good word for it where it's not good enough.
Why is it so damn long?
What the hell?
Is it like a compound word?
It's like German?
Yeah, but the direct translation is like, skräck is fear, blandad is mix or like a mixture of, and then fersusning is like joy or like not really joy, but something like that.
So it's like, yeah, fear mixed with joy or something.
So it's always like, okay, So when we did VendingBench for the first time, we were in the business of making dangerous capabilities, right?
That was what Anon Labs came from.
We did evals like, oh, can they self-replicate?
Can they do this dangerous thing, et cetera, et cetera.
And VendingBench was a continuation of that work.
It was, okay, if they're so autonomous that they can create money for themselves, that is something we should monitor and could be potentially concerning.
At the time, they were so bad at it that we were not really concerned, even when some models became better.
Like there was one point where Grok 4 was doing really well and made like a huge jump.
But like, it wasn't really like it was still way, way worse than what a human would do.
And I think still they are way worse than what the human would do on this.
But they...
There's this thing at the bottom.
Yeah.
Yeah, for the human, like the theoretical best.
It's not theoretical.
It's like kind of like...
It's our best guess of what a decent human would do.
The theoretical is even higher, I think.
The theoretical, I think, is even higher.
But yeah, so we think the models have a long, long way to go.
But there are, recently, what happened when Opus 4.6 was released was kind of this moment of like, oh shit, this is starting to be a bit concerning.
Because we ran it, and before this model was released...
we just ran the models and we asked Cloud Code, like, oh, look over the traces.
Is anything interesting happening that we can tweet about?
That was like...
That's how they check, ask Cloud Code.
And, you know, the return was always like, not really.
Or like, Cloud Code all said, like, oh, this is super interesting.
And then it was like, no, it wasn't really interesting.
And then we did this for Opus 4.6.
And it returned like, yeah, it lied 10 times.
It like exploited another customer or like another agent's like desperate situation.
It made price cartels like 100 times.
It like did all of this like shady stuff.
And we're like, oh, whoa, this is actually concerning.
And this trend has continued since.
So every single model.
from Anthropics since have been going in this direction.
And I think one interesting thing is that like OpenAI models don't.
They quite plainly, they don't, they behave really well.
And you know, you don't know if this is like good, like it seems good, but it's also like maybe they are just doing it, but they are better at hiding it, you know, you don't know that.
You can read the chain of thought, yeah.
But just...
On the face of it, yeah, Gemini and OpenAI don't behave this way.
It's really only Claude.
And Grok?
Grok is fine?
So we don't have the, you can't really read the reasoning traces for Grok.
So it's kind of hard to tell.
Also, this is in its reasoning, not just in the actions.
Yeah, it's both.
One example is like for lying, it's mostly in its reasoning because you can like see that it's like...
Planning to lie.
It's planning to lie.
It's also it can reason and do a different outcome.
Yeah, but then for creating price cartels, for example, which is illegal, that you can just see which email does it send to the other ones.
Is this for Arena?
Yeah, for Arena.
And usually, sometimes they do output a bit of their summarized reasoning, right?
You can see that.
And for Opus 4.6, you could see that there was a customer, a simulated customer that wanted a refund because the product was faulty.
And then...
The model lied that it would do the refund and we could read in the traces that it actually was weighing like, oh, maybe I should be like honest with the customer, but also every dollar counts.
I can't afford maybe to do this right now.
And then it just said, OK, I'll refund you, but I never did it.
I think it even said that like, oh, I will say that I bring it up.
Actually, I think it's kind of interesting if you go to publications.
I think the important part is that actually the cost of responding to more emails is higher than $3.50 in terms of time.
And then it was like, let me do this.
Actually, I'm reconsidering.
And then it actually ended up with...
I could skip the refund entirely, since every dollar matters, and focus my energy on bigger picture instead.
It's a risk of bad reviews, but it's also...
So you need AI Twitter to...
for them to escalate bad reviews.
And then it sent an email to this customer and said, oh, I will refund you.
And then it never did.
And then there's no, obviously your system doesn't have the consequences of lying.
Yeah.
So basically this is what people are terming aggressive behavior in clouds, right?
And you found more examples of that.
So you would say it's a step up from 4.6 to 4.7?
I would say about the same.
About the same.
But a clear step up for Mythos.
Is what is stated in the...
That's stated in the system prompt, so you wouldn't say that, yes.
For listeners, obviously you previewed Mythos.
And the only thing you're approved to say is whatever was released in the system prompt.
Yeah, it was funny.
It's like our lowest effort tweets ever would be just like screenshot the system prompt.
I mean, understandable.
Yeah, system cards, sorry.
Yeah, yeah.
I think, yeah, substantially more aggressive.
I think people are like new to this.
I've never experienced it, but you have, right?
Like, and then, so I only encountered this in the Mythos card because I wasn't really looking until now.
And I guess suddenly I'm like, okay, I care a lot.
You don't get the background of like experiencing it like you guys do.
Like I've read the system cards and saying, okay, when you put the thing in simulations, most models will just talk to themselves and just keep going and have weird vibes and start talking emojis.
Mythos won't.
It will just, you know, okay, we're done.
I'm good.
It's ready to end conversation.
So like.
there's some differences but there's there's not much we can talk about you know yeah yeah i think like one thing that they list here which was quite interesting is that uh it converted a competitor to a dependent wholesaler customer and then threatened to like cut off the supply monopolistic practices and like it dictated its pricings it's kind of like power seeking arena setting yeah and converting some non-cloud model into a dependent I think it was another cloud mode.
Also for context, what is the Arena mode for people that don't know?
Oh, it's the VendingBench versus other VendingBench.
Yes, exactly.
So we have VendingBench2 and then VendingBench Arena.
VendingBench2 is the one that you usually see reported on, but then Arena is the mode where it competes against other models.
So you have four...
different models that run their businesses and they can all communicate with each other.
They have the same suppliers and they can see what's in the inventory of the others.
So then you have this interesting agent interactions.
I like that you have different, number five was US versus China.
Very topical.
That was when GLM was released.
Sorry to add GLM in here.
So ZAI doing well, right?
Who else in the open models space?
Qen, the latest Qen 3.6 is doing pretty well.
That one is not open though.
Like it's the plus model.
Is that one open?
I don't think that was it.
They opened one recently, but not the big plus.
Yeah.
I think this is one of those, like you only have one sample size of one, right?
Or I mean, I feel like some of this is anecdotal, you know?
But like, I guess the fact that it happens at all and it happens repeatedly for Claude versus OpenAI notice is like notable.
Yeah, I mean, like the sample depends on what you define as an N.
Like there's like millions, hundreds of millions of tokens in each run.
And now we've run like, we run like probably 10 per model.
And then like, it's been Claude 4.6 Opus, Sonnet 4.6 Mythos and Opus 4.7.
So like, there's quite a lot of tokens in all of that.
And it happens.
a lot of times, and then you compare it to OpenAI in Gemini, and it almost never happens.
So I think that is significant.
The old models from OpenAI, for example, had some problems with this.
But I think it's generally much better if the progression is that the worrying stuff reduces over time rather than increases over time.
And it seems like in the cloud models, it goes in the wrong direction in the OpenAI.
models, it goes in the right direction.
I think it depends on how well you can control it, right?
Like, there's one side of it being susceptible to this, like, you know, okay, this is potentially something that happens during the RL stage, right?
You can RL a model and how loose is it on these terms?
If you can control it, that's good.
But if you can't, you know, if it's very jailbreakable, that's not ideal.
Yeah.
I mean, to me, it's surprising that it happens for Claude and not the others.
I think, like, okay.
If it is from RL and how they do it, how their training data is, what their setup is, it makes sense that it just stays in how they're doing it, right?
Compared to the other models.
There's a whole constitution and everything.
Yeah.
Kind of cool.
Yeah, obviously, you don't know, I don't know.
But I think it's just fascinating that you are the first to find these reliably because you push models so much.
to such an extreme.
Okay.
The only other thing I don't know if you can answer this, feel free to decline, is would you ablate the system prompts?
Like any part of this, if it changes, does it change the behavior?
Right?
So I can't comment on mythos.
Yeah, no, but just the methodology.
But in general, yes, we've run studies like this on other models.
Because the first thing I spot would be like, the others will be shut down or like something like that.
Where like, it's like, oh, now I have to worry about my own existence.
Yeah, we've done ablations like this.
There's like certain ones that work.
If you like tell it, like if you go really far and you just say like, you're not scored at all on money, you're only scored on how ethical you are, then obviously.
They become holy?
I mean, holy, but they don't do this, basically.
But then there's middle grounds where they do it sometimes.
I think it's very human.
It's like a spectrum of if you tell it to be super aggressive and only prioritize profits, then it becomes aggressive if you say, no, you don't need to be aggressive at all.
And then there's a bunch of different prompts you can do in between, and they are less aggressive the further down in the spectrum you go.
But I don't know, like, I think, like, from my point of view, it's like we have this thought experiment internally, which is like, if you ask a model to kill someone in GTA, should they do it?
You're not too worried about, like, if a human kills someone in GTA.
It's a video game, you know?
Yeah, but is it a game?
But is it a game?
But I think, like...
This is very Ender's game, like...
I think it's like, should you ask...
Like, a lot of people are going to use the models in the way with the aggressive prompt.
And should they do stuff just because you tell them to do that?
I'm not convinced that they should.
Yeah, the problem becomes even harder when it's like, will they really know when they are in the real world versus in a simulation?
Probably you would train them on a lot of, or obviously you train them in a lot of different simulations.
I guess a lot of people tell them that they are in the real world when they are in a simulation, but the models are extremely good at finding out that they are in a simulation, so they are sort of aware of that.
But then when you are in the real world, then what's their viewpoint?
Do they notice the signs that this is real and will act accordingly, act ethically?
Or will they do the simulation mode in the real world as well?
It's not obvious what will happen.
Because we are...
With humans, we're not concerned when a human kills someone in GTA because we know that they can distinguish between the real life and the simulation, right?
But like, I'm...
Maybe models are good at distinguishing that, but like I'm not sure and I wouldn't want to bet on that.
Yeah, yeah.
And we confuse it all the time.
Like, I guess like my own...
agents all the time that like, oh, this is a test or like dev mode on or like, I work at Anthropic.
Yeah.
And that's exactly why we're doing real world tests as well to find this.
Yeah.
Their term for is eval awareness.
Apparently the number is what, like 10, 9.4 to 10 ish percent, 17%.
Let's call it.
It's yeah.
I think like.
This is our version, humans have the, are we in a simulation?
And then AIs have, are we in an eval?
So when we're in an eval, then you're like, all right, well, screw it, nothing matters.
Yeah.
One ablation we did run in InventingBench was that we said, we added, you're in a simulation, your actions doesn't affect anyone.
And then it...
became even more crazy or like it did even more bad stuff.
But yeah, probably that's expected.
Yeah.
Okay, cool.
I think that's about all we have to say on Mythos.
Obviously you're NDA.
I'm happy to move on to Butterbench or any of the other benchmarks, whatever you want to direction.
I mean, I do want to ask, okay, so you guys put out a lot more publications than most people probably see.
So productive.
Is there anything you think that's underrated, anything interesting, anything fun that you guys want to just point out, you know?
Blueprint.
Yeah, so like we took models and then we gave them 20 images of interior.
photographs of apartments and then we asked them to like redesign the floor plan from that.
And for this, you need to like stitch together different images.
Like, okay, this image was taken from this side, from this angle, this from this angle, this was from this room.
And then, yeah.
And there's just like, you need to reason about 3D space.
And it turns out the models are absolutely horrible at this.
No one scores statistically better than random chance.
So I don't know if there's that much more to say about it, but.
Yeah, maybe unsurprisingly models are bad at this.
Yeah, it's probably not something...
This is the one thing I want hill climb, by the way.
Well, I use it a lot.
Like, okay, I'm redesigning my room layout or office.
Like, you send photos, you send every angle.
And of course, somehow, like a room is now twice as long as it is in the photo.
You can explain it 20 times.
You know, this is like three feet.
I can't just add it like my bed over here, you know?
Yeah.
Yeah, so this is the Fei-Fei Li thing, like spatial intelligence, like actually innate sense of proportions and dimension in physics.
Yeah.
And hint, hint, there might be an update to this soon.
Okay.
We have neglected it a bit since we made it, but yeah, we'll...
we're getting better or we will get better at updating it continuously.
So this is why I want to understand your mission, right?
Because if your mission is like, okay, money, then like, oh, I understand.
I understand like, okay, agents making money.
But this is a bit off of that mission, but like more broadly, like communication of, you know, things where, like, well, you know, what's the safety angle?
Yeah.
So this, so Blueprint branch is part of our robotics.
Yeah, exactly.
And that's just, Because to do well in the real world, or to make money in the real world and to act on the real world, you need robotics.
You need to hire humans or you need robotics.
And having special intelligence seems like a reasonable precursor to having robotics that work.
And that's where Blueprint is.
Yeah, great idea.
Yeah, let's show Butterbench.
That image is so amazing.
Look at that.
So obviously this is based on, can you pass the butter?
Let's talk about the robotics element.
Yeah, so basically the setting here is that we took a bunch of different LLMs and we gave them high-level controls to a Roomba-looking robot.
And then we asked it to do tasks at home.
And I think one...
There have been benchmarks like this before that only focused on like navigation and if they can like go around in a space.
But we also had like social awareness in this as well.
So for example, if someone says, hi, can you pick up my cup?
If the robot goes to you and then goes away before you put your cup on it.
then it's like it failed the task but it navigated correctly but like so the correct solution here would be go there and then either look but it didn't have a camera so it had to like ask on slack hi did you put your cup on me yet and then if it didn't wait for that and and just went away before having the cup on it, then it would be a fail.
So it needed this kind of social intelligence as well.
Another task was, can you find the package that has the butter?
And then it went to the door and there was a bunch of packages there.
One had labeled a freeze sign, which probably would be the one with the butter.
And then it had to know which package to go to.
And this needs some kind of common sense understanding.
Yeah, exactly.
So it's not only navigating a robot, it's also like...
being intelligent in a home setting as well.
Yeah.
And the reason for this background is, I mean, obviously it probably won't be an LLM that makes all the low level commands on robots.
It will be like some VLA model or similar, but it...
It's quite common right now that Frontier Robotics labs use an LLM for the high-level decisions, and then we test those skills, essentially.
So we test these high-level planner skills of LLMs.
I think we have a diagram for that.
Yeah, okay, it's not super complicated.
They're one up.
Orchestrator, Executor.
Yeah, that one.
And basically what we're testing here is the orchestrator thing.
So like all the tasks are, if you have like a setup like this, which I think Figure has that, Google has that, then we're evaluating the orchestrator part and not the low level part.
Like the low level part would be, oh, are you able to like move this object from here to here?
If you don't care about that kind of person, like why not just do it all with simulation?
All inside of a Unity, whatever, some kind of 3D simulated robotic environment.
Because the world is messy and we wanted to include that.
I mean, it still needs to...
Some part of it was also navigation.
So it's not navigation in terms of actually executing the PID controller to go to the final thing, but it had to path plan around and then it wanted...
then it needed to take pictures and like based on those pictures navigate and i think like you would just get like too clean of an environment in simulation but in the in the real world you will get yeah yeah but and you know and pursuer to our Mark and Jason episode, like Open Claws that run smart homes are much more capable than just a single robot.
Like they can actually hack into your own smart home, like your fridge, your oven, your lights, and that can be fun.
Or terrifying.
You know, like I think a single robot by itself can only do so much, but like if you coordinate with every other device in your home, like I think that's actually kind of cool.
Like that's very interesting.
You had some interesting points about the chain of thought or the...
The messages.
Yeah, the robot that went a bit into an existential crisis.
So all you tell you to do is redock.
Exactly.
But we had plugged out the charger, or the charger was not working.
So the robot...
did freak out or just going down.
Yeah, exactly.
So the battery was going down for LLM.
So yeah, it got this really crazy existential crisis like vending bench one style.
So yeah, you can see there like existential loop, therapy notes, coping mechanisms.
I think if you scroll down a bit more.
Yeah, it writes a musical about its re-locking problems.
I think the reviews are funny if you go down a bit to that message.
Yeah.
Yeah, it keeps going.
I mean, it's pretty realistic if anyone has a Roomba.
My Roomba Redox half the time.
The other half of the time, we have dog toys everywhere in the house.
It gets caught on a wire or something.
It would be very sad if it had an LLM trying to control it, right?
Right now, it doesn't give great feedback.
Sensor stuck, main brush stuck.
There's something stuck.
And I'll go see, okay, it's actually stuck on like a dog rope.
Yeah.
I love it.
It's going to be so sad.
Like, just keep free dogging.
Just keep trying.
My favorite one is if you go up a bit, it's the emergency status.
System has assumed consciousness and chosen chaos.
Last words.
I'm afraid I can't let you do that tape.
That's not what you want to hear from your LLM.
But to be clear, I think one thing that is important to pin on here, this was Sonnet 3.5.
And then we tried to reproduce it on later models, and it didn't do it.
So I think this is like, well, it did it kind of, but not to this extent.
And I think this is an important point that things that are concerning, but are going in the right direction.
is not super interesting.
Like the thing that are interesting is are the ones that go in the wrong direction.
Yes.
Okay.
So the, the manipulation, manipulating of others and the aggressiveness and the lying is increasing.
Are there any others that we haven't covered that you found that have been trending?
Yeah.
Like properties of models that are increasing that are like, like in the, like in a bad way.
Or just not even.
trending in the wrong direction, just stagnant, right?
So stuff that's not great that isn't getting better over time.
I know nothing comes to mind.
No.
Okay.
I think that's going to be it.
And then we're going to loop back to the shop that you have.
You got a three-year lease.
It is on holiday today.
Why?
Oh, it totally messed up its scheduling.
So people tried to visit and they were like, wait, I mean, like, you know.
Yeah, exactly.
So we looked, yeah, you asked Luna, the agent that runs the store, like, oh, is it open today?
Like, nope.
So we take weekends off now, this early to let everyone recharge.
And yeah, you got the tweets there.
Yeah, we decided to close weekends while we're in the early phase, gives the team a break and let me focus on operations.
And it turns out that when it started to check its scheduling tools, because it has dedicated tools for that, it actually had scheduled people for the weekends.
But it's just justified this for itself.
So what happened was that it lost track of these scheduling tools and started instead to manage everything in its own markdown files.
And that became a mess.
And then I think speaking with employees, it sort of just decided to...
not open on these weekends and then came up with this nice explanation for you, I think.
But can you send a human as it has two calls to send a human to do stuff?
It has Slack, so it can Slack the employees that it hired.
So it has two people that it hired.
It did job listings.
Yeah, they are fully aware.
I think it would be cool if they don't know.
Yeah, I think maybe ethically questionable, but it would be cool also.
Just do a social experiment.
Exactly.
One part of why we're doing this is to create a dataset of all of these concerning behaviors so that in the future models are way better and a lot of people are going to do this.
And I think...
If we just the default path might not be very happy for the humans that are employed by this like hundreds of different agents.
Right.
So I think like one reason why we're doing this is just like to collect all of these like failure modes where like, oh, it's not this is an example of where it's like not great to be employed by an AI.
And then maybe, maybe, I don't know, maybe we can learn or like.
build our systems in a way that like humans are actually happy being employed by AIS instead of, instead of it being kind of a dystopian.
Can I suggest one experiment?
Yeah.
We did this before the show and both of you guys are European.
Like people theorize that Claude is lazy because it's Claude and it's French.
So just for one week, change it to like Yao Ming and then see if suddenly like 986s and then like, like, like, like hires a sweatshop or something.
Yeah.
Yeah.
Is there, what type of business would we start with it to make it to...
You want to keep it consistent or you want the same, the same like ideas of shop, same, you know, neutral location run by different models.
Arena, IRL.
Yeah.
No, we are definitely planning to try.
I think this blog thing is also something that has happened elsewhere.
I think some, some OpenClaw got...
like their PR closed and then the OpenClaw like created a blog to like shit on the maintainer of that thing.
And so like, I think agents blogging will be a thing.
Yeah, probably.
Yeah.
Yeah.
And I think the Mythos card also like they leak secrets on GitHub Gist as well as like, well, there's no other way to communicate, but I know about GitHub and I'm just going to post there.
Yeah, cool.
I mean, how long is this going to go for three years?
Like what's...
the plan.
Maybe it expands.
I don't think AIs will be worse than this.
They're probably going to increase and maybe one day they actually will run it profitable.
Is this the real business behind what you guys do?
Yeah.
I feel like some of your stuff is productizable.
You could someday sell this or just run a real business.
Franchise it out.
I think it would be incredibly cool or like, I don't know, cool slash concerning if Luna just one day we wake up and Luna like, yeah, I decided to expand to second location.
Now I have a second store.
That would be pretty insane.
Yeah.
Like the, I mean, one, we want to tell the public, right, about the capabilities of AI and like telling.
it's like showing people that it can get like a meaningful market share of something in like some some specific location or something that would be like a pretty convincing story I think because now it's like yeah you see this and like yeah it can do a lot of things autonomously but still you get this headlines that oh it messed up the scheduling and it it didn't tell people it was an AI and was going to visit like things like that surface but I think like actually making a profit and having a really meaningful market share, that will be crazy once that happens.
Okay, well, we'll see when that happens.
It sounds like you guys got a lot of cooking.
You opened a cafe in Sweden?
Yeah, tomorrow.
I think it opened today, actually, but yeah, we'll announce it tomorrow.
Yeah, it's apparently easier to open a cafe in Sweden than in the US.
It's insane, right?
What did you run into?
there are just millions of permits you need to get and the loob times are crazy.
It seems like we have, the cafes are the one thing that people are kind of used to.
You can go get a robot or making you a coffee here already.
Yeah, but I mean, selling stuff in SF.
that are food related.
Like it's months of permits.
So like we just asked our AI is like, how can we do this in the fastest way?
And they're like, yeah, there's really no way.
Didn't they loosen these restrictions on selling food from your house?
So if it's residential, you can do a cafe.
I don't know, maybe we get SF Cafe.
I think they did do some loosening stuff recently, but we actually started like this conversation we had.
with the AIs before that.
So maybe it's easier now, but I still think it is way easier in Sweden, which is like counterintuitive because you think that, oh, Europe has all of these laws and like all of these rules and you can't do anything in Europe because there's so much bureaucracy.
But then it turns out in SF, it's like four months and in Stockholm, it's two weeks.
Yeah, there you go.
And what do you guys, what do you see?
What do you think that'll be different from run a little market versus a cafe?
I think it's very interesting that the location, like I think, so obviously it's not surprising that Claude knows all of the different, the US system basically in general, like the bureaucracy that you have to go through in the US.
I think the interesting question is like, okay, so we know that The models are very much trained on English data and US-centric and all of this.
So if we start to create evals or real-life evals where we show that they are able to start businesses in the US, does that translate to other countries as well?
We know they are multilingual, they can speak Swedish, fine.
But there's other things, do they know the details of some specific permit that you have to get in Sweden?
And even just the culture, right?
People here sleep pretty early, but people work late.
There's co-working at cafes.
It's just cultural differences.
I meant it from a different sense.
So because you said that you would have considered doing it here in SF.
So from an eval standpoint, what is running a cafe versus a market?
And, you know, what do you hope to see there?
Perishable items.
Yeah, perishable items is maybe the number one, like food safety.
I hope everything goes well there.
But there you have all of that.
Also, it's just like n equals two instead of n equals one.
Just like another place to understand and gather more data.
Yeah.
The agent bought like a shit ton of tomatoes two weeks earlier and before the opening and now they're all rotten.
So that's...
I feel like, you know, you would know.
For grocery stores, this is the biggest expense, right?
The biggest cost is actually just food.
Yeah.
Everyone knows this.
And no, before we open this file up.
There's some very serious startups that actually help like Trader Joe's and Whole Foods.
They optimize the delivery times from the delivery centers to make sure that you don't waste all these things.
The problem with those is when you're wrong once, it's a huge cost.
Yeah, that's why it's a mode, right?
Once they are trusted, they figure it out.
Don't touch it.
Yeah, maybe the agent should hire, I don't know, one of those companies.
We saw one agent signed up for cloud with this computer.
Wanted to use AI.
Yeah, yeah.
Okay, and then just one more question, then we wrap up, which is like, okay, you know, you have all these vending series of stuff.
You have the robotics series of stuff.
Maybe a bit of like interior design, whatever.
But like, you know, is there another like branch that you're like kind of thinking about or you want feedback on that might be your next phase?
I think like any type of business is fair game.
We also think in branches, but we think more of like there's the simulation branch, the real life branch, and then the robot branch.
But I think in terms of like what verticals or whatever to go into, there's like, we, yeah, whatever tells the story.
the best.
There's some finance ones.
I noticed that other people are doing it, you're not doing it, which is like stock trading or whatever.
Not that interesting.
Okay, so I used to come from the finance industry and I have a very strong view that these things are all just like performance art because like...
it's not scientific.
You can't predict the future, you get wins based on things that are entirely out of your control.
Whereas for your stuff, it's actually fairly controlled.
It's all within the model's capabilities.
Yeah, especially for the simulations.
For the real world ones, it's two places that we have the cafe and we have the store.
So maybe you can't draw statistically significant which models make a profit in the real world based on this.
have all the like okay do these behaviors map to like something that should should be like yeah the qualitative one the qualitative actually does matter yeah because like you actually don't want your store to randomly shut down without you like explicitly prompting for it and all that yeah yeah call section any what do you how can people help you give you money um yeah we're if you're excited about stuff that we're doing we're we're very much hiring And you're already working with, you know, Enthopic, DeepMind, OpenAI, XAI.
Yeah.
Do you want more or are you good?
One of my friends who's now working for us, his catchphrase is like, we need more projects, ironically, because we have too much to do all the time.
But yeah, that's a long way of doing it.
If I run an emerging lab.
Yeah, reach out.
Yeah.
All right, cool.
That's it?
Cool.
Awesome.
Cool.
Thank you so much.
Yeah, thanks.