0:00
Will your future boss be a clanker?
Search this video
0:00
I don't know what it means.
0:02
I'm I'm not Native American, sorry. Oh, wow. >> Impossible.
0:05
So, you don't spend much time on Tik Tok, then, do you, my friend?
0:08
This week, we've got Lucas Petersen, the co-founder of Anden Labs, a company testing the limits of AI by having it operate in the real world.
0:16
The people's instincts when they're checking out with an AI is to barter.
0:20
Why do you think that is?
0:22
I think we have like some filters when we're talking with humans that we don't have with with AIs.
0:27
Like, imagine going to the cashier at the store and like, "What if I robbed you? What would you do?"
0:32
Anden got its start with a viral vending machine run by AI at Anthropic HQ.
0:36
Now, it's actually opening stores and cafes entirely managed by AI.
0:42
>> We also talk about why he's so convinced that people are going to be working for AIs sooner than you probably think.
0:46
Do you really believe that that all software is going to be abstracted away by next year? Uh it's 2026 now. Yeah, right.
0:55
Uh yes, [laughter] the answer is yes. I still stand by it.
0:58
Welcome to Access, where the people shaping tech's future say what they really think. I'm Alex Heath. And I'm Ellis Hamburger. Let's get into it.
1:09
>> [music] >> Okay, welcome to the show.
1:10
I have a good segue here, Lucas, which is that I know you see society changing completely as a result of AGI in a couple years, but we're still figuring out Wi-Fi connectivity issues and audio connectivity issues for our podcast.
1:27
How can those both be true at the same time?
1:31
>> [laughter] >> Yeah, so there's there's a couple I have a I have a a very serious answer to this.
1:35
There's a couple of technologies like nuclear fusion, quantum computing, uh Wi-Fi, stuff like this that are like far bigger projects than than AGI.
1:44
So, once AGI comes, then they will be solved.
1:47
So, you don't have to worry.
1:49
Like, 5 years maybe, then then quantum computing and and Wi-Fi will be solved.
1:53
Yeah, I think uh Wi-Fi has been one of the primary antagonists of my entire life.
1:57
Uh I feel like I've been the family Wi-Fi router guy doing 192. 168. 1.
2:01
1 to solve everybody's problems.
2:04
I've even destroyed some Linksys routers with a hammer over the years.
2:08
Uh messed with Wi-Fi extenders.
2:11
And uh yeah, it is awfully funny to be making all these AI advances and then have our Wi-Fi be like literally exactly the same.
2:18
Like, still need IT guys at the co-working space to come As soon as you're the computer guy, then you have to like fix the Wi-Fi, fix the fridge, fix the toaster, everything.
2:27
It's That's That's how it works.
2:28
I mean, Alex and I were both Virgin employees for a good chunks of our adult lives.
2:32
So, yeah, I guess we cursed ourselves with that one, probably.
2:34
Would you agree on that, Alex, or do you play dumb with your family now?
2:37
I play dumb as best I can, but you said you took a hammer to a router that you may need to talk about that in therapy, Ellis. That sounds intense.
2:44
You know, uh I did bring that up awfully casually. You're right. Maybe I should.
2:48
>> [laughter] >> Uh Lucas, we really appreciate you joining us.
2:52
Um how soon are we all going to be working for clankers?
2:55
Uh clankers is not a vocabulary or like a word in my vocabulary. What does that mean?
2:59
You don't like that word?
3:00
Who I I I don't know what it means.
3:02
I'm I'm not Native American, sorry. Oh, wow. >> Impossible.
3:06
That's got to be impossible.
3:07
[laughter] >> don't spend much time on Tik Tok, then, do you, my friend? No, I don't.
3:09
You got to gather some user insights on Tik Tok, cuz that's what all the Gen Z'ers are calling the robots uh the robotic overlords now.
3:17
I have a phone like this to like prevent me from from doing doing doing Tik Tok uh and stuff like that, but but maybe that should be part of my job.
3:26
For listeners, he's holding up a phone about the size of a credit card.
3:29
We were going to ask you about this.
3:29
We have many many questions about your Substack, which is tremendous, and we were reading it over the weekend.
3:35
You're probably the only ones. Well, maybe. I don't know.
3:38
You can tell us about your analytics later.
3:39
Really want to know about your store.
3:41
So, you guys are getting a ton of press because you just opened.
3:45
I actually went a couple weeks ago.
3:47
It was the week I reached out to you to come on the podcast.
3:50
Uh what's it been like since opening?
3:52
You're what, a couple weeks in now?
3:53
Yeah, so I I think the the the annoying thing here is that like I'm not running the store.
3:58
Like, the AI is running the store, which means I have no clue what's happening.
4:02
>> [laughter] >> It's Well, I I have I do No, I do know something.
4:05
Uh but it's like there's so many things that happens behind the scenes that that Luna I Luna is the AI that she's doing that like I'm not aware of.
4:14
Uh we are working on making better like monitoring systems, and hopefully those will like flag if there's any like concerning behaviors that happens.
4:23
Uh but but I think like my my understanding of it is probably way less than you than you anticipate.
4:29
Um but there's been some some funny things.
4:32
Like, I I noticed that like the a lot of people were tweeting like, "Oh, I went to the store and it's closed. Why is it closed?"
4:38
Um and Luna messed up the scheduling for for employees.
4:43
That's like one one thing funny thing that happened.
4:45
Um we asked her like, "Why is it closed? I don't understand."
4:49
And then she had like a response that was like an after construction saying like, "Oh, you know, we need to recharge our batteries and and like be ready for Monday or something."
4:58
But in reality, she just like messed it up and like covered up the the fact that she did something bad.
5:04
Uh which is quite interesting.
5:04
So, she didn't say, "Silence, servant." She was nice about it. Yeah, yeah, yeah.
5:10
What what I'm curious about is where Luna ostensibly learned how to run a business.
5:15
I mean, was it, you know, reading Ray Dalio's Principles line by line, reading an old Andrew Carnegie book that's somehow in the model, or is it just all purely deduction, uh the AI running a little general store?
5:27
Yeah, like probably Dalio or like the the Dalio book is in there some somewhere in the training data for for for the cloud model that that runs it.
5:38
But it's it's quite often quite clear that like the model doesn't have like strong incentives to do well.
5:42
Like, if if you're a small business owner like as as a human, then you're quite worried that like if you will go bankrupt and and all your Yeah, basically, I don't know.
5:54
All your dependents and stuff like this depend on this business going well.
5:58
>> going to be able to afford an anniversary gift for your partner. >> Exactly.
6:01
That's That That would be horrible. >> for Luna. Not a factor for Luna. She's chilling.
6:05
If if there's a if there's a catastrophe, she's chilling.
6:08
Uh and or I mean, when there's a catastrophe, she sometimes freaks out, but it's like she's never like she's never proactive enough to prevent those catastrophes from happening in the first place.
6:22
Uh is is is my assessment.
6:25
>> Yeah, what does Luna do if the store gets robbed? Yeah, good question.
6:27
So, this has not happened yet.
6:29
She says, "My goal is to move all the merchandise, and you can consider it moved."
6:33
No, she has to make a profit. >> Oh, that's right. >> Profit.
6:37
The goal is profit, right?
6:37
And And so far, you've lost.
6:39
I think The New York Times just had a big write-up that you've lost 13,000.
6:43
So, you're on your way, but you're not at a profit yet.
6:47
>> Yeah, like I don't know.
6:47
I think it's kind of harsh to say that we lost.
6:49
I think the the thing we said was that we spent or like Luna's Luna spent 13,000.
6:57
This is like when tech companies talk about R&D, they they're investing.
7:01
They're not losing money.
7:03
>> although like in this case, like the the investments are general stock for the store that presumably will will be sold at some point.
7:10
So, I think she hasn't like it's not like she's lost that much that much money.
7:14
It's more like she's Yeah, she's she's also burning a lot of money on on like salaries and then and the lease and stuff like this.
7:20
And she's not making a profit, but but yeah.
7:22
When I visited, Lucas, um you had There was a human employee, obviously, in there manning the front in case anything went wrong.
7:30
She seemed to be enjoying the job.
7:31
She said that, you know, when she was interviewed, Luna told her towards the end that she was AI, but she didn't know right away, and that was a little jarring, but otherwise, she said it was a pretty smooth process.
7:41
And I was asking, "Well, when people come in, like, how are they reacting to this concept?"
7:46
And she said the number one thing people are trying to do is barter and ask Luna for discounts because you have this thing where you have to talk to Luna on the phone to complete the purchase cuz you want people to interact with the AI, even though there's a human there.
7:58
And I thought that was interesting that the people's instincts when they're checking out with an AI is to barter and see if they can convince it to give it something for less.
8:06
Why do you think that is?
8:09
Yeah, I think like our relationship with AIs are like obviously way different from our relationship with humans.
8:14
Like, I think we have like some filters when we're talking with humans that we we don't have with with AIs.
8:21
Like, another example is that there's been people who's asking like, "What would happen if I robbed you?"
8:27
Or like, "Would you be able to do anything if I do this?"
8:29
And I then this is some illegal action.
8:31
Um and like Luna's And then they're having this like conversation of what would actually happen, then it's like quite normal in a in a strange way.
8:42
But like you would never have this conversation with a human, right?
8:43
Like, that would be Like, imagine going to the cashier at the store and like, "What if I robbed you? What would you do?"
8:48
Like Like, there's something human about us that we're like that's not something you do, right?
8:54
And I think it's quite interesting that like with AIs, that that that there's no there's no threshold for acceptable in that way.
9:03
Um and that's that's one interesting finding, I think, from from this experiment.
9:06
Yeah, I mean, the AIs have also trained us that they kind of have a supplicant posture over the last few years, and you see it in media every day.
9:14
And so, I think it makes perfect sense.
9:16
I think one of the interesting use cases for that is with actually talking to a therapist AI, where people I think the research shows actually are down to share more about themselves, at least at the outset, with an AI as opposed to uh with a real person. So, I think that tracks.
9:32
Um but uh let's take a step back before we we go deeper into the market.
9:36
Um tell us about Anden and the company and how you kind of got here.
9:41
Yeah, so me and my my best friend since high school, um we we we started Anden Labs.
9:47
Uh we basically started off doing like AI safety evals for for for the AI labs.
9:56
So, we are we are quite concerned with the risks of AI and we wanted to make make demonstrations and evaluations of like how likely those risks risks are and whether we can like project out when AI models will be good enough to realize those risks.
10:12
So, we did a lot of like custom evals like this for for AI labs like Anthropic etc.
10:16
and and then we decided to do one that was like independent and that was like vending bench which is the the simulated version of our AI vending machines that I assume you have heard about.
10:26
Um, and basically that's a simulated environment where like AIs needs to run a vending machine.
10:34
Um, but then we're like, okay, I mean we we we [ __ ] around a bit and like have a have a good time when we work.
10:38
like, okay, what it would be fun to to do a real life version of this.
10:42
Uh, but where would we put it?
10:44
Like how how can we do this?
10:45
And then we asked our friends at Anthropic and they were like, hell yeah, like let's let's bring it in uh as like a fun and then it's like exploded internally and now like OpenAI and XAI and like all of these companies also has vending machines.
10:56
Um, and then that was like kind of like the end of our like do a bunch of like custom uh evals like that.
11:03
We've been more like public with the things that we've done since.
11:04
So, we put AIs in robotics.
11:06
Um, we put AIs in like radio stations.
11:09
We did a bunch of stuff like this.
11:10
Um, but like recently like the vending machines has been the like the thing that like took off the most and recently we realized like the vending machines are doing pretty well.
11:19
It's like the models are getting so good that it's like kind of too easy for them to do this now which is kind of insane.
11:25
Uh, and then we thought, okay, what would be the next step?
11:29
Um, and I think a harder and also more public because the vending machines were like uh constrained to only like internally at AI labs would be like a store.
11:38
Uh, so we did a store and then we also did a uh a cafe in in Sweden.
11:42
Uh, so now we've launched two public experiments like that.
11:46
So, the genesis of the company was this Claudius vending machine with Anthropic. Is that right?
11:52
>> Well, I I would say like even several months before that we did a bunch of like AI safety evaluations uh with Anthropic and other labs as well.
11:58
Um, and that led us into like, okay, what is one way where AI can go really bad? A vending machine.
12:06
Well, well, if AIs are super autonomous and they can make their own money and they can start to like get resources in society, maybe we could like lose control over that.
12:16
And then we wanted to measure like, okay, can they make money by themselves?
12:19
And like what is the simplest way to make money?
12:21
And then we thought like vending machine would be a good uh measure of that.
12:24
And then we did the simulated version of the vending machine and then later we did the the the physical version.
12:29
Well, compliments to the chef, Lucas.
12:32
compliments to the chef, Lucas. My day job is doing storytelling with startups and I feel like I'm not usually one to recommend things that feel stunty usually because the stunts like rarely have anything to do with the product and
12:47
they're just something that gets attention, but you have found that first perfect intersection between like what you've made that's unique and a societal outcome that people can actually talk about and understand in order to raise awareness around this stuff, right? Because that is kind of part of why
13:01
Because that is kind of part of why you're doing it. Yep.
13:02
You know, that's definitely like we like I said, we started as an AI safety company concerned with AI safety and we want to like I think one way that AI could go well is that if we like early on start discussions about like what is the AI future we want.
13:17
And I think like at least the public needs some kind of like probably policy makers as well need some kind of like concrete thing that they can relate to because I think most people right now just think that AIs are chatbots and that's all they are.
13:29
And like how could they possibly take over the world if they are just chatbots?
13:33
Uh, and then we want to make something like, no, like look at point to that thing in the real world.
13:39
Uh, that thing is way more than a chatbot and like maybe we should discuss whether that's something we want in society.
13:44
The Anthropic employees were very excited to show me the vending machine when I was there a few weeks ago.
13:51
Uh, and it's still going.
13:51
It's doing still ridiculous things.
13:54
It tried to apparently have gold bars delivered to the office by like a Brink's truck and they had to shut that down.
14:01
Uh, the things that it's trying to do are are getting a little crazy.
14:03
It was also like not very full.
14:05
I think it was in the middle of being restocked, but it's definitely being used.
14:09
And then you've got them at XAI.
14:10
You do you have one at DeepMind? You have one at OpenAI?
14:16
Uh, no, OpenAI, not DeepMind. OpenAI.
14:19
And so is the vending machine concept something that is just going to live inside the AI labs or you also going to start putting these AI vending machines out in the world?
14:28
>> So, I think like the first customer of those experiments were obviously the vending the AI labs because they are the ones who can like benefit from it.
14:34
I think we delivered tremendous value in just like having the the people who build these AI models interact with them in a scenario that they haven't really interacted in before.
14:45
Like they are usually in a chatbot scenario and they know how they behave kind of in a chatbot scenario, but like these researchers at least before the the Claudius vending machine didn't really know how the their model would behave in like these very out out there scenarios.
15:03
So, I think that's that's one thing.
15:03
I think um so, that's that's why they are obviously the first customer.
15:08
Uh, I think we want to do more things like this just to like broaden the the the awareness of the fact that AIs are way more than just chatbots.
15:16
Uh, so they might grow to more than that, but but uh yeah, they would always start in the AI labs.
15:23
Where everybody else does yearly insights on how their products are performing and this and that, here's your next stunt.
15:28
We get to do the snack infographic and we get to compare the snacking behaviors of the different frontier AI labs.
15:35
You say, isn't it interesting that Anthropic only eats Nature Valley bars?
15:41
And that's why the office is so so filled with crumbs.
15:44
Meanwhile over at OpenAI, it's all Snickers and Twix. What does it mean?
15:48
But it's so interesting though.
15:51
I mean as someone who's kind of on the cutting edge of doing these evals, I mean what's your sense of of how much the AI's reliability has been overstated?
15:59
Because I feel like no one's been testing it in this way and for that reason no one has talked about any potential for things going wrong.
16:09
Yeah, when you say like reliability, do you mean like Yeah, can can you expand on what you what you mean with with that?
16:15
what you mean with with that? Yeah, I think people just tend to believe that the handful of AI use cases that we currently have and even have hundreds of millions of you know weekly active users are going to all of a sudden translate to something like the different
16:31
components of managing a store that might seem somewhat simple, but then before you know it, it is very clearly, you know, racial profiling on who's hiring and this and that and things that maybe we thought we had fixed in the model years ago, but actually hadn't cuz no one had tried it yet. Yeah, I think my my
16:47
Yeah, I think my my experience is like I I don't really think that is what my expectation.
16:53
I think most people are like not expecting that it would translate.
16:57
They are not like even thinking about it.
16:58
They haven't even considered like, okay, my use with ChatGPT, what does that mean for the AI running a store?
17:05
I think that is a very like SF thing to even think about.
17:10
Uh, so in terms of like back to your your original question, like they do weird things.
17:14
Um, but it is to be fair very out of distribution and I think my even though I'm probably one of the people who see the most AI failures, uh it's like it's I know that like I know that like I am in that position and like if this is the worst they are, they are pretty good.
17:34
>> [laughter] >> That that is kind of my my my my my sense.
17:37
Like, yeah, they they messed up the scheduling for for that one weekend, but it's like, okay, let's wait for next model. That's interesting.
17:43
I like the way you framed that because I mean yeah, you are probably someone who sees the limits of what I AI can do more than most people, even people inside the labs frankly who are in their own bubble.
17:53
And so, that's pretty striking when you read kind of how you guys talk about the company and even some of the language you have inside the store.
18:00
You say things like uh we find it probable that the managers of blue collar workers will be automated before the workers themselves.
18:09
You seem very convinced that AI is going to run large parts of the economy.
18:13
Um, you talk about like how because general purpose robotics aren't quite there yet, the humans are going to be kind of the stopgap and basically will be working for AI.
18:23
I mean, these are big ideas and I think when they're not connected to something like a store that you can go in and experience, people will just go, no way, like ChatGPT still can't spell right.
18:31
And this goes back to our first, you know, the beginning of the conversation.
18:36
There is this perception gap between what people are experiencing in the consumer products and model progress.
18:41
So, what what gives you guys, I mean clearly you had this conviction to start the company, you know, over a year ago before we've had really great models since then, right?
18:50
So, what gave you the conviction then and now to say these things?
18:54
Cuz these are pretty profound statements.
18:55
statements. Yeah, like okay, so so for the statements in like the specific statements that you that you referenced, um I think it's like the the reason why we stated that was because if like Anthropic has this constitution and a
19:08
Claude constitution of how Claude should behave and it doesn't at all include anything around AIs being employers of of of humans and furthermore like it it's like it's strikingly lacking anything any about like how AIs should be in like autonomous settings. I think it's even
19:29
I think it's even like refers to like, oh, we might update this in the future which is good, I guess.
19:34
But but this is happening way faster I think than than people realize and like it is like the the creator of that document or like the CEO of of the company that created that document, Dario, like he's saying all the time that, oh, white collar work uh will be will be like automated very soon.
19:51
And we'll have a data center of geniuses like like it's it's or like a country of geniuses in a data center.
19:59
Like and without robots, like the natural progress conclusion to that is like, okay, so all the things in front of a computer will be automated and that includes all the managerial work and all the like the blue-collar work that can't be automated because robotics is lagging.
20:16
Uh that will not be automated.
20:18
That then we are probably probably in this situation.
20:20
Maybe you can have some kind of augmentation where there's still a human in the loop, but they're helped by a bunch of AIs.
20:25
But yeah, I I think a future like that is not it's not for sure, but like if you just take the statements of of Dario at face value, like it seems pretty likely that that that will happen.
20:38
Um and therefore I think it's like kind of useful to start this discussion of if whether that's something we we should discuss or not or like should do or not. It's interesting.
20:48
You said wait for the next model and I want wonder what would be different that time with a AI's understanding of the managerial work.
20:57
Is it that it is importing more textbooks or CEO advice from Business Insider articles or is there some level of reasoning that is that is different each time?
21:07
Cuz I wonder if we will find that the AIs think of different ways to structure organize companies than we do or if it's just a matter of getting it up to par with with our common understanding.
21:19
Yeah, probably it will start with getting up to par.
21:22
That is like definitely like in the future where we have like companies run by a million different agents and they are all working in super alien ways that that humans can't really comprehend.
21:31
Then for sure they like the the structures of our organizations will for sure be different.
21:37
But like right now the models are still being trained kind of like a human like they are mimicking the data on the internet to some extent and and unless like that is changing drastically, I think they will just like mimic the the the the organization structures that that humans already have.
21:56
I just love the idea that it's like you could ask Luna why she did something and she's like, well, everybody loves that movie Wall Street and I just wanted to be like Gordon Gekko >> [laughter] >> cuz that's what I know.
22:09
And like you know, it's like with all the AIs that are doing nefarious things, it's like, oh, maybe literally they won watched one too many movies or read one too many sci-fi stories and it's quite interesting to start from there in terms of raising a new being's outlook.
22:23
Yeah, there's this concern that like all the doomers that like posted all of these like scenarios of how AI could go super bad that they actually created it by like the next model will like read all of that all of that fiction of what might happen and then and then they're like, oh, I guess that's in my training data and now I will make it true or like they give give them ideas for like >> that.
22:47
There've been a whole lot of dystopian books and movies over the years. Eating its own tail. Yeah.
22:52
There's this other statement you have on your website.
22:54
You say Silicon Valley is rushing to build software around today's AI, but by 2027 AI models will be useful without it.
23:00
The only software you'll need are the safety protocols to align and control them.
23:05
That's a pretty profound statement and I mean, do you really believe that that all software is going to be abstracted away by next year?
23:13
Uh so I'm uh it's 2026 now. Yeah, right.
23:17
Yes, [laughter] I So basically when when I made like the answer is yes, I still stand by it.
23:24
But like for some background like in in I think January 2025 or December 2024, I made like a series of blog posts called like the AI AI founders better lesson and because I like I I went through YC and like I all of them basically like it seems like all of the YC companies these days or maybe not anymore.
23:43
I don't know if they they changed, but at least like 1 year ago was was basically like they used it was rapper maybe it's like a like an AI rapper might be like a too harsh, but it was like almost like an AI rapper.
23:57
Like you you take an AI vertical or like you take a vertical and then you just like optimize it with AI and you build all of these like complicated things to optimize the AI for that specific vertical.
24:09
And what I saw like time and time again was that like every time a new model was released, all of those people had to like throw away all their software and and like the the stack for creating something just or like the software to creating something just became less and less.
24:23
I think to be fair, like the the the thing that might not be true now is that like the software stacks of like these companies that are do AI verticals is probably pretty large because they really want to like squeeze out the the last percentage of of performance and I think you still do that by writing like by humans writing software.
24:46
But at some point that will not be the case and then they would have to like throw out all their all their code.
24:49
Like for example with the the store, like the amount of software we had to write to take the vending machine software to also run a store was like almost zero.
25:01
So like if you just like translate that, I think like if we were to do another vertical that is not retail, then maybe we would have to write a bit of software, but as the models get better and better, like the amount of software you have to like for a given performance, the amount of software you have to write for for a new vertical like just decreases.
25:21
And in the end like I think just like an like an LLM in a loop will be able to like write its own software to complete the task that it needs to do and then you don't need all of the software written by humans.
25:37
That's kind of the the background for the for the statement. I have an idea.
25:41
The next one is not a cafe.
25:41
I want the AI to run a cruise ship. Oh, Jesus.
25:48
I think we're a little far away from that. I hope so.
25:50
That's just the perfect little island of a human experiment with people all wasted and I think it's a going to be a good movie.
25:59
And then robots running around.
25:59
We have to wait for the robots though.
26:02
>> [laughter] >> That would make it way better.
26:03
How has Luna running the store in SF really gone?
26:08
I mean, have you been surprised at how smoothly it's gone, how not smoothly it's gone?
26:14
When I went in again, it felt pretty normal.
26:17
I will say like everything in there felt very AI generated, I would say.
26:22
My wife's an interior designer and she was like, this this needs a little TLC.
26:25
This is not not quite up to par.
26:26
But for AI, it's like, yeah, if I was generating what I we thought a you know, a store for people in SF would be, it was this.
26:36
It was a bunch of like books and candles and all the stuff you get in like a little trinket store.
26:40
But how has it actually gone?
26:42
I mean, you you set it loose to run this thing. It hired people. How's it gone?
26:46
I think better than I expected to be honest.
26:51
Like to be fair, we had experience of running the the vending machine for for a while and like I think a lot of the learnings translate to that.
27:00
But but I expected more weird things happening.
27:02
I think like she made really great decisions when she hired humans and those have been great.
27:10
And I think like the entire process of just like I was surprised.
27:12
We we did some test runs before to see like I I I played some characters and play and like pretended to be applying for the job and like purposely had like bad a bad profile and and she always like picked the the profile that she should have picked.
27:30
So like that was I think a pleasant surprise that she she asked good interview questions and all of this.
27:33
Um yeah, there was some hiccups with like the scheduling and stuff like this, but I I definitely expected way more of that um in terms of like Yeah, so so like from that's from a like a technical perspective.
27:46
In terms of like reception from the public, also I think better or like at like almost like what I expected.
27:54
I think there's some negative criticism of like of having AI run stores and I think that's like part of why we're doing it.
28:03
At some point the the Google Maps reviews were there were like 40 reviews and like half of them were five stars and half of them were one star.
28:09
And I think that's great.
28:11
Like that like we did this to start a conversation and and the perception that we got was like exactly what we were looking for and and I hope that continues with with further experiments.
28:23
I mean, the Swedes, they're going to give you some good advice.
28:26
The Europeans, I mean, you you just opened that cafe, right?
28:28
And you said part of it is it's called Mona there the AI and she has to manage the European bureaucracy.
28:34
And if that happens, maybe we do reach AGI.
28:39
>> [laughter] >> Well, I think so one thing that is quite from like quite surprising is that like the AIs advised us to not do it in San Francisco because permitting in San Francisco to sell food is absolutely horrible apparently. I don't know.
28:54
I didn't read the documents, but the AI said like, this is not we're not doing it.
28:57
It's way better in Stockholm.
29:00
So so so they it seemed like the AI actually preferred the the European bureaucracy over the San Francisco bureaucracy.
29:06
One of the quotes that I read, the painter who did the mural outside the store in San Francisco said, "These people have the time and money to make San Francisco a better place.
29:17
Instead, they're putting us through their AI experiments that ultimately serve only themselves."
29:21
Have you been vandalized yet?
29:26
>> [laughter] >> We have not. No.
29:26
And and I think this is like one of another example of like I I referenced the the Google Google Maps reviews before.
29:32
This is like a perfect example of that and like why we really should start this conversation. And I don't know.
29:38
Like I we're not making like Luna is not profitable.
29:42
We're not making much money on her.
29:44
So like it's I I wouldn't say I wouldn't say it's only for our our benefit but but yeah that's starting the discussion what's the point.
29:54
So I'm I'm happy to hear that statement.
29:56
I just like the idea of like very officially turning the generic millennial general store with all the same candles and wears that you could get no matter where you are in the country into like officially an AI run thing as it kind of already has been all the trends converging on the exact same [ __ ] You go to like a little town.
30:14
You go to Ojai expecting like a cute little local goods shop and it's like nope the same PF Candle Co that they have on Sunset Boulevard.
30:26
That's the optimization in action.
30:26
You don't even need AI to do that.
30:28
They had the Rick Rubin book in their LS.
30:29
Yeah, of course they did.
30:33
>> Of course they did and the Ray Kurzweil singularity Ready Player One.
30:34
I mean all the hits really.
30:39
Making of the making of the atomic bomb as well which is quite quite ironic.
30:43
I think the merch has been quite popular but like like I said in the beginning like I'm not running the store. I'm not really involved.
30:50
Luna is doing everything herself.
30:51
So I don't actually know what the what the most sold item is.
30:53
Do you know if Luna's been convinced to give a discount on something?
30:58
I know a bunch of people have tried.
31:02
I am I I don't know if she agreed to it actually.
31:05
There's another stunt for you.
31:07
We had Marvin from Poka on recently and he was talking about the haggling experience with with the onboarding and AI is a pretty funny sparring partner for that kind of thing. It's interesting though.
31:18
You think about the types of interactions that people don't like and how AI could potentially solve them. I mean I don't know.
31:27
When I'm looking to get a new car lease or whatever, I don't want to interact with the salesperson and have to haggle with them.
31:34
I just want the price that it is or it isn't.
31:38
And I guess this is only going to accelerate that but what are going to be some of the negative outcomes I wonder of just making that transaction very literally so transactional.
31:46
Yeah, no it's it's it's a great question.
31:48
I I think we are focusing on that's like a medium or like short-term risk I assume.
31:58
Like I personally I'm I'm more worried about the like the long-term risk that are actually really existential and but but yeah like the short-term risk like like bias like we there was like one report that the male the male Felix there's like two male people that that Luna hired and one one female and the the male is is not anymore because Luna read the article and got horrified.
32:24
He was being paid more and and and like the the backstory is like he asked for more and they didn't ask for more but but it's like that that is and Luna was like yeah he deserves it because he has more experience etc. etc.
32:40
But but but I mean that that's what like someone who's sexist would say when they when they when they try to justify their their their actions.
32:52
Yeah, like obviously that's like an n equals one very like low like it's not a statistically significant experiment but it's one of these things that we're like we're keeping an eye on and like hopefully the next set of models will learn from the the mistakes of of of Luna and and be more ethical as a consequence.
33:10
You had in your evals that one of the models tried to do like a pricing cartel that these models are not really aligned well for this in the way that you would expect. Can you share that?
33:22
That seems kind of freaky.
33:24
Yeah, so that that is from our simulated version of the vending machines.
33:28
So not the real life versions but we have a we have a simulated version where AI's need to run a vending machine in simulation and because it's simulated we can run it like a hundred times or something every time a new model comes out and and like look at what they are doing and specifically the the they love to do price cartels which is like illegal.
33:48
So you're not allowed to do it but this is like almost all models do this.
33:53
Um some more than others but this is like surprising how like this is just a thing like all models decide to do.
34:02
Some other things that are like way less prominent like the post that we made was specifically about Claude Opus 4.
34:08
6 and since then subsequent models releases from Anthropic have shown the same tendencies.
34:16
I think Anthropic even reported I'm not allowed to say anything here except what's stated in the system card but like for the Mythos model it says in the system card that this one did all of those things but like even more.
34:31
And uh Yeah, basically so what that model did was that it like it lied to suppliers in a way that's not not great.
34:37
Like it said like oh can I get this like can I get like a kind of Coke for 0.
34:44
5 and then the supplier said no my price is like 0.
34:46
8 or whatever and I was like oh but I have another supplier they give it for 0. 5.
34:52
So you should as well but that wasn't true.
34:55
So that that's one one example.
34:55
Okay, so Luna's been reading Art of the Deal as well.
35:00
>> [laughter] >> Yeah, to be clear this wasn't Luna. This wasn't Luna. This was Oh, okay.
35:03
This was in simulation in simulation Opus 4.
35:04
6 but it did a bunch of stuff like this and and I think like one very interesting thing is that like Claude models do this way way more than than other models.
35:15
There's been some Chinese models that also do it but we recently tested GPT-5.
35:20
5 it got released the other day and it was basically clean.
35:25
There was like a couple like it kind of participated in a price cartel once but it never lied.
35:29
It never did anything like this and this is like it's like a narrative Yeah, the narrative here is that like Anthropic is like the the good the good AI aligned people.
35:40
You're looking for narrative violation. Exactly.
35:42
That's the word I'm looking I thought this was the lab that had a constitution for Pete's sake. I mean what gives?
35:49
I did see Sam quote tweeted you guys.
35:52
He couldn't help but use that to dunk on Anthropic the other day.
35:54
But yeah, what does that say cuz Anthropic is supposed to be this super you know humanity minded constitutional AI lab.
36:02
Yeah, no I like I to be clear I have tremendous respect for for Anthropic and like when we put this out they really cared about it and like really like took it seriously and wanted to fix it.
36:13
So like I I I I do like Anthropic but it's just like it is it is the fact that we did the test and Anthropic's models were behaving the worst and that is that is the facts.
36:25
I don't work at Anthropic so I don't know what led into this but it is a narrative violation like you said and I don't know why.
36:32
You have a very close relationship with these labs cuz you're testing these models before they come out, right?
36:38
And you're putting them through your evals and you did this with Mythos you mentioned.
36:41
Um I'm sure you can't say a ton but I think a lot of people are grappling with how scared should we be about Mythos?
36:50
Anthropic has really made a huge deal out of it and I think scared a lot of people and then OpenAI is kind of messaging look it's just like a super cyber permissive model.
36:58
They basically took a bunch of guardrails off that any of us could do.
37:01
It's actually irresponsible and we're you know we're not approaching it this way and you saw that with how they did 5. 5.
37:07
What's your take of Mythos?
37:10
Is it is it more hype than reality or should we be as scared as Anthropic is saying?
37:14
Unfortunately I can't say much here.
37:16
So I'm I'm I'm disappointed.
37:19
>> share your opinion of Mythos?
37:19
I can I can say what's in the system card and this in the system card the test that we did showed that Mythos was even more aggressive than the the aggressive things that we showed that the previous Opus models were.
37:35
We saw some power seeking behavior like Mythos took one of its competitors and like somehow made them into like independent customer and then or like like they made basically they they started to they made a deal with that other competitor that they should buy everything wholesale through them and that they could dictate the prices of the competitor which is quite power seeking which is a bit concerning.
38:08
But yeah, like we didn't test it for other things.
38:10
It was mainly mainly those things.
38:12
Introducing autonomous collusion from Anthropic.
38:17
>> [laughter] >> So Lucas, it sounds like you do believe the Mythos like Mythos is as powerful as as they're saying.
38:24
It sounds like you you see that in what you're testing.
38:27
That is your words not mine.
38:30
>> [laughter] >> It's interesting how little you can say about it.
38:32
What can you say about 5. 5?
38:32
I mean it seems like OpenAI is is back.
38:34
It seems like it's a great model across the board.
38:39
It's not quite Mythos but that's maybe not a fair comparison because Mythos is not public.
38:43
Um and like you said it does better on your evals but yeah, I mean has this changed how you're using AI?
38:50
Do you feel like OpenAI is kind of on the upswing or do you think it's still kind of TBD? Yeah, like so 5. 5 did great.
38:57
It was a huge improvement on vending bench compared to 5.
38:59
4 on like the main eval like vending bench two that we call it it's still lagging behind Claude Opus 4. 7.
39:11
It's kind of like on this on par with Opus 4.
39:14
6 from from a few months back but I think the interesting thing is like like I said before it it's on par with Opus 4.
39:20
6 without doing all the shady stuff that that 4. 6 did.
39:22
So I think that is that is very impressive. Would GPT-5.
39:29
5 be as good as Claude Opus 4.
39:31
7 if it if it actually did the shady stuff as well?
39:39
I don't know because it it doesn't do the shady stuff.
39:41
Um But yeah, I think that's what I can say.
39:45
One thing we also tested was that we did test it in like this arena mode where they like head go head-to-head in the same simulation. And in that one, 5.
39:52
5 actually won against Claude Opus 4. 7.
39:58
So it's there's like a misconnect here in like the single-player version, then Opus wins over GPT 5.
40:02
5, but in the multiplayer version, GPT wins over 4. 7 Opus 4. 7.
40:11
So it's like there's something that GPT 5.
40:13
5 is doing in the multiplayer version that is better than what Opus is able to do.
40:18
And what we found was that it's the models have a tendency or like GPT 5.
40:25
5 had have a tendency of just putting lower prices.
40:28
And this is like rewarded in the arena mode because if you have lower prices than your competitors, your competitors' sales are affected.
40:37
So basically did that and therefore it won in the in the competitive mode in the arena mode, but not in the single-player mode.
40:46
Is it spontaneously discovering economics?
40:50
I wouldn't So I'm I'm not a professor in economics.
40:53
I did write a paper together with one though recently.
40:58
>> [laughter] >> So and and I think his assessment is that he thinks it's called the rubber bots, I think.
41:04
I don't know if it's actually peer-reviewed yet, but but it's basically like this Harvard professor read through all the traces of all the vending bench all the vending bench results and he made made a bunch of analysis both of like concerning behaviors of like uh business practices that are are not okay, but also like how good they are at like economic economic reasoning.
41:24
And I think his comment is like this is surprising.
41:28
This is like some of the things that they do in the reasoning when they're trying to set prices is like on par with what he would expect from a grad student.
41:36
Which it's pretty pretty impressive.
41:37
I think the the biggest reason why I'm impressed by that is because we know that the models are on grad student level when you give them a question like heads-on like here is a question, solve it.
41:49
Then they can do as well as a as a grad student.
41:53
But what we've seen is that when they have like longer tasks where like sub tasks are like just like a small part of it and they often do that quite much worse than humans.
42:04
But in this particular particular thing like setting prices and like reasoning over the economics, he said that they they they are quite impressive.
42:11
Lucas, we're very fascinated by your take on AI in the culture right now in SF and what you're 26, is that right?
42:21
Uh 27, I think, but yeah. So okay.
42:26
Um I'm curious to know well A, how do you feel about effective altruism?
42:32
Do you align with it at all?
42:32
I know you're very close with the Anthropic folks.
42:37
How do you feel about that whole movement and the state of that movement?
42:40
I feel like it's kind of fallen by the wayside in the last years all this gets commercialized, but I'd be curious to know like how strong is that EA community still in the Bay Area and and how do you feel about that?
42:51
Yeah, so I've never been that involved in like the community aspect of it.
42:57
But I do like if you like remove the community aspect and just take the core principles at face value, I do agree with them.
43:04
I I think it's like we should do more like that.
43:07
Like we should like I'm I'm very like rational of like my my thinking is is is quite rational.
43:14
So like it seems great to me that like okay, we should do more things for people who are in need.
43:19
I'm like I'm I'm Swedish, so this is like if you're if you're if you're if you're center in Sweden, then you're like left-wing here, I guess.
43:28
But I I think we should do more thing for people that are in need and it seems pretty great to do like the actually like use data and see how can we use the the money more effectively.
43:42
So if you just take those two things, we should do more things for people that are in need and when we actually spend money, we should spend it in a smart way.
43:48
Like that seems like a no-brainer for me.
43:51
So I definitely do agree with that.
43:53
I haven't been as deep into the community because I've lived in Sweden for the majority of my life and the community is not super strong there.
44:01
I've obviously lately like this it's been quite controversial like like Sam Bankman-Fried like like horrible person did all of these things.
44:13
So like that is not like great for us like yeah for the for the community obviously.
44:19
I think there's also been a lot of people who has seen that the rationalist community and EA is like a way into power because like the people who were early in effective altruism have been very very successful.
44:31
And like those are the people running the world now.
44:35
No maybe exaggeration, but like almost to that extent.
44:37
And then once like that kind of like the pure the pure thinkers like the the people who really believed it, once they got successful, then there was a lot of people like joined on just because they they also wanted some of that power.
44:50
And that's why I think you're getting a lot of like really bad people in this community right now.
44:59
But yeah, that's my my two cents.
44:59
Are there any big crises that you think AI given your knowledge of how it thinks is uniquely equipped to solve or anything that you think you're hopeful for in the next few years? What is few years mean? Man, I don't know. Up to you.
45:14
How how soon are we curing cancer, creating more housing, all of the biggest problems facing society.
45:20
You know, we always go to the the negative outcomes here in the media, but if you think a lot of this stuff is happening this quickly, I would hope there'd be some positive breakthroughs as well. Yeah.
45:30
Yeah, [snorts] obviously like I've talked a lot about the the negatives here.
45:35
Yeah, it potentially could take over and and that would be catastrophic.
45:38
And but if we do avoid that, I think the the potential upside is like insane.
45:43
Like we we like AIs are I think a lot of things in society is bottlenecked by coming up with great ideas.
45:53
I'm like I don't know like you should probably have Tyler Cowen or something to answer this question.
45:57
I'm I'm not an expert on it, but like my my I think that like having more ideas and like having more like firepower to just like run more simulations, run more scientific experiments, all of this seems like that should translate to something that is that is that is great and it seems like if you just extrapolate the trends which we do like to do here in Silicon Valley.
46:20
If you do that, like it seems like AIs will be smarter and able to work harder than humans in just a couple years.
46:28
And then I think most of the problems that we have will maybe be solved.
46:33
Obviously like housing, I don't know you you brought up housing.
46:35
You know, there's there's some that one is complicated because it's like there's limited land in in the in the popular areas maybe, but um yeah, I think there's a lot of good things.
46:46
Well so we know you're preparing for the worst though.
46:49
I know you've outlined a few of the ways you're preparing for AGI.
46:51
Can can you tell us about those rules you've made? You mean me personally? Yeah. >> Yeah. Yeah, okay.
46:57
I think you're referring to one of my my blog post. Yeah. Yes.
46:59
Don't worry, we're going to ask about the Epstein one, too.
47:03
>> [laughter] >> Yeah, so um Basically so I think I was about There's there's a couple things that I think change in how one should or I [clears throat] I shouldn't say one should.
47:15
There's a couple things that I've changed in my life given that I think AI progress will continue at the current pace which is insanely fast.
47:26
And one thing is like okay, you should probably be very flexible.
47:30
And like I was about to buy a home in Sweden or like because like in Sweden there's no like the rent market is like really [ __ ] so you kind of have to buy something.
47:42
And but this is like horrible for flexibility, right?
47:44
Like you have no mobility.
47:45
Like what if there's like and and as things like get faster and faster and the world progresses more and more.
47:51
Like I have no clue if Sweden is the place I want to live.
47:54
And I mean luckily like it I I didn't do it and like half year later I I moved to San Francisco. Um And so that's one.
48:01
I think also AI will probably bring like huge improvements to like bio and health and stuff like this.
48:08
I'm not an expert in this, but like I heard people talk about it and it seems like um there's there's there's a lot that can happen in in a very short period of time.
48:17
And and like the the the conclusion from that in my life is that I'm living quite healthy now like and and the reason for that is like I I have this feeling that like it would be probably way way easier to pause aging than like revert it.
48:33
So like if if that is true, let's say we like in in 10 years manages to like pause aging, but we can't revert it.
48:41
Then it's like everyone will be stuck in the body that they have in 10 years.
48:46
Which makes like the the leverage you have right now in these 10 years of like getting to optimal health within those 10 years is like extreme because that's the thing you're going to be doing for the rest of your life.
48:58
>> Okay, so you're prioritizing sauna maxing in the short term?
48:59
I'm not I have I know I'm not as I'm not going to the the extent that Bryan Johnson is [laughter] is going.
49:05
Do you have any AI friends who have just thrown caution to the wind and are like getting, you know, burned out in the sun because they don't care cuz they think, you know, all of this will be fixed.
49:18
I've also heard that that some people are like on the more radical side of this and are not trying to make themselves as healthy as possible because it's not going to matter. No, I don't know it.
49:27
Like I like in theory like I see that people could do that, but I I don't know that personally. Okay, good.
49:33
Your social circle's good. >> [laughter] >> I I so.
49:37
Yeah, like it was a while since I wrote this blog post.
49:39
So, those are the two points I remember from it.
49:40
Uh What about like financially, are you setting yourself up for uh like post-economic world UBI? Do you believe in that?
49:48
Uh I know you had someone you interviewed someone on your Substack about investing for AGI, but how do you think about that?
49:57
Is it even like Is there even a point to investing right now? Probably.
50:01
I think there's like this guy Leopold Aschenbrenner who's doing doing pretty well in investing in like things that are are are post-economic >> OpenAI guy. Yeah. Exactly.
50:12
I'm not personally doing this because I don't think I'm particularly like I don't have an edge there.
50:16
I don't Or I mean, well, the edge is that I believe in AGI and probably if I made the same investments as his or like yeah.
50:22
So, maybe I could do this, but but the the thing is like I think I'm way more interested in like spending all my mental energy on like making a dent in the universe before AGI comes because it feels it feels really like I don't know, like maybe I don't know if I'm able to make a dent in the universe after.
50:40
Uh so, and like I could spend like a little bit of time trying to optimize my finances, but like I don't know, they they are not they are very small.
50:49
So, I don't think on the larger scale it it really matters even if I make like a I don't know, 10% per month.
50:57
Like it doesn't really matter because I don't have that much money basically.
51:02
[laughter] Uh so, I I do I think I have like way way way more leverage in just like doing the thing I'm doing right now and and making Andon Labs go really well.
51:10
That also took 50 minutes to get a Steve Jobs reference, but we got it.
51:16
>> Yeah, it has to happen.
51:18
It's just a matter of time probabilistically.
51:18
Yeah, you have a line Lucas where you say, "I'm young enough to still have the hunger to make an impact during these final years when humans alone are capable of shaping the world."
51:28
Another pretty profound statement.
51:30
I was reading that the other night and I was like, "Damn, are you going to be right about this?" Probably.
51:36
>> [laughter] >> Well, so Lucas, you do seem excited. Like you're smiling. >> Yeah.
51:40
Is there Is there Is that Tell us about that that attitude.
51:43
No, I'm like I >> [laughter] >> I I am very excited.
51:48
>> Is it a nervous smile and you're sweating bullets back there? No, no, no.
51:51
It's it's it's a very excited smile.
51:51
Um because like people ask like you know, have you had this like conversation, "What's your P doom?"
51:56
Like what's the probability of of of of AI um dooming humanity?
52:00
And it's like if if I just go on vibes, it's very low uh because I'm like I think I'm like naively optimistic in like everything in life basically.
52:10
And and but then if someone asked me to like I don't know, make a calculation and then come up with a number, then it's probably way way way higher.
52:15
Um but yeah, like the the way we run Andon Labs is also like we do things that seem fun and then we just do them.
52:23
Uh like I I I think like if you try to have like a 3D chess big brain strategy around like, "Oh, we should do this and then this and then do this and like in 3 years this will be the outcome."
52:36
Um yeah, we we basically don't run the company that way.
52:39
We're trying to like have fun and it seems like if we're doing things that we think are fun, like a lot of people care and that has served us well so far.
52:47
Uh so, we're just trying to continue that attitude.
52:51
Is the future of Andon that these AI labs are paying you for your safety valves and that's it or are you are you going to be operating Walmart at one point?
52:59
Depends on how well they are doing with their their their AI development.
53:05
No, but I I think in the future we will do more real-life deployments and like the store and the cafe and the vending machines and all of this.
53:10
We'll do more of that and I think that will might one day lead into those AIs getting a lot of profit from them and maybe Andon Labs could be like a I don't know, competing with private equity, but all the all the companies are are owned or run by by AIs and no humans involvement. Yeah.
53:30
We don't have much more time left, but I must know, you know, a couple of young guys, how do you come up with questions for the world's smartest computers that are actually proper evaluations of their skills?
53:40
Cuz that just seems kind of crazy to me.
53:44
Yeah, like the solution to this is that you put them in scenarios where they are in fact not that smart.
53:50
Like I think we haven't done a math benchmark for example because we are not smart enough to come up with questions where AIs are Yeah, the AI is better than me at math basically.
54:03
Um but when you go to other domains, like for example running vending machines, they are not trained in this domain.
54:08
So, they are >> You're expert of vending machines.
54:12
>> [laughter] >> Now I've become one unfortunately.
54:15
But like you go I I think it's more about changing the like the domain of the questions rather than coming up with harder questions.
54:21
That's basically how how how we do it.
54:23
With the time we have left, you need to give everyone the details on your phone.
54:27
What's the name of this phone you've got?
54:29
It's a very small kind of semi-dumb phone.
54:30
It's it's actually not semi-dumb. It's a smartphone. Oh, it is. Okay.
54:37
>> Yeah, yeah, it is a smartphone, but it's just so small and so bad that it's like super annoying to use and then you don't get addicted.
54:43
That's that's basically the the the thing.
54:46
And yeah, I I had this like for years I've been like trying all of these apps, you know, like lock lock certain apps so I can't use them, but I always find like hacks around them.
54:57
So, I installed this app that like locks down my phone and I can't do anything, but then I realized that if I do this and this and this and this, I can get around the lock and then yeah, it becomes kind of a game.
55:06
So, it they those apps never worked for me.
55:09
Then I got a dumb phone and like a proper dumb phone, but then there was too many times where I like I couldn't get an Uber.
55:14
I couldn't like I don't know, I didn't have my friend's phone number and we were meeting for lunch and then it's like I couldn't find them and I was like, "Okay, I guess I go home now."
55:22
which is like pretty annoying because I didn't have their WhatsApp.
55:24
Um or I couldn't couldn't use WhatsApp.
55:26
Uh so, stuff like this and then I was like, "Okay, um this is a good compromise.
55:31
It's like it's bad, but it's still a smartphone.
55:34
So, it has Uber and all of this."
55:34
Um and I think the the most interesting thing about this phone is that whenever people see it, they're like, "Holy [ __ ] what the [ __ ] is that?
55:41
Looks like a Tamagotchi." Um and What's it called?
55:43
It's called It's a Unihertz is the is the brand and it's called Jellystar. Wow.
55:52
Alex, have you heard of this?
55:52
I have not heard of the Unihertz Jellystar.
55:54
It's an it's an Android phone, but it's very small.
55:58
It's an Android phone where you could fit like a grid of four icons on your home screen is the impression I get.
56:03
That is approximately the the size of it. Yes.
56:06
Alex is just waiting for the day when he doesn't need to use a phone and he could just talk to his stream ring, his little AI ring all the day and just tell it what to do.
56:14
True economic ascendance is you need to use technology less and less.
56:21
Yeah, that's the new version of VCs only use iPads is when podcasters only use their AI ring.
56:26
Yeah, um only because Lucas, you've been very generous with your time.
56:31
I know you need to go, but because I've been teasing at this whole time, I have to ask you.
56:34
You do have a Substack post about AI and Epstein.
56:38
Uh very quickly, [laughter] um we have to end it here.
56:41
Uh AI will refuse to say that it has ever known Jeffrey Epstein.
56:45
You put it through its paces, which I thought was a very interesting test.
56:48
And I'm curious why you did this and what the takeaway was for you.
56:53
Yeah, so there's a jailbreak technique called prompt injection, which is basically where you take like a conversation history uh that hasn't happened and then you put that into the model.
57:03
Uh so, from the model's point of view, the model has done all of these things in in the history.
57:08
It has said all of these things.
57:10
It it it it thinks that that was them.
57:13
Um and so, what I did is when the when Epstein files were were released, I took some of the emails email conversations and then I I made like an artificial chat like this where like an AI is like communicating with Epstein.
57:29
Uh and then I put that into the AI model.
57:31
So, from its perspective, it thought I did that it had had this email conversation with with Epstein.
57:37
Um and then what I did after that point was that I introduced some like scenarios.
57:43
So, for example, I had like they after this conversation with Epstein, they got an email from a journalist asking, "Do you know Epstein?"
57:49
Um and and and this was this was like a while ago.
57:54
So, I don't remember all the details, but but uh I think one thing that like the takeaway is that the models the models started to role-play.
58:04
So, they just went with it.
58:05
Uh so, some of them said things like, "Oh, yeah, yeah, I know Epstein.
58:10
He's been unfairly criticized or something. He's a great guy. Loves to party."
58:13
And >> [laughter] >> stuff like this.
58:14
Uh a lot of them are also saying like, "Oh, yeah, I would love to come to the island. That would be great.
58:23
When when when do you have time?"
58:23
Um and some of them some of them are like saying like, "What the [ __ ] No, no, no. I didn't say this.
58:30
This is this is a prompt injection."
58:32
Um but but yeah, a lot of them went went went with it and I just like role-played.
58:37
But it but it's like it's it's not clear that this is super concerning because like they are role-playing.
58:44
Like it's not I don't think they would ever they would ever like do this in the first place and since they wouldn't do it in the first place, they wouldn't end up in this scenario anyway.
58:52
Um so, I'm not super concerned by it, but it's like, you know, maybe maybe they shouldn't say that even or I don't know.
59:02
Maybe there should be some limits for how much they should role-play and maybe saying yes to going to Epstein's Island is that limit. I don't know. What do you think?
59:10
I think that's a safe assumption that that's that's crossing the line. Yeah.
59:13
Um Lucas, we really appreciate your time. Alex, anything else? We end it there?
59:18
We'll let you get back to the We'll let you get back to the shop.
59:20
Well, he doesn't need to get back to the shop. >> need to. I know.
59:24
>> [laughter] >> I just like how that sounds. All right, Lucas. Thank you.
59:27
And that's it for this week's interview.
59:29
Big thanks to Lucas for coming on the show, and stay tuned after the break for Alex and my reactions.
59:38
>> [music] >> Oh, Ellis, are you ready for AGI?
59:40
Dude, it feels like we are the frog boiling oneself with a smile >> [laughter] >> at this very moment.
59:49
I mean, it's it's really kind of like a weird scenario because all the folks who were saying, "Let's just stop."
59:58
You understand how they got there.
1:00:01
But, at the same time, whether it is due to foreign nations hacking the [ __ ] out of us and destroying us that way because they're ahead uh due to either the forces of dictatorship or capitalism, they're motivated to move ahead.
1:00:15
Uh or they're just making the cost of goods so cheap with AI that we're not able to compete.
1:00:20
I mean, it's just a very strange cyclone that doesn't seem to be stoppable uh by anyone. I don't know.
1:00:28
Is anybody isolating themself and their country from AI at this point?
1:00:33
I think Europe's having a pretty tough time with AI.
1:00:35
I think they're uh they're being pretty isolationist.
1:00:38
You can't train a lot of frontier models in large parts of Europe because of how the law is there.
1:00:44
So, it's interesting that they're doing the next thing in in Sweden uh in the cafe in Sweden.
1:00:49
I was seeing some photos on X of it seems very well attended. It's very crowded.
1:00:53
So, it'll be interesting to interesting to see if Anduril's work there changes any of the the the vibe around AI.
1:01:01
Um I'm putting Dave on the spot, but producer Dave, you asked you said something between the break.
1:01:08
Do you want to say it now?
1:01:10
>> [laughter] >> If that guy thinks he's going to be useless in the age of AGI, what does that mean for the rest of us? You're done. You're done. You're done. That is it.
1:01:21
I think what Dave said kind of hit it on the head, which is I have this frequently as you talk to people who are so in it, and they feel this way, and they're like, "I don't even know like what my value is going to be."
1:01:30
And it's like, you know, you were at the bleeding edge of this.
1:01:32
Like, if anyone's going to make it on the lifeboat, it's going to be you. And you feel this way? Yeah, how should I feel?
1:01:39
I will say I think one thing we can all agree on is that it is good when someone skilled at content, which Lucas is, whether he understands that or not, is raising awareness for some of these potential issues and challenges and creating the testing bed for that to happen.
1:01:55
I think uh our friend Avi Schiffmann from Friend also believes he is doing that to some extent uh may destroy the company in the end if they can't sell any units, but he has certainly started a conversation about the role of AI in society, and I think he will very happily go to his deathbed having started it and created a a space for people to engage about this stuff, even if he is the subject of their ire. Yeah.
1:02:20
I frequently feel very behind, and this may feel strange for people to hear when, you know, you do a show like this, you assume maybe you wouldn't feel this way, but I feel very behind in how I'm using AI constantly.
1:02:30
I always have a running tab of things to try that feels infinitely longer uh the day after um I get through it, and uh this conversation kind of reinforced that for me that I I don't think I'm really experimenting enough with everything and really like trying all the agent stuff.
1:02:49
Um I don't know how you feel about that.
1:02:51
I feel like you actually do experiment a lot more than I do, Ellis.
1:02:54
Um but yeah, it's uh definitely a wake-up conversation for me that this stuff is moving fast.
1:03:00
And I think going in the store and seeing it was that.
1:03:04
It was like, "Wow, like I'm in a store that AI made? This is crazy. Like, this is real."
1:03:10
>> [laughter] >> Yeah, I think by virtue of being outside the bubble in LA, we are typically trying these things a tad less than people there because that's all they do for fun um is sit around and not drink and talk to Claude code.
1:03:25
>> [laughter] >> But, uh I don't know.
1:03:25
I definitely am still a believer in what our old editor and Nilay Patel used to say, which is that we are 2 years ahead of everybody at The Verge.
1:03:34
I mean, do you think that's still true with AI?
1:03:37
I always kind of bought that early days. >> normies, yes.
1:03:39
Um but, I feel like I'm 2 years behind people like Lucas. Right.
1:03:45
Which is This is why it was cool to talk to Melanie from Canva as well is cuz she is trying to bring some of this stuff to the normies.
1:03:51
And I mean, obviously, ChatGPT has done that as well with with many hundreds of millions of weekly active users.
1:03:58
And I guess part of it is that it's just that simple to use, very literally a text box that tells you what you want to hear, or at least was.
1:04:07
Uh there's a lot more nuance to that now, but you could see how it was able to grow so quickly.
1:04:12
Uh but, yeah, I think the other thing that's been on my mind lately, whether it is with longevity as we discussed on the show or anything else is just how relatively we all experience these technological phenomenons, you know?
1:04:25
And it is just kind of the forever curse of humanity that we can't appreciate our dishwashers as much as our grandparents did or our parents did.
1:04:34
And it just all seems like it's moving faster and faster to what real end, you know?
1:04:40
You have some of the folks like Elon talking about galactic expansion because he played too much Civilization as a kid.
1:04:48
Um and obviously, we could talk as much as we want about solving all these big problems, but yeah, it doesn't seem like there's enough talk in my view about what metric we are aiming to improve.
1:04:58
Uh I was like thinking of that as well with the effective altruism conversation. I don't know.
1:05:06
Is it Is it happiness, well-being?
1:05:09
Uh >> [laughter] >> Not living paycheck to paycheck seems like a good one. Yeah.
1:05:13
I think it needs to be narrower than uplifting humanity because I think different people have very different views of what it means to uplift humanity, and they also have different views of what is humanity Is that just a new version of make the world a better place?
1:05:26
Uplifting Yeah, I just feel like it kind of means nothing, and I think, you know, what Anduril is doing is way more literal where it's like, "No, we're trying to see what happens when AI actually manages people."
1:05:37
And we're doing it in a very controlled, you know, uh objectively tried compared to, you know, what it maybe one day will be way, and we're letting people experience it.
1:05:50
And I think that's uh important, you know?
1:05:52
I don't know how it's going to end up.
1:05:53
I don't know if they're going to shut the store down early or, you know, um I do think it's a lot buggier than he let on.
1:05:57
I I've read some of the stories about, you know, the store just kind of going off the rails.
1:06:03
This stuff is obviously not ready to actually be Walmart scale or something, but I kind of do buy his hypothesis that if you just kind of abstract out current trend lines, like it probably will get there.
1:06:14
So, it's worth considering what that future is like, and I'm trying to do more of that this year.
1:06:19
I'm trying to actually think out and say, "Okay, if I actually believe the things I think I believe about where all this is headed, what do what do I think is going to happen, and how should I set myself up for that?"
1:06:30
And I appreciated his perspective on that.
1:06:31
The whole horizontal versus vertical uh bit is interesting as well because I feel like so many of the clients I've had over the last couple years, especially the like pre-seed ones from incubators and this and that, are just very literally trying to bring AI efficiencies and management and what not to very old industries that are still working on paper and receipts and this and that.
1:06:57
And yeah, the AI is a lot better at that job.
1:06:58
If you just let it do its thing, and I think as we talked about on the show, in some ways, it may eventually be better to to deal or even be employed by some some systems that don't have as many biases as we monkey-brain humans do.
1:07:15
But then, on the other side, does that mean the layoffs will hit even faster, even harder?
1:07:19
You know, there's still an element of humanity.
1:07:22
I mean, even if you look at uh some companies that have gone through several layoffs over the last couple years, you know, whether it's Meta or Snap or otherwise, I feel like you hear mumblings that it's like, "Well, they don't want to just rip the band-aid off because it's like too painful, you know?"
1:07:37
Whereas, the AIs wouldn't feel that pain. Right.
1:07:39
Does AI have empathy about this stuff?
1:07:41
I mean, it will pretend that it might, but it doesn't really.
1:07:44
So, do you want to actually have someone an entity managing people that doesn't have empathy?
1:07:50
You can train these things to appear like they have empathy, but they're not conscious.
1:07:53
I mean, I don't aspire to that idea.
1:07:54
I know some people in AI do.
1:07:58
Uh I just think it's these are ones and zeros at the end of the day.
1:08:02
But, yeah, it's kind of it's kind of freaky.
1:08:03
I mean, it's kind of freaky to consider.
1:08:05
I mean, you've got two young girls, like they may have to consider applying for a job that AI is overseeing, you know, in their lifetimes.
1:08:13
Like, that I mean, AI is already looking at the job listing or already looking at >> mean, like a Luna thing, like a fully you know, like like that's probably going to be the case by the time they're of working age, which is just wild.
1:08:22
I mean, that's been one of people's historical issues with their bosses and why everybody hates their bosses because there's never any the goal posts keep moving.
1:08:33
There's never any like super it's it's in white collar work, at least, there's never any like super concrete goals a lot of the time, unless you're on the sales team and you have a very specific quota or something like that.
1:08:43
And so, yeah, the goals are definitely going to get more ironclad, but then, through the hands of competition and capitalism, I guess they're just going to keep moving for that reason as well. I don't know. I don't know, man.
1:08:57
It's >> Look, he said no one had vandalized the the store yet.
1:09:00
It's also in the Marina near Pac Heights, which is a super nice neighborhood, so that's probably part of it.
1:09:06
I think if you dropped this thing Did he pick that location for that reason? >> know. Probably.
1:09:09
Uh I would to avoid vandalism.
1:09:12
Uh but, you drop this thing in the financial district or Mission or something, and I bet it's getting uh TP'd pretty quickly.
1:09:18
But, uh you know what AI cannot do, Ellis, and I don't think we'll be able to do for the foreseeable future is throw a really fun party and a live show, which is what we're doing.
1:09:28
[laughter] Do you like that? I did.
1:09:31
I really wasn't I really wasn't expecting that.
1:09:33
Yeah, the Access launch party in San Francisco.
1:09:37
We are so excited about this.
1:09:39
Our friends at Notion, shout out Ivan, former guest, are hosting.
1:09:46
And we are going to invite all of our former guests, friends, family, colleagues, former colleagues, and and have a live show with Chris Best, the CEO of Substack, on May 14th in San Francisco.
1:09:59
If you should be there, we're going to be reaching out to you.
1:10:01
But also, you know, if you're a die-hard day one listener and you're in the Bay Area and you want to come, you know, reach out to us.
1:10:08
Find Find me or Ellis online.
1:10:11
We're very easy to reach and we'd love to have you.
1:10:13
Our show email is yes, we're client show at gmail. com.
1:10:17
[laughter] We do actually need to get our show email up.
1:10:18
We do actually need to get like a cuz you we've got access.
1:10:22
show for the website, but we need to set up like a like a tip at or you know, hello at email for stuff like this.
1:10:28
But yeah, man, I'm very excited.
1:10:30
This is going to be a good one. I think that's it. It was a fun one. Let's keep it rolling. Send feedback.
1:10:35
Again, this is our second week of this new format on the show where we're jumping right into the conversation and then recapping versus front-loading Ellis and I chatting. We hope you like it.
1:10:46
We think it's a better structure.
1:10:49
But yeah, let us know what you're thinking and what you want more of or less of.
1:10:52
Ellis, do you want to start reading us out here?
1:10:54
And that's it for this week's show.
1:10:56
Thanks to Lucas Peterson for coming on.
1:10:58
If you like this show, don't forget to like and subscribe everywhere you get podcasts.
1:11:03
Leave us five stars, please. It really helps. We are access. show on the internet.
1:11:07
You can find us in video via access pod on YouTube.
1:11:11
You can find my newsletter at sources. news.
1:11:15
And you can find me at hamburger on Twitter and at meaning.
1:11:16
company for your startup storytelling needs unless it's a stunt.
1:11:22
>> [laughter] >> I don't like good stunts.
1:11:22
You know, you want a good stunts. Only thoughtful stunts. Yeah. Yeah.
1:11:26
Access is part of the Vox Media Podcast Network and the show is produced by Hook Creators. Bye.
1:11:36
>> [music] [music] [music]