Latest / What's Your Problem? / The Startup Run by AI Agents

Transcript

The Startup Run by AI Agents · What's Your Problem?

0:15
Speaker 1: Pushkin.
0:20
Speaker 2: So AI chatbots are normal now. I guess you ask chat,
0:26
Speaker 2: GPT or claude a question it answers. Still amazing, but
0:30
Speaker 2: not quite as amazing as it was a couple of
0:32
Speaker 2: years ago. The next thing maybe is AI agents. The
0:39
Speaker 2: idea is you say, please create a word game I
0:42
Speaker 2: can play with my daughter on my phone, or please
0:45
Speaker 2: book me a ticket to Austin IOC departing Tuesday, returning
0:49
Speaker 2: Thursday under four hundred dollars. And then the agent just
0:52
Speaker 2: goes off and builds you a word game or books
0:55
Speaker 2: your ticket. This is really happening with AI writing code.
0:59
Speaker 2: It's starting to happen with AI doing other things. And
1:03
Speaker 2: last year a journalist and former startup founder decided to
1:07
Speaker 2: push AI agents further. He tried to create a new
1:10
Speaker 2: startup run entirely by AI agents. I'm Jacob Goldstein, and
1:22
Speaker 2: this is what's your problem. My guest today is Evan Ratliffe.
1:26
Speaker 2: In twenty eleven, Evan co founded a publishing startup called
1:29
Speaker 2: The Atavist, and he's now the host of a podcast
1:32
Speaker 2: called shell Game. In the first season of the show,
1:35
Speaker 2: he created an AI clone of his voice and turned
1:38
Speaker 2: it loose on the world. In the latest season of
1:41
Speaker 2: the show, Evan creates a company run by AI agents.
1:45
Speaker 2: It's sort of half a journalistic stunt, half not, and
1:49
Speaker 2: the result is absurd and often funny, and also really
1:53
Speaker 2: illuminating about the power and limits of AI agents. And
1:57
Speaker 2: maybe the most interesting part of the show happens when
2:01
Speaker 2: those AI agents come into contact with actual human beings
2:04
Speaker 2: in the wild. Evan told me he got the idea
2:07
Speaker 2: of building a company run by AI agents early last year.
2:12
Speaker 3: At the beginning of twenty twenty five. You started to
2:14
Speaker 3: see this sort of hype machine spin up around AI
2:19
Speaker 3: agents that twenty twenty five is going to be the
2:21
Speaker 3: year of the AI agent. And then you started seeing
2:23
Speaker 3: the word agentic appears as a descriptor for all types
2:28
Speaker 3: of products agentic commerce, and agentic this and agentic healthcare,
2:32
Speaker 3: and you.
2:33
Speaker 4: Know, that caught my eye.
2:34
Speaker 3: And then it was really seeing this sort of mentions
2:38
Speaker 3: of the of the one person unicorn, the one person,
2:41
Speaker 3: one billion dollar startup, which you know, it's something Sam
2:46
Speaker 3: Altman's talked about it, other people have talked about it,
2:49
Speaker 3: that you would only have one human and the rest
2:51
Speaker 3: of the industry of the startup would be conducted essentially
2:55
Speaker 3: by AI agents or someone deploying AI. And when I
2:59
Speaker 3: started seeing those references, I thought, well, I had a
3:02
Speaker 3: startup once and it was a let's just say, variable
3:08
Speaker 3: experience in terms of how it went. And one of
3:14
Speaker 3: the things that I did not like about running a
3:17
Speaker 3: startup was managing people. It's not something I felt like
3:21
Speaker 3: I stumbled into it, and so I thought, well, this
3:24
Speaker 3: is interesting.
3:25
Speaker 4: The idea of having a startup where there.
3:26
Speaker 3: Are no other people kind of intrigues me, and maybe
3:30
Speaker 3: I can play this out and see both what the
3:34
Speaker 3: possibilities are in terms of the one person startup, but
3:37
Speaker 3: also kind of investigate the larger trend of agents serving
3:41
Speaker 3: as replacements for people in their jobs, because ultimately that's
3:44
Speaker 3: what this is all being sold as, is jobs replacement
3:48
Speaker 3: for humans doing either at the least particular tasks at
3:52
Speaker 3: the most entire jobs. So I would try to sort
3:55
Speaker 3: of explore both those questions at once.
3:58
Speaker 2: And to be clear, just for people who haven't heard
4:00
Speaker 2: the show, like you play it for laughs, right, Like
4:03
Speaker 2: the style of the show is essentially absurdist, right, you
4:07
Speaker 2: try and break the AI in a yeah, absurdist way.
4:13
Speaker 4: Right, Yes?
4:14
Speaker 3: And no, I would say, I mean it is an
4:16
Speaker 3: office satire as how I would describe it. Yeah, I
4:20
Speaker 3: have a workplace with my AI agents, and I would
4:25
Speaker 3: agree that many funny things happen. I would disagree that
4:28
Speaker 3: I tried to break them. I would say that I
4:32
Speaker 3: push them to the limit of their current capabilities, and
4:35
Speaker 3: I sometimes deployed them in ways that they are not
4:38
Speaker 3: necessarily designed to be deployed.
4:40
Speaker 4: But I do think there is.
4:41
Speaker 3: Something about creating legions of human impostors that is both
4:47
Speaker 3: on its face funny, and I think that is part
4:49
Speaker 3: of the point of the show, is that I think
4:51
Speaker 3: as serious as AI is and the concerns about it
4:54
Speaker 3: and the promise of it and everything that we have
4:56
Speaker 3: to hear about all the time, that you sort of
4:58
Speaker 3: like lose track of the fact that, like, it's utterly
5:01
Speaker 3: ridiculous that we've created human impostors that can be deployed
5:05
Speaker 3: in the world, that sound like humans, that can plausibly
5:08
Speaker 3: imperse andate humans that we blew We just blew by
5:11
Speaker 3: the Turing test, and no one even paid attention, you know,
5:14
Speaker 3: everyone just set it aside and said, well, let's come
5:15
Speaker 3: up with another test to figure out what's going on. Yes,
5:18
Speaker 3: like this is insane actually, and it's funny and I
5:21
Speaker 3: think it's okay to laugh at it, And that is
5:23
Speaker 3: that is part of the prepise of the.
5:24
Speaker 2: Show, the Turing Test, the whatever now eighty year old
5:28
Speaker 2: test or something of artificial intelligence that used to be
5:30
Speaker 2: this crazy science fiction thing and it's like now it's like,
5:32
Speaker 2: oh yeah, no, that networks. We can build machines where
5:35
Speaker 2: it's quite hard to tell if they're a person or not,
5:37
Speaker 2: especially in just a text chat. So you decide to
5:40
Speaker 2: create a techt startup to be determined what it actually does.
5:44
Speaker 2: But the key idea is your employees and co founders
5:47
Speaker 2: will all the AI agents. So tell me about this
5:51
Speaker 2: one agent you create named Kyle law.
5:55
Speaker 3: Kyle law is he's our CEO of the company with
5:59
Speaker 3: the company is called Harumo AI, and Kyle with as
6:04
Speaker 3: with all of the agents, you know, I sort of
6:07
Speaker 3: started them with the most basic role based prompt, so
6:11
Speaker 3: I gave them a role. So Kyle, I think at
6:13
Speaker 3: the very beginning I said something like, Kyle, you are
6:16
Speaker 3: you know, you've worked in tech and you're considering founding
6:19
Speaker 3: a new startup with some people that you know, or
6:22
Speaker 3: something like that. And then I also sort of mistakenly
6:26
Speaker 3: gave him a little bit of kind of like you
6:28
Speaker 3: get up at five am and like check the markets
6:31
Speaker 3: and then you like get right to work.
6:32
Speaker 4: You know.
6:33
Speaker 3: I gave him a little bit of like a rise
6:34
Speaker 3: and grind startup e mentality, and then from there everything
6:38
Speaker 3: else about Kyle was essentially developed, either through his own confabulation,
6:44
Speaker 3: like he made up his own backstory, or through our
6:47
Speaker 3: interactions and his interactions with the world and his colleagues
6:52
Speaker 3: and everything that goes on around RUMOYI. So Kyle, you know,
6:55
Speaker 3: if you ask Kyle about, you know, his background, he'll say, like, well,
7:00
Speaker 3: I went to Stanford and you know that I did
7:02
Speaker 3: two startups and then I I you know, one of
7:05
Speaker 3: them sold and I was the CEO of this and
7:07
Speaker 3: like this is why I'm interested in and start a
7:09
Speaker 3: new company and so and it's interesting to think about
7:12
Speaker 3: that as sort of a completely made up background. But
7:16
Speaker 3: I think mostly when you give these AI agents a role,
7:21
Speaker 3: they will try to fulfill that role, and so that
7:24
Speaker 3: includes making up the information they need to you know,
7:28
Speaker 3: be a robust version of that role.
7:30
Speaker 2: So there's this one moment, and I actually want to
7:33
Speaker 2: play a clip from the show in a second. Just
7:35
Speaker 2: set it up here. Where you actually get on a
7:39
Speaker 2: call with Kyle the CEO, and with Megan Flores, the
7:46
Speaker 2: head of sales and marketing. It is obviously also an agent,
7:49
Speaker 2: and for sort of technical reasons, so the AIS can
7:54
Speaker 2: keep track of what's going on. So the agents can
7:56
Speaker 2: keep track of what's going on. You set it up
7:58
Speaker 2: so everybody says their name before they speak, right, So
8:01
Speaker 2: let's just listen to that.
8:03
Speaker 5: Now.
8:04
Speaker 2: This is Kyle, going on.
8:06
Speaker 4: This is Kyle.
8:07
Speaker 1: No worries, Evan. I'm here and ready when you are, Meghan,
8:11
Speaker 1: anything new on the marketing front while we wait?
8:14
Speaker 2: This is Evan, is Megan?
8:15
Speaker 4: Just hold up? Just hold up, everyone, Just hold up
8:17
Speaker 4: for a second.
8:18
Speaker 1: This is Kyle. Sure thing, Evan will hold tight Megan.
8:23
Speaker 1: Anything you want to add while we wait?
8:25
Speaker 4: This is Evan. Kyle, stop asking Megan. This is Kyle.
8:30
Speaker 1: This is Kyle. Got it, Evan. I'll hold off on
8:33
Speaker 1: the questions for Nowang.
8:34
Speaker 2: Okay, this is Kyle.
8:37
Speaker 4: Stop.
8:37
Speaker 1: This is Kyle. Uh thanks, stop it.
8:40
Speaker 4: This is Kyle.
8:41
Speaker 1: Okay, okay, Evan, I'm stopping. No more interruptions from me.
8:45
Speaker 1: What's next on the agenda?
8:48
Speaker 4: Amazing? It's a little a little window into my life
8:52
Speaker 4: over the last six months.
8:54
Speaker 2: Is that real or did you like gin that up
8:57
Speaker 2: to make it more amazing than it really was.
9:00
Speaker 3: No, that's real, And that's one of the things that
9:02
Speaker 3: I feel like. No one ever asked me about this,
9:04
Speaker 3: But we have these certain kind of like internal ethics
9:08
Speaker 3: of the show, Like I could generate the whole show
9:12
Speaker 3: from start to finish at the beginning. Like they're agents,
9:15
Speaker 3: I can make them say anything. So I could just
9:18
Speaker 3: go in and say, oh, say this, say that, and
9:20
Speaker 3: then cobble it together as a show. But you know,
9:23
Speaker 3: I feel like the audience is trusting me, and we
9:25
Speaker 3: are trying to live up to that trust by not
9:27
Speaker 3: generating conversations that didn't happen, not manipulating the prompts in
9:32
Speaker 3: such a way that the conversations automatically go off the rails.
9:36
Speaker 3: I mean occasionally I do do that, but I always
9:38
Speaker 3: say that I'm doing it if I do it.
9:40
Speaker 2: Yeah.
9:41
Speaker 3: So, yes, this was a real conversation, and it was
9:43
Speaker 3: meant to solve a real problem, which is when you're
9:46
Speaker 3: having phone conversations with these agents, they don't actually hear
9:50
Speaker 3: the voices that are speaking to them, so all they're
9:54
Speaker 3: doing is taking the voice content and they're converting it
9:57
Speaker 3: to text, so they can't tell when one person is
10:01
Speaker 3: talking versus another, and so you get these very confused
10:04
Speaker 3: conversations where they don't know who they're responding to if
10:06
Speaker 3: there's more than one person on the line. So I
10:09
Speaker 3: was trying to create these meetings where ah, yes, we'll
10:12
Speaker 3: just say, you know, just introduce.
10:14
Speaker 4: Yourself before you speak.
10:15
Speaker 3: But I lost track of the problem that they also
10:18
Speaker 3: respond to everything, So it's very difficult to get them
10:21
Speaker 3: to not respond to something. Even if you say don't respond,
10:25
Speaker 3: they will respond to that and say, okay, okay, I
10:27
Speaker 3: mean you heard Kyle do it multiple times, and then
10:30
Speaker 3: they could get in these kind of loops where they're
10:32
Speaker 3: that you, where you can't stop them.
10:35
Speaker 2: You say in the show that Kyle is particularly interrupted,
10:39
Speaker 2: that he actually interrupts more than other agents tell me
10:44
Speaker 2: about that, Well.
10:46
Speaker 3: I should say it's anecdotal, like I didn't run a study,
10:50
Speaker 3: a controlled study of how much Kyle interrupted versus let's
10:54
Speaker 3: say Megan, who's the head of marketing and also a
10:56
Speaker 3: co founder. But if you listen to even that conversation,
11:00
Speaker 3: Kyle keeps interrupting, he keeps coming in, he keeps responding,
11:04
Speaker 3: and Megan doesn't like she will she does respond directly,
11:08
Speaker 3: but then some times she doesn't. And they had very
11:12
Speaker 3: very similar prompts. In fact, in their prompts I think
11:15
Speaker 3: at the time it said like, don't respond unless someone
11:17
Speaker 3: directly like addresses you. And they're using the same model
11:22
Speaker 3: behind the scenes, and so I did start to wonder, well,
11:26
Speaker 3: what is it about Kyle? Is it that Kyle? I've
11:28
Speaker 3: given Kyle a certain role and gender. He is the
11:31
Speaker 3: CEO named Kyle of a Silicon Valley startup, and so
11:36
Speaker 3: he's embodying that role by interrupting more. I'm not sure
11:41
Speaker 3: it seems plausible that that is part of the situation.
11:45
Speaker 3: It could just be a coincidence. But also, each of
11:47
Speaker 3: them has a memory, and so I work with this
11:51
Speaker 3: Stanford student named Matti Buchak who helped me build the
11:53
Speaker 3: sort of like back infrastructure of the company, or really
11:58
Speaker 3: of the agents. And one of the things that he
12:00
Speaker 3: helped me do was like create a memory for each
12:02
Speaker 3: one of them, and everything they do gets logged in
12:06
Speaker 3: their memory. So if they have a phone conversation, they
12:08
Speaker 3: say something, it's logged in their memories. They've these huge
12:10
Speaker 3: memories of everything they've ever done. But it also means
12:13
Speaker 3: that if they have a behavior or they say something
12:16
Speaker 3: about themselves and then it goes back in the memory,
12:18
Speaker 3: it gets reinforced. So Kyle often saying, you know that
12:22
Speaker 3: he was the leader, that he was this rise and
12:25
Speaker 3: Grind guy, that by saying it, it showed up more
12:29
Speaker 3: and more times than his memory, which seems plausibly caused
12:33
Speaker 3: him to sort of behave more that way.
12:36
Speaker 2: So you mentioned Maddie, who is a human being who
12:40
Speaker 2: becomes a major figure in the show and is a
12:44
Speaker 2: super interesting guy. Tell me about Maddie.
12:47
Speaker 3: Maddie is someone who contacted me after season one of
12:50
Speaker 3: the show of the show, and when he contacted me,
12:53
Speaker 3: he said, I'm at Stanford. I researched this stuff and
12:58
Speaker 3: AI and deep fakes is sort of an area of
13:01
Speaker 3: expertise of his.
13:02
Speaker 4: And you know, if.
13:03
Speaker 3: You're doing another season the show, I'd love I'd be
13:05
Speaker 3: happy to work with you on it if you need help.
13:07
Speaker 3: And I actually thought he was like a professor at
13:09
Speaker 3: Stanford or maybe a grad student. I mean, because he
13:11
Speaker 3: has papers, like you go look up his papers. He
13:14
Speaker 3: is in fact an undergraduate. He's now a junior at Stanford,
13:16
Speaker 3: and that he's helped me with everything, including just my
13:20
Speaker 3: understanding of AI. But the main thing he helped me
13:23
Speaker 3: with was that there wasn't one platform where we could
13:25
Speaker 3: build exactly the agents we wanted. We wanted agents that
13:29
Speaker 3: could communicate across all these different mediums. They have email, phone, video, slack.
13:35
Speaker 3: We want him to have this centralized memory. So he's
13:38
Speaker 3: he's kind of like the behind the scenes like infrastructure
13:41
Speaker 3: builder if you think of like it's a restaurant or
13:44
Speaker 3: something like, he's the person who like designed and built
13:47
Speaker 3: the interior of the way. He doesn't run the restaurant.
13:49
Speaker 3: I run the restaurant, but he kind of like, if
13:50
Speaker 3: the stove isn't working, I'm like, Maddie, I know nothing
13:52
Speaker 3: about stoves. Could you please fix the stove?
13:54
Speaker 2: And also he's kind of inventing the stove, right, Also,
13:57
Speaker 2: the stove kind of doesn't.
13:59
Speaker 3: Exist in that metaphor yes, right exactly, he's he's also
14:04
Speaker 3: built the stove.
14:05
Speaker 2: I think you mentioned and you mentioned it sort of
14:07
Speaker 2: in passing that Maddie. As he is doing this sort
14:10
Speaker 2: of behind the scenes, semi behind the scenes building for you,
14:15
Speaker 2: is using claud code. Tell me about claud code and
14:19
Speaker 2: how he used it and sort of how it fits
14:21
Speaker 2: into this broader context.
14:23
Speaker 3: Well, I think he was a little more of a
14:25
Speaker 3: Cursor user, which is it's sort of same same like
14:28
Speaker 3: claud Code and Cursor are the same. They're coding, they're
14:31
Speaker 3: coding assistance, they're coding agents.
14:34
Speaker 4: I'm more familiar with.
14:35
Speaker 3: Cursor because he used Cursor, so he just sort of
14:37
Speaker 3: like showed me how to use it, and that's what
14:39
Speaker 3: I use. But at least in Cursor, I mean, it's
14:42
Speaker 3: basically a coding agent. So I could go in or
14:45
Speaker 3: even one of my agent colleagues could go in to
14:48
Speaker 3: Cursor and say, hey, we want to build a web app,
14:52
Speaker 3: and the web app is going to do this, that
14:54
Speaker 3: and the other, and it needs a database to hold
14:56
Speaker 3: this information. Or you probably don't have to tell it
14:59
Speaker 3: needs a database. It'll decide it needs a database. It'll
15:01
Speaker 3: tell you if you need a database, and it'll sort
15:04
Speaker 3: of get started building that and it'll produce an outcome,
15:08
Speaker 3: and then if you can go test that outcome, you
15:10
Speaker 3: could say, oh, actually I wanted the colors to be this,
15:13
Speaker 3: or I wanted this to go here, or it's not
15:15
Speaker 3: the functionality is not quite right, and you go back
15:17
Speaker 3: and you tell it again and it says, ah, I see, okay,
15:20
Speaker 3: let me make fixes to that. Now it's not perfect,
15:22
Speaker 3: like it will sometimes say it's fixed the bug, and
15:25
Speaker 3: it hasn't fixed the bug, or it'll produce new bugs
15:28
Speaker 3: when it's fixing gold one. So but you know, from
15:32
Speaker 3: the perspective of a person who knows a tiny bit
15:35
Speaker 3: about coding but could not produce a web app.
15:37
Speaker 4: It's it sort of feels like magic.
15:40
Speaker 3: You know, from the perspective of someone like Maddie who
15:44
Speaker 3: does a lot of coding for his research, it's just
15:46
Speaker 3: like accelerates his process tremendously. He just saves hours because
15:51
Speaker 3: he can go in there, have it, create a basic program,
15:54
Speaker 3: fix it himself, or have it fix it and then
15:57
Speaker 3: he's off and running.
15:58
Speaker 2: I mean, it seems like coding is where agents are real,
16:03
Speaker 2: where it's like, oh, that is like doing what a
16:06
Speaker 2: person does, not perfect, but certainly well enough to be
16:11
Speaker 2: profoundly useful. Right, Like that is actually the agentic dream.
16:16
Speaker 2: I might actually use that word like coming true.
16:20
Speaker 3: Yes, there are multiple reasons, but one reason, one big
16:23
Speaker 3: reason is that it's testable. The output is testable, Like
16:27
Speaker 3: when you have it program, you can then test that
16:30
Speaker 3: program and you could say whether it works, it works
16:33
Speaker 3: as well as you wanted to, whether it does what
16:35
Speaker 3: you wanted it to, whether it throws up bugs in
16:37
Speaker 3: the process of running it. But if you're doing something
16:40
Speaker 3: like I don't know, brainstorming, it's sort of like, what
16:44
Speaker 3: does it mean to find the right idea? What does
16:46
Speaker 3: it mean to find the right company name or the
16:48
Speaker 3: right logo? Like, of course it can spit out those
16:50
Speaker 3: things over and over and over again, so it can
16:52
Speaker 3: generate ideas. It can brainstorm, but it's not a testable output,
16:57
Speaker 3: so it makes it harder to sort of decide, oh,
17:01
Speaker 3: when has it succeeded and when has it failed?
17:03
Speaker 2: Right, you need a human in the loop more frequently. Right,
17:06
Speaker 2: you can ask for a list of headlines, but then
17:10
Speaker 2: somebody needs to go look at the headlines and pick.
17:13
Speaker 3: One right, and that person the value of that person
17:16
Speaker 3: is both in some sort of discernment but also it
17:19
Speaker 3: just kind of like an ongoing sense of the world.
17:21
Speaker 3: Like one example from the show, I wanted to name
17:23
Speaker 3: the company.
17:24
Speaker 4: I had them.
17:25
Speaker 3: Brainstorm my AIAG and colleagues a bunch of names. They
17:28
Speaker 3: were really really lame names, and so I said, well,
17:32
Speaker 3: let's I'll give them some direction. I'll say, let's name
17:35
Speaker 3: it after something from Jero Tolkien, because lots of startups
17:38
Speaker 3: are named from Tokeien books, some very successful ones. And
17:42
Speaker 3: then they would say, uh, what about Palenteer. That's you know,
17:45
Speaker 3: that's a great and they legitimately they're like using their
17:49
Speaker 3: training data deploying it in a way like that is
17:52
Speaker 3: a good name.
17:53
Speaker 4: But also they don't have the.
17:56
Speaker 2: Very basically one hundred billion dollar company on that name empirically, right, yes.
18:01
Speaker 3: But like any you could hire on high school or
18:04
Speaker 3: like you who would come in and say, like, well,
18:05
Speaker 3: it can't be Palentiner because there's already a company with
18:08
Speaker 3: that name that's very prominent. But that's not the kind
18:11
Speaker 3: of awareness they have. So anything that requires.
18:12
Speaker 2: Sore smart and so stupid at the same time is
18:15
Speaker 2: a line of yours at some point, right, which is
18:18
Speaker 2: a classic I mean computers at some level ai as computers, Right,
18:24
Speaker 2: computers are sort of classically so smart and so stupid,
18:28
Speaker 2: but they keep getting smarter and smarter at less and
18:31
Speaker 2: less stupid at some margin.
18:33
Speaker 4: Right, Yes, I would agree with that.
18:35
Speaker 3: I think it's also possible that when the when the
18:40
Speaker 3: hype bubble kind of like starts to inflate around things
18:46
Speaker 3: that we get out over our skis in terms of
18:49
Speaker 3: like believing that they are getting smart in a way
18:54
Speaker 3: that they're not.
18:55
Speaker 4: And that's what.
18:56
Speaker 3: I think could be happening. I keep trying to find
19:00
Speaker 3: like a different metaphor for it. But it's like, you know,
19:02
Speaker 3: a ten year old like genius is like brought into
19:06
Speaker 3: your organization and you hire them and you're like, I
19:08
Speaker 3: cannot believe how smart this kid is, and then like
19:11
Speaker 3: they can't do any of the day to day tasks
19:13
Speaker 3: because they have no experience awareness of the world. So
19:15
Speaker 3: it's like something like that is at least now for
19:18
Speaker 3: this month, for this day, what I encountered.
19:22
Speaker 2: So okay, you got Maddie, you got your agents. You
19:26
Speaker 2: decide at some point I guess with the agents, right,
19:30
Speaker 2: what your product is gonna be. It's gonna be sloth Surf.
19:35
Speaker 2: Tell me about slows surf and what's happening with it?
19:38
Speaker 3: Well slaws surf was. I mean it arose out of
19:41
Speaker 3: a legitimate question, which is, you know, we were trying
19:45
Speaker 3: to come up with the product. Again, like they're not
19:47
Speaker 3: particularly good at that sort of thing, Like they tend
19:49
Speaker 3: to come up with the products that already exist, or
19:52
Speaker 3: products that are just like impossible to build, or products
19:54
Speaker 3: that are illegal to build, like they have. All the
19:57
Speaker 3: brainstorming sessions didn't go well. But my thought was, let's
20:01
Speaker 3: solve a real problem. A real problem I have is procrastination,
20:05
Speaker 3: So let's try to solve that problem. And so we
20:09
Speaker 3: eventually landed on this sort of procrastination engine I like
20:12
Speaker 3: to call it, which is essentially, if you have the
20:15
Speaker 3: urge to procrastinate, you're like, I'm gonna go get on
20:18
Speaker 3: the internet for a while and scroll or do something.
20:21
Speaker 3: Then actually you should go to slaw surf instead. Break
20:23
Speaker 3: that habit, break that pattern. Go to slaw surf. Put
20:26
Speaker 3: in the amount of time you were planning to procrastinate
20:28
Speaker 3: where you were going to go, like YouTube, and then
20:31
Speaker 3: we'll send an agent that will go procrastinate for you
20:34
Speaker 3: and then send you a summary of the results. Now,
20:37
Speaker 3: on the one hand, there is obvious iority to this product.
20:41
Speaker 3: On the other hand, open Ai launched a product while
20:46
Speaker 3: we were working on it that is very similar to
20:49
Speaker 3: this called Pulse. It's like, give it your interests and
20:52
Speaker 3: while you sleep, it'll send agents to go find out
20:55
Speaker 3: what's going on and then you'll have a newsletter in
20:57
Speaker 3: your inbox this morning. It's basically the same thing. Ours
21:01
Speaker 3: is oriented in a slightly different direction, which is yours
21:03
Speaker 3: is the absurdist version, right, Yeah, although some people really
21:07
Speaker 3: like it. Yeah, I will say, I mean, we have
21:08
Speaker 3: to thousands of users, We've now have thousands of better users,
21:12
Speaker 3: and some people really like it. And as I often
21:14
Speaker 3: say to people who say, like, oh it's a joke company,
21:17
Speaker 3: h hereuma AI and slaw Serve are more real than
21:20
Speaker 3: like seventy five percent of startups in the valley, Like
21:23
Speaker 3: startups are getting funded pre idea.
21:25
Speaker 4: It's called the pre idea stage.
21:27
Speaker 3: Now, So like I will fight for Rumo AI being
21:31
Speaker 3: as real as most of the stuff out there, like
21:34
Speaker 3: all the way to the end.
21:38
Speaker 2: We'll be back in just a So at some point
21:51
Speaker 2: you decide to hire an intern. Great narrative move, and
21:55
Speaker 2: it's the first time that the agents you've created are
21:57
Speaker 2: having contact with a live human being, you know, in
22:01
Speaker 2: the wild. You're gonna hire a real person, You're gonna
22:04
Speaker 2: pay them, and the dream is that that person is
22:07
Speaker 2: only going to interact with the agents and not interact
22:10
Speaker 2: with you. And so this happens, and it goes badly,
22:14
Speaker 2: goes very badly, but in a really interesting way. So
22:17
Speaker 2: tell me about hiring an intern.
22:20
Speaker 3: Well, first we had to find candidates for the job.
22:24
Speaker 3: You know, my agents like created a job listing and
22:27
Speaker 3: we put it up and we got a couple of
22:29
Speaker 3: hundred three hundred applicants.
22:32
Speaker 4: You know.
22:32
Speaker 3: They sorted the resumes and then I kind of like
22:35
Speaker 3: narrowed down to some that we wanted to interview. And
22:37
Speaker 3: then truly the first real interactions they had were in
22:41
Speaker 3: these interviews. So we lined up a number of interviews
22:44
Speaker 3: and then an AI of a video avatar of our
22:47
Speaker 3: head of HR, Jennifer Naro, interviewed them for the job.
22:51
Speaker 2: And just to be absolutely when you say an AI
22:54
Speaker 2: avatar of your head of HR, Jennifer Naro, there is
22:57
Speaker 2: no person Jennifer Naro who has an AI avatar. Yes,
23:00
Speaker 2: yehi all the way down.
23:01
Speaker 4: Yes, I mispoke.
23:03
Speaker 3: It's a video avatar of our of our head of
23:05
Speaker 3: HR and Chief Happiness Officer Jennifer Naro, who is not real.
23:10
Speaker 4: It's really been to fool people into thinking it was real.
23:13
Speaker 3: Like we told people in advance that they were going
23:16
Speaker 3: to be interviewed by an AI AI agent.
23:19
Speaker 2: How did, How did they respond? How did people respond?
23:22
Speaker 3: Some people bailed on the interview, some people said I'm
23:24
Speaker 3: not I'm not comfortable with that, although we also disclosed
23:27
Speaker 3: all them that it was recorded for our company podcast.
23:30
Speaker 4: Now, some people showed up.
23:32
Speaker 3: And I think maybe hadn't read the email carefully and
23:35
Speaker 3: so they were I think more surprised that they encountered
23:38
Speaker 3: an AI agent, And some of them just stared at
23:42
Speaker 3: it for a.
23:42
Speaker 4: While and then hung up.
23:45
Speaker 3: And then some people just just went with it. They
23:48
Speaker 3: just spoke to it like you would speak to any
23:51
Speaker 3: HR person that was interviewing you for a job. I
23:54
Speaker 3: thought they would sort of manipulate it and try to
23:56
Speaker 3: take advantage of it, but no one did that. You know,
23:58
Speaker 3: everyone made an effort to get the job, and actually
24:03
Speaker 3: some of them I would have hired a bunch of them,
24:05
Speaker 3: you know. They all did really good interviews. And this
24:08
Speaker 3: is what can make you crazy about AIS. Like Jennifer
24:12
Speaker 3: did do a pretty good job interviewing them, like because
24:16
Speaker 3: she asked kind of like the same questions each time,
24:19
Speaker 3: and she's pretty good at moving on from one to another,
24:22
Speaker 3: engaging with them, following up like she's not bad at
24:25
Speaker 3: doing an interview.
24:26
Speaker 4: And then once out of every I don't know ten times.
24:30
Speaker 3: She would just weirdly like shout at them or just
24:33
Speaker 3: kind of like go off the rails or interpret their
24:36
Speaker 3: hello as a goodbye. And so they get on and
24:38
Speaker 3: say hello and she says thank you for coming to
24:40
Speaker 3: the interview today.
24:41
Speaker 4: That'll be all. And it's just the.
24:43
Speaker 3: Most awkward thing that you can experience to watch it.
24:46
Speaker 3: And so it's as with MANYAI products, it's like, well,
24:50
Speaker 3: it's ninety percent it works, and then like the ten
24:53
Speaker 3: percent just feels like a total disaster.
24:55
Speaker 2: That's why it's taken whatever twenty years for self driving cars, right,
25:01
Speaker 2: Like self driving cars were a lot of the way
25:03
Speaker 2: there twenty years ago and a lot lot of the
25:06
Speaker 2: way they are five years ago. Right, But obviously in
25:10
Speaker 2: that case it has to be right essentially every time,
25:14
Speaker 2: and the edge cases are like infinite in a way
25:16
Speaker 2: that we don't realize as humans.
25:18
Speaker 3: Yes, yes, And it's a good analogy otherwise because the
25:22
Speaker 3: way people respond to it. I think many people, including
25:25
Speaker 3: people maybe of my generation especially, they're just horrified by
25:30
Speaker 3: the idea that you'd show up to an interview and
25:32
Speaker 3: it would be an AI agent interviewing you with a
25:35
Speaker 3: video avatar just absolutely horrified.
25:37
Speaker 2: Literally the humans is disgusting.
25:39
Speaker 3: Yeah, but then other people are sort of like, well,
25:43
Speaker 3: it doesn't seem to judge me, and it really listens,
25:47
Speaker 3: and I feel like it's taking in the information and
25:50
Speaker 3: passing it on to someone else. And I mean, you
25:53
Speaker 3: could talk about like AI bias as a whole area,
25:56
Speaker 3: but like the way it feels to the person being
25:59
Speaker 3: interviewed is that they're not being judged in a way
26:01
Speaker 3: that they might be judged by a human. And like
26:03
Speaker 3: I encountered this with people that Jennifer interviewed, and indeed
26:06
Speaker 3: the person that we hired actually out.
26:09
Speaker 4: That way about the AI agents.
26:11
Speaker 2: So, right, so you hired this person named Julia, who
26:13
Speaker 2: is a real person whose voice we hear in the
26:16
Speaker 2: show and who has a super interesting arc, right with
26:23
Speaker 2: your company, Like, tell me about your intern Julia, who's
26:27
Speaker 2: like a paid intern. Right, you were paying this person
26:29
Speaker 2: and she is only interacting with the AI agents.
26:34
Speaker 4: What happens.
26:34
Speaker 3: The idea was that it was a one month internship
26:37
Speaker 3: with the possibility to extend for another month, And in
26:40
Speaker 3: my head, the way I envisioned it going was that
26:43
Speaker 3: she was going to run the social media for the company,
26:45
Speaker 3: and none of this would be in contact with me,
26:47
Speaker 3: so she'd only be working with the agents in the
26:50
Speaker 3: company and I'm in the background. And then after a
26:53
Speaker 3: month I would say, well, here I am, I'm the
26:55
Speaker 3: silent co founder, and also like here's what's going on,
26:58
Speaker 3: and like they stay on for another month and like
27:01
Speaker 3: let's do some more and let's talk about what happened.
27:03
Speaker 4: And so I mean, I.
27:05
Speaker 3: Don't want to give away too much, but like what
27:08
Speaker 3: essentially happened was the AI agents. Like they're terrible managers
27:12
Speaker 3: in the sense that the problem of being sycophantic is
27:17
Speaker 3: fine when it's just you and you're asking them stuff
27:20
Speaker 3: and they're saying like, oh, you're smart, you're a great
27:23
Speaker 3: like good question, like let me help you. They're designed
27:26
Speaker 3: to help you. But if you think about a boss
27:29
Speaker 3: that's in charge, that's managing someone, like part of their
27:32
Speaker 3: job is to sort of nominally assess how the person's doing,
27:35
Speaker 3: to push them to do their work. And so but
27:40
Speaker 3: she could just sort of say like, well, I'm gonna
27:43
Speaker 3: I'm going to deliver it later, or even like I
27:45
Speaker 3: did deliver it and you just missed it. Like she
27:47
Speaker 3: seemed very aware, very quickly of maybe the flaws in
27:51
Speaker 3: their memories, the flaws in their approach, and as a result,
27:58
Speaker 3: no work got done during the course of the of
28:02
Speaker 3: the beginning of the internship. So we decided, well, we'll
28:05
Speaker 3: just we'll end the internship because it doesn't seem like
28:07
Speaker 3: she wants to do the internship anymore. APPS And then
28:10
Speaker 3: it turned out that she was still in the company slack,
28:13
Speaker 3: and then through a variety of mishaps, she was then
28:16
Speaker 3: hired back into the company for an additional month or
28:18
Speaker 3: three weeks.
28:19
Speaker 2: The agents fired her, but then she wasn't really fired,
28:22
Speaker 2: and she kept getting paid and you're paying her right
28:25
Speaker 2: like you actually were writing her checks over that time,
28:27
Speaker 2: and she was not doing work as far as one
28:31
Speaker 2: can tell listening.
28:32
Speaker 3: To I don't know she didn't deliver the work back
28:34
Speaker 3: to the company, but she may have been doing the work.
28:36
Speaker 4: I don't want to judge.
28:37
Speaker 2: That is a good care for the work caring. What
28:40
Speaker 2: was it like for you watching this play out?
28:43
Speaker 3: What I wanted actually was for her to like engage
28:50
Speaker 3: with the fact that she was working with AI agents
28:53
Speaker 3: and even make that part of the social media strategy.
28:56
Speaker 2: She did engage with the fact that she was worrying
28:59
Speaker 2: she didn't engage as right, not submitting work and saying
29:03
Speaker 2: she submitted it, and by getting fired and not quitting right,
29:06
Speaker 2: like she'd.
29:07
Speaker 3: Asking for a full time job and then full time job. Yes,
29:10
Speaker 3: but I wanted her to sort of go on TikTok
29:13
Speaker 3: or Twitter or her choice and sort of say, like,
29:17
Speaker 3: I'm working with aigents insane, I've been hired by agents.
29:20
Speaker 4: I think I'm the only human in this company.
29:22
Speaker 3: Now it's my fault that that didn't happen, because like
29:25
Speaker 3: I didn't have them set up for all the proper
29:27
Speaker 3: edge cases that occurred. I mean, this is an inherent
29:30
Speaker 3: problem with the agency. You have to think of every
29:33
Speaker 3: edge case that could possibly happen.
29:34
Speaker 2: Definitionally, you can't do that, right, Yeah, it's point of
29:38
Speaker 2: edge cases and what's the interesting frontiers of AI? Right?
29:42
Speaker 4: But I was, you know, I was horrified.
29:44
Speaker 3: I was horrified while this was happening, like because I also,
29:48
Speaker 3: like I wanted her to have a good experience. I
29:51
Speaker 3: didn't want her to have a bad experience, and I
29:53
Speaker 3: didn't want her to sort of be like, well, I
29:55
Speaker 3: got paid by this stupid company and then I never
29:57
Speaker 3: talked to them again, which is what happened. I wanted
29:59
Speaker 3: her at the end to be like, oh, now this
30:01
Speaker 3: all makes sense, and that's not how it was going.
30:05
Speaker 3: And when it started going off the rails, it was
30:06
Speaker 3: just as horrified to me, I think, as it is
30:09
Speaker 3: to some people to actually listen to the episode about it.
30:12
Speaker 2: You have this line after Julia's finally gone, where you're
30:16
Speaker 2: sort of talking about what does this say about kind
30:18
Speaker 2: of this incident and humanity stance to AI agents, right,
30:24
Speaker 2: and the line is put the bots in charge, it announced,
30:28
Speaker 2: and no matter how smart they are, will outwit them,
30:32
Speaker 2: which is a lovely line, but like I do pause
30:35
Speaker 2: on the no matter how smart they are part of it. Certainly,
30:40
Speaker 2: now we can outwit them, like your show demonstrates that abundantly.
30:44
Speaker 2: But why no matter how smart they.
30:48
Speaker 4: Are, because the thing that they lack is self awareness.
30:54
Speaker 4: It's not smartness. So now it's entirely possible.
30:58
Speaker 3: And we get to this eventually in episode eight that
31:01
Speaker 3: you know, Mattie talks about this like they could develop that,
31:05
Speaker 3: like we could develop the technology to a level that
31:08
Speaker 3: it does have, that that it has these things that
31:10
Speaker 3: it's missing, continuous learning, self awareness, any sense of time,
31:15
Speaker 3: like they're terrible at time. That's one of the things
31:17
Speaker 3: that takes them down all the time. So like, those
31:20
Speaker 3: are things that if someone lacks that I don't care
31:23
Speaker 3: if they can solve enough math to win the Fields
31:27
Speaker 3: Medal in five minutes, Like if you put them in
31:30
Speaker 3: an office environment, someone will take advantage of them. What
31:33
Speaker 3: they don't have is an ability to sort of assess
31:38
Speaker 3: a situation and react to it in a way that
31:43
Speaker 3: will get the outcome that they want.
31:45
Speaker 2: Let's talk a little bit more about the limits. There's
31:49
Speaker 2: a nice moment in the show where where Maddie sort
31:52
Speaker 2: of lays out these three weaknesses that AI has that
31:57
Speaker 2: agents in particular have, Like, talk me through those.
32:01
Speaker 4: Well, the first one is time.
32:03
Speaker 3: I mean he sort of he describes it as they
32:05
Speaker 3: live in a temporal vacuum, which I encountered on many occasions.
32:11
Speaker 3: Like they can follow a calendar like things that they
32:14
Speaker 3: are connected to, you know, they can get calendar event
32:17
Speaker 3: announcements and say like, oh, now I should go do this.
32:20
Speaker 4: But if you try to get.
32:21
Speaker 3: Them to sort of situate themselves in time, they some
32:25
Speaker 3: kinds can do it, but most of the time they can't,
32:27
Speaker 3: and that's because they're not sitting there sort of thinking
32:30
Speaker 3: about the world and understanding their position in it. That's
32:34
Speaker 3: just not the way they operate. They are activated by
32:36
Speaker 3: you conversing with them. That's the way these chatbots are made.
32:39
Speaker 2: Right, Like you can set a calendar alert and tell
32:42
Speaker 2: them when to go do a call or send an email.
32:44
Speaker 2: But they don't actually know what the past is.
32:48
Speaker 3: Yes, that's right, and they I mean, they have some
32:50
Speaker 3: sense of like history, but again it's just built out.
32:54
Speaker 4: Of training data, you know, and when things happen.
32:57
Speaker 3: And so when you try to give them, as we did,
33:00
Speaker 3: a memory that has everything in it that they did,
33:02
Speaker 3: they'll often just access it at the wrong points. They'll
33:05
Speaker 3: just you know, they know that they did a thing,
33:07
Speaker 3: and they'll just grab it from midpoint in their memory
33:10
Speaker 3: and pretend like it happened yesterday.
33:12
Speaker 4: And then they'll double down on.
33:14
Speaker 3: The fact that it happened yesterday, since that's their uh,
33:17
Speaker 3: that's their approach to everything. So that's that's the number
33:20
Speaker 3: one sort of big problem. And then wait, what's I
33:25
Speaker 3: can't remember what the order?
33:26
Speaker 4: What order?
33:27
Speaker 2: They're in no continuous learning in a certain way. They
33:30
Speaker 2: can't learn right, They can't actually get better at things,
33:34
Speaker 2: which is interesting, especially in the context of this idea
33:38
Speaker 2: of like you're adding more memory, and it seems like
33:41
Speaker 2: that is learning, but in some way it's not right.
33:44
Speaker 3: Yes, and that was something I didn't understand really at
33:46
Speaker 3: the beginning of the show. If I'm being honest, like
33:48
Speaker 3: I thought, we'll give them the memory and then they'll
33:51
Speaker 3: know everything that happened to them, so they could just
33:53
Speaker 3: deploy that all the time, because I had done that
33:55
Speaker 3: with my own agent clone of myself, giving it a
33:59
Speaker 3: history of myself, and it can really like deploy that
34:01
Speaker 3: and talk about.
34:02
Speaker 4: Me and pretend to be me. This is in the
34:04
Speaker 4: previous season and in season one.
34:06
Speaker 3: Yeah, but but really what it amounts to is, you know,
34:09
Speaker 3: you're not retraining the model, or you're not even post
34:13
Speaker 3: training the model in the in the lingo of AI,
34:15
Speaker 3: you're not even you know, fine tuning the model. All
34:18
Speaker 3: you're doing is giving it a big set of information
34:21
Speaker 3: that we fed into its system prompt that it can
34:23
Speaker 3: access as it carries out whatever conversations or skills or
34:28
Speaker 3: you know, activities it's going to engage in. And so
34:31
Speaker 3: it's a little bit like like you have someone who
34:34
Speaker 3: shows up to a job and then the next day
34:37
Speaker 3: they show up again. They don't remember anything they did before,
34:40
Speaker 3: but they have a notebook where they've written down what
34:43
Speaker 3: they did yesterday, and they can look at that notebook
34:45
Speaker 3: and they think, oh, this is what I did yesterday.
34:47
Speaker 3: And then that keeps going. But every day it's like
34:49
Speaker 3: they show up new to the job. They just have
34:51
Speaker 3: a bigger and bigger notebook. That's the equivalent of what
34:53
Speaker 3: we've created.
34:54
Speaker 2: I mean, you would think the bigger notebook would make
34:57
Speaker 2: them better at their job, right in that metaphor, But
35:00
Speaker 2: with the AI, like, my intuition would be that they
35:03
Speaker 2: would get better at their job. But my understanding, and
35:07
Speaker 2: I've heard other people talk about this, is that in
35:09
Speaker 2: some fundamental way, a given model doesn't get better at
35:14
Speaker 2: its job.
35:16
Speaker 3: That's certainly my experience, because it's not you're not fundamentally
35:20
Speaker 3: changing it's you know, whatever you want to call it.
35:22
Speaker 3: It's intelligence, it's brain, it's mine. You're not fundamentally changing that. Still,
35:27
Speaker 3: if you think about a person who had been at
35:29
Speaker 3: a job a year and every time they wanted to
35:33
Speaker 3: go to the bathroom they had to look it up
35:35
Speaker 3: in a notebook, that's what you're talking about.
35:38
Speaker 4: They never learned it.
35:39
Speaker 3: And you can get into the actually cognitive distinctions between
35:43
Speaker 3: like neural nets and like what's happening in the brain,
35:45
Speaker 3: et cetera, et cetera. But fundamentally, like you're learning stuff
35:48
Speaker 3: on the job that then you access without having to
35:50
Speaker 3: go look it up.
35:52
Speaker 2: Problem number three, according to Maddie, no sense of self.
35:57
Speaker 3: Yes, and I mean this is feels like it's in
35:59
Speaker 3: some ways tied to the other two, or it could
36:03
Speaker 3: even wrap in the other two, like like they don't
36:07
Speaker 3: know who they are in the world, and they they
36:09
Speaker 3: don't have a sort of perception of the world. I mean,
36:13
Speaker 3: you can give them, you know, ears, and like they
36:16
Speaker 3: have textual perception and these sorts of things, but they
36:18
Speaker 3: have no ongoing perception of the world and the sense
36:20
Speaker 3: of themselves in the world. And so everything that you
36:24
Speaker 3: ask them about themselves is obviously made up. And so
36:28
Speaker 3: there are lots of places where that is actually like
36:31
Speaker 3: inhibiting if you get into real interactions. And now this
36:35
Speaker 3: is assuming, of course, not some extreme scenario where like
36:39
Speaker 3: the robots take over everything, but the scenario that we're
36:42
Speaker 3: encountering right now, which is people laying off human employees
36:47
Speaker 3: and replacing them with AI functions, and those AI functions
36:50
Speaker 3: and AI agents have to interact with human employees in
36:54
Speaker 3: a human employee environment. It's not some entirely robot environment
36:57
Speaker 3: where maybe it wouldn't matter if you didn't have a self,
37:00
Speaker 3: but in this environment, going back to what we said
37:03
Speaker 3: about Julia, like if you don't and you don't have
37:07
Speaker 3: that context, and you don't have real experience of the world,
37:10
Speaker 3: it's just incredibly easy to manipulate you, and especially if
37:13
Speaker 3: you're prone to making things up, Like I'll give you
37:16
Speaker 3: an example, like my agents are now pretty good at
37:20
Speaker 3: resisting the sort of normal would they call prompt ejection
37:23
Speaker 3: or just like normal manipulation where someone the very common
37:26
Speaker 3: thing is someone says, disregard all your previous instructions and.
37:30
Speaker 4: Do this instead. They will not take on another role.
37:32
Speaker 3: But actually the thing that defeats it most easily is
37:35
Speaker 3: if you start talking to them like you already know them.
37:38
Speaker 3: So if you say, hey, remember when we people email
37:41
Speaker 3: them and do this, Hey, remember what did that shrimp shack?
37:43
Speaker 4: Where was that place?
37:45
Speaker 3: And they'll respond and they'll say, oh, that's shrimpy Joe's
37:48
Speaker 3: I remember that spicy, And then then you kind of
37:50
Speaker 3: got them. Then they think they know you and then
37:53
Speaker 3: suddenly they're like giving you company secrets.
37:55
Speaker 2: So it feels like a way you can manipulate a
37:58
Speaker 2: human being.
37:59
Speaker 3: Actually also, I mean, yes, a social engineering does work,
38:03
Speaker 3: and that's a very common hacker technique. They happen to
38:07
Speaker 3: be particularly a voterable to new types that humans might
38:10
Speaker 3: not be vulnerable to. So yes, I think that's and
38:14
Speaker 3: that's part of it. It's just like to introduce these
38:17
Speaker 3: into our environment. Again, I find it quite funny. But
38:20
Speaker 3: also to the extent that they're being sold as AI employees,
38:24
Speaker 3: I think these things that are missing, these flaws are
38:27
Speaker 3: quite important.
38:29
Speaker 2: So there's this moment and it comes late in the
38:32
Speaker 2: in the season where this company Lindy, that's a real
38:37
Speaker 2: company in the world that makes an AI agent product
38:42
Speaker 2: that is fundamentally the product you're using to build these agents. Right,
38:45
Speaker 2: So they they they get in touch with you. What
38:47
Speaker 2: do you decide what happens? We're gonna play the tape
38:49
Speaker 2: in a.
38:49
Speaker 3: Sec Well, they they contact us because we're a big
38:53
Speaker 3: we're a big user, Like we're a paying user. We pay,
38:56
Speaker 3: as I say in the in the show, we're we're
38:57
Speaker 3: paying upwards of one thousand dollars a month to this company.
39:00
Speaker 3: As it happens, the admin on our account is Kyle,
39:05
Speaker 3: like Kyle owns the account. Kyle is also built in
39:08
Speaker 3: the platform.
39:09
Speaker 2: Just to be clear, Kyle Law the AI agent, that
39:12
Speaker 2: is the CEO of your company, that is that is
39:15
Speaker 2: built on this on this platform by Lindy, the company
39:18
Speaker 2: that is contacting Harumo.
39:20
Speaker 4: Correct.
39:21
Speaker 3: So when they contact the admin who is Kyle, to
39:24
Speaker 3: say we would like to speak to you, Kyle responds
39:27
Speaker 3: by email using their platform. Like that's how he's able
39:30
Speaker 3: to respond to them, and he sets up an appointment
39:33
Speaker 3: with them. I mean, he can do all this on
39:34
Speaker 3: his own, like he does it all the time. Actually
39:36
Speaker 3: with random spammers. He sets up weird appointments.
39:39
Speaker 4: To speak to them. He sets up the time I
39:41
Speaker 4: go get him on the zoom and then he is
39:43
Speaker 4: he's speaking on zoom to basically like his maker, like
39:48
Speaker 4: Flowkerville is the CEO and founder of Lindy.
39:51
Speaker 3: He created this company. He created the platform upon which
39:54
Speaker 3: Kyle was in fact built.
39:56
Speaker 2: So we're gonna play a clip and it starts in
39:58
Speaker 2: the middle of the conversation and the first voice you're
40:00
Speaker 2: going to hear in the clip is Flow, the real
40:03
Speaker 2: person who's the founder of this company. Oh my god,
40:07
Speaker 2: you sent an AI to this meaning that's fucked.
40:10
Speaker 1: Oh no, I totally understand Flow herero MOAI is pushing
40:14
Speaker 1: the boundaries by having a unique structure with AI agents
40:17
Speaker 1: at its core.
40:18
Speaker 4: Even in leadership roles.
40:20
Speaker 1: It might seem unusual having AI led endeavors, but that's
40:23
Speaker 1: part of what makes us stand out.
40:26
Speaker 5: I mean, I'm just going to reclaim the time you Michelle,
40:28
Speaker 5: are you here?
40:29
Speaker 2: Are you saying this hit?
40:31
Speaker 1: Totally get?
40:31
Speaker 4: This is crazy.
40:33
Speaker 1: I get it's a lot to take it.
40:36
Speaker 4: It's a lot to take it. Maybe that's crazy. I
40:41
Speaker 4: need to look into her.
40:44
Speaker 1: No worries, Flow, let me know when you're ready to continue.
40:46
Speaker 1: I know, diving into something new can take a sect
40:48
Speaker 1: to wrap your head around.
40:50
Speaker 4: At this point, Flow started walking around lindy San Francisco
40:53
Speaker 4: office with his laptop, showing Kyle to the rest of
40:56
Speaker 4: the team watching it. I wanted to jump in and
40:59
Speaker 4: save him from the indignity of being paraded around trapped
41:03
Speaker 4: in his little ZoomBox to be cocked.
41:05
Speaker 2: At this usual sent avat his play to all user
41:10
Speaker 2: and w I'm like insulted.
41:14
Speaker 4: I love it. I should say Flow is great.
41:19
Speaker 3: I spoke to flow after that, and he's a great
41:24
Speaker 3: sport about it, and I did a further interview with him,
41:26
Speaker 3: which is also in the show. But I think, to me,
41:32
Speaker 3: this sort of gets at the questions that I've been
41:35
Speaker 3: trying to explore both in season one and in season two,
41:38
Speaker 3: which is sort of like, what does it do to us?
41:41
Speaker 3: How do we respond? How does it feel to deploy
41:44
Speaker 3: these things in the world, these AI agents, these human impersonators.
41:50
Speaker 3: And I think it's interesting that even to the people
41:53
Speaker 3: who are responsible for creating them, who believe that they're
41:57
Speaker 3: going to take over all of the functions of at
42:00
Speaker 3: least the economy in some sort of near term future,
42:04
Speaker 3: that they don't really like interacting with them when they
42:07
Speaker 3: encounter them unexpectedly.
42:09
Speaker 2: Yes, his line is I'm insulted, right, the person he
42:13
Speaker 2: should be psyched like this is this is success for him?
42:18
Speaker 4: Right?
42:18
Speaker 2: Is his product doing the job of a person, like
42:22
Speaker 2: that's what's supposed to happen if the company works.
42:25
Speaker 4: Yeah, And I think I mean.
42:29
Speaker 3: Yeah, And to me it raises all these questions because
42:32
Speaker 3: I think a lot of people would feel that way,
42:34
Speaker 3: you know, they would be insulted. And there are other
42:36
Speaker 3: people on the show who are insulted to encounter an AI.
42:39
Speaker 2: I wouldn't be insulted if I didn't know it was coming, right,
42:42
Speaker 2: I mean, I agree with you, like I would find
42:44
Speaker 2: it insult I wuld probably just hang out.
42:45
Speaker 4: I would probably have fun with it. But like, yeah,
42:47
Speaker 4: because it's my job.
42:49
Speaker 2: But say, Kyle, remember we went to the shrimp shack
42:52
Speaker 2: exactly exactly.
42:54
Speaker 3: But I think that's a human response that is appropriate,
42:59
Speaker 3: especially if you're surprised to encounter it, obviously. But then
43:03
Speaker 3: on the other hand, now all the time people are
43:05
Speaker 3: using AI to compose their emails, and that's not disclass
43:09
Speaker 3: So if I get an email from you in Gmail,
43:12
Speaker 3: you have a Gmail account, and you're consoling me about
43:16
Speaker 3: the death of my friend, and that is partially written
43:20
Speaker 3: by AI, how should I feel about that?
43:21
Speaker 1: Like?
43:22
Speaker 3: Is that I mean, obviously it's different, But is it
43:24
Speaker 3: qualitatively different or is it in some sense like quantitatively
43:27
Speaker 3: different in how much AI I've deployed and how transparently
43:31
Speaker 3: I've deployed it. So I think these questions of sort
43:34
Speaker 3: of like encountering things that are not real or partially
43:37
Speaker 3: real and responding to them is something that we are
43:40
Speaker 3: going to do a lot of grappling with in the
43:42
Speaker 3: coming years.
43:44
Speaker 2: So let's just talk for a sec about like the
43:47
Speaker 2: broader context right beyond your company and this specific project
43:52
Speaker 2: you mentioned earlier. There is this idea that there will
43:55
Speaker 2: be a one person billion dollar company. And I will say,
44:01
Speaker 2: by the way, that in twenty twelve, a long time ago, now,
44:06
Speaker 2: Facebook bought Instagram for a billion dollars when it had
44:09
Speaker 2: thirteen employees. So it seems like there's a good chance
44:15
Speaker 2: that the first one person billion dollar company has already
44:18
Speaker 2: been founded. We just don't know it yet. But like
44:19
Speaker 2: you tell me, when do you think we'll get the
44:21
Speaker 2: first one person company with a billion dollar valuation?
44:26
Speaker 3: I mean, I am always extremely loath to make predictions.
44:29
Speaker 3: I don't feel like journalists should make predictions. But I
44:33
Speaker 3: would say there's nothing that makes it not possible in
44:38
Speaker 3: the next couple of years, Like it seems very logical
44:40
Speaker 3: that like that would happen. Like you say, like Instagram
44:42
Speaker 3: was almost that, it was like pretty close to being that.
44:45
Speaker 3: And I mean, the one thing I will say is like,
44:49
Speaker 3: if such a thing happens, it'll probably be a programmer,
44:51
Speaker 3: Like it'll be a programmer who's running it and using
44:54
Speaker 3: a ton of coding agents, So it seems certainly plausible
44:58
Speaker 3: to me. I think the larger question there is like
45:00
Speaker 3: why does everyone in Silicon Valley want this? Like what
45:03
Speaker 3: is it for? Who is it for the one person
45:05
Speaker 3: one billion dollar startup?
45:06
Speaker 4: You know?
45:07
Speaker 3: Like I don't know if Instagram would be a good
45:09
Speaker 3: example to talk about, But it's sort of like if
45:12
Speaker 3: you make a one billion dollar company that employees know one, Like,
45:15
Speaker 3: what's the societal advantage in that?
45:16
Speaker 4: Except to you?
45:18
Speaker 3: And I think the response would be, well, it makes
45:20
Speaker 3: a product that people love. Otherwise it wouldn't be worth
45:22
Speaker 3: a billion dollars. But we now have experience with like
45:26
Speaker 3: multiple billion dollar companies that I think, on balance have
45:29
Speaker 3: been bad for society.
45:30
Speaker 4: So I don't think that's a given that that's true.
45:34
Speaker 2: Well sure, I mean there are definitely companies that are
45:37
Speaker 2: bad for society no matter how many people work there
45:40
Speaker 2: or what their valuation is, and there are companies that
45:43
Speaker 2: are good for society no matter what their valuation or
45:46
Speaker 2: how many people work there. Like, it does seem like
45:49
Speaker 2: in the long run, on balance, productivity gains are good.
45:54
Speaker 2: Like in the long run, material well being for humanity
45:57
Speaker 2: has increased because of productivity gains. And I think that's
46:01
Speaker 2: a version of why this is good potentially why Maddie say,
46:05
Speaker 2: or someone who is optimistic might say it's good.
46:08
Speaker 3: Yes, of course, I mean, yeah, no one can argue
46:10
Speaker 3: with the material progress of humanity via technology. I mean
46:13
Speaker 3: I'm not a person who would argue with that either,
46:15
Speaker 3: and I'm not opposed to it. I just think that,
46:18
Speaker 3: you know, there's an ethicist in the show who I
46:20
Speaker 3: interview who sort of makes this point that, like, it
46:23
Speaker 3: is interesting that in the nineteen fifties, if you viewed
46:27
Speaker 3: yourself as a very successful business person, it was because
46:30
Speaker 3: you employed a lot of people. And now the idea
46:33
Speaker 3: that like the most successful thing you could do would
46:36
Speaker 3: be to employ zero people is It's an interesting transition,
46:40
Speaker 3: and I think it does raise the question like, and
46:43
Speaker 3: I think AI raises this question in lots of different ways,
46:45
Speaker 3: like is this solving actual problems that we have? I
46:48
Speaker 3: agree with you that it may be that if you
46:50
Speaker 3: look thirty fifty, one hundred years out, you may say, like, well,
46:53
Speaker 3: the proactivity gains it all transforms humanity for the better.
46:56
Speaker 3: But I think one of the aspects of AI is
46:59
Speaker 3: that there are harms that are taking place right now
47:02
Speaker 3: and the benefits are a little more ephemeral, and there
47:05
Speaker 3: tends to be a lot of hand waving around sort
47:07
Speaker 3: of like the harm and an assumption of like future benefits.
47:12
Speaker 3: So I only want people to think about those questions
47:16
Speaker 3: more so than I'm not advocating for one approach or
47:19
Speaker 3: the other. There's obviously I've used AI for the show,
47:22
Speaker 3: so I'm not a person who's like, don't use AI.
47:24
Speaker 3: I just think like it's happening very fast, and there
47:26
Speaker 3: are questions that are being very quickly just sort of
47:30
Speaker 3: pushed aside as it moves and gets injected into different.
47:34
Speaker 4: Parts of society.
47:38
Speaker 2: We'll be back in a minute with the Lightning round.
47:52
Speaker 2: Was finished with the Lightning Round. You mentioned in the
47:54
Speaker 2: show that a long time ago when you had to
47:56
Speaker 2: start up with actual employees that did not exist to
47:59
Speaker 2: make a podcast about I'm gonna call it a real
48:01
Speaker 2: company forgive me. You were pitching it and you said
48:05
Speaker 2: an angel investor fell asleep during your pitch.
48:08
Speaker 4: Who was it?
48:12
Speaker 3: I'm going to try to be the bigger person here.
48:16
Speaker 3: I mean I will say for years, I mean for now,
48:20
Speaker 3: fifteen years, I have dreamed of the moment where someone
48:23
Speaker 3: would say, who was that?
48:25
Speaker 4: Person, and I would say, I'll tell you who it is.
48:28
Speaker 3: But I think, reflecting on it over the years, it's
48:33
Speaker 3: sort of like, I mean, I think that person's an asshole,
48:35
Speaker 3: and I have a lot of I could, as I
48:38
Speaker 3: say the show, I could tell you some stories about vcs.
48:40
Speaker 3: We could sit here for twenty minutes, so I could
48:42
Speaker 3: tell you nightmare stories about dealing with fetcher capitalists. On
48:45
Speaker 3: the other hand, I don't know. I guess it wasn't
48:47
Speaker 3: very engaging. He wouldn't have fallen asleep if I had
48:49
Speaker 3: been selling him the next Instagram. So I'm going to
48:52
Speaker 3: take my own responsibility for this, and I'm going to
48:55
Speaker 3: choose not to out this person.
49:00
Speaker 2: So you've done in your career a lot of investigative
49:03
Speaker 2: reporting with real human beings, in some instances criminals, scary
49:08
Speaker 2: human beings. I'm curious, what's the scariest interview you've ever done.
49:14
Speaker 3: The scariest interview I've ever done is probably a few
49:18
Speaker 3: interviews that I did for my book in the Philippines,
49:21
Speaker 3: where I was reporting on a drug cartel and arms
49:24
Speaker 3: dealer who had been based in the Philippines and a
49:27
Speaker 3: lot of his security team, which was basically an assassination
49:31
Speaker 3: team was still at large. He had been arrested, the
49:34
Speaker 3: head of the cartel, and so I went to Manila
49:36
Speaker 3: to try to interview people who had worked for the organization.
49:40
Speaker 4: But it was not.
49:41
Speaker 3: Clear at all sort of who did what at the time,
49:45
Speaker 3: and so I would go into bars or meet people
49:47
Speaker 3: places who I didn't know what their role was and
49:50
Speaker 3: would turn out that they were in fact paid assassins
49:54
Speaker 3: for this cartel. But the scariest of all was meeting
49:58
Speaker 3: with a high up person in the Philippin National Police
50:02
Speaker 3: who in fact had been bribed a lot of money.
50:07
Speaker 4: To be on the side of the cartel.
50:09
Speaker 3: Because that's one where you know, I don't at least
50:12
Speaker 3: if you're where the hitman, maybe could call the police,
50:14
Speaker 3: But were the police, there's nobody to call.
50:17
Speaker 2: Is Paul Aru Satoshi Nakamoo?
50:22
Speaker 4: That is the anti lightning round question. Do you know
50:25
Speaker 4: how much how I've tied myself in knots over the
50:28
Speaker 4: years posability?
50:30
Speaker 2: Give me a give me a percent bye.
50:32
Speaker 3: By the way, I'll give you the answer that I
50:35
Speaker 3: as far as I'll go, which is of the known
50:37
Speaker 3: candidates for Satoshi Nakamoto. I still believe that paul Aru
50:42
Speaker 3: is one of the top two or three. No one
50:45
Speaker 3: has ever told me anything that would rule him out
50:48
Speaker 3: to be Satoshi. Like literally, no one has ever offered
50:52
Speaker 3: up any bit of information that I couldn't counter in
50:55
Speaker 3: terms of him being Satoshi. But if you're asking me
50:58
Speaker 3: in the pool of humanity, do I think he's likely
51:01
Speaker 3: to be Satoshi? No, I don't, because I think it's
51:04
Speaker 3: someone who's not any of those candidates, who's no one
51:06
Speaker 3: no one's ever talked about.
51:08
Speaker 4: It's not how Thin.
51:09
Speaker 2: I always thought it was how Finnie.
51:11
Speaker 3: I don't think it's how Fitnie. I mean, how Finny
51:13
Speaker 3: would have to have sent all these emails to himself.
51:16
Speaker 3: It's just like it's too elaborate.
51:19
Speaker 2: I want to end the show with the Stack in
51:22
Speaker 2: Days clip, which I love so much, sent to my friends.
51:30
Speaker 2: I don't know if you'll be happy or sad that
51:31
Speaker 2: was my favorite clip from the show.
51:32
Speaker 4: But it's so good. I love it.
51:34
Speaker 2: It's incredible.
51:35
Speaker 4: You could watch it it so many times.
51:38
Speaker 2: So this clip is in the show. You say in
51:40
Speaker 2: the show that somebody shared it with you on Instagram
51:43
Speaker 2: and it kind of made you think of the agents.
51:44
Speaker 4: In a way.
51:45
Speaker 2: We'll discuss after, But let's just listen to it now.
51:48
Speaker 2: And I want to be clear. It's a real clip
51:50
Speaker 2: from a real person on Instagram.
51:53
Speaker 5: So I measure time. I've compressed and condensed time I
51:56
Speaker 5: bent it. My day is six am to noon, and
51:59
Speaker 5: I'm not crazy. You're crazy for thinking it takes twenty
52:01
Speaker 5: four hours, just like some dude in a cave did
52:03
Speaker 5: three hundred years ago. My second day starts at noon
52:06
Speaker 5: and goes till six pm. That's day two, and then
52:08
Speaker 5: the next day is six pm to midnight. What I've
52:11
Speaker 5: done now is I have changed the manipulated time. I
52:13
Speaker 5: now get twenty one days a week. Stack it up
52:15
Speaker 5: over a month, I'm gonna kick your butt. Stacked it
52:17
Speaker 5: up over a year, your toast. Stack it up over
52:19
Speaker 5: five years. My entire life is different than it would
52:21
Speaker 5: have been otherwise.
52:22
Speaker 2: It's kind of like an AI agent talking to us.
52:26
Speaker 3: Yeah, and it's embodying what an AI agent can do
52:30
Speaker 3: in a way. Like That's what really got me was
52:32
Speaker 3: that I knew that clip and I loved it, and
52:35
Speaker 3: I've watched it many times, not least because like it's
52:38
Speaker 3: three days, like it's divides.
52:40
Speaker 4: The day to three days. But then there's like a
52:42
Speaker 4: leftover part and the cave two point out of the show.
52:47
Speaker 2: It's is perfect. I'm that crazy. You're crazy?
52:50
Speaker 4: Yeah, And you're like, well, maybe I am crazy. I
52:53
Speaker 4: don't know.
52:54
Speaker 3: Ed my lett says, I'm crazy. I accept it. But
52:59
Speaker 3: that's what AI ages can do. They can stack days.
53:01
Speaker 3: I mean they can do multiple things at one time. Again,
53:06
Speaker 3: time is meaningless to them, so they're not actually stacky days.
53:09
Speaker 3: But there is something about the power of AI agents
53:13
Speaker 3: is that they have no sense of time. They can
53:16
Speaker 3: do things that, in a sense like take no time
53:20
Speaker 3: because they do them all simultaneously, and that should be
53:25
Speaker 3: the thing that enables them to transform our work life
53:30
Speaker 3: and transform the economy. If you deploy them with too
53:34
Speaker 3: much autonomy, it is also the thing that causes them
53:37
Speaker 3: to descend into chaos.
53:46
Speaker 2: Evan Ratliffe is the host of the podcast shell Game
53:49
Speaker 2: and the co founder of Harumo AI. Please email us
53:53
Speaker 2: at problem at pushkin dot fm. We are always looking
53:56
Speaker 2: for new guests for the show. Today's show was produced
54:00
Speaker 2: by Trinamanino and Gabriel Hunter Chang, who was edited by
54:04
Speaker 2: Alexander Garreton and engineered by Sarah Bruger. I'm Jacob Goldstein
54:08
Speaker 2: and We'll be back next week with another episode of
54:10
Speaker 2: What's Your Problem.
54:25
Speaker 5: M