Latest / Founder Mode / When AI Agents Go Rogue

Transcript

When AI Agents Go Rogue · Founder Mode

• 0:00
  But I think the thing that is, is funny now because I'm seeing
• 0:03
  like posts all over the Internet and friends are calling me and a
• 0:06
  lot of people are trying these different tools that are now
• 0:08
  like, oh, AI kind of went crazy on me or it didn't do what I
• 0:11
  thought it would do is kind of no different than what we've
• 0:14
  seen over the past couple years of like you give something to
• 0:17
  ChatGPT and it kind of answers a little wrong or you give
• 0:19
  something to Claude and it kind of answers wrong.
• 0:21
  But when you're asking a question, you're able to kind of
• 0:24
  like correct it in real time. You don't notice the wrongness
• 0:27
  in the in the same way. Hey, welcome to founder mode.
• 0:45
  Hey, Kevin. You know, it's funny.
• 0:47
  This is kind of, I hit the record button.
• 0:49
  You, you were doing something else.
• 0:50
  And, you know, suddenly now here we are back, back in the studio
• 0:53
  doing the thing. No agenda this first time ever.
• 0:56
  We don't know what we're going to talk about.
• 0:57
  It's kind of fun. No, I have a bunch of things I
• 0:59
  always like to start with. Like, what's the trickiest thing
• 1:02
  you've had to solve in the last? Week.
• 1:04
  Well, that's good. It's kind of like Elon's text to
• 1:06
  what? No.
• 1:07
  Elon's is different. Elon's is.
• 1:08
  My accomplished. Last week, what did you get done
• 1:10
  this week? I actually gave you a softer
• 1:12
  question. Yeah, I didn't ask what you
• 1:14
  achieved. This is the trickiest thing to
• 1:15
  get solved. I don't want to talk about what
• 1:16
  I've achieved. I'm not proud of it, but it's
• 1:18
  Monday, so I think it's good. We're we're recording this on a
• 1:20
  Monday. So yeah, the, the, the thing I
• 1:22
  had to solve this week, you know, you know full well what it
• 1:24
  is. And it's funny because I live in
• 1:27
  a household where I am the non-technical person.
• 1:29
  And then my good friend Kevin is, is technical and, and my
• 1:32
  wife is technical. And so at all times I'm hearing
• 1:34
  about how dumb I am basically because I can't figure out
• 1:37
  technical things. But I will say to your credit,
• 1:39
  you're much kinder than my wife is about the things that I can't
• 1:42
  figure out. And so so the the challenge of
• 1:45
  the past week was, you know, we're all wrestling with these
• 1:47
  new agentic kind of operating systems, right?
• 1:50
  You've got perplexity computer just launched.
• 1:52
  People are using Co work and called code to do things.
• 1:54
  You've got the open claw people. I'm not going to disclose which
• 1:57
  of those things I'm in because I don't want to create new vectors
• 2:01
  for infiltration into my system. But I'll just say that I'm using
• 2:04
  one of these, you know, AI, chief of staff, agentic kind of
• 2:08
  frameworks, and I'm having a really hard time getting it to
• 2:11
  work the way I want to. And so that's been the biggest
• 2:13
  challenge and also the biggest opportunity.
• 2:15
  Like I'll just frame it the right way, which is like, I got
• 2:17
  to do some learning. I got to learn a little bit more
• 2:20
  about what's happening at the command line and what the what
• 2:22
  the code is doing when it's, you know, the what the instructions
• 2:25
  are telling the computer to do. So I can get to a place where I
• 2:27
  can use it. And I'm fortunate to have here
• 2:29
  with me today my fractional harness engineer, Kevin
• 2:32
  Hendrickson. Kevin, what it was the hard
• 2:34
  thing that you were working on this week?
• 2:36
  This harness engineering, as they call it, or harness, you
• 2:38
  know, whatever is this idea that like we're all used to using
• 2:41
  chat bots and you ask it a question, it does something.
• 2:44
  And I think the newer frameworks all have two things.
• 2:48
  One is this sort of like overarching sort of long term
• 2:51
  memory, long running memory. And then I think the other is
• 2:53
  this combination of like, you know, scheduling or kind of
• 2:57
  triggers, right? And so it's like things that can
• 2:59
  happen and start so events and triggers or if you just think of
• 3:03
  it simply as scheduling because the thing just wakes up on some
• 3:06
  work periodic basis and can do work.
• 3:08
  It's like, I've sent my kid to the store to go get me milk and
• 3:11
  they brought back orange juice. You're like, like that just
• 3:14
  feels way more broken when like you had this mental model of
• 3:18
  like you were going to get milk and then you got orange juice.
• 3:20
  And like if you just were at the school store with them and you
• 3:24
  said get milk and orange juice, you're like, oh, that's simple.
• 3:26
  Just go and like just swap them. But if they've already like went
• 3:29
  to the store, bought it, checked out, came all the way home and
• 3:32
  then like you open the fridge and it's wrong.
• 3:34
  Like it feels way worse because like, and I think that's the
• 3:37
  best way that I've thought like to sort of explain it to kind of
• 3:40
  people that aren't in the notion of like why it feels so much
• 3:43
  worse that AI got it wrong. And then if you add the next
• 3:45
  layer, which is like what we're doing at the company where we're
• 3:48
  trying to, you know, essentially sell AI or sell this idea, it's
• 3:52
  even more painful. Or you're in your role where
• 3:55
  you're like selling things that have applied AI to a business
• 3:58
  problem and then that doesn't come out perfect or the way that
• 4:02
  you, you know, each kind of anticipated.
• 4:04
  It it's still being evaluated as if a human was doing it.
• 4:06
  Yeah, yeah, and it's actually worse.
• 4:09
  No, the bars actually held higher because when a human does
• 4:12
  it like again, it's back to the kid.
• 4:13
  Like if my kid gets. Yeah, you're not going to the.
• 4:16
  Food wrong. You're like, I get it, they're
• 4:17
  still learning. But if you tell a computer to do
• 4:20
  something or you tell AI. It's this idea of of
• 4:23
  deterministic versus non deterministic.
• 4:25
  And you're, you're assigning it like a higher level of
• 4:29
  consistency because you're like, oh, this computer is definitely
• 4:31
  going to get it right. Computers just do things.
• 4:33
  It's one or. Zero.
• 4:34
  I see all these, you know, benchmarks.
• 4:35
  Oh, it's PhD level this it can complete the Mcats and the Lsats
• 4:40
  and it's like smarter than the average human.
• 4:42
  And, and when it fails to get the, the milk and visor orange
• 4:45
  juice instead, you know, and there's the distinction of like,
• 4:49
  what model are you using? Is it the latest?
• 4:51
  Is it the smartest is to do a lot of thinking.
• 4:53
  Did it do it really quick? Like there's so much nuance
• 4:55
  there. And it becomes very difficult
• 4:57
  for a layperson like me to use the tools to their maximum
• 5:01
  potential because of this. Like harness engineering,
• 5:04
  whatever you want to call it, is like picking the right model to
• 5:06
  do the task and setting up the system and orchestration and
• 5:09
  architecture is actually more important than the just like,
• 5:13
  hey, go do the thing. And part of this is guardrails,
• 5:15
  part of this is defining like how to check results.
• 5:18
  But I think it doesn't solve like the problems that you and I
• 5:21
  are describing, where it's like you've asked it to do a like
• 5:24
  more complex workflow or you've asked it to wake up on some
• 5:27
  recurring tasks. Like I just.
• 5:29
  I'm not there, yeah. But I do think that's the next
• 5:32
  version of this or next kind of like whatever frontier of this
• 5:35
  is to like, how can we make it so that they can self correct?
• 5:38
  How can it be easier for people to detect that it's going the
• 5:41
  wrong way? So yeah, it doesn't.
• 5:42
  You don't end up with the wrong sort of beverage in your fridge
• 5:45
  after you've sort of ran to the store and already paid for it,
• 5:47
  yeah. Yeah, and I think it connects
• 5:49
  really well with the last time we did one of these episodes
• 5:51
  where it was just you and I and we're talking about AI agents as
• 5:54
  the as the new employees, right. And it and the question becomes
• 5:58
  as a founder, like, you know, are you an early adopter and you
• 6:02
  start playing with all these tools where you have like your
• 6:04
  AI chief of staff and it is imperfect, but you're willing to
• 6:07
  put in the work to kind of fix it when it breaks and deal with
• 6:10
  the kind of, you know, early adopter challenges.
• 6:13
  And like then on the kind of late adopter, late to the party,
• 6:17
  like when does it go mainstream? When can anybody use it?
• 6:19
  When is it, you know, ready for for neophytes like me?
• 6:22
  I think that those are some of the questions I'm thinking
• 6:25
  about. Yeah.
• 6:26
  And I think the best way that, again, I keep using this as an
• 6:28
  analogy, is like, have you had an intern before?
• 6:31
  Have you hired a junior assistant before?
• 6:34
  And you need to keep reframing that.
• 6:36
  And you set the expectations correctly.
• 6:38
  Because I think that's where people get into trouble is
• 6:40
  they're like, oh, I've hired this new AI employee and I
• 6:43
  expect it to be perfect or I expect that it to not make
• 6:46
  mistakes or I expect that I've already corrected it once.
• 6:49
  It won't make that mistake again.
• 6:50
  And we know that humans do that all the time.
• 6:52
  You can give this same requirements.
• 6:55
  Please show up for work on time. Please do this task.
• 6:59
  And like that can be a reoccurring kind of wrongness.
• 7:02
  And I think AI is no different, partly because it is this sort
• 7:05
  of indeterministic, sort of like the, the environment can be
• 7:08
  exactly the same, the request can be exactly the same, but it
• 7:11
  can actually give you different results for both good and bad
• 7:14
  reasons. And so, yeah, I, I think it's,
• 7:15
  it's a, it's a cool thing. I think the, you know, to me,
• 7:18
  it's just like this sea of opportunity because I see like
• 7:21
  the potential for it and I see when it works well.
• 7:24
  And it's, and again, I run into people not just like you, but
• 7:26
  lots of other people that are trying to use this and also see
• 7:29
  the potential, but are struggling to sort of connect
• 7:31
  the dots. And to me that's like, well,
• 7:33
  there's businesses in there, there's things to go be solved.
• 7:35
  And it's like, just, are you doing that in in a domain?
• 7:37
  That's interesting. And I think the other thing that
• 7:40
  to switch topics a little bit that I know I've fallen into
• 7:42
  trap wise is this notion of like the dopamine loop of seeing it
• 7:46
  do things and not being able to stop working on it, right?
• 7:49
  Like I, somebody described this as a video game where you're
• 7:52
  constantly leveling up and you're like getting more gold,
• 7:56
  more resources. Like I can't stop playing.
• 7:58
  And there is a version of that where I've caught myself many a
• 8:02
  times, like one more prompt, one more tweak.
• 8:04
  Yeah, just get there. And then you know, it's a bit.
• 8:06
  Like a drug. Yeah, and I'm four hours past my
• 8:09
  quote UN quote, you know, hard bedtime.
• 8:11
  Yeah, I think the thing that I've yeah.
• 8:13
  And I mean to me, it's funny, my Co founder JJ sent me a document
• 8:16
  this morning. It was like A8 page document
• 8:18
  that he wrote. And I was like, oh, did AI write
• 8:20
  this? And he's like, no, I wrote some
• 8:21
  of it and you know, obviously used AI to clean it up and
• 8:23
  things like that. And the first thing I did was I
• 8:25
  dump it into AI and I'm like, go read this.
• 8:27
  And he's like, dude, you didn't read it Like you're not even
• 8:29
  thinking. And I'm like, I think that's a
• 8:31
  fair critique that I'm definitely not thinking.
• 8:33
  I don't think first, like I sort of like, can AI short circuit
• 8:37
  like getting my frame around? What is this 8 page document?
• 8:40
  What am I going to learn? What should I be kind of mindset
• 8:42
  should I be in as I go to read it in detail?
• 8:44
  And I think it's and it's changed the way I work for sure.
• 8:46
  And the way I think, it is harder to go in and just read an
• 8:49
  8 page document straight up or just read an article in the
• 8:51
  newsletter you know or write. A yeah, I mean, I have AI have a
• 8:54
  similar story from this past weekend.
• 8:56
  It was on like I think I don't remember it was on Friday or
• 8:59
  when it was, I think it might have been late Friday evening.
• 9:01
  There was, you know, somebody on my team, I'm interacting with
• 9:03
  them. And there was like an issue that
• 9:05
  I wanted to resolve quickly and I sent, I sent an e-mail that I
• 9:10
  had AI help me write asking for time to chat about a topic that
• 9:15
  was important. And then they replied and it was
• 9:17
  clearly an AI generated reply. And then I gave their reply, you
• 9:21
  know, like back and forth. And like, I didn't do any of the
• 9:23
  writing. Like I told AI what I wanted to
• 9:25
  have happened. Like, hey, I need this meeting.
• 9:26
  Here's why. Here's the thing I'm working on.
• 9:28
  And I, you know, like you, I'm like, hey, I'm working this
• 9:30
  weekend. So can we get on a call this
• 9:32
  weekend to talk about it? And, you know, sent the follow
• 9:35
  up and didn't get a response. And then, you know, Monday
• 9:38
  morning hits like getting a meeting invite now for Tuesdays.
• 9:40
  It's like, it's fine. That's when that person's
• 9:42
  available. I'm not, that's not the issue.
• 9:44
  It's just the AI talking to the AI and the AI reading the
• 9:47
  documents, the AI doing all the things, you know, when do we,
• 9:49
  when does the human insert itself into the loop?
• 9:51
  I think is another really interesting conversation.
• 9:53
  And in the context of these AI chief of staff systems, right
• 9:56
  now, it's so early that it feels like the human needs to be very
• 9:59
  involved. So I, I think the human in the
• 10:01
  loop thing and, and again, part of this is like, again, it's the
• 10:04
  new employee thing, man. It's like, go do this thing.
• 10:06
  Let's review it together. Go do this thing.
• 10:08
  Let's review it together. That's just a way better way of
• 10:10
  thinking and prompting versus saying, right, hey, go do this
• 10:14
  thing. And I'm assuming you're going to
• 10:15
  get all of understand all the nuances and not going to make
• 10:18
  mistakes. Like today I had a weird
• 10:19
  Internet thing. I'm actually running on my cell
• 10:21
  phone right now and I had kicked off an an agent and it drafted
• 10:26
  an e-mail and came back and said, oh, by the way, I couldn't
• 10:28
  connect to any of the systems you asked me to go connect to.
• 10:31
  But here's the summary of what I think I would have seen.
• 10:35
  And I was like, do you want me to send the e-mail?
• 10:37
  But. At least.
• 10:38
  But at least it checked with you and didn't just send the.
• 10:40
  E-mail well, I have it set up so that I never sends anything.
• 10:43
  It only does drafts and it actually can't send and so like,
• 10:46
  but the fact was it was kind of funny because in its world send
• 10:49
  is basically let me copy the draft and consider it done.
• 10:52
  But the cool thing here is like it actually was very, very
• 10:54
  wrong, right? Because it's like, no dude, like
• 10:56
  the task is useless if you're not the whole task was to go
• 10:59
  look at the system and review it.
• 11:00
  But the fact that, you know, my Internet was down or weird on my
• 11:03
  computer, it couldn't do the task.
• 11:05
  And so, yeah. Yeah, the hallucinations are
• 11:07
  real and they're and they're kind of persistent across
• 11:09
  various surfaces. I think that as you were
• 11:11
  talking, it just made me realize that like if November 2022 in
• 11:15
  chat TPT was like really the phase one for most people in
• 11:18
  terms of AI exposure. This next phase is about, you
• 11:22
  know, you don't have to sit there in front of the chat
• 11:24
  interface and wait for the response.
• 11:27
  Like what we're really talking about is like now there are
• 11:30
  agents, AI agents that go and do things and they come back later
• 11:33
  with something for you. Could could it be good or bad?
• 11:36
  I don't know. It depends.
• 11:37
  It depends how you've set things up and what what you're using.
• 11:40
  But you don't have to sit there. Yeah.
• 11:42
  And so that that's like, it feels like there's a progression
• 11:44
  here. And then, you know, what comes
• 11:45
  after the like agent that can go off and do things like when do
• 11:48
  we when do the guardrails get get further removed?
• 11:51
  And the agent is allowed to send that e-mail for you, Kevin,
• 11:54
  because you trust implicitly that it's never going to go and
• 11:58
  pretend it did the research, hope that it's right, and send
• 12:01
  the e-mail. Yeah, I think it's no different
• 12:04
  than like what's the task that it's kind of been given to do?
• 12:07
  And then what's how much are you trusting it?
• 12:09
  Right. So there's obviously computers
• 12:11
  today that land airplanes, auto land, right.
• 12:14
  And those have a number of checks and balances that have
• 12:17
  been well tested. And people have like agreed that
• 12:20
  like, and they're saving human monitoring it, right.
• 12:22
  And it has they have the ability to monitor it in a way.
• 12:25
  And so I think yeah. So I think there's there's
• 12:27
  different cases there, but I do think that like the AI stuff is
• 12:29
  going to go through a similar transition.
• 12:31
  There's going to be things where you're like, hey, it's fine to
• 12:33
  go like troll through this data or do discovery and send an
• 12:36
  e-mail to my internal team. And then there's, I mean, you're
• 12:38
  seeing it today. So I'm just thinking about like
• 12:40
  last, last thing to chat about before we wrap up, you know, all
• 12:43
  of this stuff we're kind of dancing around the what is the
• 12:47
  future look like. I had another conversation right
• 12:50
  before this that I think it would be good to surface with
• 12:52
  you, which is, and we talked about this before and, you know,
• 12:56
  maybe it's not for founders, but it's more for like everybody
• 12:58
  else. Like if you're not willing to
• 13:01
  like play with all these rules and try to figure stuff out,
• 13:04
  what do you do 5-10 years from now?
• 13:06
  Like, you know, for example, and maybe that's the wrong question,
• 13:09
  but like the answer I gave to somebody that asked me that
• 13:11
  question a moment ago was like, it used to be you tell you tell
• 13:14
  like a kid who's thinking about college, you know, go get a
• 13:16
  computer science degree, go become a developer.
• 13:18
  That doesn't seem like good advice anymore.
• 13:20
  Or maybe there's an edit to it and become an AI engineer.
• 13:22
  I'm not sure what that means anymore.
• 13:24
  But you know, I'd love to get your take on this because my
• 13:26
  answer was like, go learn a trade, like be a trades person.
• 13:29
  Like they're still going to need to be people to maintain all the
• 13:32
  robots and machines. And like when whatever form they
• 13:35
  take, like it's the like I have a plumber coming over tomorrow
• 13:38
  and like I got somebody working on my like there's stuff that's
• 13:40
  just never going to go away entirely because there's not
• 13:42
  going to be the plumbing bot that can do everything.
• 13:44
  And so like, or at least not for a very long time.
• 13:47
  So what's your take on like the way that the nature of work
• 13:51
  diverges? Yeah, so I, I do think there's
• 13:53
  this thing where it's like, are you going to be building the
• 13:55
  robots or are you going to be having the robots tell you what
• 13:56
  to do like we've heard. That my favorite Kevin line.
• 13:58
  I used it in that conversation. I think then there's the, you
• 14:01
  know, do you want to do things, you know, that don't scale
• 14:03
  right, You know, in person clinicians, you know, in person,
• 14:07
  you know, kind of local services.
• 14:09
  I think the other vector that's entertaining is like live
• 14:12
  entertainment. Like it's things that are IRL,
• 14:15
  like hosting events and like doing entertainment, going to
• 14:18
  live Broadway shows, going to live comedy shows, Like there's
• 14:20
  going to be this resurgence of things that are like human and
• 14:23
  authentically human. So I think those are some
• 14:25
  interesting things. If that's if you're more of a
• 14:27
  creative type or you're into that sort of like vibe, I think
• 14:29
  there's some interesting opportunities there.
• 14:30
  Again, it's, it's part of like where you are in life, right?
• 14:32
  Like, you know, I think if you're sort of past the working
• 14:35
  age or you're deep into your job, like, you know, the, the,
• 14:39
  the area under the curve of like potential work left to do might
• 14:42
  impact how you think about this. If you're like a new college
• 14:45
  grad. We posted this job for one of
• 14:47
  our companies. We got 2166 kind of applications
• 14:52
  of those 1000 were qualified, meaning that they had the
• 14:55
  skills, the location and the right to work in the US.
• 14:58
  And then of those we sent out around for about half of them
• 15:03
  got it. You know, we sort of like
• 15:04
  invited them to sort of participate in the interview
• 15:06
  process, which was started with a project.
• 15:08
  And then of that, about 10% responded and completed the sort
• 15:12
  of like task. And then of that, we probably
• 15:14
  interviewed another 10% of those, so 6 to 8 people.
• 15:17
  So you think of like the funnel from like 2000 people.
• 15:19
  And again, the 2000 number probably includes AI bots.
• 15:22
  But by the time we got to the 400 and definitely the 50 that
• 15:26
  sort of submitted it, they were very uniquely human or they
• 15:28
  human was deeply participating in the loop of the interview
• 15:31
  process. But all that goes to say is that
• 15:33
  like the demand for this kind of work and the people that want to
• 15:36
  go do it is like grown faster than I ever would have imagined.
• 15:39
  Like this is 10X what we got. Literally a year ago we posted a
• 15:43
  similar job, got 200 applications.
• 15:45
  This year we got 2000, yeah. And so I think that part changes
• 15:49
  my view at least of how I think about where people are going to
• 15:51
  go with what they want to do. I think on the other side of it.
• 15:54
  Would that be your advice to like teenager looking at
• 15:57
  colleges and what they're going to go do is like it's that.
• 16:00
  No, I think it's, I think it's even more meta than that.
• 16:01
  It's be able to find things to go work on where you can work on
• 16:05
  a team, find a solution. Because really it's just like
• 16:07
  problem definition. First principles.
• 16:11
  Super first principles, iterative experimentation, and
• 16:14
  then can you have a human to human conversation about it?
• 16:16
  Because like we opened up some of the videos and it was very,
• 16:19
  very hard to understand people, right?
• 16:21
  They weren't great at communicating.
• 16:22
  It wasn't even like a English it.
• 16:25
  Was it wasn't that they couldn't do the work, it was that they
• 16:27
  couldn't communicate about the work.
• 16:28
  Couldn't because it was like literally send me a 5 minute
• 16:30
  video describing what you built that was the problem.
• 16:33
  Yeah, well, and that's a great, that's a great test because
• 16:36
  communicating not only to your future bosses, but to the AI is
• 16:40
  critical. Like if you're a bad
• 16:41
  communicator you can't tell the AI what to do.
• 16:43
  If you can't tell me what you built, you can't tell A-Team
• 16:45
  member or somebody else. And so I think those are like.
• 16:47
  That's a problem. And and was that?
• 16:49
  Personal language barrier stuff. Or was that just like even
• 16:52
  setting aside like maybe English wasn't the first language It
• 16:54
  was, it was beyond that. It was beyond no like this was
• 16:57
  not an English as a first language issue, it was way more
• 16:59
  about like. These are all English speaking.
• 17:01
  Intelligently walk me through again and hub project that you
• 17:03
  built and so there's two pieces there one either you didn't
• 17:06
  build it or you didn't understand what you built
• 17:08
  because it was built by AI or you don't have a way of, you
• 17:11
  know, putting together enough thoughts to like walk me through
• 17:14
  your architecture. Hey Jason, how did you wash your
• 17:16
  car? Well, I got the bucket.
• 17:18
  Like it's there's a literal way of going to describe those kind
• 17:21
  of things that multiple ways to get there.
• 17:23
  But like the fact that somebody couldn't describe what they had
• 17:26
  done. That's a, you know, a yellow
• 17:29
  flag probably edging on red flag for where we are today.
• 17:31
  Like you're still gonna have to talk to humans in our in our
• 17:34
  jobs. Evan, that was fun.
• 17:35
  I like it when it's just us riffing.
• 17:36
  No agenda, no gas, just two guys trying to figure out AI before
• 17:39
  it figures us out. All right, Kevin, what were your
• 17:41
  top five? Yeah, AI agents feel way more
• 17:43
  broken than chat bots when they get things wrong because, you
• 17:46
  know, you see the mistake, the task is already done or it's
• 17:49
  kind of already shipped. I mean #2 the right mental model
• 17:52
  is really, like, think of this as a new junior employee #3 I
• 17:55
  mean, people going deep on AI are really working more, not
• 17:59
  less, self included. So #4 the dopamine loop's real.
• 18:02
  And, you know, just treat it like a video game.
• 18:04
  You can't stop playing. But again, you know, you do need
• 18:06
  to start a hard stop, maybe put some screen time on.
• 18:09
  And then in a world of AI generated everything,
• 18:11
  communication skills are, quite frankly, the ultimate
• 18:13
  differentiator. You've got to be able to explain
• 18:15
  what you built and what you didn't build.
• 18:17
  Yeah. Number five hits hard, you know,
• 18:19
  2000 applicants and most couldn't walk through their
• 18:21
  GitHub project. That's crazy, Yeah.
• 18:23
  Just talk to humans. It's still job.
• 18:25
  Yeah, if this episode got you thinking about how you're using
• 18:27
  AI or how you're not, share it with a founder or friend who's
• 18:30
  trying to figure that out, too. And if you haven't already,
• 18:32
  subscribe to Founder Mode. We're going to start doing more
• 18:35
  of these solo episodes alongside our interviews, so leave us a
• 18:38
  review. It helps new listeners find us.
• 18:39
  That's it for today. Because the future isn't about
• 18:42
  what AI can do for you. It's about whether you can
• 18:44
  explain what you did together.