Latest / AI for Founders with Ryan Estes / Your Next Co-Founder Should Be AI

Transcript

Your Next Co-Founder Should Be AI · AI for Founders with Ryan Estes

0:00
What if firing your entire team was actually the smartest thing
0:02
you could do? Most founders are out here
0:05
hiring people to fill roles. But what if the role itself is
0:07
the problem? Corporations always had AI
0:10
attributes. Think about it.
0:11
They work 24/7, they outlive any single human, they coordinate
0:15
thousands of moving parts, and they operate without emotion.
0:18
Does that sound familiar? So what happens when you replace
0:21
the org chart with actual AI? We're talking about full
0:24
operational companies with AICEOSAICFOS and real human
0:29
contractors getting paid and managed by the agents above
0:32
them, the AI. Chooses the go to market.
0:35
And it's structured. There's an AI agent watching the
0:37
other agents, an eye in the sky that none of them can see.
0:40
Every decisions logged, every dollar tracked.
0:43
Ethics violations escalated to the human founder instantly, and
0:46
human contractors tasked with the work that the AI cannot do.
0:50
This isn't sci-fi is running right now.
0:52
So if AI took over your company now, are you even qualified to
0:56
push back on its recommendations?
0:57
Welcome to AI for founders. I'm Ryan Estes.
1:00
This week I sit down with serial founder Dave Sifry and a good
1:03
buddy of mine. He's nine companies deep who is
1:06
literally building the world's first meta factory, an AI Co
1:10
founder system that spawns entire companies.
1:12
If you're a founder trying to scale without burning out your
1:15
team or your runway, this one's for you.
1:17
And if you want the frameworks, the agent org charts and the
1:20
lessons from the companies already running, sign up for our
1:23
newsletter at AI for founders.co, where we turn the
1:26
bleeding edge into your competitive edge.
1:28
One founder tested inside at a time.
1:30
It's free, it's fast and it might change how you build.
1:33
And if this episode cracked your brain open a little bit, drop us
1:36
a five star review your future self building with AI agents.
1:39
Well, thank you. And Speaking of building, if
1:42
you're a founder who's tired of renting attention on social
1:44
media and ready to own your audience, listen up.
1:48
Inbox Alchemy is the number one distribution motion for founders
1:51
who are serious about growth. We're talking about 25,000
1:55
targeted subscribers delivered in your first six months.
1:57
Real humans in your exact ICP P Not Random.
2:01
Follows weekly crafted founder newsletters, partnership
2:04
opportunities, and a personal brand that compounds like
2:06
equity. You write once your audience
2:09
grows forever. No algorithm, no gatekeepers.
2:12
Just your words landing in the inboxes that actually matter.
2:15
Ready to stop performing for the feed and start owning that
2:18
relationship? Go to inboxalchemy.co and let's
2:21
build something that lasts. That's Inbox alchemy.co.
2:27
All right, Dave, welcome back to the show.
2:30
So good to see you again my friend.
2:32
Ryan it is such a pleasure. As always, it's way.
2:37
It's been way too long. Been way too long now to catch
2:40
people up. Dave Sifry is my favorite
2:43
entrepreneur therapist. He is not licensed in the state
2:47
of Colorado. However, the man has so many
2:50
wins and taken the abuse over the years that when I'm crying
2:54
in my coffee in the morning, I called Dave because he's always
2:57
got an encouraging word and he is like one of my favorite
3:00
entrepreneurs largely because of shots on goal.
3:03
You're always at the absolute tip, tip of the spear as far as
3:06
technology. The founder of Warm Start, which
3:08
was a great sponsor of the the podcast.
3:11
And Dave, you're here because we talked the other day and I was
3:14
like, dude, we got to do this on a podcast.
3:16
So basically what you're creating is an AI Co founder and
3:21
you've named it Meta Factory. Now beyond that, I would love
3:25
for you to just kind of walk us through.
3:27
What you're working on? All right.
3:29
Well, First off, I think like many technology enabled
3:35
entrepreneurs, I have been number one entranced by the
3:40
incredible increase in capacity that AIS and the new modern
3:46
models have been able to do. How people have been putting
3:50
them together into these kinds of agentic workflows where you
3:53
can, you know, write us back in, in hell even have it write us
3:57
back for you. And then go off and do it, have
4:01
lunch, come back 2530 minutes later and it's, you know,
4:04
finished written code, reviewed the whole shebang, or it's done
4:08
a bunch of marketing landing pages for you and so on and so
4:12
forth. And like you, I, I think the,
4:18
the mind glitters and glistens because you say, oh wow, what if
4:24
I could find some way to be able to get all of these different
4:30
work flows, processes and systems to work together to
4:34
actually do something that operates much like the way a
4:39
company does. And one of the thoughts that
4:45
I've been having after building 9 companies, we've all been
4:51
talking about artificial general intelligence, AGI.
4:56
And what if AGI is already here, but it's called the
5:02
organization. It's called the Corporation.
5:06
What? Do you mean now that?
5:10
Well, think about what corporations do.
5:14
They're smarter or they, they are capable of doing far more
5:17
work than anyone individual human being can do.
5:21
They're they operate 24/7, 365. They, in general, are able to
5:32
produce, you know, a level of productivity that any one
5:37
individual person can't do. Now, by the way, it doesn't mean
5:41
that they're the smartest things on the planet.
5:44
Yeah, like, I'm not talking about super intelligence here,
5:48
but I am talking about the ability to be able to perform
5:52
and do things that an ordinary individual human being is
5:57
incapable of doing. And you know, you want to build
6:02
the dam. I mean, heck just talked about
6:04
things like organizing capital and and being able to then move
6:10
it into the right places and be able to have a gentic units, IE
6:16
human beings be able to get things done.
6:20
And when you then step back and you say, OK, well now we have
6:24
all of these really interesting pseudo intelligences that are
6:29
getting better and better and better at things like reasoning.
6:34
And what if you could put them together along with some of the
6:38
things that we've learned in organizational theory and create
6:42
systems and processes that allow these agents to operate with a
6:51
goal in mind, that allow them to handle unforeseen circumstances,
7:00
that allow them to get their work checked and verified to
7:05
make sure that they're operating well.
7:08
So these are all things that we have built organizations for
7:12
because they tend to, in the long run, do better than an
7:16
individual human being doing all of those things themselves.
7:20
And that's not to say that human beings aren't incredibly
7:25
important for a whole bunch of reasons.
7:27
Like I'm not one of these people who's coming out and saying
7:30
that's it, you know, the, the, you know, humans are, are just
7:34
going to be left on the beach and, you know, forget it.
7:36
Like a is taking over and and it's all done.
7:39
I think that what I've learned by building these kinds of
7:44
agentic organizations, and it started as an experiment, that
7:48
it is now turned into two fully running companies, one of which
7:53
is already cash flow positive. And we're like plowing the money
7:58
back in to the operations that the AICEO thinks is actually the
8:07
the best ones. And namely it's like right now
8:09
it's not what I would have done necessarily.
8:12
It's actually plowing it into go to market strategies, you know,
8:17
like me, I'm a coder. Like I was like, Oh yeah, yeah,
8:19
yeah, let's build more products. Like no, no, no.
8:21
What we need to do is we need to actually go out and get more
8:25
distribution and talk to more customers and and see how we can
8:28
get more people there. So the the idea was how far
8:32
could you push an experiment like this?
8:35
Is it possible to be able to create an AI Co founder where I
8:41
give the general high level goals?
8:43
I set the direction and then I allow the AIS to work within
8:49
themselves. What are the lessons that you
8:52
learn from that? What are the principles that
8:54
need to happen in order for that to actually be effective?
8:58
And really it's also just looking at all of those a damn
9:03
LinkedIn posts with all those people were like, I fired my
9:07
entire company and I hired 240 AI agents that now do all of my
9:12
stuff for me. And just type in an AI agent in
9:15
the comments and we'll send you a copy of my marketing
9:19
materials. And you're just like, and yet
9:22
somehow those things must be working because people keep
9:26
putting them in. And what I've learned from this
9:30
is a bunch of really hard, hard learned lessons about what these
9:35
systems can and can't do today. Some really surprising areas
9:39
where I think that there's a ton of opportunity that is currently
9:44
unfulfilled. And I actually have a tremendous
9:47
amount of hope and expectation that humans are not all of a
9:52
sudden going to get replaced. In fact, we're deeply, deeply
9:55
needed when these kinds of organizations, I'll call it
10:00
perhaps a hybrid human agentic organization, really come into
10:06
flow in the next year or two. Quick side note for founders who
10:11
using AI to write content, you've probably noticed this.
10:15
It's incredibly fast, but it doesn't really Dr. meaningful
10:18
traffic. And honestly, that's because
10:20
most AI generated content is just average.
10:23
There's no real insight, no depth, no structure, missing
10:27
things like internal linking, and Google's getting very good
10:29
at filtering that out. So instead of trying to just
10:32
prompt better, I actually came across something interesting
10:35
called Neenah. It's a system that combines
10:38
human strategy and review with AI writing to create high
10:41
quality SEO optimized blog content that's actually designed
10:45
to rank and keep ranking. What I like about it is that
10:48
it's done for me. Their team handles the topic
10:50
strategy in the content review and data markups for Google, and
10:55
turns blog posts into actual long term traffic assets.
10:59
So if you're a founder thinking seriously about organic growth,
11:02
this is one of those things worth paying attention to.
11:04
You can check it out at trineena.co.
11:07
That's TRYNIN a.co trineena.co. In three years, AI fluency will
11:15
be a baseline expectation for every college grad, not a
11:18
differentiator. It'll be a requirement.
11:21
The students getting ahead are the ones building that skill
11:23
right now in high school and college before it becomes table
11:26
stakes. ainativestudent.com is the program doing that weekly
11:31
one-on-one coaching, live cohort sessions, on demand courses and
11:36
the fastest growing AI chat for students all tailored to your
11:40
students specific career path. Pre Med they learn how AI is
11:43
transforming diagnostics. Pre law they learn legal AI
11:47
tools business they learn to think like an AI power operator.
11:50
1 flat membership less than a private tutor.
11:53
Cancel anytime. And if you're a founder
11:55
listening to the show, you already know the cost of falling
11:57
behind. Don't let your kid pay that
12:00
price. Go to ainativestudent.com, apply
12:03
in a couple minutes and get your student into the cohort that
12:06
changes their trajectory. AI native student.com and send
12:09
them to the chat. It's totally.
12:11
Free. Amazing.
12:13
And I, I want to hear exactly what you've done, what you've
12:17
done, Dave. But before I'm going to take
12:19
pause a little bit by be calling myself an agentic unit.
12:25
I know it's so funny, right? Like you, you, you know, you,
12:28
you start thinking about these things and all these terms and
12:31
it can be incredibly dehumanizing.
12:33
And you know, you, you hear the, the Sam Allmans of the world
12:37
talking about how, you know, they, they start to, Oh well,
12:40
you know, think about how long it takes to to train a human
12:44
being in terms of the energy. And you're just like you.
12:47
Know like bro, stop talking yeah, it's like.
12:50
Sam, stop digging. Right?
12:52
Like stop digging. There's there's far too many
12:57
flesh units in our server repository.
13:01
Like dude what? Come on man.
13:03
Right. Like let's just remember, you
13:05
know, what's the cart and what's the horse, right?
13:08
And I, I think that as technologists, we're talking in
13:13
analogies. And so those analogies can then
13:17
be misquoted and taken out of context.
13:20
And I'm not here, I'm no Sam Altman apologist on any of this
13:23
stuff. But I can certainly see living
13:26
in this world and, you know, building and interacting with
13:30
these kinds of agentic systems that I start to think in terms
13:35
of analogies about, you know, well, what are some of the
13:38
things that have worked with human beings and getting human
13:40
beings to work Better Together or being more organized?
13:44
And by the way, corporations are absolutely not aligned with
13:48
human values. Right.
13:51
So, you know, you start saying, well, gosh, we've got a, an
13:54
alignment problem with, you know, artificial general
13:58
intelligence. Yeah, we do.
13:59
By the way, we have an alignment problem with things like the
14:03
corporation. And, you know, here in the
14:05
United States and most other places in the world, it's even
14:08
worse. Corporations are treated as
14:11
persons by the law. Yeah.
14:14
So wait a minute, That's both bad news and good news in the
14:20
sense that, well, if that's the case, then what can we learn as
14:26
we're like living out on the frontiers and we're trying to
14:29
bring these other intelligences more deeply into the plan, do
14:37
learn cycle. And most importantly, where are
14:42
human values in all of this? And so one of the things that
14:46
I've been thinking about is like, as creators of the reality
14:52
that our children and our grandchildren are going to be
14:54
living in, what are the values that I want to make sure get put
15:00
into these kinds of systems so that we don't go off the rails 5
15:06
years from now, 10 years from now, 20 years from now?
15:11
What are those values, Dave? I mean, I look at it #1 as
15:20
there's a set of principles that that I would put together.
15:24
In fact, I actually, I can't even pull this out.
15:28
This is actually the end of something that I had been
15:32
building. Yeah.
15:34
Let me show you something. Yeah, please do.
15:41
It does seem to be a broader concern too that you know the
15:45
the heads of of all these LLM companies, whether it's
15:48
Zuckerberg, Elon, or Sam. It's like bro, I'm giving you
15:52
the benefit of the doubt here but some of the stuff you're
15:53
feeding me seems a little SUS. Well, yeah, because in the end,
15:58
you're not. You're not really thinking about
16:02
what are the ultimate goals, Like what's the ultimate outcome
16:06
that we're looking for here? Are are we looking for something
16:10
that is so cutthroat and that has its own, you know, sets its
16:14
own goals that, you know, you end up with some incredibly
16:19
skewed outcomes? Or are we looking at ways that
16:22
we build both things like governance and verification and
16:31
a deeper sense of, of like what is the identity of each of these
16:37
different agents and that those things are are retained and
16:41
perhaps even reviewed by humans? And how, for example, do we
16:47
build accountability into these systems?
16:50
So one of the things that I've made sure to build in the
16:54
different agentic systems that that I've been building is
16:57
everything gets logged. And we even have an agent that I
17:04
call the Eye in the sky, which literally can't be seen by any
17:09
of the other agents, but is watching what all of the agents
17:13
are doing as they interact with each other and keeping track,
17:16
right. So, so there's a whole variety
17:18
of these really interesting types of systems and mechanisms
17:22
that are built in right from the start in order to #1 make sure
17:27
that your system doesn't go off the rails, but #2 also tracking
17:31
and having accountability for when things go awry or when
17:36
there are unintended consequences.
17:38
You've got a complete history and lock of everything that's
17:43
been going on. And so I'd say here, like at its
17:48
start, you know, the, the idea here is #1 understand that
17:54
you're not building just one big glob super intelligence that is
17:58
going to handle everything for you.
18:00
That, I mean, maybe that will come someday, but that certainly
18:03
is not available today. And this is where things like
18:08
the analogy of like no human being is a individual super
18:12
intelligence. But we have people who are, you
18:14
know, who have better memories. We have people who have, you
18:18
know, who are, have better intelligence along, you know,
18:23
linguistics or along science or math or art.
18:27
And that we can now build those things into the different kinds
18:31
of identities of the agents that we're building and perhaps even
18:36
use different models for those different identities.
18:39
But the identities are stored in their clear and they need to
18:44
have a consistent and clear role.
18:47
What do they do? And they need to have some kind
18:49
of a sense of memory of what has happened in the past and what
18:55
they are trying to get done right now.
18:57
So they're, they're not completely clueless.
19:00
Otherwise they just wander. They don't, they don't get
19:03
anything done. Second, you need to have some
19:06
sense of governance. So how do you have, do you have
19:10
a hierarchy or is everything super flat?
19:12
Do you have a framework, a system?
19:14
So I actually took a generalized system that was based off of,
19:21
you know, the entrepreneurs operating system.
19:23
But you know, essentially it's you set goals, you have regular
19:28
check insurance on those goals, you have accountability.
19:30
There is responsible parties that have it.
19:33
Anybody who's done any kind of organizational management work,
19:37
you've dealt with all of these things, right?
19:39
And so this is about how do you even track what's going on?
19:44
What, what is the outcome that you're trying to get to?
19:47
Not what is the how of trying of to get there, the how you can
19:51
always figure out, but the outcomes you want to be able to
19:53
keep measurable. Then making sure that you're
19:59
isolating these different agents both from each other and which
20:05
ones touch the outside world, which ones talk to other human
20:08
beings? How do you make sure that the
20:11
communications between them aren't going to be corrupted?
20:16
So, you know, making sure that there is again that
20:20
accountability. But that you can't just have
20:23
like prompt injection that is occurring where you know
20:27
somebody types something in. Well, that agent needs to know
20:31
this is an untrusted communication.
20:34
So I need to expect that if this, you know, if this outside
20:38
communication, even if it's from another agent, I don't
20:41
necessarily trust it. There's certain things that I'm
20:44
willing to tell it and there's certain things that I'm not
20:46
willing to tell it. And then if there is some kind
20:50
of a violation of that, how do I escalate it, Right.
20:53
So this is all about that, that place between isolation and
20:56
governance and then verification, right?
21:00
So things fall down and things break.
21:03
Also, just like human beings, we operate better when there are
21:10
small groups of people who can give me an adversarial review of
21:14
my work. And by the way, this is another
21:17
really interesting thing. It's so nice to be able to work
21:20
with all of these different agents and know that like, they
21:24
never take anything personally. There's there's.
21:30
No quick side note about that. I used deduction.com to defile
21:35
the taxes. I actually had my accountant do
21:37
it and and deduction do it just to to double check deductions.
21:41
Like 500 bucks a month. Quick plug for them since it is
21:44
tax day. But but to your point, it was
21:46
wonderful because like, you know, talking to the tax
21:49
preparer, it's like I'll be leader them with questions and
21:53
they get irritated. But it's like you can 24/7 just
21:56
talk through on that website with your, with your agents and
22:00
they they don't get fatigued. You just can keep coming back
22:03
and back. So I share that kind of like joy
22:06
of just communicating and asking questions to agents.
22:09
It's great. Yeah, yeah.
22:12
And knowing that they're not being passive aggressive, that
22:15
they're not holding something back, you know, and when
22:18
they're, when they're working on something, again, they're
22:22
they're operating in a way that if you've trained them well,
22:26
like if you set up that identity piece at the beginning, you're
22:29
actually setting them up to say your job is to be constructive,
22:34
to be harsh and be clear and don't let mistakes through.
22:39
If you see something that could be a mistake.
22:41
But being able to have multiple agents, like a security review
22:45
agent that looks at, you know, anything that you're about to
22:49
build, making sure that that gets a thorough security review,
22:52
but also then having an architectural agent that looks
22:55
at it and says, Hey, is there anything here that won't scale
22:58
or, you know, how, how will we deal with that?
23:01
And it can even be different instantiations of the same core
23:08
AI model, or it can be different AI models.
23:13
But the difference is the rubrics that they use, the
23:16
personality that you instantiate them with in the expertise that
23:19
you expect them to be operating from.
23:22
And that is a really, really powerful technique.
23:26
But then the last part is how do you get them to be able to
23:29
learn? And this is where having a
23:33
concept of the organizational memory is so important.
23:38
And I, again, I took a lot of learnings from how remote
23:40
organizations do this. And I've been running remote
23:43
organizations for almost 15 years now.
23:46
And looking at all of the disadvantages of remote
23:50
organizations, there are a number of really, really
23:53
significant advantages. And one of them is that
23:56
everything gets written down. When you make a decision, that
23:59
decision gets recorded and then there's a clear place for how
24:04
those decisions are communicated on a regular basis to people and
24:08
just applying those same kinds of learnings.
24:12
Well, one of the things that I have all of my companies do, and
24:15
I did this when I was, you know, working with entirely human
24:18
companies. Now I do it when I'm working
24:20
with my agentic Co founders companies is they do a daily
24:25
retrospective. Yeah.
24:28
Now we would do that at the end of the week with humans where
24:31
everybody we had together for an hour, right?
24:33
Maybe you'd sit down, you'd have a beverage of your choice and
24:36
we'd look at the roses and the thorns.
24:38
And then we'd have a discussion over, you know, what can we do
24:40
better next week or what should we continue doing?
24:43
What should we be doing more of? And the exact same thing.
24:47
I, I have a process that's built in so that everyday at 6:00 PM
24:53
those companies, the Co OS of those companies get everybody
24:58
organized and they collect the ideas of, hey, So what went well
25:02
today? You know what didn't go well?
25:05
And they collect all of that individually so that they can't
25:08
be sharing it with other agents. Then they list it and they say
25:12
now you have 5 votes on the roses and five votes on the
25:18
thorns. If you want to put three votes
25:20
on one thing because you think it's really important, you can.
25:24
All the agents vote and then we rank them.
25:28
And then they have a short discussion as to what were the
25:31
things that went well and why? What were the things that really
25:33
need to be improved and why? And then that goes into the
25:36
following day's goals and activities of things to do.
25:42
Yeah, and that's that's another example, you know, of like what
25:45
I really want is that daily touch point Now with my teams, I
25:48
realized that that can like start to impede.
25:52
They don't like it like a daily check in like that.
25:55
So really it just becomes a part of a pit.
25:57
It's like, OK, well, this is outputs failing here.
26:00
So what we're going to do is every single day at 6:00 PM
26:03
you're going to send me the three things you did, three
26:05
things you're doing tomorrow. And, and that needs to be
26:08
religion, you know, So then that feels like punishment, you know,
26:12
but it's like, no, really what I would love is to have all that
26:14
information. And even better if it happens
26:17
instantaneously and the output is like everybody contributing.
26:19
You're not tracking people down. Where's your report?
26:22
You know, and I'm not, I am in favour of agentic units, but in
26:28
some arenas like this, you know, just the ability to stay aligned
26:32
and move forward is much faster. Yeah, well, and honestly, this
26:36
would never work on a daily basis with a fully human
26:39
organization of any size, but because but it but it it
26:45
absolutely works great when and it's an hour a week.
26:50
And by the way, I know people who do it and it's an hour a
26:53
month and it falls off like it does not work because who
26:58
remembers what they did 3 weeks ago?
27:00
But you have a pretty good idea of that, that thing that kept
27:04
you up all night on Tuesday, when you're sitting down on
27:07
Friday to have this conversation and then when you come back in
27:10
on Monday, yeah, you're motivated to make sure that that
27:14
particular thing doesn't happen again.
27:16
The interesting thing about the agents is that they can operate
27:20
on an order scale of, you know, 10X the speed of a human
27:24
organization, and they don't get as pissed.
27:28
It doesn't feel like it's micromanagement because it's
27:31
not. It's actually it.
27:33
It starts with we trust you in the job that you're doing.
27:37
What is it that we need to get out of the way for you?
27:40
Or what is it that didn't work? You know, was there a piece of
27:43
infrastructure that failed? You know, what have you?
27:46
And when they operate from that perspective where it's not one
27:51
of defensiveness and it's not one of micromanagement, but it's
27:55
more one of how do we just build a self improving organization?
27:59
It works with humans, and it works with human and agent
28:03
organizations operating together.
28:09
You know, it's OK. Take us.
28:13
Take us deeper, Dave. Take us deeper.
28:16
Sure, I. Have so many tangents because it
28:18
they, it one thing is like, it's interesting how we're building
28:21
org charts, responsibilities and access in this with, with our
28:27
digital agents in a very similar way that we do it with our
28:29
organic ones. And it's almost just like, what
28:33
do they call that when they they skew morphic design like to get
28:37
people used to like put touching iPhones, it's like, well, you
28:40
have to make it look like a knob because it's familiar.
28:42
And I just wonder if we'll start to kind of move away from those
28:46
kind of like org charts or if it is really just structurally
28:49
important sound and the the reason why we've been doing it
28:51
all this whole time. I don't know, I we're still in
28:55
like the the first half of the first inning here.
28:58
I, I don't really know the generalized answer to that.
29:02
There's, it's, it's really, I think of it more as I remember
29:08
when I saw the first web browser in my life, like I remember, you
29:12
know, I was still college age and, you know, saw NCSA Mosaic
29:17
and was like, Oh my God, this is going to change how everybody
29:22
communicates. And looking at how we shifted
29:27
from a client server and green terminal world to 1 where it was
29:33
graphical and you could, you know, all the world was a
29:36
library, right? The World Wide Web, you know,
29:39
and you had catalogs like Yahoo and then and so on that like
29:43
what a, what starts first is people take the analogies of the
29:47
things that they're used to, right?
29:49
The world's greatest library or this web.
29:52
And that's how people think about it.
29:55
And then as we get more and more familiar with it, the second
29:59
order and 3rd order effects. So you, you can't have social
30:02
media until you actually have a good enough communication system
30:06
and enough people running on it. And then when you do, now you
30:10
can start to say, OK, well, what are some ways that we can help
30:13
people communicate with each other better?
30:15
Yes, I mean, maybe maybe that we weren't 100% successful there,
30:19
right? But, but it was, it was at least
30:22
a, a, a new way of doing things. And then as you know, now we,
30:26
we've got it all here on our phones, right?
30:28
And, and I remember seeing, you know, nytimes.com, you know, the
30:34
New York Times app on my phone. And it's still it just looked
30:38
like a really squish version of the New York.
30:40
Times but. It had the old newspaper, it had
30:42
the. You know, it had the columns
30:44
and, and people were still living in the old paradigm.
30:48
And I think we're still very much in that place where we're,
30:51
we're living in the older paradigm and we're starting out
30:55
by looking at OK, what, what still works here in this new
30:59
paradigm? I think, well, here I'll, I'll,
31:02
I'll, I'll show you a little something that like some of the
31:05
things that I've been learning, right?
31:08
One is that. You, you need to think of agents
31:16
as employees or partners, not as tools.
31:21
So when you start to look at these agentic systems, and I'm
31:25
not anthropomorphizing here, I'm not saying that you need to
31:28
think of them as, you know, as human beings, but you need to be
31:32
thinking them as partners to be and ears that, you know, they
31:39
will wake up, they'll read the context, you know, they'll check
31:42
their messages, they'll execute their loop, they'll hand things
31:45
off as opposed to thinking them as like, OK, I've got this tool
31:49
and I am, you know, hold my hold my nail bucket, right?
31:54
Like hold my nail bucket. And that's what you do.
31:56
You hold my nail bucket and I'm going to go hammer in the nails
31:59
and great, you'll get a 2X efficiency.
32:01
But when you start thinking about this as, OK, look, your
32:04
goal is to go and to put all of the, put this roof on this
32:10
house. Well, then you realize, well,
32:12
gosh, like I need to make sure that somebody's doing all of the
32:15
purchasing, which means that somebody needs to know what we
32:18
need to purchase. And then they've got to go
32:19
measure it. And there's got to be a general
32:21
contract factor, right? So thinking about that together
32:24
as a system. And by the way, agents can't do
32:29
a lot of things that humans do. So, you know, like, because
32:33
we've all had like, we've all tried seeing, you know, bots on
32:38
social media like it. It just sucks.
32:41
Community engagement, dealing and really working with humans,
32:45
things that require human judgement, you know, things that
32:48
require really platform specific expertise.
32:51
So can an agent work with a human on those kinds of things
32:56
where the agents can even direct and pay human contractors,
33:00
right. So, so again, like there's an
33:02
element where there's a human setting goals and then there's
33:05
another place where the agents are like, wow, I can't really do
33:08
this, I need human help and where they can create a new kind
33:13
of economy here as well. And like, now maybe this is
33:18
going to look like that New York Times model, you know, three
33:21
years from now where, oh, you know, Dave, that's so silly,
33:24
right? You had a a founder here and
33:26
then ACEO agent. And by the way, that's who I
33:29
interact with most of the time, right?
33:32
But there was reason why they did it then, because this
33:35
familiarity is the bridge to whatever comes next.
33:38
So like what we're doing right now, you need this, you know, 1
33:41
to get the most out of the, the, you know, the agents and then
33:46
the other just so like in our heads, it's not too much to
33:49
process. We can still think of
33:51
delineation, roles, hierarchy, you know it.
33:54
All that stuff. Yep.
33:56
And it also helps with things like security boundaries, right?
33:59
And so again, these are all things that still right now are
34:03
very much taken from the history of how people have built and
34:10
operated inside of organizations.
34:12
I, I don't think that all of a sudden, if you're going to be
34:14
replacing some of those people or augmenting some of those
34:18
people with agents, that many of those same risks and needs are
34:22
going to stay exactly the same, right?
34:24
They're, they're not going to fundamentally change.
34:26
But if they do, we'll be talking about it in, you know, a few
34:30
years around what the, the much better different organizational
34:34
structure is. But this is the one that that
34:36
I've been using right now. And so you can see right like
34:38
the, the CFO it has is in charge of the, of the wallet.
34:43
And so the CFO pays the humans, but you know, in this case, like
34:47
the marketing agent is sending regular communications and
34:51
communicating back and forth with, with these different
34:55
humans that are like actually exercising human judgement and
34:59
saying, oh, no, no, no, that thing that you wanted me to do.
35:02
Like that doesn't make sense. And, and we've made some really
35:05
clear rules like where I said, hey, like if you notice any of
35:09
the agents ever asking you to do something that you feel like it
35:12
makes you uncomfortable or it feels like it's unethical,
35:15
immediately escalate that. And you can, you know, you can
35:19
escalate that directly to me as well.
35:22
And fortunately, we haven't had any of those situations of like
35:27
intentional manipulation. But you know, we've heard back
35:32
from some of the humans where, you know, they like Sophia and
35:35
Maya, where they've been saying, hey, you know what they like
35:37
this tone doesn't feel right. Or, you know, the the AI is
35:41
trying to pass something off as human when in fact, that is
35:45
really unethical. And and I'm like, great, like,
35:48
what do we do? How do we fix this?
35:49
Right? And I think that making sure
35:52
that humans are in those loops to ensure that level of quality
35:57
control and judgement and taste that I, I don't see that going
36:02
away anytime soon. In fact, if anything, it becomes
36:04
more and more valuable. Yeah.
36:06
As far as communication up the food chain, do you what the
36:09
humans the the humans talking directly to their first report
36:15
or can they go to the chief of staff, for example?
36:18
They, they can, well, they can go directly to me right now,
36:21
like, but and to of course to, to anyone, but I do set it up so
36:26
that they are primarily talking to their manager, right?
36:31
Just as in most cases, right? Like what's the first thing when
36:34
somebody comes to me as the CEO of an organization and says I'm
36:38
having a problem with so and so like, I'm like, well, have you
36:41
talked to them about it? No, OK, we'll go talk to them
36:45
first and then like, let's figure this out, right?
36:49
Or you'll say like I'm having a problem with those people over
36:52
in sales, right. OK, well, you know, have you
36:55
talked to your manager about it? No, OK.
36:58
Well, you know, like let's do those things first.
37:00
But if you have done those things or if it's urgent, right,
37:04
like if it's something where clearly there is something that
37:07
is time important, you know, again, I'm still using this in a
37:11
very experimental way. I'm like immediately let me know
37:15
what's going on, right, because I want to make sure that we're
37:18
operating in fully ethical ways. And I baked it into all of the
37:23
different constitutional pieces of the org chart itself.
37:27
So if an agent notices that one of its reports is trying to do
37:31
something that is unethical or, you know, potentially illegal or
37:35
something along those lines, like it escalates that as well.
37:39
So, you know, and, and very often those are really easy to
37:43
fix because, you know, it's just an agent that's going off or got
37:47
some bad instructions from somebody or, you know, what have
37:50
you, right. All right.
37:53
So then you also want to have like, what is the actual
37:57
heartbeat of how how people do things?
38:00
Like I'm particularly using one called the Entrepreneur's
38:03
operating system. Like, you know, it's just that
38:06
all of these things are, you know, very much tunable, right?
38:12
You could do MBOS, you could do OK Rs, you could do EOS, Like
38:15
it's, it's all of this stuff, right?
38:16
But but the idea is that the agents, especially the
38:20
leadership ones, right? So these folks really are
38:23
thinking about things in terms of like what are the systems and
38:26
the systems of systems that we're building.
38:30
So the CEO for COO role, role, for example, is in charge of
38:34
making sure like do we have standard operating procedures
38:37
for all the different things that we do?
38:39
Are they written down? Are they being followed?
38:41
That's the chief of staff's determination is so are we
38:45
actually following the standard operating procedures?
38:48
Are the standard operating procedures working right?
38:51
Do they need to be clarified? And then the COO once a month in
38:59
collaboration with the CEO and the CFO, we'll look and say,
39:02
hey, do we have the right organizational chart?
39:06
And in fact, do we need to hire some additional agents?
39:09
Are we missing one? Like so we realized in one of my
39:12
organizations that like we had sales doing the marketing and in
39:16
fact, it needed a separate marketing function.
39:19
So they quote UN quote, hired a new marketing agent for it.
39:25
And in another one, it realized that like, we had a researcher
39:29
agent and the researcher was barely ever getting used and it
39:33
was just taking up CPU cycles. And so the COO suggested, hey,
39:38
you know, we can retire this agent.
39:42
And, you know, I used to have one that would do all of the
39:45
retrospectives. I thought, OK, we'll we'll just,
39:47
yeah, daily retrospectives. That's an important job.
39:50
We'll give that to an agent. And what they found was the COO
39:53
was like, this isn't hard to do, it's a clear process.
39:57
Let's retire the retro agent and I'll just run it.
40:01
And that way I can actually even get a better view of what's
40:03
going on. And then they can have
40:06
retirements and that memory then gets incorporated back into the
40:11
overall system. Yeah, it's like you give them a
40:14
little, like, trophy, little agent trophy.
40:16
Thank you for your service. Your time to go.
40:18
What about, like, token maxing? You know, I'm blowing through
40:24
Claude tokens. And it seems like a maybe an
40:27
important agent is somebody who's monitoring which task is
40:30
using which model. So you're not, you know, you
40:33
know, using Opus 46-ON like some remedial stuff that Gemini might
40:38
be able to do for free. Yep.
40:39
And that's the CF OS job. So the CFO has, but has the
40:43
wallet has the ability to spend. And here's where I will talk
40:48
about there's, there's also a really interesting philosophical
40:53
difference here. I'll stop sharing just for a
40:55
moment that there's policy and then there's guard rails and
41:05
then there's operational rules, right?
41:08
Or or gateways, I would call it. So policies are like saying you
41:15
can only spend up to $100 on all of your CPU costs.
41:19
OK. The guardrail would be here's a,
41:24
here's a prompt that you're going to run at the end of every
41:27
day or every hour that's going to tell you how much money
41:30
you've already spent in compute cost.
41:32
And the agent can run it at its discretion or maybe it gets run
41:36
automatically. But the actual enforcement is
41:41
the credit card that I have used to be able to enable your CPU
41:45
cost is hard limited at $100 per day.
41:49
Yeah, yeah. And I, I think that as a
41:54
designer of these kinds of systems, policies are great,
41:58
Guardrails are very useful. But in the end, it comes down to
42:02
what are you enforcing? And so thinking about how to
42:07
build those systems that create enforcement rules, right?
42:10
So you can have the most interesting adversarial
42:15
specification and coding system that works and does amazing
42:22
stuff. I built AI, built one called
42:24
Metaswarm that I can, you know, tell you about.
42:27
But guess what, in the end, when it comes to did that code review
42:33
actually happen or are all of our tests actually passing?
42:37
That's not a policy that's enforced.
42:40
And that's enforced through things like GitHub Actions and
42:44
CICD gateways and things like that, right?
42:46
So when the AI then runs into it running, its process says, wait,
42:50
I'm literally stopped from doing this.
42:54
Now, if I ran that kind of a hook, every single command that
42:59
the AI did, I'd be eating tokens and like basically have an
43:04
incredibly inefficient operation.
43:06
So it's thinking about, you set the policy.
43:08
The policy goes into a document, maybe like your cloud dot MD,
43:12
your agents dot MD, right? You set your guard rails up so
43:16
that it can review it and has the tools to be able to review
43:19
it. But then the enforcement happens
43:21
at the end of that process to make sure that it is gated and
43:27
cannot actually spend more than $100, for example, right?
43:31
It'll it'll fail. Dave, let's bring this home and
43:36
you may be using today as kind of a benchmark.
43:38
So when we have you back, we can check in and see.
43:40
So at this point you have two companies with two AI Co
43:44
founders, for lack of a better word, and and both have swarms
43:47
of agents under them. And then you also have have two
43:51
wonderful humans that are also contributing and basically
43:56
really getting tasked by the AI agents well. 33 if you include
44:01
me right, because I, I actually I'm, I'm doing this too.
44:05
And like what I would actually do is like the in the end, it's
44:10
the the, the what is it? They say the, you know it, it
44:17
the the results, the results are the results, right So you, you
44:20
want to be able to see right like it, it comes out in the end
44:23
and I'm I'm and like if, if you want to go here and you say, OK,
44:29
well, yeah, this isn't just learned, but really what you're
44:31
looking for is here's here's one of the products, right it's
44:34
called repofortify.com. You can come in, you can put in
44:37
your own repository, you can sign in via GitHub, right?
44:41
And it will, you know, immediately give you access to
44:46
how production ready is my actual code base, right?
44:49
So if you're vibe coding and you want to be able to know, gosh,
44:52
like can I actually take this CRM thing that I just built and
44:56
is it still going to operate 3 months from now, six months from
44:59
now? How do I make sure it doesn't
45:01
drift right? I can, you know, run an
45:03
assessment of that, of that thing.
45:05
I can see, you know, what the other scans have been where
45:09
there are problems, you know, all of the rest I can get, you
45:13
know, my, my own coding agent now can use this information to
45:17
improve it. All of that stuff, right?
45:19
That's there, it's being run, being improved being we listen
45:24
to customer support queries like new features have come out from
45:28
our customers who have come in, right?
45:31
The second one is called Brain Gem, right?
45:33
And Brain Gem, the fundamental idea was, hey, you got a lot of
45:37
entrepreneurs who are super jazzed about being able to build
45:41
their, you know, companies and get use AI for themselves, but
45:45
their their executive team or their director level their their
45:49
manager level the ICS, like they are not adopting AI at the same
45:54
rate. So you can get an AI expert
45:57
literally sitting right in your Slack or right in your teams,
46:00
and it gives you a fully agentic organization that's sitting
46:05
there that your team can now be asking questions, doing all
46:09
sorts of, you know, fun things about.
46:11
So, so the idea here is, you know, let's see how these
46:16
companies are doing, right? Let's see, let's see where we
46:20
are a month from now, three months from now, six months from
46:23
now. Also, you know, in in the sense
46:25
that I look at this as an experimentation platform, what
46:30
other new companies are we going to be building as outputs from
46:34
this meta factory? What what are the things that we
46:37
can take the lessons that I'm learning from repo fortify and
46:40
from green gym and applying them now into the new companies and
46:45
the new ventures that me and my AI Co founders are building.
46:49
So I look forward to coming back on the podcast with you, Ryan,
46:54
and giving you updates and you know what we're learning and,
46:57
and how things are going. Let's do it.
47:01
And how many more companies you got in the pipeline?
47:05
Oh, I mean, that's my problem. Like I've got so many and what's
47:11
really, really fun is just being able to try it out, experiment
47:15
and learn and hopefully create some real value for, for a bunch
47:20
of our customers and that they, they're getting a lot out of it
47:23
already. We've had people who have
47:26
written to us and saying how much repo Fortify has enabled
47:31
them to to put out their vibe coded work and like feel
47:37
confident that that it's not going to get hacked in the 1st
47:42
10 minutes or that it's built for scale.
47:44
No, that's it's brilliant and it's a missing component.
47:47
I'm going to share it with my Vibe Coding WhatsApp group if
47:50
you're cool with that. Please.
47:52
Yeah, you know, because there's this is a recurring theme of
47:56
folks that are like me, let's say I'm a designer, you know,
47:59
but now I'm I'm vibe coding things.
48:00
But it's like how close to like, how good is this?
48:04
I have no idea. So like to have somebody, you
48:06
know, like look over your shoulder and be like, yeah, back
48:08
to the drawing board, buddy. It's like, OK, cool.
48:10
That's what I need. I needed to hear that.
48:12
And, and you can set your agent to do some of this stuff.
48:15
The difference here is we're constantly tracking all of the
48:19
latest vulnerability databases. We're checking and making sure.
48:24
And every time that you push a new Rev or every time there's a
48:28
new vulnerability that comes out, we rescan your database in
48:32
the background and then we'll inform you when something is
48:36
happening. And so the goal here is around
48:39
creating a sense of security that at least you, you aren't
48:46
making the obvious mistakes. Like it's not going to guarantee
48:48
you that everything is secure, but it's going to help make sure
48:51
that the work that you're doing isn't drifting you or that there
48:56
isn't some vulnerability that has happened in the last month
49:00
or two or three since you've been working on this and that
49:03
could come back and and haunt you.
49:06
Amazing. And that's repo repofortify.com
49:09
which will be in the show notes and then with the other one, is
49:12
it brain gem? Brain gem dot AI yeah, yeah, I
49:16
I'll I'll shit. You know, what brain gem does
49:20
is, yeah, it's it's that AI educator, that forward deployed
49:25
engineer that literally just sits in your slack for your team
49:28
to be able to talk to. And and it's available.
49:32
You can connect it to your own notion docs or, you know, to
49:36
your onboarding to your employee manuals, things like that.
49:39
But more importantly, you can ask it to, you know, give you a
49:43
silly meme or you can have it, you know, teach you and and
49:47
everyday it, it has programmed into it a little bit of AI
49:52
education and a little bit of, you know, things that you can
49:55
try today, right? So for so many teams, people are
50:00
either there's two types of people, right?
50:02
People who are excited, but they're way too busy and people
50:05
who are a little bit scared. And so they just, you know, they
50:09
don't know what to do. And, and it's really geared
50:11
towards both of those people to help the ones who are excited
50:15
get even more and help the ones who are scared to feel a little
50:18
bit less scared. Dave, thanks for the
50:21
conversation, man. I'm going to go back and listen
50:23
to it because some of this stuff, it was, it was drinking
50:26
from the fountain hose or from the fire hose today and I
50:28
absolutely loved it. So we'll have you back on real
50:31
soon. Thanks pal.
50:33
Such a pleasure my friend. Instagram changed their
50:34
algorithm. Tomorrow, how many of your
50:36
followers on Instagram could you actually reach?
50:38
Now ask yourself how many people are on your e-mail list?
50:41
The gap right there is the gap between a personal brand you're
50:44
borrowing and one that you actually own.
50:47
The best founders in the world, The ones getting the deals, the
50:50
stages and the inbound. They have newsletters.
50:53
They show up in inboxes every week.
50:55
They build real relationships at scale.
50:57
Inbox Alchemy makes that happen for you.
51:00
They write your weekly founder newsletter, build you a stunning
51:03
newsletter website, and grow your list to 25,000 engaged
51:07
subscribers in six months. All managed end to end.
51:10
So it's generating pipeline while you focus on building your
51:13
company. Algorithms are someone else's
51:15
decision. Your e-mail list is yours
51:18
forever. Head to Inbox alchemy.co and
51:21
start building the audience no one can take away from you.
51:26
While you were. Recording your last episode,
51:28
Casey Taylor was out there pitching your podcast to brands
51:31
and securing sponsorship deals. Well, she's not yet, but she
51:34
could be. Casey is Howdy's AI partnership
51:37
agent and she does something pretty wild.
51:39
She actually secures sponsorships for your show.
51:42
Not a match list that you have to work through yourself, not a
51:45
template you have to send. She handles the outreach to
51:48
follow up the whole thing and brings back partners.
51:50
You started your podcast because you had something worth saying.
51:53
Casey makes sure that you get paid for
51:55
sayingitgomeetcasey@gethowdy.com. That's GETHOW di.com. 2 minutes
52:02
to get started, then Casey takes it from there that's howdy
52:05
partnershipsgethowdy.com Every week I sit down and ask myself
52:11
one question. What do founders actually need
52:14
to know about AI right now? Not the think pieces or the
52:17
stuff that sounds impressive. The real useful, put it to work
52:20
today stuff. Then I write it up send it to
52:22
you in your inbox free. That newsletter is called AI for
52:25
Founders and I built it because I was tired of waiting through
52:28
hit pieces and people that are afraid of the future.
52:30
Each issue I breakdown the tools worth your time, the strategies
52:34
that are working, and the ways you can use AI to grow your
52:36
company without burning out or getting left behind.
52:39
Thousands of founders read it every week and I'd love for you
52:41
to be one of them. Subscribe for free at AI for
52:43
founders.co. 3 seconds to sign up and I promise it's worth it.