Latest / Great Houses / 32. Cursus Honorum of the AI Future Part 4

Transcript

32. Cursus Honorum of the AI Future Part 4 · Great Houses

0:01
Gregory Treat: All right.
0:01
Hello everybody, and welcome to episode 32 of The Great Houses Forum.
0:06
We are continuing our discussion of the cursus honorum of the AI future.
0:13
So, uh, for those of you, we took a break last week.
0:16
Well, it's not really a break as, as we'll, as we'll soon see.
0:20
Um, last week we talked to Zach Fava, um, about his, um, his company, which is
0:27
providing an, an education piece for, um, for the household AI that I think will,
0:32
will come out of a lot of this stuff.
0:34
So today we're gonna return to the, to the series.
0:36
Um, so again, our goal, my goal in this series is to set out what I think is
0:41
coming and present that in a way that will help you think about this for, especially
0:45
for purposes of training your kids to participate in, in this at a high level.
0:51
And again, I think that the, the technology is, is largely here, or
0:56
at least we can see where it's going.
0:58
What, what I think a lot of people, the … What I think is not obvious to a lot of
1:02
people is how that technology will require us to reorganize things socially, right?
1:08
So the first thing that I, uh, we walked through in this series was the idea that,
1:11
that LLMs are gonna bridge domains, doing to essentially all computer work what
1:16
CAD did to design work, which is just kinda to create a, a common labor pool,
1:20
is, you know, it's, it's gonna mean that people from formerly different industries
1:24
will now be able to compete against one another, to work with each other.
1:29
That creates lots of innovation.
1:30
It creates lots of new opportunities.
1:32
It creates, especially at the high end, uh, higher pay scales, and all of this
1:37
starts to look like a unified hierarchy.
1:40
Okay?
1:41
So, uh, that new domain will have a limited number of specific skills
1:45
that will allow their user to be productive and useful in a vast
1:47
array of different industries.
1:49
Again, so all of those industries are now competing with each other for talent.
1:52
Um, the, and, and the, the, there's, a lot of times there's silos
1:57
between skilled labor pools, and those silos are gonna come down.
2:01
So- And the way that I've been framing this is the skills of
2:03
code, content, and command.
2:05
Uh, you know, code is basically the, the skill of logic or architecture,
2:09
uh, you know, m- mental architecture.
2:11
This, you know, people talk about, uh, uh, graph design or,
2:15
or, or graph engineering, right?
2:16
This is thinking at a higher and higher level about what is it that
2:21
you want, what are the outcomes, what are the constraints, what
2:23
are the specs that you want?
2:25
Um, because when… The, the more vague you are about that, the, the
2:29
more it increases your token cost.
2:31
Okay?
2:32
Then there's content, understanding, you know, matters of taste, what will
2:36
appeal to people, how, how to use AI, uh, effectively for those things, um,
2:41
being able to explain, to, to notice.
2:44
I mean, basically, it's just basically noticing, oh, these are all the things
2:47
that I don't wanna show up in my work.
2:49
and then command is, uh, the ability to, to make decisions, to, uh, to own
2:54
responsibility for liability purposes.
2:57
And again, the, one of the interesting things about the, the command idea
3:01
is, again, it, it dramatically lowers the token cost, right?
3:05
If you give… And I'll, I'll, I'll c- may come back to this
3:07
point several times through this.
3:09
Uh, if you give a- an AI, you say, "Well, well, make this decision," that is much,
3:16
much, much more expensive token-wise than saying, "Give me 10 detailed plans
3:21
for what you could do, you know, for, for, for different decisions we could
3:25
make when faced with this option, and then I will pick between them."
3:28
The… If you just say, "Give me the options," and c- even build
3:31
out quite detailed plans for how you'd go about that, the AI can do
3:35
that, you know, quite effectively.
3:36
So last time we reviewed three points.
3:38
Um, you know, humans are not obsolete, AI reshapes every industry, and the
3:42
skillsets, the, the point that I want to make, it most, most notably in
3:45
the agricultural industry, they all have military applications, right?
3:49
There, it's, it, it's, it's going to be more and more difficult to, to have
3:55
these separate boxes where … And, and, and I think, again, the, the
4:00
point that I'm making here is not, not that we will all be part of the
4:03
military, but that, that there, that will result in social changes, right?
4:08
There is part of our culture, part of the way that, that we have expressed
4:12
ourselves and, and, and in many ways the, the origin of a lot of our
4:15
cultural pathologies is the way that we have walled off our scholars from
4:20
our warriors for the past 150 years.
4:23
And, and it looks to me like in, you know, 10, 20, 30 years that may
4:28
not be technically possible anymore.
4:30
Um, the, the, the skill of being good at using an AI and robots to do something
4:35
… Now, that doesn't mean that the, the cultures, the subcultures that we have
4:38
created will disappear in 30 years.
4:40
But it means that similar to, to the CAD space where you had, you know, 3D
4:44
printer people and CNC people, kind of hard manufacturing production people,
4:49
and then all of these, uh, you know, arts and design, prop manufacturers.
4:52
They were all thrown into the same space, and their values kind of ha-
4:57
got to hash it out with each other.
4:58
I think that's gonna happen, but again, on a, a vastly larger scale here.
5:03
Okay?
5:04
So today we are going to discuss safe jobs and, uh, w- you know, it's … A
5:09
bunch of people are talking about safe jobs and why that is the wrong question.
5:13
And then, uh, my, the second point that we'll, we'll talk about is, is something
5:17
about the, the economics of a prime.
5:19
What, what one person who holds all three AI skills that, you know, that code,
5:24
content, and command, uh, what, what that person is worth and how that, that
5:28
pricing, that capacity redesigns the firm.
5:31
And again, as we go through this m- my, my hope is, and we'll, we'll, we'll
5:35
turn … You know, we, we will spend some episodes spec- specifically talking
5:39
about this at the end of the series.
5:41
But h- my hope is is that as we're going through this I'm presenting the
5:45
information in a way that will help you make educated decisions about how
5:49
to train your children, what to do, what not to do, um, 'cause I think
5:52
that, that in the broad strokes these are, these are the trends that, that,
5:57
um, that we're gonna be all facing.
5:59
And obviously if we can be on the, the positive side of the trends than on the
6:02
negative side of the trends that's, that's the benefit of our, of our children.
6:05
Okay?
6:07
So, uh, the internet thinks that there are some, some AI safe jobs, AI-proof jobs.
6:13
Uh, so Resume Now 2026 says nurse, anesthetist, uh, anestheticist, which
6:19
I, I, I just think that's totally wrong.
6:20
Um, emergency physician, judge, pilot, surgeon.
6:23
We're gonna kinda go through those, those lists and critique them,
6:26
and then I'm gonna tell you what I think is really going on, right?
6:29
you know, one of the things that I wanna, I wanna point out is that
6:33
AI and robots are very powerful.
6:35
Um, you know, if, if we're talking about the skilled trades or clinical
6:37
medicine, or therapy, right?
6:39
People say, "Well, therapy is this AI-proof, uh, presentation." I mean,
6:43
I just wanna observe, you know, AI plus robots is very powerful, right?
6:48
Even on … Even in, like, the, the super soft, uh, skills
6:52
like, like a therapy job or a, a pastoral, you know, counseling job.
6:55
I mean, think about how many people were drawn into
6:57
ChatGPT-4, I think it was, right?
6:59
And there's all these people that, like, ha- ha- created this close
7:02
emotional connection with an AI.
7:04
Now, the AI was not good enough to manage that well.
7:07
There were some problems.
7:08
That's why they had to, to, um, to kinda decommission that, that model because
7:13
it was, it was causing them troubles.
7:15
Um, but these are very powerful models.
7:18
Now, now e- especially in the trades, I think the key issue, uh, that we will face
7:23
for the foreseeable future is cost, right?
7:26
It's not gonna be cost-effective to have an autonomous plumbing robot.
7:30
We'll talk about that in a second.
7:32
Um, but, but, you know, costs espe- uh, for robots I think are
7:36
going to change dramatically.
7:37
I think that the cost of inference is gonna get cheaper and cheaper.
7:40
I mean, we're, we're already seeing higher and higher capacity models
7:44
being squeezed down so that they can function on smaller and smaller bits
7:47
of hardware, especially when you can define the job that you wanna do, right?
7:50
If you want a general purpose AI, right, that's, that's a harder thing.
7:54
But if you can say, "Hey, we want something that can think about these
7:57
types of tasks," man, um, I anticipate that g- just getting cheaper and cheaper.
8:02
So-
8:05
Um, a-and again, the, the, the, the question that most people are
8:11
asking on this topic is wrong.
8:12
They're asking, are humans obsolete?
8:15
And what I want instead for you to ask is, does mastery of core AI skills,
8:20
right, code, content, command…
8:21
And, and, you know, one of the other points is there
8:23
may be other skills, right?
8:24
I'm, I'm … I, I actually can think of three or four other skills, but
8:27
for, for rhetorical purposes, I'm sticking to three for, for this
8:31
presentation and for the, uh, the, uh, the essays that I'm gonna release
8:34
on Substack that, that pair with it.
8:36
Um, but regardless if it's, if it's five, three or five or 10, like again,
8:43
this, this represents a dramatic shift in, in the, the number of skills.
8:48
When people, people can potentially master all three skills or all five skills.
8:52
Certainly, if there's 10 core skills, you will be able to intimately know
8:56
people who have all 10 skills and kinda manage them as a local team, right?
9:02
Um, and the other point is that, um, just because a, a, a profession is
9:10
AI-proof, like what does it mean?
9:12
What does AI-proof mean?
9:13
Does that mean that the people in that prof- like, in that profession will
9:17
still be doing the same things that they were doing five, 10 years ago?
9:20
Um, I, I don't think so.
9:22
I think that, that AI is going to dramatically transform all of
9:27
these professions, um, and, and
9:32
So the fact that there will still be people that are called plumbers and
9:35
people that are called therapists and people that are called this and
9:38
that does not mean that AI won't dr- uh, be as impactful on those
9:43
industries as the computer was, right?
9:47
So again, my contention is that a practitioner that can afford the
9:50
upfront capital for some combination of AI and robots and hire or train
9:54
the skills to use them effectively will dominate their competition.
9:57
Okay?
9:58
So, uh, you know, the- the- the first example I wanna p- put in
10:01
front of you is plumbers, right?
10:03
So we're gonna start with the trade.
10:04
So- so the question is, i- are- are plumbers gonna be turning wrenches?
10:08
And- and for now, yes, probably.
10:11
Uh, it's- it's not cost-effective to have a- a plumbing robot, especially not
10:15
an autonomous plumbing- plumbing robot.
10:18
Um, I- I think that's- that's changing.
10:20
It could change even- even as shortly as- as within a year.
10:23
Um, but I think the thing to think about, the question to ask
10:28
is what will happen when certain plumbers start to integrate AI?
10:32
What will that look like?
10:33
Okay?
10:34
So we have one plumber who, you know, he comes out, he walks the
10:37
property, he quotes a number.
10:39
Uh, everybody understands that that number is kinda subject to all of these things.
10:44
Um, and then- and then he, um, he relies on his reputation.
10:49
And in many- in many times, it's- it's more about I'm here right now, and I
10:53
will help you, and if- if you don't agree, then I'll just leave, and you'll
10:55
kinda go to the- the end of the list.
10:57
Um, a- a newer plumber will have a camera drone or they have these, you know,
11:02
cameras that go on- on long lines, uh, that they can put through, and they can
11:05
just see everything inside the pipe.
11:07
They can show their client what the pipe looks like.
11:09
There's- there's YouTube.
11:11
I mean, this … The thing about this is this is content, right?
11:13
There are YouTube channels devoted to basically plumbers showing people what
11:19
was wrong with their, uh, with their plumbing system and then fixing it, uh,
11:23
which is kind of a- kind of a crazy thing.
11:24
People watch this on YouTube.
11:26
Um, you know, as- as AI, kind of the more personal assistant AI-type stuff,
11:30
uh, comes online, imagine that he has a conversation with a client which the AI
11:35
is listening to, and when he … when- when the client says yes, the- the AI just
11:39
generates a quote, generates a proposal.
11:42
Um, it's already probably pulled a profile on, uh, on
11:45
this person, uh, on this client.
11:49
So I think, you know, the thing is, is that- that all of these
11:52
things are running on code.
11:53
So, so there is someone … If you … Once you see a- a plumber
11:56
like that, there is someone in his life, maybe it's him, maybe it's not,
11:59
that is deeply familiar with code.
12:02
Um, there is- there is someone in his life that is deeply familiar with
12:06
content, that has that content skill.
12:08
And then again, I think the, uh, the thing that's gonna make it so that there's still
12:11
a plumber for the foreseeable future is it's going to be wildly cheaper to have a-
12:17
a human kind of peering at the different options that the AI is pulling up.
12:21
And the AI says, "Well, it could be this," you know.
12:23
"Here's the images that we're interpreting.
12:25
We're gonna circle this.
12:26
Could be this option, could be that option." And the plumber just says, "Yeah,
12:29
it's that option." Um, even if … Even when we get to the point where, like,
12:32
the frontier models could accurately make that calculation, it's not clear to me
12:39
that, um, that it'll be cost-effective for that to happen anytime soon.
12:43
Okay?
12:43
Now, um, and- That's probably not true of medicine, for instance, right?
12:50
For medicine, uh, the, the, these procedures are so costly that maybe it
12:56
is cost-effective to have a, a robot that, that makes no decision … that
13:00
makes all the decisions, and we just kinda churn the, the token cost.
13:03
On the other hand, right, if you've got, if you've got somebody that is,
13:08
is operating on your body, you probably have a, a greater concern with, with
13:13
what's going on there because you don't, you know, you're, you're, they're
13:16
You want someone to take responsibility, okay?
13:21
Um-
13:24
And, and again, I think the, the, the point to make here is, is that the,
13:27
there are gonna be people who master these skills and people who don't
13:30
master these skills, and the people who master these skills will be able to
13:33
out-compete their, their competition, especially if they can scale, right?
13:37
There's a, a chiropractor, I think he's actually… Uh, is he a chiropractor or
13:40
a, a… Oh, he's a, he's an eye doctor.
13:42
I think he's in the DFW area.
13:44
Um, and he has fired all of his staff, and that mean- AI
13:50
has replaced all of his staff.
13:52
It does all of his coding.
13:53
Um, and, and in his mind, that means he has more time to spend with a client.
13:58
So he spends a full hour getting to know people, talking to them, just
14:02
kind of… He, he doesn't have to constantly be scratching down notes
14:06
'cause the AI takes care of all of that.
14:08
He can, you know, at a certain point in the conversation, he can look
14:10
over at his iPad and see what, what questions do I still need to, to ask,
14:14
and it'll list them, those out for him.
14:16
But he is able to be more approachable, more personal,
14:20
more focused on the client, okay?
14:25
Now, I wanted to, um, I wanted to talk about therapy.
14:29
Uh, you know, talking about these, these, the kind of the talking professions.
14:33
And I think it's interesting, uh, we, we will have to see
14:39
how human tastes change, right?
14:42
Uh, 'cause, you know, the, the, w- if, if the thesis is AI and robots,
14:46
you know, a counseling session does not get better because you can
14:48
fly a drone or run a robot, right?
14:50
Uh, it's, it's not clear immediately how, how the, the way that we think about
14:56
AI will Will be helpful to a therapist.
15:00
Um, and, and, you know, obviously we can design really compelling personas.
15:04
I, I mentioned ChatGPT earlier.
15:06
Um, but for a lot of people, that would poison the product, right?
15:11
The client wants a human.
15:13
They're, they're gonna demand a human.
15:14
That's their expectation.
15:16
Um, but even in things like therapy, uh, you know, AI is fantastic at building
15:21
profiles of specific people, what they like, avoid, fear, respond to.
15:25
Uh, you know, gathering that, that sort of public data on one person used to be
15:28
a capacity, on the civilian side anyway, that, that only Google possessed, and now
15:32
everyone with access to AI will have that.
15:35
And so in theory, you could have a, a session that begins with the clinician
15:39
already knowing the loves, aversions, and history that no intake form will, would,
15:43
would have captured, and so the client, you know, can, can have that experience.
15:46
Now, uh, you know, professional ethics is, is, is always a question
15:52
in these types of, of conversations.
15:53
The, the, um, many boards of ethics in many states have already said, "Hey, you
15:58
can't look at social media profiles." I, I, I kinda wonder how long that'll,
16:03
that'll last, um, especially given, you know, how much you're, um, how, how
16:10
much people are, are relying on AI for all of their decision-making, right?
16:15
You know, one of the things that, that, that I talk about in this is whoever
16:18
has access to your core AI chats, right, which, which is not just the things
16:23
that you're typing into the AI, but the AI's reasoning about you, right?
16:27
'Cause, 'cause AIs have to build profiles on you to predict what
16:32
you want so it can serve you.
16:33
So i- whenever you're interacting with an AI, it's building a profile
16:36
on you, um, and that profile is contributing to its answers.
16:41
It's kind of referring to those things internally, okay?
16:44
Um, and so who has access to that, right?
16:47
Do you, do you want, you know, your boss to have acc- I mean, i- in many
16:51
cases, especially in, in sensitive, uh, sensitive, uh, professions, the
16:57
company will demand access to that.
16:58
They will require access to that because it's a, it's a sensitive profession.
17:02
Do you want your therapist to have that?
17:03
What, what if you could know that your therapist would be wildly more
17:07
effective if, if they had access to that?
17:10
Um, and, you know, if … And the other question is, like, what
17:14
happens if it just gets banned?
17:15
Well, if it just gets banned, uh, by the ethics boards, that
17:19
technical capacity doesn't go away.
17:20
It just probably migrates to people who are going to, you know, use it
17:24
without, without any restraints.
17:26
Um, you know, what does that, what does that look like?
17:28
Is that really the, the, the direction that we wanna go?
17:32
Um- And so again, I think even in, in therapy, even in these like
17:38
intense human-driven things, as AI reshapes all of society, it's
17:44
gonna reshape that, that experience.
17:46
Even though I, I don't think it's gonna replace the human.
17:48
I mean, one of the things, you know, again, therapy shows up on
17:53
the list of these AI-proof jobs.
17:55
I actually think there's a, there's a world in which once human taste, uh,
18:00
accommodates for it, uh, where, where we get rid of a lot of human therapists,
18:05
where many people are relying on, on, on AI therapists and kind of running
18:09
their decision-making through them.
18:10
Uh, I think that would be probably terrible for society.
18:13
Um, I, I, I, I wonder how long that, that could last, but I think it's a thing that
18:17
people will, will try at a certain point.
18:20
Um-
18:23
And there's just a, you know, there's just a lot of, there's a lot of questions
18:27
about what, what happens when you give machines access to these tools?
18:32
What will happen when a, when an AI has access to a camera that's looking at
18:39
you, when it can track your eye movement?
18:41
You know, there was a, there was a scandal, um, a couple years ago with
18:45
the, the Apple Vision Pro where it, um- It, it was tracking people's eye
18:52
movement sp- and it had to because you can actually type with your
18:55
eyes in a, in an Apple Vision Pro.
18:57
But they … It also allowed you to have an avatar, like a, like a, like a
19:01
chibi avatar, like a animated computer CGI avatar, and the eye movement
19:07
of the eyes of the avatar was so accurate that people were able to steal
19:12
people's passwords if they had access to the feed of your, uh, your avatar.
19:17
Um, so that is an incredibly detailed, um, you know, camera that's watching you,
19:24
that's, that's tracking all these things.
19:25
For sure that could track your, your facial mo- your facial, um, features, your
19:30
expressions, maybe your, your skin tone.
19:33
Um, you know, as, as we get more and more, um, you know, health, health type
19:38
things where they can, they can have something on your wrist that tracks your
19:40
heart rate and maybe even blood pressure, blood sugar, all of those things.
19:44
What will a machine be able to do with that level of information on you?
19:49
Right?
19:49
What will it be like?
19:50
How, how will it be able to influence your thinking?
19:53
And, and I mean, you know, maybe for good, right?
19:55
People are trying to, um, people are trying to use this for good.
20:00
People are trying to, uh, have this be a positive experience.
20:04
But I just think, you know, we, we need to understand that if you could
20:08
have an AI that set your kind of mental temperature, right, that could, that
20:13
could put you in a calm state, right, uh, what would you pay for that?
20:17
Uh, uh, you know, in many ways this is kind of like, like
20:20
light-based therapy, right?
20:22
Uh, it's sort of
20:23
But it will be, it will be e- equivalent, I think, to a drug.
20:26
It will have a, an overriding impact on you, um, that will change what
20:32
you're doing and, and, and how you're experiencing your life, uh, as soon as
20:37
you give an AI access to a camera that's looking at you and a screen that it can
20:40
put in front of you to show you images and maybe, maybe some sound component.
20:44
Um, will that be, you know, something like hypnosis?
20:48
I don't know.
20:49
Um, but I think all of these things are … You're, you're not gonna be
20:55
able to, to keep the bounds between, uh, you know, a therapeutic counselor,
21:01
you know, who's licensed and all these things, and w- when the only thing you
21:05
need, right, 'cause, 'cause a therapeutic counselor or even, you know, a hypnotist,
21:08
right, or, or, or a facilitator, those people it takes years of training.
21:12
There's, there's kind of a huge regime around it.
21:15
People, people are, are suspicious of it.
21:18
But what happens when if you put a phone in somebody, in front of somebody's eyes
21:23
and you put earbuds into your ears and you've got that same level of capacity?
21:26
What then?
21:27
Um, so I think that will, I think that will dramatically
21:31
influence how these things go.
21:33
Okay?
21:35
Now, uh, one of the professions that I'm, you know, particularly interested in is,
21:39
is the law, is, is the legal profession.
21:41
Um, and I'm, I've listened to a whole bunch of pitches about how AI is
21:44
gonna impact my field and, you know, my current stance is as follows.
21:49
So I think the, the, the human lawyer is irreplaceable in at least two rooms.
21:53
So one is when a client is entering into an engagement with a law firm.
21:57
I think for the foreseeable future, the, the people that come wanting
22:01
a law firm… Right now there's a question about people who come wanting
22:04
legal services or compliance services.
22:06
Will they want something different?
22:07
Maybe, maybe they will.
22:09
Uh, but if you're looking for a law firm, you are looking for a,
22:12
a human person that's gonna make you feel safe and trustworthy, and
22:16
like they've got experience and they're gonna be good at their job.
22:19
And then the second thing is, again, also for the foreseeable future, anytime
22:23
that a judge is gonna rule, right?
22:24
I mean, judges are gonna want lawyers.
22:26
People, I don't think, appreciate how much work it is.
22:29
Basically, what a lawyer does for a judge, two lawyers do for a judge, is they make
22:33
all decisions into the, into a binary.
22:35
The kind of binary that you can just do a, a, a yes or no, party
22:39
A, party B kind of a decision.
22:41
Um, and that, that takes a lot of work, it takes a lot of negotiation, and I
22:45
don't think judges are gonna be willing to not have lawyers do that for them.
22:49
I, I, I very much doubt that they will be interested in, in having an
22:53
AI do that for them anytime soon.
22:55
Okay?
22:57
Uh, but outside of those two rooms, you know, I think the, the, the
23:01
workflow is going to, going to, to, to become the, the AI-empowered
23:06
individual or small team, right?
23:07
So client marketing is content, drafting is code and to some extent content.
23:12
Uh, discovery review, scheduling, follow-up, all of these are things
23:16
that AI will do better, faster, cheaper, and with increasingly
23:19
less and less human involvement, aside from the AI operator, right?
23:23
Aside from somebody who is, who is checking on the AI and making
23:26
the decisions and, and perhaps they need some amount of legal
23:30
expertise, but their primary interface with the law is through AI.
23:35
Um, there's a really interesting component for, for the law, which is,
23:39
uh, jury selection and profiling, right?
23:41
Um- I'll talk about, a little bit more of that in a second.
23:45
Uh, another thing is, is argument selection, profiling the bench, right?
23:48
Um, there are commercial systems out there that already claim to
23:51
predict a judge's ruling on a motion to dismiss at roughly 85% accuracy.
23:56
Um, as those systems become cheaper, it will become increasingly
23:59
difficult to avoid using them, right?
24:01
Um, so there was a TV show on CBS called Bull that, uh, had the
24:06
actor who I think of as the, the actor that plays, uh… He's, he's,
24:10
he's one of the actors from NCIS.
24:11
That's how, how, how I, uh, re- always remember him.
24:14
But he plays Bull.
24:15
Um, and it's, it's based on a company, um, Courtroom Sciences,
24:20
which is again, in, in, in Dallas.
24:23
Um, and it's been shaping big firm trial outcomes since the 1970s.
24:27
Uh, now their, their cost is, you know, tens of thousands of dollars a day.
24:30
I think their cheapest package is like $25,000 a day.
24:33
Um, and so you have to be able to afford that, right?
24:37
Um, that, that was a, that… You know, $25,000 a day is a
24:41
pretty incredible moat, right?
24:43
Um, and now any solopreneur, any solo practitioner who has mastered
24:49
increasingly common AI tools is gonna be able to do that, right?
24:53
Um, e- even the judge himself, right?
24:55
You know, that, that kind of argument profiling, judicial profiling.
24:59
The, the… Now, again, I, I think judges will resist this.
25:01
They will, they will want to, to resist AI.
25:05
But what do you do when, when, uh, you're getting arguments that
25:09
are, that are tuned to you, right?
25:11
When someone's looked at the law and then they, they, um- They go through and they,
25:16
"Well, we're gonna, we're gonna present this argument in, in a way that this
25:19
particular judge will like the best."
25:22
And, and I think AI is gonna be fantastic at that sort of thing.
25:25
Once you … You're like, "Here's the outcome that I want. Here is the case
25:28
law and the authority. All right, now go do all of the research that
25:31
you can do on, um, on this judge, and come up with a way of making
25:36
it sound attractive to him." Okay?
25:39
Um, and, and, you know, this will include for sure all public information, and then,
25:44
you know, for law firms and clients that, that, that have, uh, have means, probably
25:47
substantially non-public information.
25:50
Um, and how do you … The question is how you defend against that, right?
25:53
Well, with ev- as with everything, the way you defend against
25:56
AI is with another AI, right?
25:58
So, uh, you're gonna need a huge amount of AI integration, uh, where, where
26:04
you'd have an AI that's listening, that's kind of on the judge's side,
26:07
that's listening to all the arguments and telling the judge, "Hey, we think, uh,
26:12
we think this, this argument is designed to influence you in this direction."
26:16
Right?
26:17
Um, a- a- and again, it's, it's unclear what a judge is gonna do, right?
26:21
Like, if you've come up with a compelling argument, an, a- an argument that's
26:24
compelling to them personally, but also, you know, stands up, is, is, um, is,
26:30
you know, gonna, gonna be able to hold on, um, is gonna be able to stand up
26:35
to, to scrutiny, maybe survive appeal, why wouldn't you decide that, right?
26:40
And I'll note that, you know, France has already banned this practice, okay?
26:44
Um, so when we think about the changes that AI is gonna make, um, I, I, I
26:51
wanna make a, a broader, a broader claim, and this, you know, would
26:53
include things like, like AI and, or like, uh, medicine and ag that we
26:57
talked about last week or last time.
26:59
So most of these examples are command seats, right?
27:02
They are a verified human holding judgment and accountability over a
27:06
high-stakes decision at, at kind of a decisive moment, some- something
27:10
where there's going to be a, um, uh, potentially liability for it, right?
27:15
Um, and, and again, I, I, I think that there's, there's going to be p- places
27:20
in the stack in, in almost every industry where that human is going to survive.
27:25
Um, but that's, that's because those are the places where
27:31
that skill is most needed.
27:33
And so obviously, you know, you have to have a human there to, to do that
27:35
skill, but that human is going to be in that seat using AI to mediate
27:41
the, the, the command skill, right?
27:42
And occasionally the content skill, okay?
27:44
And that means that the, the day-to-day operations of these
27:48
professions is, is gonna change, okay?
27:53
So, um, so that's, that's where I think, uh, that's where I
27:56
think we're g- people are going.
27:58
I think the, the, that is kind of the, the final thing that I'll say about
28:02
just specific, um, specific professions.
28:06
Um, but I, I, as, as I was preparing for this week, I just wanted to
28:10
kind of conclude that and, and, and finish off with my thoughts.
28:14
Every time when I, when I look at these things, I look at these listicles and
28:18
people say, "Hey, there's, there's gonna be AI, AI-proof professions." Y-
28:23
yes, that means there, there's, that we're not gonna replace those jobs
28:27
entirely with AI, but all of those people are going to be using AI, okay?
28:32
All of those people are gonna be doing content, right?
28:34
Increasingly, they're gonna be doing their own marketing, and that
28:37
means that, that they're gonna be using AI for content and persuasion,
28:41
proof, the story of the work, right?
28:42
I was talking earlier about, you know, and there are plumbers whose content
28:45
is so good that they release it on YouTube and people watch it, right?
28:49
Um, all … A bunch of back office functions are also gonna become just,
28:54
you know, pure codable functions.
28:56
Um, you know, the, any accounting and billing, uh, scheduling, drafting.
29:02
I mean, again, the, the, the, the pieces of accounting that are gonna stay are
29:05
going to be, like, actual accounting, like making, making decisions about
29:08
characterization, making decisions about what, what kind of money is this, right?
29:13
That are kind of more soft skills.
29:15
Um, and then, and then at the end of the day, you're gonna have a lot of
29:18
command decisions, and not just command decisions where the AI can't, right?
29:23
Um, but command decisions where it's very, very, very expensive for the AI
29:29
to make this decision, so we'd rather have a human in the loop t- 'cause that
29:34
will dramatically, and by dramatically, I, I know I'm being told 10,000
29:37
times reduce the token costs, okay?
29:41
So now we're gonna move into what, what does this look like for the firm, right?
29:46
So there was a guy named Ronald Coase, um, or Coase, I can't re- I can't
29:50
remember how to pronounce his name.
29:51
But he, he talked about, you know, why firms exist.
29:54
So they say the, the market has coordination costs, and the
29:57
firm grows to the size at which internal cor- coordination is
30:00
cheaper than, than external coordination or outsourcing, right?
30:04
Um, so every department, if you've got a big, you know, you know, large, um,
30:10
uh, a large corporation in the modern day, uh, is that, that, that … Those
30:14
are all coordination costs, right?
30:15
You have legal, design, operations, communications, compliance, and those
30:19
people are hired because someone with money believed that routing that work
30:24
through outside contracts was going to be more costly, at least in the long run,
30:28
than employing the people directly, okay?
30:31
Um, and so we're gonna … You know, I, I think that it's important to know
30:34
that that co- that coordination problem is real, and it's not going away.
30:38
Um, and, and there's still some open questions about how much AI can absorb
30:43
about that, uh, that coordination problem.
30:45
But, uh, so we'll, we'll … We're gonna talk about two examples.
30:48
We're gonna talk about the zero human company, and then we're
30:51
gonna talk about Microsoft.
30:52
So, um- You know, one of the things that, that I talked about last week was
30:57
how, or last time, was how electronic motors, uh, changed everything, right?
31:01
So you had the, the old steam-based motors or, or, um, you know, they would
31:06
have large water-powered, uh, motors where you have a, a mechanical drive
31:10
shaft that would kind of go over the top of the, the factory floor, and everyone
31:14
would mechanically attach their little machine there to, to that big drive shaft
31:18
that was, that was over the top or maybe running through the middle of the floor.
31:21
Um, and then when electric motors came on, in- initially they just
31:24
attached a new electric motor, um, or a new internal combustion motor
31:28
to that same drive shaft, and that turned out to not, not work very well.
31:32
What they needed to do was give, you know, an electric motor at each individual
31:36
station, and then ultimately, um, they, you know, every, every, every tool,
31:41
power tools have a, a, an electric motor that's specific for each one of them.
31:44
So, um, I think that, that, that we are going to see a similar process
31:49
in, in terms of how we redesign the corporation, um, how we redesign
31:54
these business organizations.
31:56
You know, the corporation, um, was, was basically designed i- a- after the
32:01
invention of the printing press and it was kinda designed to take, take advantage
32:05
of the fact that you could print multiple identical sheets and keep track of things
32:10
and move, um, you know, push paper from, from desk to desk, and that represented
32:14
an incredible efficiency boost, right?
32:17
So in January 2026, Brian, uh, Roemmele, uh, launched what he calls
32:22
the world's first zero human company.
32:25
No human employees clocking in, AI agents running the day-to-day on a
32:28
12-year-old MacBook in his garage, right?
32:29
That's where he started.
32:31
So now again, zero humans means zero humans on payroll because
32:36
there's one guy, Brian Roemmele, who is still in the picture.
32:39
He calls himself the chairman.
32:41
So, um, he, he initially kind of framed it as a, as a C-suite of models.
32:46
He said that Grok was the CEO.
32:48
It was scanning X. It was holding meetings every 15 minutes and issuing directives.
32:53
Uh, he had Claude Code be his chief engineer.
32:56
He gave it full OS access.
32:57
It was writing scripts and downloading tools and requesting budgets.
33:00
Um, and you know, there's these hilarious interactions and, you know- One, one
33:05
of the other points that I'd make is this, there's clearly some content
33:08
still going on here, so this is all mediated by, by Brian's Twitter account.
33:12
Um, it's not like he's posted, you know, uh, cryptographically verified
33:16
chats of these interactions or anything.
33:18
So they fought with each other.
33:19
Uh, so Groq denied Claude a $150 PDF tool and demanded $1,700 for marketing.
33:26
Um, and then, uh, then Rommel announced a new hire, uh, which was
33:30
Claude Bot, uh, and, and which gave it open source hands for the agents.
33:34
And then he created this thing called Joule Work, which was basically like
33:37
an energy budget, so he's paying the agents in, uh, in energy to
33:42
run their, their computing cycles.
33:44
Uh, and, and, and this is, this has continued on, you know, uh, I don't
33:47
know exactly what the business model is or how he's making money, but
33:50
you know, last month he said he had expanded to thousands of live agents.
33:53
He has a simulation layer to stress test decisions before real capital, um, and
33:58
sales and agent payment rails are made.
33:59
Um, but again, so this is all, like, fun.
34:03
This is cool, right?
34:04
And it's, it, it's, it's kind of a fun little rabbit hole to, to dive down.
34:07
This is a zero human company, but it's not a zero human company, right?
34:11
There is still one guy, um, and you know, he's… You ask what he's doing, and he's
34:16
doing these same three skills, right?
34:18
He is coding the high level architecture.
34:20
He is, you know, setting out the internal logic of, "I want this agent
34:23
to do this, and I want that agent to do this." He's deciding what agents exist,
34:28
how they talk, what tools they get, which processes to prune or restart.
34:32
He is still producing the content on X. I mean, I'm assuming that the, that the
34:35
agents are helping with that, but he is, you know, involved in the posting
34:38
process, and he is still making the, the key decisions, what to spend, who to hire,
34:43
when to kill a runaway loop, and, and he is authorizing the, the next redesign.
34:47
And, and, you know, again, this is not what chairman means,
34:50
right, in, in normal parlance.
34:52
Um, a lot of times the chairman is somebody who kind of
34:55
understands the overall direction.
34:57
They're good with the finance piece, but they're not somebody who is intimately
35:01
involved in, in technology architecture.
35:04
Um, what, what he is is he's, he's a, he's a prime.
35:07
He is a, he's an AI paladin.
35:09
He is somebody who is really, really good at this technology and, you
35:13
know, he's someone who can design the architecture, rhetorically defend it,
35:16
and then take responsibility for it.
35:17
Okay?
35:19
'Cause his job is basically redesigning the company.
35:22
Um, so he, you know, he brings on Claude Bot, he brings on…
35:28
He, he, he decides we're gonna, we're gonna pay you in, in energy.
35:31
Um, he wired in, uh, remote offices over LM Link and hiring idle home
35:36
computers from, uh, from other, other people on the, on the internet.
35:40
And then in March, uh, he was making DDR memory for, for his agent swarm.
35:44
And so, you know, he, he is constantly evaluating where he,
35:48
he's given this, this design, this architecture, and the AIs live within
35:51
the architecture that he set up.
35:52
They're not going… They might propose changes, um, but they're
35:55
not going to really be able to think about, well, what, what don't we know?
36:00
And ma- many of these AI agents also, you know, they, they don't do
36:04
as well with brand-new tech, right?
36:06
'Cause there hasn't been-- They weren't trained on an internet where
36:09
this was common knowledge, right?
36:10
That's been, been one of the interesting challenges of, of AI,
36:13
is it, it knows, it, it has a deeper intuitive understanding of stuff that
36:20
was common knowledge on the public internet when it was trained, right?
36:24
Um, or if there was a cutoff date for the training data, whatever the
36:26
cutoff date for that training data is.
36:28
Stuff beyond that, that it has to reason about it, it, it's
36:32
much worse about reasoning it.
36:33
Okay?
36:35
Um, and, and again, uh, and this, this is, it was, it was in kind of reading
36:40
about Brian, what he's doing that, that I, I learned this, that there is a, a,
36:44
a token cost to making decisions, right?
36:47
If you, if you ask an AI to make a decision, if you give it open-ended
36:51
autonomy and you say, "Just keep, just keep looping until it acts," uh, it
36:54
can cost tens of thousands of times more inference than just laying, saying
36:58
to the AI, "Give me your top 10, 5 options or, or 10 options," and even,
37:02
you know, getting, charting a course in detail, and then I'll pick, right?
37:06
Um, so I wanna emphasize this again.
37:09
Autonomy isn't free, right?
37:11
Like, the fact that the, that the model can do this and, and maybe even
37:16
at approaching, you know, five nines percentage points make the right decision,
37:21
well, that, that doesn't mean that, that that's gonna be cost-effective
37:25
in most industries anytime soon, okay?
37:27
Um, and, and, and even, you know, there's al- there's costs in terms
37:32
of how much, how much energy you have to feed it, how much you have
37:34
to pay for that kind of compute.
37:36
There's also time because there is just a, a time component that the, that the
37:41
models have to work through, right?
37:43
Uh, there, it, it's gonna be a long time in many industries, I think,
37:47
before the, the models can make a decision a- in a feedback loop and
37:52
be fast enough for that to, to work.
37:55
Um, in many cases, I think h- just having a human in there that says, "I'm gonna
37:58
look at these three to five options, and I'm gonna make a decision quickly,"
38:02
uh, that's going to, uh, that's going to make workflows economically viable
38:07
long before we get there with the, with, with a pure autonomous model, okay?
38:11
Um, so again, chairman is not the right word for what Brian is doing.
38:18
I think the right word is paladin or prime or something, something
38:21
that acknowledges, you know, he, he is making all the key decisions.
38:25
Um, and, and I think that, that, um- You know, one of, one of the reasons
38:30
is because, like, you can't just freeze the agent org chart and, and, and say,
38:36
"Well, now instead of having humans doing these different jobs, we're gonna
38:39
have AI do these different jobs," right?
38:41
AI reasons differently, it processes memory differently than humans.
38:45
There will be enormous efficiencies to find if you move, um, if you're
38:49
able to say, "Well, somebody needs to redesign it," okay?
38:52
And I think that Brian, you know, imagines that at some point he'll create a stable
38:55
workflow, that the zero human company will be able to run, run without him.
38:59
That, that seems… That, that's how he comes off on Twitter, right?
39:02
Um, and, and I think that, that there's gonna be applications where that's true.
39:06
Things like a, a fabrication run, you know, a manufacturing run, you know?
39:10
Uh, or an aquaponics greenhouse.
39:12
Some kind of limited, defined space.
39:15
Something that, uh, that, that could be run on some level by
39:18
a CNC machine or robot arms.
39:20
Um, you know, famously Elon's building this, this massive
39:23
chip fab in, uh, in Texas.
39:26
God bless him.
39:27
Um, and the whole thing… You know, it's, it's funny because on the
39:30
one hand we're used to chip fabs.
39:31
We're like, "Oh, this is gonna be great for the community."
39:33
I mean, I think it will be.
39:34
I think it'll… It's a good thing that it's here, but it's not going to
39:37
generate jobs in the same way that, that, that factories would in the past.
39:40
I mean, the, the, the factory is going to be as autonomous as anything
39:44
has, that, that's ever been, right?
39:45
Anything inside a box, right, whether that box is a CNC machine or a building, it, y-
39:51
you're gonna be able to get to the point where you can optimize it and, um, yeah,
39:55
the, the, the inference is good enough.
39:57
It's making decisions fast enough.
39:58
It doesn't need to be improved.
40:00
But what about for a company, right?
40:03
What about when you're competing in the markets?
40:04
What about when you're solving those massive integration problems,
40:07
the, the coordination problems?
40:08
What about when you're providing liability backstops?
40:12
At this point, I, I just don't see how there will ever be a time
40:16
where that doesn't require humans.
40:18
It does- where, where humans thinking, the, especially our capacity to think
40:21
outside of the box and say, what, what level of abstraction is it most
40:26
efficient to try to solve this problem in?
40:28
You know, there's, there's that old, that old line about, uh, you
40:31
cannot solve a problem at the same level that it was created, right?
40:34
You have to go, you know, deeper or higher, whichever de-
40:36
depending on kinda how you're, how you're thinking about it, right?
40:39
So what does this look for, like for the, the modern tech giants?
40:42
What does this look like for Microsoft?
40:44
Um, so Microsoft, uh, you know, let's think about Microsoft today.
40:49
They, they still run as a command hierarchy.
40:51
They have Windows, they have Microsoft 365, Azure, Gaming, LinkedIn, GitHub,
40:56
and they're, that's, they're all under one CEO, and they have one kind of
40:59
equity narrative that they, that they tell, uh, to the market at least.
41:03
Um, externally they, they have a, a multi-model story.
41:06
They have Foundry, they have Copilot.
41:08
Um, they, they are really pushing more that, that they want you to be able to
41:12
swap the model and stay on Azure, right?
41:15
Uh, that's their, their, their, their data, data repository, their
41:19
data management system, right?
41:21
Um, now internally it's, you know, they're, they're, they're
41:23
wanting you to use Copilot, right?
41:25
Um- But it's interesting, you know, the, the thing that people… The
41:29
reason why people buy Microsoft, um, is not necessarily because
41:33
it's the best software, right?
41:35
Um, uh, it- it- it's because there's a, a, a lack of risk there, right?
41:42
Um, you know, the, the, there's this old phrase, you know, no- no- nobody ever got
41:45
fired for buying Microsoft Azure, right?
41:48
Like, there… Now, now people might get fired for f- implementing it improperly,
41:53
for setting it up wrong, right?
41:54
You know, one of the things about these deep Microsoft projects is
41:57
they're, they're more customizable f- in, in many ways, and that, and
42:03
that means that you really can set them up in a way that's, that's…
42:06
You can screw up the setup.
42:07
Um, but it's such a robust product.
42:10
It's so familiar.
42:11
So many people use it, and especially if you're in any kind of, like, compliance
42:15
type of a, uh, type of a thing.
42:17
Um, you know, if, if you, um, make a decision to, to buy Microsoft
42:25
and something goes wrong, it's the fault of the implementer.
42:27
It's the fault… You know, basically the person who makes
42:30
decisions to buy Microsoft Azure, that's like a VP of a company.
42:34
That's somebody in the C-suite.
42:35
The person who sets it up is, you know, a frontline guy, um, and, and those
42:42
are, those are two different levels of, of firings taking place, right?
42:45
If you were…
42:46
If you… But if you're the, if you're the VP and you say, "Well,
42:49
we're gonna use this untested new software," right, and something
42:53
goes wrong, well, that's your fault.
42:55
You decided you wanted the untested new software, okay?
43:00
And that brings us to the, uh, the forward deployed engineer, right?
43:05
Because I think that this is, this is all gonna change, right?
43:08
So, uh, Mark Cuban, um, has, uh-- was on a podcast, and he, he talked
43:13
about, you know, the, the demand side.
43:15
He thinks that software is gonna die as a shared project, uh, product when
43:18
every firm can afford customization to its own utilization, and there
43:22
are tens of millions of US companies that will need someone to build that.
43:26
So the, the job title that people are, are throwing around on the internet
43:29
is the forward deployed engineer, FDE.
43:31
And this, as, as in so many other things, is a Palantir framing.
43:35
It's a small team with an end-to-end owner, kind of like a startup
43:38
CTO embedded inside the customer.
43:40
Um, so Frontier Labs and the big vendors, including Microsoft, are already hiring
43:44
them in volume, and this is, this is, this is the new, the new model, right?
43:48
The, the deployments die at integration, compliance, and the ops handoff
43:52
and, and that's what the, what the FDE is supposed to, uh, to solve.
43:56
So an FDE wears three hats.
43:58
One, they're a software engineer writing production code in the customer's stack.
44:02
Uh, two, they are a business consultant discovering what
44:05
the domain actually needs.
44:06
And three, they're a product manager fielding field pain,
44:08
uh, back into the vendor's core.
44:11
And this sounds, friends, a lot like code, content, and command to me, right?
44:16
the, uh, the forward deployed engineer is, is the guy that, that, that
44:19
bridges that last mile between an AI product and, and, uh, enterprise value.
44:24
Um, and I think what's important to understand here is that, that the FDE
44:28
looks like someone who is designing custom workflows with a big company
44:32
like Microsoft providing a set of compute capacities or underlying app
44:35
primitives that are designed to work in an increasingly modular stack.
44:39
And, and, and this is how, uh, Microsoft is gonna compete with
44:43
the zero human company, right?
44:45
Because Microsoft is, is seeing, I'm sure, everything that Bryan and other
44:48
people like him are putting out.
44:50
They're seeing what, you know, Jack Dorsey is doing over at Block and saying, "How
44:53
do we, how do we compete with that?" Okay?
44:57
So, but as power moves from the product organization that ships a, a horizon
45:02
feature, um, to the people that, who can make Azure and Copilot do a
45:06
specific thing for that customer's idiosyncratic needs, they're…
45:10
That's custom work, right?
45:12
Custom work means liability, means responsibility, and so there's gonna
45:16
be i- uh, pressure to say, "Well, actually, we want, we want these
45:21
forward deployed engineers, but we kind of want something that bears the
45:24
risk, uh, between them and us," right?
45:27
Um, so the old chart, the old org chart assumes that the, the moat
45:31
was, was intellectual property created at Microsoft headquarters
45:34
and, and consumed at the edge.
45:36
And the FDE model assumes that the moat is, is hardware primitives
45:40
and, and the ability to, to, to make this thing do what the, what the
45:44
customer specifically needs, okay?
45:47
And so when, when we've got forward deployed engineers that are, that are
45:54
arranging the core services that, uh, that Microsoft provides, th- that,
45:59
again, looks more and more culturally.
46:01
That, that forward deployed engineer is going to think of himself as competing
46:05
with, with Brian Romel, and, and, and he's gonna treat Microsoft more and
46:10
more like a c- a hardware vendor, right?
46:12
Microsoft is providing the compute.
46:14
Microsoft is providing a bunch of off-the-shelf components, but he is
46:19
kind of in business for himself, okay?
46:22
So if this is the new structure, what does a sane reorganization look like?
46:27
And I, I think a good answer is to look at Visa.
46:30
So in the 1960s, the new technology, uh, that people were
46:33
dealing was the credit card.
46:35
Um, you, you had, you know, the, the new plastic credit card with the ridges.
46:38
You could, uh, sometimes you could clear that electronically.
46:41
Sometimes people would still use these old, uh, these, these paper cards, um,
46:47
where you would, you would run the, the card through, uh, you know, um, contact
46:52
paper or th- things that would, would show up, um… Ah, what am… Carbon paper.
46:55
That's the word I'm looking for.
46:56
And, and then you have an imprint of the card.
46:59
That's why the, that's why the numbers were raised in a lot of these old cards.
47:02
Um, and it's important to think about the fact that, that before this,
47:05
payments were, were local and closed.
47:07
So you would have a, a department store charge, uh, card that
47:11
would work at one mer- merchant.
47:12
A gas card would work at one brand.
47:14
Uh, Diners Club and American Express were travel and entertainment charge
47:18
cards, um, and, and they, they worked in a lot of places, but they were aimed
47:23
at kind of expense accounts elites.
47:25
They were not, they were not a, a mass market product, and they certainly were
47:29
not shared between competing banks.
47:31
Like, that, this, this was the thing that was so interesting
47:34
about the Visa card, right?
47:36
So uh, you know, it started off as the, the, uh, BankAmericard- And, and
47:42
you have national licensing, right?
47:45
So you have an ordinary consumer revolving credit.
47:47
It had one brand, it was issued by many banks, and, and a
47:51
incr-incredible number of people had to honor every franchisee's card.
47:56
Um, so this, this created a nationwide common market for digital
47:59
money, and that was something that the old systems could not do.
48:02
And, and it was massively successful, but it created a bunch of, uh, of problems.
48:07
So Bank of America had not accounted for the new competition between their
48:11
members that they had created, right?
48:13
Because i- and, and the way that that competition would work.
48:16
So by the late 1960s, the technology was, was chaos.
48:19
There was, there was weak authorization and clearance, and clearing.
48:22
There was a ton of fraud and, and just a bunch of losses, and then, you know,
48:27
the, the, the franchisees would kind of get together and, and blame each other.
48:30
And, and there was nothing wrong with the system.
48:32
The technology worked.
48:33
It worked exactly as advertised and in many ways was, was better.
48:36
Um, but the, the system wasn't working because you needed a
48:41
new social arrangement, okay?
48:44
Um, so… And one, I think the, one of the things that they, they just
48:49
didn't really anticipate was how much competition they had to allow for, right?
48:55
Um, they… People were now able to compete with one another for clients,
49:02
for customers, um, and that was, that was one of the, the aspects
49:07
that the technology facilitated.
49:08
Because again, it was, it was not just a common market for money, it was a common
49:11
market for credit card customers, right?
49:14
So everybody could now be wired into the same, the same thing, and it
49:18
wasn't, it wasn't a big hassle to change, for the customer, to change
49:22
from a card issued by this bank to a card issued by that bank, okay?
49:27
So they, they had to change their management structure,
49:30
their social structure.
49:31
They called it, uh, chaordic structure, uh, so chaos plus order.
49:35
So they, they said, "We're all gonna compete on what are the rules, what is
49:38
the brand, how do we do authorization, clearing and settlement?" And, and
49:42
they're competing on, on kind of the financial products, what deals
49:45
that the cards offers, and they're competing on the customers, okay?
49:48
And I'll note this is also a very liability efficient structure.
49:53
So virtually all complaints and decisions were owned at the member level, meaning
49:57
Visa itself was allowed to focus on the piece they did best, which was
50:00
the, the payments technology, right?
50:03
So if we turn, if we say, "What would this, this look like for, for Microsoft?"
50:09
Well, this looks a lot like, you know, Microsoft becoming
50:12
kind of a, a neo-feudal, um, a neo-feudal structure, right?
50:16
So Microsoft is the sovereign.
50:18
It's got identity, it's got graph memory, it's got compliance, um,
50:21
and, and it's still, you know, for, for certainly for a generation,
50:24
it's going to be a career safe rail.
50:26
No… Again, no one is gonna get fired for buying Microsoft, and that is, is
50:30
kind of the core of their offering.
50:32
Then you're gonna have the forward deployed engineers
50:34
and the integrator houses.
50:35
These are your barons.
50:36
They're, they're vassals who can, uh, in principle switch overlords
50:40
or run, run, run for multiple people, but they'll probably have
50:43
one core offering that they…
50:45
One, one core, uh, sovereign that they serve.
50:48
Um, and increasingly, the enterprise customers are, I think, going to be serfs.
50:52
They're, they're bound less by the license fee than by audit, career
50:55
risk, um, the way the data works.
50:58
And, a- and even if they were to switch, it- it's going to… It's always
51:01
going to be easier to switch to a new vendor, to a different baron, if you
51:04
will, uh, than it will be to switch to an entirely new system, right?
51:09
This is, this is also true, one of the, one of the dynamics with AI
51:13
that, that, that w- that they're gonna have to deal with is the potential,
51:17
the new potential that, that comes from AI vendors being able to cheaply
51:21
compete with their clients, right?
51:22
So Chamath, uh, has put out, uh, some stuff about, you know, if you
51:27
deploy OpenAI or Anthropic deep into your firm, you are letting
51:30
the fox in the henhouse, right?
51:32
So, you know, PwC and Accenture trained PricewaterhouseCooper, um,
51:37
ha- spent tens of thousands on, on Claude, and it's starting to look
51:42
like the, the, the labs are, you know, deploying competitors to them, right?
51:47
So the AI vendor is using your usage to get better at replacing you.
51:50
Um, the same thing is true, uh, in, in pharmaceuticals.
51:53
So Anthropic, uh, sells Claude Science to drug makers, um, and, you know,
51:59
big names like Sanofi, Novo Nordisk, uh, were among the named clients.
52:02
And then in June 2026, it announced that it would enter the pharmaceutical
52:06
market and develop drugs itself.
52:08
Um, this, this, this struck a lot of people as, as unfair and, and I think
52:13
it created a big backlash, but, uh, but this is what's happening, right?
52:17
This is what's happening.
52:18
So, so how do you solve that?
52:21
And, and I think the way you solve that is by, you know, to lean into
52:24
the, uh, into the, the feudal, the neo-feudal metaphor, is to give the
52:27
baron skin in the game, all right?
52:29
Um, so you grant the forward deployed engineers a real chunk of the
52:33
profits, maybe even a chunk of the equity, um, i- i- i- in these kind
52:37
of surf clients that are, I think, gonna be more and more captive.
52:40
If they can't figure out how to make AI do it for themselves, right, um, then,
52:46
then they're just gonna be more and more unable to move away from this, right?
52:51
And, and so what you wanna do is say, "Hey, we don't want you to compete
52:54
with us. We want you, we want you to be invested in our success."
52:59
Um, and, and so you give them, you give them some of the upside.
53:02
So this is gonna be the Visa logic again.
53:04
Push liability and upside to the edge that actually holds the relationship,
53:07
um, uh, 'cause a salary deployer who owns nothing is, is, is gonna,
53:12
you know, defect endlessly, right?
53:14
But a baron, somebody who see I, I own this territory, these are my clients,
53:19
this is, this is my thing that I'm doing, um, they're, they're gonna
53:23
defend the surf against being eaten, including by their vendor, right?
53:27
Um, and, and I think again, th- there's, there's… One of the
53:30
key questions in here is who has access to all, to the underlying
53:35
chats, right, and the underlying processing that the agent is doing.
53:38
And if Microsoft can, you know, say, and they will need to say credibly, "Hey,
53:43
we're gonna have systems where cr- you know, cryptographically, we actually
53:48
don't have access to any of that stuff.
53:50
Um, we're gonna push that to the forward deployed engineer.
53:53
He'll have access to it.
53:54
He'll be able to grant access to, um, the, the executives of the company," that,
54:00
that seems like a, a really interesting, uh, a really interesting possibility.
54:04
So to conclude Uh, what, what is the market value of, of an FDE, right?
54:10
So this is … And, and again, I think, I think this is too small of a name, right?
54:14
Like I talked about, um, i- i- two weeks ago, I talked about how builder
54:18
is a name that's gotten thrown around, and I think this is just, that's
54:21
just too small of a name, right?
54:23
Um, so what is the market value of a forward deployed engineer, right?
54:28
And, and, and again, keep in mind that, that the paralegals and IT professionals,
54:31
and the technical support assistants, and the forward deployed engineers
54:34
that have been software engineers for their whole careers, they will
54:38
all be competing against each other.
54:40
They will all be priced a- a- against each other.
54:43
I mean, that forward deployed engineer is worth a ton, right?
54:48
He, he becomes the bottleneck.
54:50
He becomes the, the one thing standing between this firm a- and greatness.
54:54
And that start … You know, when, when you start asking the questions like,
54:56
like how, how much is that person worth, well, i- i- … And this is especially
55:01
true as we're, we're looking at education and training costs just dropping near
55:05
to zero as, as training and education becomes an AI mediated thing as well.
55:10
A- and you get this interesting point where let's say you got a cohort of
55:13
people, you know, one in 20, um, and maybe even one in 50 or one in 100.
55:19
Um, if you get one person that can, that can develop into a true prime who can
55:23
enter any industry and compete at the highest level, that more than repays
55:26
the training of a whole class, right?
55:28
It, it, it will likely even pay for maybe even a lifetime support
55:32
for the rest of the cohort.
55:33
Not that we should necessarily do that, right?
55:36
It's not … That, that's a separate social question.
55:38
Um, and, and … But, but I think the fact that we could do that, the fact that
55:42
we can say, "Hey, we can push on people. We can even potentially take risks in
55:47
training people because it's so important to get as many of the, the AI palims."
55:53
The fact that the numbers work for that is really int- in- interesting.
55:57
So, uh, when the upside of one AI operator dwarfs the cost of the pipeline
56:01
that produced him, that, that, that allows you to take these sorts of risks.
56:06
So, and again, I believe that societies that can maximize the
56:10
production, it's not just, you know, individual operators or companies.
56:13
I think it's, it's societies as a whole that can maximize the
56:16
production of trustworthy primes.
56:17
I think they'll outcompete societies that cannot.
56:19
Okay?
56:22
Um, and so as we're, as we're, you know, concluding, again, code, content,
56:27
command, this is one skill set.
56:30
It's gonna apply in, in military applications.
56:32
It'll have civilian applications.
56:34
It's gonna apply to the trade, to medicine, to therapy, to law.
56:37
I don't see any occupation that, that escapes competition, at
56:41
least within the job, right?
56:43
This is not to say that humans are obsolete, but somewhere in the stack,
56:46
there's gonna be humans, and AI-assisted humans are going to out-compete,
56:51
um, all of their competitors.
56:54
And that… A- and when, then when we realize, hey, a, a Prime is, is
56:58
going to create enough value to make it worthwhile to train entire cohorts
57:04
of people just to get one, right?
57:06
Um, a- a- and what, what that means is, is that people will invest in this, right?
57:11
It means, you know, I was having a conversation with somebody the other
57:13
day like, "Well, but y- you know, you only, if only people with money will be
57:17
able to do this, how, how f- we're not gonna have enough, we're not gonna have
57:20
enough people to be the AI paladins."
57:22
I'm like, "I think money will flow to people with the skill set.
57:25
Money will flow to people that might have the skill set," right?
57:29
Um, so that's n- that's the framing.
57:34
This is how important this, this skill set will be.
57:37
This is how important the- these people who will be quickly, I think, a new class,
57:43
a new group of people who have this, a, a very similar set of core experiences,
57:48
who have a similar set of core interests, who are going to, you know, be, be
57:52
in a position to wield political power to their collective interests.
57:57
Um, and, and so I, I think that is the, the, the thing that, that is coming.
58:03
That is the future that I see.
58:05
And so the question will be how, how will society as a whole, uh, respond to it?
58:09
Again, at the, at, at the individual level I think there will be people
58:12
that, that, that master AI and out-compete, uh, their peers.
58:15
At the corporate level there's gonna be, you know, the zero human companies
58:18
or, you know, baron, the barons of, uh, of, of Microsoft, right?
58:23
Um, and then next week we're gonna talk about the, the Roman cursus honorum, um,
58:29
and, and how that's going to apply to, um, to how we could potentially redesign our
58:36
society or redesign chunks of it, right?
58:39
Uh, one of the things is we wouldn't have to redesign … When I say
58:41
society, that doesn't mean, you know, America, 300 million people.
58:44
You could do it on the scale of, of a, of, of a small, a small city.
58:49
May- a large town kind of a thing, deal.
58:51
Um, but, but, you know, which, which, you know, is what Rome was compared to w- you
58:56
know, for, for much of Roman history when they were running the cursus honorum,
58:58
you know, Rome was, was the size of a, of a, of a small or e- or mid-size city.
59:03
It wasn't, you know, it didn't, I don't think it hit two million
59:05
until, until nearly or right around the time that it was an empire, so.
59:09
Um, because one of the other points that I wanna make is it's not just
59:15
gonna be important to produce these people that have the skill set.
59:19
When you're giving all this trust, when you're giving all this power
59:24
to someone, um, you're gonna wanna know that they're high character.
59:28
You're gonna wanna know that there's been some kind of moral formation and, you
59:32
know, the answers that the people have been giving for the last, you know, 100
59:35
years, well, well, moral formation is just not, not predictable, not doable.
59:39
I don't think that's true, and I think it'll become necessary to have
59:43
people that have a certain, uh, a certain degree of moral formation.
59:46
So we'll, we'll get into that next time and, uh, thank you
59:50
so much for, for listening.
59:52
This is, uh, the end of my prepared remarks, and now we'll move into, uh, Q&A