Well, good morning,
Chamath. Thank you
for being here at
Dreamforce with us.
Yeah, let's just jump
right into it. You
know, there's been
a lot of talk at
Dreamforce here this
week about the transition
to the agentic
enterprise and, you
know, how the
architectures are changing.
I mean, just as
these agents become a
lot more commonplace,
you know, what are
your thoughts on,
like, what specific
things around, like,
the data and the
tech architecture
that needs to evolve?
Well, I mean, what
are we learning? I
think it's always a
good place to start.
One thing that I
think we're learning
is that there is
an emergence of
what I would call
agentic sprawl.
So, you know, when
we first, when the
only endpoint that
was usable was a model
and it was sort of
a reactive Q&A bot,
that was an extremely
small surface
area. Right. And so
the risk around that
surface area was
very easily managed.
But we've had two
iterations since
then. We've had
these basic agentic
primitives, right?
And so you have
things like Codex
and CloudCode, which
wrap and model
with these tools.
There was a certain
amount of expansion
that you can clearly
see in the revenue.
But now we've had
this next iteration,
harnesses, whether
it's in law or
customer service or
finance or sales.
So the real question
now is you have this
intersection of people
and agents doing all
kinds of different
things inside of the
company. And I think
that that sprawl, if you
will, has to be managed
because I think you're
going to see that
there's going to be a
lot of downsides that,
if not managed well,
create problems
for enterprises.
Largely, and this is
always the case, around
governance and risk.
Right. And specifically
around the risk
part, you know, it's
interesting because a lot
of these agents started
off as running on behalf
of users, especially
like you said, when
they started off early,
it was simple stuff.
Simple stuff. And
now as the models are
getting more intelligent,
where you can have
these agents that are
more long-running, they
have multiple tasks,
this notion of the
agents themselves having
their own identity,
there's a lot of
talk around that that's
going on. I was just
curious, how do you
think about that
versus broadly how
agents are running today?
Look, let's take a
couple of very precise
examples. Let's
look at maybe the
legal function in a
company and then the
developer function
in a company. Today,
let's just say that
you are an organization
that uses no AI and
no agents whatsoever.
What do developers do?
If they want to make
a change to something,
they create a PR,
right? And then there
is a peer review
system, effectively,
that tools support that
allow another person
to look at the change
that you're proposing,
understand it,
validate it,
document it, and
then push it
into production.
You're tasked with
getting an M&A
transaction done on
behalf of a company
or a business
development deal done.
You're a lawyer. You
redline a contract.
Maybe you sit with
your peers. Maybe
you sit with your
superiors. You review
a whole bunch of
cascading sets of
gives and takes.
Right. And you produce
a final document,
all fine and good.
And we have all
believed that humans
roughly have the right
intent and that, you
know, people have the
right reward function,
right? which is maybe
shareholder value
or earnings per share
or their own bonus
or their promotion,
whatever it is,
we're roughly aligned
inside of a company.
Now you have
hundreds of agents
that are doing,
as you said, long
-run tasks semi
-autonomously or in
some cases completely
autonomously.
And then one day
somebody shows
up and knocks on the door
and said, why
did this happen?
And now you can't go
to Chamath or Rohan
and say, why did
you do this? It's
an agent. It's not
sentient as much as
people may want to
believe that it is.
It doesn't have a
first and last name.
And you're supposed to
then tell a regulator
or tell a lawyer or
tell a trial judge
or tell the government
or tell a partner.
How do you explain
it? So I think that we
have this real issue
now where you have to
find ways of creating
the equivalent governance
that we had with an
org chart. But an
org chart in a world
of agents make no
sense. There's no concept
of truly enforceable
hierarchy unless things
are written down,
unless there's true
governance and guardrails,
unless there's
tripwires, unless there's
checkers checking
the checkers. Right.
And so that's a
cascading amount of
infrastructure of
which zero exists.
So when I think
about actually
using AI in a legitimate,
long-run,
serious company,
those are the issues
that we have to solve
today. And if you don't
solve it, what you
really have is this
comical view of like
token maxing. It's all
chaos. All the agents
running around like no
guardrails, nothing.
The other way to
think about this, and
I don't mean to be
sort of like, this
is a little bit
tongue in cheek, but
you know, how was
the weekend invented?
The weekend
did not exist.
Meaning the
concept of you work
five days and you
take two days off.
Why did that come
into existence at
the turn of the
industrial revolution?
It was a way for
industrialists to
manage the disparate
religious groups
that would work
together in a factory,
okay? That's actually
how it started.
Jewish people would
work on certain
days, Christians
would work on certain
days, and there
was a separation
of days. So that's
how it started.
But if you look
at productivity,
as productivity
has gone up,
it's not that
we worked more.
And if you look at
productivity, the long
-run GDP over hundreds
of years has roughly
stayed the same. So what
has exactly happened?
That as we have had
productivity since
the beginning of the
five-day work week,
right? One huge
productivity leap. We
didn't force people
to keep working seven
days. We said five
days is good enough.
And it's just
shrank. And GDP has
stayed the same,
irrespective of
whatever technology
we've created,
right? Electricity,
the industrial
revolution,
nuclear power,
software, SaaS,
cloud, mobile.
GDP has stayed roughly
between 2% and 4%.
So I think that there's
going to be this
interesting
question, which is
societally, what happens
now when you have
this leverage? Do we
actually work more?
And I think that there's
a very reasonable
expectation that
maybe we're working 30
hours or 25 hours
because the human
tendency is to basically
token max, find a way
to do their job, and
then have free time
to do things that
they value otherwise.
And I think we're going
to have to grapple
with that as a society
because I think
that's probably the
underlying truth of what
happens in a world
without governance and
all these other
capabilities. I mean,
that's really fascinating,
you know, and it's
interesting. One of
the things we've been
having a discussion
around is, you know,
to the point around,
like, when you have
these agents that become
a part of your teams,
what is the organizational
structure of
the future, right? Do
they just, you know,
become parts of teams
as we know today?
Is it a lot more
fluid and, you know,
there's this whole
notion of… What does
hierarchy mean?
What does it mean to
be a manager and
get promoted to a
senior manager? Does
an agent get promoted
to senior manager
and then director?
Does an agent become
a vice president
overviewing people
and agents? Is that
even possible? What
does any of that mean?
So the organizational
construct and the
organizational
behavior of companies
needs to go
through a pretty
profound phase of
change management.
And by the way,
that's the other
thing that we
see. So like, you
know, just running
this business that
I started two
years ago, 80, 90,
man, we work
in very large
US government, aerospace,
energy, financial
services,
the big boys.
Yeah. And the most
difficult thing
that we help them
grapple with is
not the desire to
embrace AI or do
agentic things.
They want to deliver
better products and
services to customers.
It's the change management
of the people inside
the organization.
And the change
management is harder than
it needs to be because
a lot of the time
what we're fighting
is all the myths and
disinformation that
exists outside of
the job about all the
doomerism, about all
of the negative
changes that will come.
And we don't have
enough advocates
pointing to the
positive use cases.
And so folks
naturally have a very
reasonable reaction,
which is to
say, I'm very
skeptical about this.
And so we have to
work through that
change management
very slowly. But if
we can find a way
to just show more of
the glass half full,
I think the change
management becomes
easier. And then
we can design the
org of the future.
And I do think it's
very reasonable to not
have to expect AI to
deliver 12% GDP and
instead say, you know
what, 3% or 4% long
-run GDP, we can just
live where we want
and drive our kids
to school and maybe
we're working 30 hours
and that's good too.
No, absolutely. You
know, it's interesting
you mentioned about the
skepticism piece and,
you know, having the
focus on like, hey,
look, there's a lot of
good things that happen.
And especially with
the younger generation,
I feel like having that
perception, I think
is going to make this
change a lot easier.
But, you know,
changing gears a
little bit, you
know, talking about
a lot of things
that have changed
over the last few
weeks, you know,
the Sascocalypse
narrative.
It's over. Yeah.
What do you think the
industry got wrong
there? And then what
should they be focused
on in the future?
you know, Warren
Buffett has this famous
quote, which is
in the short term,
they're a voting
machine. But in the
long term, they're
a weighing machine.
discontinuities
in the market,
the pendulum swings
violently in one or the
other direction. So a
simple example of this
would be during the
great financial crisis
leading up to it, it
was a voting machine.
And what it was voting
was about the long
run consequences of
what looked like a
lot of very, very
poorly priced debt. And
at the core of the
great financial crisis
was that, over lending
and mislending.
And when you had
this massive trade,
it pivoted very
quickly to a weighing
machine. And the
companies that really
thrived were the
ones that had long
-run, highly
predictable businesses.
this same pattern
replayed itself, where
initially people that
have a responsibility
to manage other
people's capital
felt like they needed
to make a decision.
And what it looked
like was the model
makers very aggressively
moving up market
observing workflows,
observing behavior,
and frankly, for lack
of a better phrase,
completely copying
it and trying to
cannibalize their
customers' businesses.
And whenever you have
something new versus
something established,
and this is not
right, but it happens
all the time in many
things, you are more
trusting or more aware of
the new thing. Correct.
And you take the
established thing for
granted. So in SaaS,
I think they looked past,
for example, Salesforce's
predictability,
ServiceNow's
predictability, Adobe's
predictability, and they
said, well, yes, you guys
have been here. Yes,
you've been consistent.
Yes, you've been
reliable. But we're
just going to throw all
that out the window.
And we're going to
pick the new guy.
And that then
generates momentum. And
again, it's a voting
machine. So when things
start to vote in
one direction, a lot
of people outsource
their diligence and
they just start to
vote the same way.
And now we're going
through a reset.
And in the reset,
companies that
have extremely
established long-run
businesses, I
mean, I saw Mark,
maybe it wasn't Mark,
but it was attributed
to Mark, put out
your long-run 2030
revenue forecast was
like $63 billion.
And I was like, good
***. I mean, that
is a gargantuan
juggernaut business
of just like
monstrous proportions.
And so I think what
it shows is that
the trade was, like
in most of these
reflexive trades,
pretty illegitimate.
They're very
emotional and
it tends to
be fear-based.
And then in the long run,
you kind of, you
know, the pendulum
swings back to a
more normal, balanced
way of viewing
risk. In my opinion,
established enterprises
that have deep
long-run
relationships have two
things that are
extremely valuable. On
the surface, the
superficial asset
that people will
point to is the data.
You know, other
people will make
theirs headless.
You'll make it
easier and easier
to interact.
But I think the second
thing is the critical
thing, which is there
will be people making
these decisions. And
so there is trust. And
the relationships.
And the cycle time of
trust is not something
that humans, I think
in the short term, will
outsource to agents.
I think you'll take
experiments. I think
you'll guardrail it.
But think about how
unlikely you are to
trust another human.
And so the idea
that then you
outsource trust in
how you are supposed
to do business
to effectively a
black box, I find
pretty unlikely.
Now, that's
very well said.
Maybe one final question.
You know, what's one
thing that's conventional
wisdom today that
you think when you
look ahead of the next
three to four years
won't play out the
way people are talking
about? I think what
people are meaningfully
underestimating
is that we have
assumed that there
are model and harness
families that are
integrated. So if you think
about it like – think
about like taking
a brain, embodying it
in a body, and then
training it to be a
lawyer or a doctor or
whatever. That's
effectively the natural
evolutionary path of
AI thus far. Right.
What I think people
do not understand
or appreciate
is that you can
actually tease
these things apart.
So, for example,
if you look at a
model from a frontier
lab plus their
product harness, it
turns out you can
tease them apart
and rip them apart.
You can use an anthropic
harness with an
open AI model. You
can use an open source
model with anthropic
and open AI models.
And then when you
mix and match against
different tasks,
you have completely
different behaviors
and cost structures.
So I think what
people assume is
like they assume
that there's a very
straightforward
trade where open
source cannibalizes
closed source. I
don't think that's
what's happening.
I think you're developing
this really rich
matrix of capability.
We're learning
that mixing and matching
against different
tasks create different
behavioral curves
for price, for precision,
and that's going
to create an incredibly
profound set of
capabilities that is
very poorly understood
today. And part
of it is that the
data is not widely
known, nor tested, nor
available. But I
think as you see that,
that's going to be a
huge thing where the
fear of vertical integration
will be debunked.
That'll be the next
myth to debunk.
I think that's the
time we have. Thank you
so much, Ahmad. Really
appreciate you being
here and it's great to
see you. Thanks, Ron.