Latest / Elon Musk Podcast / AI IPOs and the Jobs Apocalypse

Transcript

AI IPOs and the Jobs Apocalypse · Elon Musk Podcast

0:00
Open AI and Anthropic are gearing U for initial public
0:04
offerings that could value their companies at nearly $2 trillion
0:09
combined. But to secure that capital,
0:11
their leadership has abruptly reversed their public warnings
0:14
about an artificial intelligence jobs apocalypse.
0:17
Yeah, it is a highly convenient rhetorical shift.
0:21
I mean, they move from predicting mass unemployment to
0:24
describing this technology as a, you know, a simple productivity
0:27
multiplier exactly when they need to court public market
0:30
investors who demand societal stability.
0:33
Welcome to the debate. We're looking at a fundamental
0:35
disagreement about the future of white collar labor.
0:38
Right. The core question is whether the
0:40
current wave of generative models represents A structural
0:44
threat that will displace knowledge workers permanently,
0:48
or if this is a familiar automation cycle that will
0:50
primarily Dr. productivity and reinstate human labor in
0:54
entirely new tasks. I argue that the underlying
0:57
capability curve of this technology ensures severe labor
1:01
displacement, regardless of how corporate narratives shift to
1:04
appease financial markets. And I argue that the jobs
1:08
apocalypse narrative relies on a flawed, highly linear
1:12
understanding of Labor economics.
1:14
The economics of task automation actually favor a strong
1:18
reinstatement effect, where human labor doesn't just
1:20
disappear, but simply shifts to new, higher value activities.
1:25
To understand why the displacement effect will
1:27
dominate, we really have to look at the raw capabilities coming
1:30
out of the Frontier Labs. Take Anthropic Smithos model.
1:34
It recently demonstrated the ability to autonomously discover
1:38
thousands of 0 day vulnerabilities in software,
1:41
including hundreds in a single browser like Firefox.
1:44
Right, a zero day vulnerability just being a software flaw that
1:49
the original developers don't know about yet, meaning they
1:52
have zero days to fix it before it gets exploited.
1:55
Exactly, and traditional cybersecurity tools, they just
1:59
scan for known signatures or recognize patterns of past
2:02
attacks. Mythos is doing something
2:05
completely different. It acts like a human detective.
2:09
It reads the code, understands the underlying logic, and
2:12
actively hunts for novel loopholes that have never been
2:15
documented. It is not a tool that makes a
2:18
human worker slightly faster, it is a synthetic employee.
2:23
We are talking about a system that applies superhuman
2:25
persistence to average human intelligence.
2:28
When a technology can hold vast amounts of context, run in
2:31
parallel without fatigue, and execute tasks independently, it
2:35
fundamentally collapses the dimensionality of white collar
2:38
work. Hold on, Before we accept that
2:40
jobs are just going to collapse, walk me through the economic
2:43
framework there. Because that perspective treats
2:46
a job as a single indivisible unit of Labor that just gets
2:49
deleted with a keystroke. Economists like Darren Asimoglu
2:53
and Pascal Restropo have developed a much more accurate
2:56
task based framework. OK.
2:59
A job is just a bundle of tasks. When capital takes over certain
3:03
tasks, a displacement effect occurs.
3:05
We agree on that mechanism, but historically this is
3:08
counterbalance by a reinstatement effect.
3:11
Technology creates new tasks where humans retain A
3:14
comparative advantage. But how does that remaining
3:16
percentage of a job actually expand to fill a 40 hour week?
3:21
I mean, give me a tangible example of how reinstatement
3:24
supposedly saves the knowledge worker.
3:26
Think about an accountant. When spreadsheet software was
3:30
introduced, calculating sums by hand and double checking ledgers
3:33
used to take U you know, 90 of their week.
3:37
Software automated all of that, but the accountant wasn't fired.
3:40
Because the output changed. Exactly that remaining 10 of
3:45
their job, the financial strategy, the client advising
3:49
the complex forecasting expanded because they could model
3:52
scenarios instantly. Clients wanted more models, more
3:56
forecasts, and more strategy. You scaled their aggregate
3:59
productivity tenfold rather than eliminating the worker.
4:03
The executives you mentioned are realizing this dynamic applies
4:06
to artificial intelligence as well.
4:08
That works for an accountant because financial strategy is a
4:12
high level cognitive task that remains valuable.
4:15
But let's look at this through Alex Imas and Thorsten Shukla's
4:18
concepts of job dimensionality. Right, the number of distinct
4:22
tasks. Yeah, dimensionality is simply
4:24
the number of distinct complementary tasks a person
4:28
performs to do their job. Think of a junior financial
4:31
analyst or an entry level copywriter.
4:34
Like a Wiss Army knife, they pull data, format sreadsheets,
4:38
write daily summaries and e-mail clients.
4:41
If a software model masters just the data pulling and the
4:44
formatting the main blades of the knife, the job ceases to
4:47
exist. You don't pay a full salary for
4:49
the Corkscrew. You are assuming the demand for
4:52
the output is perfectly inelastic though.
4:55
When productivity rises and the price of an output drops,
4:58
customers almost always buy more of it.
5:01
Think about software developers after cloud computing lowered
5:04
the cost of deploying software. The demand went up.
5:07
We didn't need fewer developers. The demand for software exploded
5:10
and companies hired aggressively.
5:12
But we can see the displacement happening in the revenue numbers
5:16
right now. Anthropic's revenue has grown 80
5:19
times over against their own internal projections.
5:22
That kind of exponential revenue growth happens because the
5:25
models are successfully replacing specific cognitive
5:28
tasks entirely, not just assisting workers in a way that
5:32
creates induced demand. Asimo Glue makes a critical
5:34
distinction here that we really need to address.
5:37
He separates SOSO automation from platform technologies.
5:40
SOSO automation replaces human labor in tasks where the
5:43
productivity gains are marginal. Like a kiosk?
5:46
Right. A self checkout kiosk displaces
5:48
a cashier, but it doesn't lower the cost of groceries enough to
5:51
stimulate job creation elsewhere in the store.
5:54
It is narrow. Contrast that with a platform
5:56
technology. A typewriter made producing
5:58
legible text faster, but the personal computer was a general
6:01
purpose platform it. Created new fields.
6:04
It automated typing, yes, but it also created software
6:08
development, database administration, and IT
6:10
networking. The question is whether current
6:13
artificial intelligence is drifting into narrow labor,
6:17
replacing so so automation, or if it is a general purpose
6:21
platform that will create entirely new complementary human
6:25
tasks. But that platform analogy only
6:28
holds up in a vacuum. The reason this acts as a
6:31
displacement engine rather than a platform for new employment is
6:34
the autonomy I mentioned earlier.
6:36
Platform technologies of the past required a human operator
6:39
to extract value. Sure you need a user.
6:43
A personal computer sits completely dormant without a
6:45
user typing on the keyboard. The current models operate
6:48
autonomously. They are assigned a goal and
6:50
they execute it. That completely severs the
6:53
requirement for human complementarity.
6:54
Integrating those autonomous actions into a functioning
6:58
business still requires human oversight, strategy and
7:02
management. You are looking at the raw model
7:04
capability in isolation and ignoring the friction of the
7:08
real economy. Well, if we look at how these
7:11
companies are actually behaving in the real economy,
7:14
specifically how they talk to governments versus investors,
7:17
the platform narrative falls apart.
7:19
There is a severe disconnect in how these companies communicate
7:22
risk. We brought up Anthropic Smithos
7:24
model and its ability to find software vulnerabilities.
7:26
Right. Anthropic took that model to
7:28
government entities and essentially treated it as a
7:30
lethal cyber weapon. They established an
7:32
international containment effort called Project Glasswing.
7:35
The mechanics of that were fascinating.
7:37
They gave a select group of organizations exclusive access
7:41
to the models outputs for a short embargo window just to
7:45
build defenses before the vulnerability became public
7:48
knowledge. Exactly.
7:49
They treated it like a global security crisis.
7:53
But then the chief executive stood on a stage with the CEO of
7:56
JP Morgan Chase and pitched the exact same software to financial
8:00
services clients as a manageable patching exercise.
8:04
They called it a transitory period.
8:06
The rhetoric shifts depending on the audience.
8:08
It shifts depending on who is in the room.
8:10
When they speak to national security regulators.
8:13
It is an autonomous weapon when they speak to enterprise buyers
8:16
and public market investors. It is a safe, reliable tool for
8:20
parsing credit memos. The corporate narrative actively
8:23
obscures the true disruptive capacity of the technology.
8:26
That divergent in messaging is directly connected to the
8:30
realities of public equity markets.
8:32
These companies are preparing to raise unprecedented amounts of
8:36
capital. A combined valuation approaching
8:38
$2 trillion requires a story that institutional investors can
8:42
actually underwrite. They need stability.
8:45
Pension funds and retail buyers are not going to invest in the
8:48
systematic destruction of the middle class, nor will they fund
8:52
an unmanageable security threat. The messaging has to emphasize
8:56
stability and economic growth. But beyond the public relations
9:00
strategy, there is actual economic data supporting the
9:03
stabilization argument. What data points actually offset
9:07
a model that can run an entire enterprise workflow
9:10
autonomously? Look at the physical constraints
9:13
and the capital accumulation required to run these systems.
9:17
Goldman Sachs produced a perspective acknowledging that
9:19
per quarter of current work hours may be automated, but they
9:23
also highlight the capital accumulation required to build
9:26
the infrastructure. The data centers.
9:28
The physical build out of data centers has already created
9:30
hundreds of thousands of new construction and engineering
9:33
jobs. We are talking about acquiring
9:35
the land, securing gigawatts of electricity, installing cooling
9:39
systems, pouring the concrete, and the structural engineering.
9:42
When investment expands the production base, demand rises
9:45
for complementary human tasks. Capital accumulation physically
9:49
offsets digital displacement. The.
9:51
Math on that assumption doesn't add up because it treats manual
9:55
construction labor and white collar cognitive labor as
9:57
perfectly fungible. Fungibility implies you can swap
10:01
one unit for another seamlessly. They are different markets, yes.
10:05
The senior project manager or the financial analyst whose job
10:08
dimensionality has collapsed to 0 is not going to pivot to
10:11
pouring concrete for a data center.
10:13
The displacement is targeted strictly at the knowledge
10:15
economy, and the reinstatement you are pointing to is happening
10:18
in the physical infrastructure economy.
10:20
Those are completely different labor pools.
10:22
You are destroying a middle class office job and replacing
10:25
it with a specialized construction job that the
10:27
displaced worker has no training for.
10:29
They are different pools, but the macroeconomic effect is net
10:32
job creation or at least stabilization, and even within
10:36
the white collar sector we are seeing the creation of entirely
10:39
new roles. We need people for auditing,
10:41
compliance, prompt engineering and managing the outputs of
10:44
these models. Let's test that idea against the
10:47
internal realities of corporate governance.
10:50
Traditional corporate governance simply cannot withstand the
10:53
current technological velocity. Chief executives are struggling
10:58
to run companies on standard quarterly or annual planning
11:01
cycles when the underlying model capabilities evolve on a monthly
11:05
basis. Right, it's too fast.
11:07
You cannot plan a three-year software rollout when the tool
11:11
you are evaluating is obsolete before the procurement contract
11:14
is even signed. But that planning friction is
11:16
exactly why adoption is slower than the theoretical capability
11:20
curve suggests. Organizations absorb change
11:23
slowly. Legal departments have to review
11:25
the tools IT has to secure them, and human resources has to train
11:29
people. The employees are bypassing the
11:32
organizations completely though. If software engineers are
11:35
demanding access to agent to coding tools, or threatening to
11:38
quit and join competitors who provide them, the organization
11:41
loses control of the adoption rate.
11:43
It becomes a shadow IT problem. We're talking about systems that
11:47
don't just auto complete a line of code.
11:49
They spin up independent instances, read the company's
11:52
entire code base, write the feature, test it, and submit a
11:55
pull request, essentially filing the digital paperwork required
11:59
to merge new code into the main software branch.
12:01
When the workforce forces the adoption of highly autonomous
12:04
tools just to stay competitive, the automation of internal
12:07
processes becomes rapid and inevitable.
12:09
It bypasses any controlled top down reinstatement of human
12:13
labor. The organization hollows out
12:15
from the bottom up. That bottom up adoption argument
12:17
assumes the only thing a business produces is technical
12:20
output. It completely ignores what
12:23
consumers actually value on the other end, which is the human
12:26
premium. You mean the preference for
12:28
human interaction? Yes, there was a highly
12:31
publicized experiment where a technology executive tried to
12:34
automate his own personal communications.
12:37
He set up an artificial proxy to answer his messages, clearly
12:41
labeled as an automated assistant trained on his past
12:44
emails. How did that go?
12:46
The people communicating with him rejected it completely.
12:49
They felt it was dehumanizing and inauthentic.
12:52
The transaction wasn't just about exchanging information, it
12:55
was about the social proof of human attention.
12:58
People want to know that another human being cared enough to
13:01
spend their finite time formulating a response.
13:04
That was for personal networking and executive communication,
13:08
which is inherently relational. That doesn't protect the entry
13:11
level data processor or the logistics coordinator.
13:15
No one cares if a human or machine optimize their supply
13:18
chain route. It points to structural shift in
13:21
the economy, though, as routine cognitive work is automated
13:25
human employment shifts toward the relational sector.
13:28
We're talking about healthcare, high stakes negotiations,
13:32
bespoke services, education and specialized consulting.
13:36
Where the interaction is the product.
13:38
Exactly. In these fields, human to human
13:42
engagement is the actual product being purchased.
13:45
The empathy, the shared understanding and the
13:47
authenticity cannot be automated because the friction of dealing
13:51
with another human being is what creates the value.
13:54
That assumes the economy can support a sudden influx of
13:57
workers transitioning into the relational sector.
14:00
If the shock of cognitive automation happens as quickly as
14:04
the capability curve suggests, the friction of transitioning
14:08
millions of knowledge workers into bespoke relational roles
14:12
will cause severe economic pain. The transition will be
14:15
difficult, yes. A data entry clerk cannot become
14:18
a bespoke consultant overnight. The sanitized public offering
14:21
narratives are designed to hide this reality from investors.
14:25
The compounding capabilities of these models constitute A
14:28
structural shock that our white color professions are completely
14:32
unprepared for. The speed of the displacement
14:34
effect will overwhelm the slow, messy process of human
14:38
reallocation. The tension is certainly there
14:41
between the raw power of the models and the friction of
14:44
integrating them into the real economy.
14:46
The open question we are left with is whether our economic
14:50
institutions can actually foster the creation of those new human
14:54
centric tasks fast enough to bridge the gap.
14:57
We have to see if the market can incentivize the reinstatement
15:00
effect before the displacement effect takes a permanent toll on
15:03
the labor force. If you're not subscribed yet,
15:06
take a second and hit follow on whatever app you're using.
15:09
It helps us keep making this. We appreciate you being here.
15:12
Also, check out our YouTube channel for more business and
15:14
tech updates. There's a link in the
15:16
description.