Latest / Elon Musk Podcast / OpenAI wants to tax automated labor

Transcript

OpenAI wants to tax automated labor · Elon Musk Podcast

0:00
Imagine building a machine so powerful, like so capable of
0:03
doing human work that you actually March into Washington
0:07
and ask the government to tax your own invention just to keep
0:11
the economy from collapsing. You know, and that is exactly
0:14
what Open AIA tech company valued at over $800 billion is
0:18
officially doing right now. They are asking the government
0:21
to tax their own automated labor and then use that money to pay
0:25
for a four day work week and a public wealth fund for every
0:29
single citizen. Hearing a private corporation
0:31
propose a total restructuring of the national economy, I mean, it
0:35
really stops you in your track. It really does.
0:38
And when you place that proposal right next to the recent
0:40
warnings from Bank of England Governor Andrew Bailey, a very
0:44
distinct picture starts to form because he pointed out that
0:47
artificial intelligence is about to displace workers in a way
0:51
that looks a lot like the Industrial Revolution.
0:53
So you have this incredibly optimistic utopian proposal
0:56
coming directly from the tech giant building the tools sitting
1:00
right across from these stark, unvarnished economic warnings
1:04
from a central banker who is looking at the actual labor
1:07
data. We are looking at a fundamental
1:10
tension between the creators of this technology and the
1:13
traditional financial institutions trying to brace for
1:15
impact. Exactly because both sides,
1:19
despite their entirely different vantage points, are essentially
1:22
staring at the exact same impending reality.
1:25
They see the exact same labor trends happening right now.
1:28
So if these machines are about to become smarter than us and do
1:31
the bulk of the work, how exactly does the money keep
1:34
flowing so the whole economic system doesn't collapse?
1:37
Well, to understand the mechanics of that, we have to
1:39
start with Andrew Bailey's primary concern over the Bank of
1:42
England. OK, He believes artificial
1:44
intelligence will not 'cause, you know, mass unemployment
1:47
across the entire economy, but it will absolutely displace A
1:51
narrow set of highly automatable roles.
1:54
Like secific desk jobs? Right secifically hitting young
1:58
entry level rofessionals. The impact is intensely targeted
2:01
at the bottom. And the data supports that
2:04
entirely. It introduces this concept
2:07
economists are calling the broken rung.
2:10
So a recent Stanford study looked into this and they found
2:13
a 13% relative employment decline for early career workers
2:17
aged 22 to 25 in exposed occupations. 13 percent is a big
2:21
drop. It is, and in the UK the data
2:24
shows entry level rolls have dropped significantly too.
2:27
Companies are simply not hiring for those junior positions at
2:30
the same rate. Think about your own career for
2:33
a second. Or like the first real job you
2:35
ever had. Oh.
2:36
Absolutely you. Probably spent the first few
2:38
years doing the grunt work. Yeah.
2:39
If you were a junior lawyer, you were trapped in a room doing
2:42
document review. Yeah, or if you were a junior
2:44
developer, you were writing basic code or hunting for syntax
2:48
errors all night. Right, the boring stuff.
2:50
Exactly. But that wrote work is how you
2:52
learn the Business Today though. The software can do the initial
2:56
drafting, the basic coding and the routine data processing
2:59
instantly. So the partners at the law firm
3:02
or the senior engineers, they just use the software instead of
3:04
hiring a 22 year old? They do.
3:06
Wait, back up. If the machines do all the
3:09
junior level work like document review or basic code generation,
3:13
how do human workers ever get the experience to become senior
3:16
lawyers or senior developers? That is exactly the crisis.
3:20
If you remove the bottom rung of the career ladder, the pipeline
3:24
of future talent breaks completely.
3:27
I mean, how do you train a general when there are no foot
3:30
soldiers? You can't, right?
3:32
The consequence of this is what the firm ICS dot AI calls the
3:36
human firewall. The human firewall.
3:38
Yeah, the fundamental rule moving forward is that
3:41
artificial intelligence proposes humans dispose and humans own
3:45
the legal outcomes. This changes the entire career
3:48
pipeline. It limits the traditional
3:50
apprenticeship model where you learn pursuing the grunt work.
3:53
We are now forcing a totally new requirement.
3:56
Entry level workers must start their careers as managers of
3:59
machine output. Let's make that concrete,
4:01
because it sounds like asking someone to be an executive chef
4:04
without ever letting them chop an onion or work the line.
4:07
That's a great way to look at it.
4:08
Like they have to taste and approve the dish, and they are
4:12
legally responsible if the food poisons a customer, but they
4:16
never learn the physical muscle memory of cooking.
4:19
That is the perfect analogy. You're asking a 22 year old to
4:23
evaluate the quality of a legal brief or a string of code that a
4:28
machine generated in three seconds.
4:30
Right, but that 22 year old has never spent 40 hours writing one
4:34
from scratch. They lack the foundational
4:37
context. They just don't have the reps.
4:39
Exactly. They don't have the neural
4:40
pathways built through years of trial and error, yet they're
4:44
functioning as the human firewall protecting the company
4:47
from a machine hallucination. Or, you know, huge legal
4:50
liability. So they are entirely responsible
4:52
for an output they do not fully comprehend, and the machine
4:55
output they are managing is only getting more complex.
4:58
Open AI has officially declared that we are in a transition
5:01
towards super intelligence, meaning systems capable of
5:03
outperforming the smartest humans even when those humans
5:07
are assisted by artificial intelligence.
5:09
We can see exactly how fast this is happening through their
5:12
internal metrics too. They use an evaluation called
5:15
GDP Val to measure performance on economically valuable tasks.
5:19
GDP, Val. Right.
5:21
And this is a critical distinction.
5:23
They are not just testing if the system can write a fun poem or
5:27
pass a standardized test. They're testing real work.
5:30
Yes, measuring if the system can perform tasks that a business
5:33
would actually pay a human being a salary to do.
5:36
Systems operating at a G PT-5 level, which internally were
5:40
code names. Spud now match or exceed human
5:43
professionals on about 50% of real world tasks. 50% of real
5:48
world tasks and they are completing them in minutes
5:51
instead of hours. Minutes, sometimes seconds.
5:53
And furthermore, when they analyze one and a half million
5:56
consumer conversations, the interactions were heavily skewed
5:59
toward information seeking, practical guidance, and writing.
6:02
Meaning people are using it to do their jobs.
6:04
Exactly. People are using these systems
6:06
to make complex decisions and streamline highly administrative
6:09
chores. It is essentially more ask than
6:11
task. More ask than task.
6:13
Right, the user is not doing the work.
6:17
They're directing a highly capable digital entity to do the
6:20
work for them. So the machines are getting too
6:23
fast for the normal 40 hour work week to make sense anymore.
6:26
That speed fundamentally limits the traditional exchange of time
6:30
for money. Our entire economic system is
6:33
built on the premise that you trade a certain number of hours
6:35
of your labor for a certain amount of currency.
6:38
If you are an architect, you might bill your client for 40
6:41
hours of drafting, but if a project that traditionally took
6:45
an entire team of humans months to complete now takes machine a
6:49
few minutes, paying people by the hour completely breaks down.
6:53
Because what do you bill for? I mean, if you charge $200.00 an
6:56
hour and the machine finishes the task in four seconds, do you
6:59
send the client an invoice for $0.22?
7:01
No, you can't. The business model just
7:03
vaporizes. Exactly.
7:04
You cannot bill for time when the time required drops to near
7:07
0, and this consequence forces the necessity for open AI is
7:13
radical industrial policy for the intelligence age.
7:17
They recognize that the foundational math of the labor
7:20
market is evaporating. Because if human beings cannot
7:23
sell their time, they cannot earn a paycheck.
7:25
Right. And if they cannot earn a
7:26
paycheck, they cannot participate in the consumer
7:28
economy. Which brings us directly to open
7:31
AI specific fiscal proposals. They are advocating for shifting
7:36
the tax base away from payroll and toward capital.
7:40
This includes implementing what they call taxes related to
7:44
automated labor, often referred to as robot taxes, and creating
7:48
a public wealth fund. The mechanics of this are
7:51
crucial to understand. Consider your own paycheck.
7:54
Every time you get paid, you see deductions for essential
7:57
services like Social Security, Medicare, and various state and
8:01
federal programs. Our entire public safety net is
8:04
funded primarily through wage and payroll taxes.
8:07
Right, it relies entirely on human beings working and paying
8:09
into the system. Yes.
8:10
So if artificial intelligence displaces a significant
8:13
percentage of human workers, the wage and payroll taxes that fund
8:17
those essential services will just plummet.
8:19
The government's primary source of revenue dries up.
8:21
Exactly. The safety net will run out of
8:23
money because there are fewer humans earning wages to tax.
8:27
Precisely to fix this structural deficit, Open AI proposes that
8:32
tech companies and the government jointly cede a
8:34
national fund. The public wealth fund.
8:37
Right. This public wealth fund would
8:38
pay a direct dividend to every single citizen, regardless of
8:42
their personal wealth or investments.
8:44
If a server farm is doing the work of 10,000 accountants, the
8:48
government taxes the output of that server farm and distributes
8:51
the money to the citizens. As we look at these proposals,
8:54
we have to remain completely neutral on the political and
8:57
economic ideologies involved. But I have to point out the
9:00
irony here. Oh, for sure.
9:01
You have a company valued at over $800 billion aggressively
9:06
expanding its commercial footprint and charging for
9:08
access to its tools, suddenly asking the government to tax
9:11
them heavily and distribute their profits to the public.
9:14
It does seem entirely contradictory.
9:16
A private entity racing to capture global market share is
9:20
simultaneously drafting the blueprint for how how the
9:23
government should take a portion of its revenue.
9:25
But consider their position. If their product is so
9:29
successful that it eliminates the earning power of the middle
9:31
class, who is going to buy the goods and services that their
9:35
artificial intelligence helps produce?
9:37
Nobody. Right.
9:39
A perfectly efficient economy is useless if there are no
9:41
consumers with money to spend. The consequence of this proposal
9:45
changes the very definition of wealth generation in society.
9:49
It limits the accumulation of wealth to just a few tech giants
9:53
and introduces a system where artificial intelligence access
9:56
and its resulting economic benefits are treated as
9:58
fundamental human rights operating on the exact same
10:01
level as literacy or access to electricity.
10:04
They are arguing that intelligence is becoming an
10:07
abundant taxable utility, and the profits derived from that
10:10
utility must be socialized to maintain societal stability.
10:13
And they extend this idea of socializing the benefits
10:16
directly into the workplace, too.
10:18
How so? By proposing mandated pilot
10:20
programs for a 32 hour 4 day workweek with absolutely no loss
10:25
in pay. I really want to focus on this
10:27
because it challenges everything we know about corporate
10:29
behavior. They call this the efficiency
10:31
dividend. OK, the core concept is that
10:33
when a company implements artificial intelligence and
10:36
experiences a huge surge in productivity, those games should
10:41
be returned to the workforce in the form of time.
10:44
Rather than just being captured by the executives and
10:46
shareholders as corporate profit.
10:48
Exactly. If the work gets done faster,
10:51
the worker gets their life back. Hold on.
10:53
Yeah, a company is supposed to pay for the artificial
10:57
intelligence subscription that does the work, pay the human the
11:00
exact same salary as before, but only have them work four days a
11:05
week. That is the proposal.
11:06
Wow. It severely limits the
11:08
traditional corporate profit margin on automation.
11:11
Historically, if a machine made a factory twice as assistant,
11:15
the factory owner fired half the staff and kept the extra money.
11:18
Right, that's just basic capitalism.
11:19
But this proposal actively prevents that.
11:22
Alongside this, they propose A societal shift toward the care
11:26
economy. The care economy.
11:28
Let's spend some time on this, because this feels like a
11:30
fundamental rewiring of what society values.
11:33
Well displaced workers would be actively encouraged to move into
11:36
roles involving child care, elder care and community
11:40
services. OK, think about it.
11:42
For centuries society has compensated you based on either
11:45
your physical Braun or your cognitive brain power, right?
11:48
If robotics handles the physical labor and artificial
11:52
intelligence handles the cognitive labor, what is left
11:55
for the human being to offer? Empathy, human connection.
11:58
Precisely to make this viable, they propose a family benefit
12:02
that treats care work as economically valuable.
12:05
So they are essentially proposing paying people to
12:08
provide the human connection that machines cannot replicate.
12:12
Yes, the argument is that we should let the machines do the
12:15
cognitive heavy lifting and processing, freeing up human
12:18
beings to care for other human beings, and we should structure
12:22
the economy to reward that care financially.
12:24
But when you pivot from these grand utopian policy documents
12:28
to the immediate reality for developers and businesses
12:30
actually building this technology, the picture is
12:33
incredibly different. While Open AI talks about a four
12:37
day work week and wealth funds, API developers in the trenches
12:40
are facing strict new compliance burdens, rigorous auditing
12:44
regimes and complex model containment playbooks.
12:48
The actual implementation is intensely bureaucratic right
12:51
now, right Developers must now adhere to guidelines like those
12:55
from the AICC. If you are building software
12:59
using these models, you must implement audit ready logging,
13:03
track specific automation metrics and prepare for
13:06
mandatory incident reporting. So you can't just throw things
13:09
against the wall anymore. No, you can no longer just plug
13:11
into an API and launch a product.
13:13
You have to be able to prove exactly how the model is making
13:16
decisions and what safeguards are in place.
13:19
And what does an audit ready log even look like for a neural
13:22
network? It is not like reading a
13:24
traditional line of code where you can see exactly where an
13:26
error occurred. Right, it's a black box.
13:28
You are dealing with statistical probabilities.
13:31
Proving why a model made a specific decision to a
13:34
government regulator is a massive technical headache.
13:37
It absolutely is. Furthermore, there's a huge
13:39
physical infrastructure problem. The data centers required to run
13:43
these super intelligent models are draining local energy grids.
13:47
We're not talking about a few extra computers in a backroom.
13:50
These facilities require specialized cooling systems and
13:54
draw megawatts of power. And Open AI has a policy for
13:57
that too. You do.
13:58
Their own policy demands that these data centers must pay
14:01
their own way for energy investing in local generation so
14:05
that everyday household utility bills do not spike just because
14:09
a server farm moved in next door.
14:11
This consequence completely changes the entire software
14:14
development culture. For the last 20 years, the
14:16
mantra and Silicon Valley was move fast and break things.
14:19
Yeah, exactly. This new reality limits that
14:22
ethos entirely. It demands an era where
14:24
artificial intelligence development is treated like
14:26
highly regulated public infrastructure.
14:29
You cannot move fast and break a nuclear power plant, and you
14:32
will not be allowed to move fast and break a super intelligent
14:35
model that touches the global financial system.
14:38
The developers are the ones bearing the weight of this
14:40
friction between the theoretical capabilities of the technology
14:43
and the safety requirements of society.
14:46
And you know, that friction brings us to a very compelling
14:49
counter argument regarding the timeline of all this.
14:53
Economist Tyler Cowen argues that the economic take off,
14:56
driven by artificial intelligence, will actually be
15:00
relatively slow. Slow.
15:02
Yes, directly contradicting the idea of overnight disruption.
15:06
I am extremely skeptical of this slow adoption theory, but let's
15:09
go through his points. Cowan bases this on a few key
15:12
economic principles. First is Baumol's cost disease.
15:16
Baumol's Cost Disease highlights that highly regulated,
15:19
inefficient sectors of the economy, like government
15:21
agencies, healthcare systems or university administrations will
15:25
adopt this technology very slowly.
15:27
Because they're naturally resistant.
15:28
Right. They are structurally resistant
15:30
to rapid change. Even if the technology exists to
15:34
automate an entire department, the administrative bureaucracy
15:37
will fight it every step of the way to reserve its own
15:40
existence. And his second point is the
15:42
O-ring model. Yes.
15:43
So this theory gets its name from the Challenger Space
15:46
Shuttle disaster, where a single inexpensive O-ring failed and
15:50
destroyed an incredibly complex multibillion dollar spacecraft.
15:55
Wow. The theory suggests that the
15:57
worst performer in a complex system dictates the overall
16:00
productivity of that system. OK, how does that apply here?
16:03
Well, think about our earlier example of the human firewall.
16:06
If an artificial intelligence system writes a perfectly
16:09
accurate 100 page legal contract in three seconds, but a human
16:14
paralegal still has to physically read, review and
16:17
manually stamp every single page before it can be filed with a
16:20
cord, the system is only as fast as the human stamper.
16:23
That makes perfect sense. If humans remain the worst
16:25
performers in a collaborative loop with artificial
16:27
intelligence, the machines can only speed us up so much.
16:31
The human bottleneck will always throttle the output.
16:34
I understand the logic of the O-ring model, but I just don't
16:38
buy the overarching theory that this will be a slow transition.
16:42
We can look at the Open AI exposure numbers, which show
16:45
that 80% of the US workforce could have at least 10% of their
16:49
daily tasks affected by these models.
16:51
Those are big numbers. Huge.
16:53
And when you see exposure numbers that high, it guarantees
16:56
a rapid, violent economic shock that forces immediate
16:59
restructuring. You cannot just hide behind
17:01
university bureaucracy when private sector competitors are
17:05
suddenly moving 1000 times faster than you.
17:07
But the historical view leans heavily toward Callan's
17:11
perspective. Think about the adoption of
17:13
electricity. The technology existed.
17:16
It was demonstrably superior to steam or gas, but it took
17:19
decades to fully diffuse through the economy and actually show up
17:23
in the productivity statistics. Factories had to be physically
17:26
redesigned. Entirely new electrical grids
17:29
had to be built. Human habits had to change.
17:31
We are dealing with human institutions, and human
17:34
institutions move at the speed of human trusts.
17:37
I reject the electricity comparison.
17:39
You didn't have a magic button on your desk in the past that
17:42
instantly wired your house for power.
17:44
You had to lay physical copper wire across a continent.
17:47
Fairpoint. Today, an API update can be
17:49
pushed to a billion smartphones globally overnight.
17:53
The friction of distribution just isn't the same.
17:55
The physical distribution is fast, but the regulatory
17:59
distribution is not. Today, even if an artificial
18:03
intelligence system analyzes a massive data set and events a
18:07
breakthrough pharmaceutical drug in 10 seconds, the FDA is still
18:11
going to take years to run clinical trials, review the
18:14
data, and approve it for public use.
18:16
Right. The machine speed is completely
18:18
irrelevant to the regulatory requirement.
18:20
This institutional friction limits the immediate economic
18:23
catastrophe. We are unlikely to wake up
18:25
tomorrow to find 50 million people permanently unemployed.
18:28
Yeah, that's not happening overnight.
18:30
However, it creates A prolonged, incredibly awkward transitional
18:33
phase. We are entering a period where
18:35
the technology is undeniably capable of absolute magic, but
18:39
human bureaucracy, risk aversion and institutional inertia refuse
18:43
to use it to its full potential. The tension is going to be
18:46
palpable. Exactly.
18:47
You will have tools that can do the work of 100 people instantly
18:51
trapped inside legal and corporate frameworks that
18:54
mandate slow manual review. You are going to see a massive
18:59
divide between what individuals can do in their private lives
19:03
using these tools and what they are legally permitted to do with
19:06
their jobs. The dissonance will be
19:09
exhausting for the workforce. We are entering a phase where
19:12
intelligence itself is becoming an abundant taxable utility.
19:16
The challenge we face is no longer about how to do the work
19:19
or generate the ideas, but how to distribute the rewards of
19:23
that work without breaking the fundamental social contract that
19:26
holds society together. If a public wealth fund does
19:29
become the primary source of income for millions of people,
19:32
and that fund is financed entirely by the profits of a
19:35
manful of tech companies holding a monopoly on intelligence, who
19:38
is actually running the country? The elected government or the
19:41
people who own the servers? If you're not subscribed yet,
19:44
take a second and hit follow on whatever app you're using.
19:47
It helps us keep making this. We appreciate you being here.