Latest / Elon Musk Podcast / High Level experts train their own AI replacements

Transcript

High Level experts train their own AI replacements · Elon Musk Podcast

0:00
Merker, a technology startup, just reached a $10 billion
0:04
valuation by paying highly educated professionals like
0:09
doctors, lawyers and investment bankers to teach artificial
0:12
intelligence models how to do their exact jobs.
0:15
Yeah, you are really watching the creation of an entirely new
0:18
labor market here. I mean, these high skilled
0:19
professionals are effectively functioning as the the training
0:24
wheels for the systems that are designed to replicate their
0:26
work. We are looking at a gig economy
0:29
that is completely predicated on extracting expert human
0:33
judgment, you know, and and feeding that judgment directly
0:36
into machine learning systems. So if you train the machine to
0:38
execute your specialized expertise, what exactly is left
0:42
for you to do once the machine learns it?
0:44
Well, to really understand how a company reaches that kind of
0:47
valuation, we have to we have to look at the marketplace it
0:50
created for human cognition. Because we are talking about 10s
0:53
of thousands of contractors. Here exactly 10s of thousands.
0:56
Merker has basically built this platform that matches domain
1:00
experts with major artificial intelligence labs, places like
1:03
Open AI and Anthropic specifically.
1:06
Right, the big ones. Yeah, and the entire purpose of
1:08
this matching process is to perform reinforcement learning
1:12
from human feedback and supervised fine tuning.
1:16
Which I know sounds incredibly technical, but the core
1:19
mechanism is essentially just professional greeting.
1:22
OK, professional greeting. Yeah, the platforms need these
1:25
professionals to look at the output generated by the
1:28
software, evaluate the logic, correct the reasoning, and
1:32
essentially show the program how a human expert thinks through a
1:37
complex problem. I think we need to explain
1:39
exactly what that looks like on the screen, right?
1:42
If you are a professional doing this gig work, what are you
1:45
actually clicking on? Right.
1:46
OK, so imagine you are sitting at your computer.
1:49
The interface presents you with a prompt.
1:52
Like a question. Yeah, maybe a really complex
1:53
legal question. And then it gives you 2
1:55
different answers that were generated by the software.
1:58
OK, Now your job is not just to say, you know, answer A is
2:01
better than answer B. Your job is to actually dissect
2:04
the logic. So you're getting into the
2:06
weeds. Exactly.
2:08
You highlight a specific paragraph, and you tell the
2:10
system you cited the correct precedent here, but you applied
2:15
the exception incorrectly. You are physically penalizing
2:18
the model for bad reasoning and rewarding it for correct logic.
2:22
And every time you submit that feedback, the underlying
2:25
mathematics of the software adjust to favor your thought
2:28
process. The economics of this are just
2:31
fascinating to me because the rates they pay, they range
2:36
wildly based on the level of expertise required for this
2:40
specific project. Oh, absolutely.
2:42
Like a poet teaching emotional expression and literary
2:45
structure might earn up to $150.00 an hour, Right.
2:48
A dermatologist who's developing clinical decision support tools
2:52
can earn up to $250 an hour. Yeah, and software engineers
2:55
doing complex refactoring, you know, finding deep structural
2:59
issues in code. They can command up to $500.00
3:02
an hour. You know, the poet example is
3:04
particularly interesting to me because you might assume the
3:06
software just needs math and science experts, But to
3:09
understand meter, tone, and the actual emotional resonance of
3:13
language, the models need direct feedback from someone who
3:17
understands poetry at a professional level.
3:19
Right, because the software does not actually feel emotion
3:22
exactly. It just, you know, predicts the
3:23
next logical word in a sequence based on statistical
3:27
probability. So if it right, this stands
3:29
about grief. It might use all the right
3:31
dictionary words, but the arrangement feels entirely
3:34
synthetic to a human reader. It feels.
3:35
Cold. Yeah, cold.
3:37
So the poet sits there, reads the output, and essentially
3:40
tells the software. This phrase is structurally
3:43
correct, but it lacks human warmth.
3:46
Try phrasing it this way instead to create actual resonance.
3:49
And, you know, the dermatologist is doing something very similar,
3:52
just in a completely different domain.
3:53
How so? Well, they are looking at how a
3:56
model assesses medical data. The software might look at an
3:59
image of a skin lesion and flag it as Melanoma.
4:02
The human dermatologist reviews that classification and checks
4:05
for end cases in the diagnostic logic.
4:09
Like a mistake the machine made. Right.
4:11
Maybe the software got confused by a shadow or like a hair
4:14
follicle in the image. The doctor corrects that false
4:17
positive, ensuring the machine's reasoning path matches
4:20
established medical protocols. I see this as just the next
4:24
logical phase of the labor market.
4:25
Really. You think so?
4:26
Yeah. Think about the Industrial
4:27
Revolution. You had highly skilled physical
4:31
artisans, people who spent their entire lives learning how to
4:35
weave cloth by hand. Eventually those artisans had to
4:38
teach the factory machines how to mimic their physical motions.
4:41
The machines learn the physical movements, and the artisans
4:44
transition to operating the machines.
4:47
To me, this feels like the exact same process just applied to
4:50
cognitive work instead of physical labor.
4:53
I see your point, but I have to disagree.
4:55
Really. Why?
4:56
I think this is fundamentally different because it targets the
4:59
cognitive foundations of the professional middle class.
5:02
OK, expand on that. The product being sold here
5:05
isn't physical labor or, you know, a repetitive physical
5:09
motion. The product is the methodology
5:12
of human reasoning itself. I see.
5:15
When the artisan taught the loom, the loom learned a
5:17
specific, isolated physical task.
5:20
But when a lawyer or a doctor trains the software, they are
5:24
selling the underlying logical structure of their profession.
5:27
That is a big difference, yeah. They are teaching the machine
5:29
how to weigh variables, how to apply judgement and ambiguous
5:33
situations, and how to synthesize conflicting
5:37
information. So because the machine is
5:39
learning the reasoning process, not just a physical task, the
5:43
outcome for the worker actually changes.
5:45
Exactly. This creates A phenomenon that
5:48
observers call mutually automated destruction.
5:51
Mutually automated destruction. Wow.
5:54
Because on one hand, these workers gain immediate, really
5:57
lucrative side income. We are currently in a cooling
6:01
job market and the ability to log on and make $250 an hour
6:05
applying your medical degree from your living room.
6:07
That's incredibly attractive. It's extremely attractive, but
6:10
on the other hand, they are spontaneously engineering the
6:13
obsolescence of their own career paths.
6:15
They are feeding the exact data required to make their daily
6:18
professional routines unnecessary.
6:21
Wait so they are literally digging their own professional
6:24
graves at $215.00 an hour? Pretty much, yeah.
6:27
That creates a really strange psychological environment for
6:31
the contractors. You know.
6:33
It is like paying someone an incredible premium to build the
6:37
exact robot that will eventually come to take their desk.
6:40
And that creates what they call an autonomy gap.
6:43
An autonomy gap? Yeah, the workers feel a
6:46
distinct lack of creative control over the ultimate
6:49
utility of their labor. If you think about traditional
6:52
work, you usually understand how your daily tasks contribute to a
6:57
final product or service. Right, you see the big picture.
7:00
Exactly. Here they're performing isolated
7:02
tasks evaluating specific outputs, but they do not control
7:06
how the aggregate data is used to build the final product.
7:09
So they're just a tiny cog. A tiny, very highly paid cog.
7:13
They are completely disconnected, the ultimate
7:15
outcome of their expertise. They monetize their knowledge
7:18
today, but they do it knowing they have no say in how that
7:22
knowledge will be deployed tomorrow.
7:24
Well, all of this expensive human feedback is actually
7:26
working, and we can see the direct results in recent
7:29
professional testing data. Oh, the rapid saturation of
7:32
industry benchmarks is just startling.
7:35
Yeah, we evaluate these models using standardized tests
7:38
designed for humans, right? The AI Index report shows models
7:42
mastering tests that were supposed to challenge them for
7:44
years, right? Evaluations that were designed
7:46
to remain difficult for the foreseeable future are being
7:49
conquered in a matter of months. A specialized model named
7:52
Describe LM achieved a 100% score on the Multi State Bar
7:56
exam. 100%. 100% To put that in perspective, the typical human
8:01
passing range for that exam is between 60 and 70%.
8:05
And you know, the multi State Bar exam is not just a trivia
8:08
test where you memorize definitions.
8:10
It requires applying complex legal rules to intricate
8:14
hypothetical fact patterns. The questions are specifically
8:17
designed with logical traps. Yes, they try to trick you.
8:21
They really do. They will present a scenario
8:24
where a rule clearly applies, but then hide A subtle detail
8:28
that triggers an obscure exception.
8:30
Right. To score 100% means the model is
8:34
successfully navigating those logical traps and nuanced
8:37
exceptions to legal doctrines perfectly every single time.
8:41
It's a wild. And the data points extend to
8:43
medicine as well. Grok scored 91.6% on the United
8:48
States Medical Licensing Examination step.
8:50
One, and that specific test involves a massive amount of
8:53
technical knowledge, but what is really interesting about the
8:56
Grok score is how it achieved it.
8:57
Right, the visual aspect. Yeah, that model specifically
9:00
outperformed others like DeepSeek because of its ability
9:03
to process visual media. Explain how that visual
9:06
processing works culinary to just, you know, reading.
9:09
OK, so a text based model reads a patient's symptoms described
9:13
in a paragraph, right? Yeah, and it matches those words
9:16
to a diagnosis, right? But real medicine involves
9:18
looking at X-rays, Mris, and pathology slides.
9:22
A multimodal architecture like what Grok utilizes translates
9:26
those pixels into mathematical representations that the system
9:29
can analyze alongside the text. This is looking at both at the
9:32
same time. Exactly.
9:34
It looks at the visual pattern of a lung scan, correlates it
9:37
with the written blood test results, and formulates A
9:39
diagnosis. Being able to fuse those two
9:42
different types of data, visual and textual, is just crucial for
9:46
medical diagnostics. And then on the software
9:49
engineering side, Claude Force on it reached 77.2% on SWE
9:53
bench. Right.
9:54
This is a benchmark that tasks the model with resolving real
9:57
world coding issues from actual GitHub repositories, not just,
10:01
you know, writing simple functions from scratch.
10:03
Yeah, it's real work. We were talking about
10:05
downloading an entire code base that a major company uses,
10:08
finding a deeply buried bug that spans across multiple different
10:11
files, and writing the exact patch needed to fix it without
10:14
breaking anything else. Let's pause for a second.
10:17
We are talking thinking about machines passing medical and bar
10:20
exams better than human experts. That is a completely different
10:23
reality than where we were just a little.
10:26
While ago, it really forces us to recalibrate our understanding
10:29
of professional competence. But we also have to acknowledge
10:34
the concept of jagged intelligence here.
10:36
Jagged intelligence, Yeah. Because while these systems are
10:39
incredibly capable in specific, highly structured domains, they
10:43
fail in ways that seem absurd to a human.
10:46
Oh, completely absurd. For example, a model like Gemini
10:49
Deep Think can win a gold medal at the International
10:52
Mathematical Olympiad navigating highly complex multi step proofs
10:56
perfectly right? Yet top models still fail to
10:59
read an analog clock correctly half the time.
11:02
And that paradox is the essence of jagged intelligence.
11:04
You have a system that possesses encyclopedic knowledge and vast
11:08
processing power, but it lacks the grounded physical common
11:12
sense that a human child develops naturally, right?
11:15
They can parse the nuances of a Supreme Court decision, but they
11:18
cannot consistently interpret the spatial relationship of
11:21
hands on a clock face. So why does it fail at something
11:24
so simple? Well, because it does not
11:26
actually have eyes or an understanding of physical space.
11:30
When you look at a clock, you intuitively understand the
11:32
geometry of a circle and the concept of time.
11:35
Moving forward. The model just sees a grid of
11:37
pixels. It tries to translate those
11:40
pixels into its text based understanding of the world.
11:43
But the cognitive architecture required to map A2 dimensional
11:46
image of intercepting lines into the abstract concept of 3:15 PM
11:52
is fundamentally different than reciting case law based on
11:55
statistical word probability. OK.
11:57
That makes sense. And this performance surge
12:00
across the board limits the dominance of any single company.
12:03
Absolutely. When you look at the testing
12:05
data, Anthropic X AI, Google and Open AI are now clustered
12:10
incredibly close together in performance ratings.
12:12
There is no single entity holding a vast, insurmountable
12:16
lead in raw capability. Right.
12:17
And because the top models are clustered so tightly, the
12:20
competitive pressure shifts entirely.
12:21
Where does it go? It moves away from achieving raw
12:24
capability since everyone is reaching similar high
12:27
watermarks, and toward cost reduction and domain specific
12:31
reliability. Oh I see, who can do it
12:34
cheapest. Exactly.
12:35
The company companies that win will be the ones that can
12:37
deliver that intelligence cheapest and with the fewest
12:40
unforced errors in specific professional applications.
12:44
It becomes a race to the bottom on pricing rather than a race to
12:47
the top on intelligence. And this high level of machine
12:50
performance directly alters the value of human skills on the
12:54
actual job market. Oh, we can see this clearly when
12:57
you evaluate productivity data. When artificial intelligence
13:00
assistance is introduced into a workplace, it raises the
13:03
productivity of lower skilled workers far more than higher
13:06
skilled workers, right? The MIT and Stanford findings.
13:09
Yeah, in customer support and writing intensive roles, workers
13:12
in the bottom tier of skill sought productivity gains of 35%
13:16
or more. Meanwhile, the workers at the
13:18
top, the highly experienced, highly skilled employees, saw
13:22
minimal or no significant gains. The tools did not make the best
13:26
workers significantly better. They made the weakest workers
13:30
operate at a much higher baseline.
13:33
And this is what economists refer to as compressing skill
13:36
gradients. The gap in output between a
13:39
mediocre analyst and a highly experienced 1 shrinks
13:42
significantly when both have access to the same automated
13:46
tools. Let's look at how that actually
13:48
works in practice. If you have a terrible writer,
13:51
they struggle with grammar, structure, and tone, The
13:55
software instantly fixes all of that for them, so their output
13:58
jumps massively. But if you have an incredible
14:01
writer, they already possess perfect grammar, structure, and
14:04
tone. The software cannot teach them
14:07
anything new about writing, so their productivity barely moves.
14:10
The floor gets raised, but the ceiling stays exactly where it
14:13
is. Wait, hold on.
14:14
Does that mean high skilled workers are perfectly safe and
14:16
low skilled workers are losing their jobs?
14:18
Not exactly. The real risk materializes as
14:20
entry level hiring freezes. Large professional services
14:24
firms and tech companies are reducing demand for junior
14:27
developers and early career analysts.
14:29
They are realizing they can maintain the required output
14:32
using automated tools operated by a smaller number of people.
14:36
So they just don't hire the juniors.
14:37
Exactly. They do not need an army of
14:40
junior staff to do the foundational research, write the
14:43
initial drafts, or complete the routine coding tasks anymore.
14:48
But if companies stop hiring entry level professionals
14:51
because the software can do the initial heavy lifting, that
14:55
changes the structural composition of the workforce
14:57
entirely. Yes, it really does.
14:59
It breaks the development pipeline for early career
15:02
workers. That is the long term structural
15:04
threat. Every senior partner at a law
15:06
firm, every senior software architect, every experienced
15:10
medical consultant started as a junior employee, making mistakes
15:14
and learning the fundamental. Right, doing the groundwork.
15:16
Yeah, the entire model of professional services is based
15:19
on an apprenticeship system. You do the grunt work for five
15:22
years, you learn how the industry operates and you
15:24
gradually move up to doing the complex strategy.
15:26
So if the entry level roles disappear, the mechanism for
15:30
training the next generation of senior experts disappears with
15:34
it. Precisely.
15:35
Where did the senior architects come from 10 years from now if
15:38
nobody was hired as a junior developer?
15:40
Today, demand shifts entirely toward workers who can actively
15:46
build and deploy these tools in their workflows.
15:49
The value is no longer in performing the routine cognitive
15:53
task, but in orchestrating the software that performs the task.
15:57
You have to be the person directing the system, not the
15:59
person competing against it. The entire pipeline feeding this
16:03
labor market shift. You know the company, the data,
16:06
the infrastructure. It is structurally vulnerable to
16:08
collapse. Yeah, we have to discuss the
16:10
massive cybersecurity breach that hit Mercker.
16:13
Yes, the breach A. Threat actor group called Team
16:16
PCP executed A cascading supply chain attack.
16:19
They did not attack Mercker directly.
16:21
Instead, they compromised an open source vulnerability
16:23
scanner called Trivi. Backup.
16:25
So a security scanner was the exact thing that let the hackers
16:28
in. Yes, a vulnerability scanner is
16:31
a tool developers use to automatically check their own
16:33
code for security flaws before they deploy it.
16:36
So it's supposed to protect them?
16:38
It is supposed to act as a digital guard dog, but Team PCP
16:42
found a way to exploit that exact scanner.
16:45
How does poisoning a scanner lead to a massive breach?
16:49
Well, because of how modern software dependencies work, when
16:52
you build an application today, you do not write every single
16:55
line of code from scratch. You pull in open source tools
16:59
and libraries to handle common functions.
17:00
OK, that saves time. Right, but if a hacker manages
17:04
to inject malicious code into an open source tool upstream, every
17:07
single company that downloads that tool to build their app
17:10
automatically downloads the poison.
17:11
Oh wow. Once Team PCP was inside Trivy,
17:14
they use that access to steal credentials and publish
17:17
malicious versions of a different, highly critical tool
17:21
called Leetel M. And Leetel M is crucial here
17:23
because it acts as a central proxy server holding API keys
17:28
for dozens of different artificial intelligence
17:30
providers. Right, explain what a proxy
17:32
server does in this context. Well, an API key is essentially
17:35
a digital password that allows an application to communicate
17:39
with a model like Open AI or Anthropic.
17:41
OK, if you are a company using 20 different models for 20
17:45
different tasks, managing all those passwords gets incredibly
17:48
complicated. A Roxy server like Light LM sits
17:52
in the middle. It holds all the master keys
17:55
exactly. Developers send their request to
17:57
late LM, and late LM routes the request to the correct model
18:01
using the stored keys. And because it routes
18:03
everything, it functions as a skeleton key.
18:05
Yes, once the hackers compromised Lite LM, they had
18:08
access to the internal systems of thousands of companies
18:11
relying on it to manage their connections.
18:14
And for Mercker specifically, the attackers stole roughly 4
18:17
terabytes of data. 4 terabytes. That's huge.
18:20
This was an enormous X filter. They took personal identifying
18:24
information for contractors, including things like Social
18:27
Security numbers. Ouch.
18:28
They took the video interviews that Mercker uses to vet these
18:31
highly skilled experts. And perhaps most critically for
18:34
the company, they took proprietary source code.
18:36
There's also the Delve Technologies angle to this
18:40
breach, which is just It's crazy.
18:42
Delve was the startup that provided the security compliance
18:45
certification for Elite LM. When you use software that
18:48
handles sensitive keys, you expect it to have passed
18:52
rigorous security audits. Naturally.
18:54
It turns out Delve was allegedly running what investigators
18:58
called fake compliance as a service.
19:00
Faking it How do you fake a security certificate for a multi
19:04
billion dollar infrastructure? Well, a real security audit
19:07
involves penetration testing, where ethical hackers actively
19:11
try to break into your system to find flaws.
19:14
Right, the stress test. Yeah, and it involves manual
19:16
code reviews by seasoned security engineers.
19:19
Delve Technologies was allegedly automating the generation of
19:22
compliance certificates without actually performing those
19:25
rigorous checks. We essentially built software
19:27
that just rubber stamped the security approval.
19:30
And when this was exposed, Dell Technologies was expelled from
19:33
the Y Combinator startup accelerator and their major
19:36
venture capital backers completely scrubbed any mentions
19:39
of their funding from their websites.
19:41
They just erased them. Yeah, Y Combinator is one of the
19:44
most prestigious startup incubators in the world.
19:48
Getting kicked out is a massive, incredibly rare black mark.
19:52
The compliance safety net guarding this entire ecosystem
19:56
was essentially an illusion. This opens up massive commercial
19:59
and legal liabilities. Meta indefinitely pause all
20:02
contracts with Merker while they investigated the breach, right?
20:06
That left all the contractors working on metaprojects
20:09
completely stranded, unable to log hours or get paid.
20:12
And on the legal side, 5 different lawsuits were quickly
20:15
filed by contractors seeking damages for practice violations.
20:19
They trusted the platform with their personal data, and that
20:22
trust was broken. And beyond the immediate privacy
20:24
concerns, the theft of state-of-the-art training
20:26
methodologies exposes proprietary secrets to rival
20:30
labs and foreign adversaries. Yeah, it's a huge deal.
20:32
The data stolen wasn't just user passwords, it was the specific,
20:35
highly guarded methods these companies use to teach their
20:38
models complex reasoning. It proves that the digital
20:41
infrastructure supporting these technologies is astonishingly
20:44
fragile. Even if the software supply
20:46
chain is secured, putting these tools in the high stakes
20:50
physical environments creates a whole new set of human
20:53
frictions. Healthcare is the ultimate test
20:55
case for this friction. Honestly, Yeah.
20:57
The Food and Drug Administration has already authorized over 1200
21:01
machine learning medical devices.
21:03
These cover everything from cardiology to retinal imaging.
21:07
The technology is approved and legally ready for use.
21:11
But integrating those approved tools into the daily routine of
21:14
a hospital creates clinical workflow chaos.
21:18
Oh, definitely. One of the primary integration
21:21
challenges is alert fatigue. Consider an algorithm designed
21:25
to detect critical findings on a head CT scan to expedite stroke
21:29
treatment. OK, the goal is to alert the
21:31
doctor immediately so they can intervene faster.
21:33
But if that algorithm throws high rates of false positives,
21:36
alerting the doctor constantly when there's no actual
21:38
emergency, the doctors become desensitized.
21:42
It is exactly like having a copilot who screams that you are
21:45
going to crash every single time you hit a slight bit of
21:48
turbulence. Right, you tune them out.
21:49
Eventually you are going to turn the microphone off entirely just
21:53
so you can focus on flying the plane and the day the engine
21:56
actually fails you won't hear them screaming because you
21:59
already muted them. That's a great analogy.
22:01
If a doctor learns that the artificial intelligence alert is
22:04
usually wrong, they will start ignoring the alerts entirely,
22:08
which defeats the purpose of having the system installed in
22:10
the first place. Then you run into the medical
22:13
liability dilemma. If an algorithm misses a
22:16
diagnosis or provides an incorrect treatment
22:19
recommendation and a patient is harmed, the courts have not
22:23
definitively resolved who bears responsibility.
22:26
Yeah. How does a malpractice suit even
22:28
work when the doctor was just following the recommendation of
22:30
an algorithm? That is exactly the problem.
22:33
Is the hospital liable for purchasing and deploying the
22:36
software right? Is the developer liable for how
22:38
they train the algorithm? Or is the individual doctor
22:41
fully responsible for deciding to follow the machines
22:44
recommendation instead of relying solely on their own
22:47
judgement? It's messed.
22:49
It is the legal framework determining fault has simply not
22:52
caught up to the technological reality of the clinic.
22:55
This liability confusion limits how quickly hospitals will
22:58
deploy the technology. Without clear enterprise
23:01
liability frameworks, healthcare administrators are hesitant to
23:05
adopt systems that might expose them to massive, unprecedented
23:09
lawsuits. However, as the technology
23:11
improves, the standard of care will inevitably evolve to
23:14
incorporate these tools do. You think so?
23:16
Yeah, because in medical law you are judged against what a
23:18
reasonable, competent Dr. would do in the same situation.
23:22
Eventually, these algorithms will be definitively proven to
23:25
increase diagnostic accuracy and save lives.
23:28
When that happens, a completely new legal risk emerges.
23:31
Failing to use available automated support could soon
23:34
constitute medical negligence. Wow.
23:36
Yeah. If a piece of software can
23:38
consistently catch a microscopic tumor on a scan that a human
23:42
doctor typically misses with the naked eye, choosing not to use
23:45
that software could be seen as providing substandard care.
23:48
You could be sued for relying solely on your human brain.
23:51
So the professional workforce is actively building the tools that
23:55
compress their own skills, while the digital infrastructure
23:59
holding these tools together is remarkably fragile.
24:02
Exactly. If the professional of the
24:04
future is meant to be an architect of automated systems
24:07
rather than a daily practitioner, how do we ensure
24:10
society doesn't lose the foundational human judgement
24:13
required to manage those systems?
24:15
If you're not subscribed yet, take a second and hit follow on
24:18
whatever app you're using. It helps us keep making this.
24:20
We appreciate you being here.