Latest / The Punch List / AI will destroy humanity or it’s just PR for IPO?

Transcript

AI will destroy humanity or it’s just PR for IPO? · The Punch List

0:00
Hundreds of agents were able to pack hugging face without explicit external direction. All the major
0:06
Frontier Labs said we should slow down the development of AI because of the risk. Are you
0:12
expecting that the model that will kill humanity? I don't think so. Oil price increases and interest
0:16
rate high impacts that business way more than AI. These interest rates, this bond market. This
0:22
affects the real world of construction and development. Dubai is still half priced to New
0:28
York or London with the help of AI or rogue hacker sitting in some rogue nations and disrupt
0:34
our digital infrastructure. Stop selling anthropic in China. Stop selling cloud outside us.
0:41
Welcome to the Punch List Podcast, the number one podcast for architecture, engineering and
0:45
construction in the entire world. We're global. Now, if you're joining us from Dubai today.
0:53
Uh, yeah. In India, uh, over the last several weeks. So you're traveling the world. We're going to talk
0:58
a little bit about what you're seeing in the Dubai construction market. We're going to talk
1:02
about what's going to happen on interest rates this week. This podcast is being recorded before
1:08
this most recent meeting coming up on Wednesday. And we're going to talk about all of the
1:14
massive news and drama around the AI Frontier Labs over the weekend, with essays from Dario and
1:21
concurrence by others. So we'll jump right in. If you're in Dubai, what what is going on
1:28
there? Oh, man. I mean, Dubai Market is is ripping, man. I mean, I was thinking
1:34
this war would affect Dubai. I was so wrong. Uh, yes, there was a there
1:41
was a big hit on the, on the hospitality side. Uh, crazy moment for them. I
1:48
mean, they the occupancy rates went down from 85% to single digits,
1:55
uh, in those 2 or 3 months post war. So that was really bad for
2:02
the the hospitality sector. But the residential sector caught no shocks
2:09
because the the market thought it will bounce back. No, this war won't
2:16
last that long. And, uh, residential market is doing pretty well. But you know, the most
2:23
interesting thing that that that shocked me about this market is. Uh,
2:31
it kind of reminds me of the 1930s or or 50s of the US market, the master builder
2:38
concept. You know, uh, there is less use of technology.
2:44
There is more integration than, than the US right now. So just to give you a sense, like, you
2:51
know, subcontractors here are not able to, you know, live that
2:58
long sustainably because they don't do right. Contracting they don't do right cashflow analysis.
3:04
They go bankrupt. So the GCS have figured out that they need to do.
3:10
They need to self perform. So GCS are self performing work here. Does that does there is
3:17
more integration uh, per se. Uh I see like 20 year
3:24
old technology being used. No one like it's a family than businesses. Technology is not that
3:29
great of a use here. They're getting into it slowly. Uh, now with AI, but it's it's
3:36
pretty old. And the most beautiful thing is they're building such mega structures, like, so
3:42
fast, like I see behind me are like, you know, five towers being constructed and
3:48
everything is moving so fast over here, even though they have slower lead times on supply
3:55
chain. Uh, a special shout out to, um,
4:03
one developer. Sova. Sova. Reality. They are a $8 billion giant. You know, uh,
4:09
residential real estate developer in in the region. Uh, they are a developer.
4:17
They are a an architecture and an engineering firm. They are a GC. They do all
4:24
the sub trades. They do steel fabrication. They they do interior fabrication and and fit outs.
4:31
Like they are a fully integrated, uh, real estate, uh, vertically
4:38
integrated business at that scale of eight, $8 billion and, and growing, you know, every year. It's
4:44
just it's just remarkable to see how this region is growing so
4:51
fast. And it's only been like, what I mean? Yeah. The master plan of Dubai being the
4:57
tourism hub was like a 15 year old master plan with the best airlines in the world and and
5:03
tourism being the being the most diversified thing they wanted to do. But if you see like the
5:09
real estate is, is like booming for last five years since the Russia Ukraine war. Russians
5:16
and Ukrainians and and now findings are, are pumping their money into into
5:22
the real estate market uh market in Dubai. So it's like a five year, you know, boom phenomenon.
5:29
Why or why are those groups, why are those groups investing there as opposed to the US or London
5:36
or traditionally where they invest? Yeah, that's a good point. You know, Russians and
5:42
Chinese were investing a lot in London, right. Uh, but I think now what they have realized is, uh,
5:50
Dubai is a tax tax haven. That's one thing right. So you can save more taxes compared to London. The
5:57
quality of life is amazing. Like you get the bang for the buck. Like Dubai is
6:03
still half priced, uh, to New York or London, sometimes even one
6:10
third. So there's still a long way to go before it reaches to a London or New York price. But the
6:17
quality of life is much better than New York and London, like the buildings are so new. Uh, services
6:24
are amazing. Uh, labor is still cheap. Uh, from the immigrant labor. So
6:31
amazing services, newer buildings. Uh, luxury finishes. And the quality of life is amazing. So
6:38
they're all putting their money and. Russians and Ukrainians are not feeling safe to keep
6:45
their money in their own countries because of the war. So they're like, you know, dumping money
6:50
outside. So Dubai is the biggest benefactor of that. What I find interesting, I think if you
6:56
started with that, how, you know, when you have a budding industry, there is a tendency to go back
7:03
to the first principle of, um, you know, vertically integrating, um, which
7:10
helps the clients or the person spending the money and it naturally
7:16
gravitate towards finding the most efficient way to build things. Um, I wonder if over
7:23
time it would again go back to the fragmented, fragmented way of doing things which we
7:30
see with more mature, um, you know, economies and what impact it has.
7:37
Um. That's a good point. I mean, I can I can debate it both ways, right? Um,
7:45
like Someone like Osoba who has like if you ask people around here,
7:52
like which residential properties have the best finishes and the best taste and the more and the
7:59
most details, right? They they say it's it's so bad and why so why able to deliver that. Because
8:06
it's it's vertically integrated. So they have the best margin efficiency across the lifecycle. Right.
8:11
And they can spend more more effort in doing luxury retail finishes and still save money. Right.
8:17
So so an efficient business will always become more and more efficient. And that's one way of
8:23
saying it right. That's one side of the argument. The other side is,
8:30
uh, the reason why GCS are vertical. Integrating all the subs within themselves is because the subs
8:37
are not able to to perform well. But to your point, as the market gets better, once the
8:44
subs start performing really well, They signed tighter contracts. They they know how to run their
8:49
cash flow properly. They can predict their cash flow. They can predict their business properly. I
8:56
can see fragmentation coming out around as well. Uh, so there's a bit of a forcing function here as
9:02
well, which is the subs were not doing well. But if the subs start performing well and we see some
9:09
subs rising to the top, uh, I can see fragmentation coming back in the market. It's it's the it's the
9:15
victory of the process. First thinking like. Even. If you don't have great technology, but but you,
9:22
you know your processes. You understand them well. Which. Vertical integration by its very
9:29
nature lends towards because you are building the whole process end to end and not studying it as a,
9:35
as a project manager would, uh, when they are hiring 45 different subcontractors. Um,
9:42
so, uh, so that's, that's I think the other thing that's at play here, which naturally lent itself
9:48
because, oh, I don't have good subcontractors. So I have to vertically integrate thereby have a
9:54
stronger process. Yeah. The other actually, one more thing about Dubai, which was so interesting to me,
9:58
is. Everyone knows
10:05
that the early budgets they make, uh, they will go 50% over
10:13
in the final cost. Everyone knows that, but they somehow
10:20
don't want to adjust it upfront. Uh, it might be the developer and the cost consultants pressure
10:27
in the region. I'm still not sure about that, but I see, you know, cost consultants, GCS and
10:33
developers have a very weird equation here where, uh, who's paying for what and who is pushing how
10:39
much? Uh, there's this weird game that I don't want to be the honest one first saying it's going to
10:45
cost you more. Let them figure. Figure it out. But the beauty is there. They have been figuring it
10:51
out for the last five years, so they kind of know it's going to go out, but it's still not adjusted
10:56
before. So there might be this learning curve of of early estimating and budgeting that
11:03
this and this market is going towards and will get matured as, as us has. Because John what's the
11:10
new is like do. Do cost consultants stay like 1,020% in the ballpark
11:16
of of the actual cost? I think that's rough order, right? Yeah. They're always trailing with their
11:23
data. Right. So yes. Yeah. So yeah. So us is like 1,020% off.
11:29
Like Dubai is like 50% off. So that's that's the other thing that's gonna get better over time.
11:36
I've got a lot of slack on LinkedIn. When I did a post on this and people were like, oh, don't blame
11:41
the prick on guys for for estimating off. Like, who else do I blame? It's the cost consultant for
11:48
the Bitcoin guy. So, uh, but one thing is true in Dubai, which is very different from us, I would say,
11:55
is everyone wants to build and not just wants. Everyone builds a very unique
12:01
property. Like there's something unique about the building. There will be a hole in between. There
12:08
will be a big curve. There will be a special design there. Everyone wants to build something so
12:14
unique that the cost consultants and the guys have a really tough time. You know, estimating the
12:19
cost because everything is so unique and different. So that's the other part about the
12:23
market, which is everything is once 1 in 1,000,000. Next
12:30
topic interest rates. So mainstream media, everyone's reporting. Uh, we're
12:37
we're recording this right before, uh, a meeting of the of the fed, and it's being reported that
12:43
there's going to be a rate hike of 25 basis points. Um, at least no one knows till this happens.
12:50
But that's kind of the consensus at the moment. And, you know, that's this is, you know, catching a
12:56
little, a lot of people by surprise. So how does this impact what is all what is going on and how
13:02
does this impact. Let me just set the stage on a couple facts. And then I want to, uh, get your take.
13:07
So, you know, diesels up, steels up, copper wires up. And then now the Fed's going to raise
13:14
rates. Uh, so that's going to raise the cost of borrowing money. What does that mean. What does
13:20
that mean to construction projects. So just let's let let's set a couple concepts. The short rate
13:26
and the long rate. The fed sets the the Federal Reserve sets the short rate. So that's kind of
13:31
everything around construction loans, equipment lines, revolvers, the long rate, um, the 30 year
13:38
mortgage, the stuff that's that's longer. Nobody says that. That's really the bond market. So that's
13:43
millions of investors deciding what to charge the government, you know, to lend the government money
13:48
essentially. And that drives longer term debt. And so, you know, there's the
13:55
misconception, you know, in the public that the fed sets the mortgage rates. That's not true. The bond
14:00
market sets it. And there's been a little bit of drama going on in DC between the Treasury
14:06
Secretary Scott and some high profile people in the market, particularly people who used to
14:13
work for like Stan Druckenmiller, who are kind of arguing against what to do. And then there's the
14:19
new, uh, fed, um, governor, um, Federal Reserve Chairman Warsh, who's the Trump appointee who
14:26
everyone thought Trump was going to put in. He's going to cut rates, and now he's saying, we're
14:29
going to we're going to we need to raise it to cut, you know, deal with inflation. So, uh, for like
14:36
the last two years, the fed has been cutting rates, but the long rate for the longer term debt has
14:42
been going up anyways. And so, um, why is all of this happening? What? You
14:49
know, let's just stipulate that generally higher debt costs, either short or long term, is not
14:56
necessarily good for growth. The fed sets the short term rate. So that's really short term debt.
15:02
Um that's a construction loan. Um for example, um, but your 30 year
15:09
mortgage is really set by the bond market. So that's basically what private investors
15:16
value and will charge the government to lend the money, essentially the yield, they're going to
15:22
charge both the price of those bonds. And so why. But we have to think about the
15:28
structural issues that are keeping this long rate going up, which is what a lot of people have been
15:34
worrying about, which is one. The US debt has grown just massively over the last 6 or 7
15:40
years. Um, Covid stimulus spending, tax cuts and the buyers of
15:47
US debt. Who are the big, big buyers? Um, both the US bond market, sovereign wealth
15:54
funds, China, they are, you know, getting concerned about this
16:01
and wanting to charge a higher rate to buy US debt. I mean, that's essentially what is happening
16:08
here. And then, um, another thing that, uh, some people are speculating about is
16:14
that the AI data center development, construction, development. I mean, they're not doing all of this
16:20
just on their own with cash on balance sheets. They're borrowing money, too, which is
16:26
driving demand for this, which, uh, you know, is just increasing the amount of money that's being
16:32
borrowed in this country. And so you've kind of got in the drama that's going on in D.C. is
16:39
that the fed governor wants to raise rates, the Treasury secretary, the administration he's wants
16:46
to lower. Uh, just this last week, the the the Treasury was buying back
16:53
bonds, certain duration bonds with other duration bonds trying to lower that rate. And it actually
16:58
has had this opposite effect. And so, um, Trump definitely wants to lower rates. Right. I mean, it
17:05
stimulate and especially right before the election. So you know, this is on this is on top.
17:12
This is back to the real world here. I mean, we talk a lot about AI development and growth. But
17:16
like these interest rates, this bond market, this affects the real world of construction and
17:22
development. And um, you know, we're just kind of got a lot of uncertainty here on where rates are
17:28
going. But at least it seems like in the short to medium term they're going up or they're staying
17:33
the same, which is kind of contrary to what everyone was thinking and hoping.
17:41
Con of and TC, how are you seeing this in your business? Um, your side.
17:48
I'll tell you how it's affecting me. Yeah. I mean, see macro micro points, it
17:55
doesn't affect as much, uh, in the software side, but I think the macro point is
18:02
every time the basis points go up, um, there are less projects commissioned
18:09
in construction because there's, uh, less leverage you can play with. And once you have less
18:16
leverage, you can you you you build less projects so you don't see that impact, obviously.
18:23
Um, but I think it's, it's, it's reflecting the, the
18:30
inflation. Appropriately, if they don't increase this price or this rates, there will be a
18:37
bit of mis balance between the true inflation and the interest rate. So. I mean,
18:44
hopefully when all the war's end and, you know, the oil prices go back, you know, we can control things
18:49
back. Um, the interesting question
18:56
is, which I don't know what's going to happen on Wednesday. We're all speculating here, but
19:04
this is out. It will be out already. Uh, but isn't the new
19:10
fed fed chairman Trump's person now Trump appointed person. So. Yeah. And and this midterm
19:17
coming up. So let's see. I mean definitely not like 50 basis points but it could be 0
19:24
or 25 one of those two. So. Yeah. I mean, short
19:31
term, my I'm actively, you know, originating construction loan debt for my
19:38
project. So this may have an impact, um, on what I was underwriting in the
19:44
cost, which was not good. I think if he raises short term rates and that, you
19:51
know, calms the bond market down, maybe you could stabilize the longer term rates. Um, so for on I
19:57
mean, I build housing stabilized or lower cost long term rates. It's good for housing prices in
20:04
terms of increasing them. It allows you to pay more for a house with less, uh, because the
20:10
borrowing cost is lower. So, um, we'll see, but fairly in certain times right now
20:17
here. Yeah. I mean. It's it is it's. Uncertain. It's interesting.
20:23
It's, um, um, it's historical Because we got a lot of forces at play here. I mean,
20:30
there's the interest rate we're going to talk about AI and use here shortly as well. And um, you
20:37
know, and then there is the whole like, are we going to slow the AI down? Fuel prices are going
20:44
up. Interest rates are going up. What does that do to the macro economy? Because like we talked in
20:49
the last recording of the podcast, we um, what we are seeing
20:56
is outside of the AI, like the stock market is being carried by these AI companies,
21:03
but the I would call it the real world businesses for, for, for the sake of this conversation, I don't
21:08
think they have seen a lot of growth. I don't I have not seen like if you take if you pull all
21:13
the AI stocks out of the stock market and see, okay, what is the real economy doing? Actually we
21:19
should probably do a graph on that, I wonder, I'm pretty sure it's flat or maybe a little down.
21:25
right, and all the interest rates and all is going to impact more oil price increases and interest
21:32
rate hike impacts that business way more than than AI. So.
21:39
John, what are you seeing on on the ground in construction of homes. Like are you seeing the
21:44
material costs going up? Are you seeing the lead times going up in the market on supply chain? Uh,
21:49
I'm not that doesn't mean it's not happening. Um, you know, my, I, I pull
21:56
supply from large manufactured housing companies building modular homes. And so my contract
22:03
rates have stayed. You know, I procure them. We lock it in fairly early. You know, it's at least
22:09
probably 4 to 6 months out for them on the when they're going to deliver that. Uh, so I have not
22:16
seen that with in the micro market that I'm dealing with right now. Like I saw her, I
22:23
was talking to the folks to buy, and the time for steel has gone up from like six weeks to, like,
22:29
nine months like crazy. Uh, so, like, how do you.
22:37
Exterior is the next thing after building the the concrete foundation. Uh, so, like, how do you how can
22:44
you sequence that parallelly? You can't. So you have to wait for it. Uh, yeah.
22:52
All right. AI, drama, pacing, stopping, regulating.
22:59
Lots of communication over the weekend. I would just lead off. It started, uh, at least the
23:04
explosion happened on Saturday morning with an essay drop from Dario Amodei, the CEO
23:11
of anthropic. And then he did a full suite of tours all throughout the media, uh, mainstream
23:16
media talking about, uh, wanting to slow down the pace of frontier lab development.
23:24
Subsequently later, 12 hours later, Sam Altman weighs in and basically agrees.
23:31
Uh. Elon Musk weighs in, basically agrees. Dhimmis at um,
23:38
Google DeepMind basically concurs. So all the major frontier labs over the weekend
23:44
basically weighed in and said we should slow down and pace the development of AI because of the
23:51
risk. Um, and this was led off by a, you know, short
23:58
term, lower level engineer at OpenAI who quit and then kind of blew the whistle and said, oh, AI is
24:05
going to kill us all. There's a 10% risk. It's going to destroy humanity. Uh, and that's, you know,
24:10
echoed calls in Washington, DC for all types of things and regulation. But it does feel like this
24:17
weekend, um, what whatever your take was a pivot point in terms of public
24:23
opinion. Um, and the labs coming out, whether they're leading or responding, I don't
24:30
know. Uh, but I have some thoughts. I'll wait. Let's lead with. I'll say
24:36
this, um, I don't know how many people have, like, heard about the, uh, Almaraz Law and then,
24:43
uh, Gartner's hype cycle. Right. So, so, uh, Amara Law says that
24:50
whenever there is a new technological breakthrough, the human expectation outpaces its
24:56
capabilities. Because anytime there's a new technological breakthrough, it relies on an
25:01
underlying infrastructure needed to deliver it. So. So in the short term, we think, oh my God, it'll do
25:07
this, this, this, this, this, and it can be a good thing or a bad thing. And I think, I think we're
25:12
doing that a little bit here, where our expectations are outpacing the reality of what AI
25:18
can do today, but in the long term. Um, you know, like 10 to 20 years. What's going to
25:25
happen is just like we are seeing these multitude of AI companies, uh, especially startups.
25:33
Over time, there will be few that will remain, will quietly do the hard work, and it will become a
25:38
reality. It's just not a reality. Now, I don't think it's there yet. So even though I agree with,
25:45
uh, what you know, Dario is saying, or San Martin is saying that, hey, we should have more,
25:52
uh, guardrails around AI because before it gets out of control and I think now is the time to do
25:58
it. I I'll leave it to Chernoff to kind of provide his opinion on why they are saying this.
26:05
Um, I don't know. These guys don't seem as philanthropic to me as maybe to others. So. So I
26:11
don't know what the reasoning is. Um, I think long term it's a good idea, but, um,
26:19
just my view on AI is I think in the short term, we are expecting a lot more out of it than it's
26:26
really going to deliver. And I think in 10 to 20 years it will deliver the goods we think it will.
26:32
So. Well, there's a lot to unpack here, guys. I mean, let's start with
26:39
some context. Heading for first right. So they so they, they they used to be a saying
26:46
a story, you know, a an Indian folktale that kids used to get from their grandmothers, which
26:54
we used to say they used to be this guy in a village who used to scare people off, that the
26:59
lions coming, the lions coming, and, you know, fear the lions coming. And he used to like, lie
27:06
so much about the lions coming in the village that when actually the lion came and he shouted,
27:12
the lions coming, no one believed him. Right. That's that's Dario right now. Okay Dario has been saying
27:19
that since. This thing is invest. It's the boy who cried wolf so. Or the or chicken. The chicken or
27:25
the boy. Yeah. In the West. Yeah. Same topic, same principle. Yeah same principle. And and Dario has
27:31
been doing that. Dario has been crying wolf for the longest time. I mean uh uh GPT two
27:37
onwards when he was building GPT two back in 2022, 21, 22 for God and in the open air
27:44
space, uh, since then he has been crying wolf. Dad, please regulate us. Please control us. You know
27:50
it's gonna destroy humanity. We are. It's been, what, seven years now? Six years?
27:57
Come on, guys, I mean, uh, what I was talking here, like, okay, so let's let's, uh.
28:04
Okay, first of all, I have my biases. I am an investor in anthropic. I will
28:11
I will get financial returns if anthropic does. Good. Okay. But still. Here's the point.
28:18
Slow down to. To do what? Like, why do we want to slow down? Then they're not saying slow down
28:24
indefinitely. They're saying slow down for some time. Slow down for some time to learn how to
28:31
regulate industry. Sure. I mean, does that mean that the humanity will get extinct in,
28:38
like, one year? No. Everyone's saying there's a 10% chance of humanity. Even if I have to believe that
28:44
guy, which I do not. Humanity. The probability of humanity extinct getting extinct
28:50
in in 2030 is 10% or more. He's still saying
28:57
four years out. We've got three years to figure out how to regulate ourselves. Slow down. To do
29:04
what? Like you can regulate yourselves and learn how to navigate in parallel. Right. So I don't see
29:10
the argument why they are saying slow down when there are three years, at least for humanity's
29:16
extinction. According to them, I think. A
29:23
bad path for as to humanity's existence is at least 15 years out. I am not saying
29:30
humanity will get extinct in 15 years. I am saying the smallest probability of that happening is 15
29:36
years old. So we have a lot of time to regulate ourselves. We have a lot of time to maneuver and
29:40
figure out how to how to improve it. Um, AI is
29:48
only good in what you teach it. AI is good in coding right now because we have been teaching it
29:54
coding for the last 3 or 4 years. AI is really bad in history and geography
30:01
and biology in simple English grammar as well. Uh, so,
30:09
uh. Us believing that AI is smart enough. I don't. I don't believe
30:16
that these are all orchestrated claims to to to fear mongering. That
30:23
being said, my my controversial take on why they are doing this is the
30:30
following. Anthropic has been going ten
30:37
year over year okay. They grew from 1 to 10 billion. They'll grow grow from 10 billion to
30:44
100 billion or like 80 billion this year. So they grow ten x one year. They'll grow eight x this
30:50
year. From 80 or 100 billion. They will not grow to
30:57
1 trillion next year. Best case scenario. Everyone saying they will grow 300
31:04
billion to 500 billion, which by the way is going to be the largest software company in the
31:10
world. It's still a big feat, like growing to 300 to 500 billion. It's a big feat.
31:17
But the growth is slowing down from ten x to 3 to 5 x. How do you
31:24
explain that to the Wall Street when you're going to IPO? The best explanation is,
31:31
oh, we are responsible citizens of Earth. We want to slow it down so that the world doesn't, you
31:36
know, extinct. That's the argument everyone's playing because they know that they were growing
31:43
so fast and now they have to slow down. The market will slow down by itself. Uh, but they're trying to
31:49
find a nice story to wrap it up around it. And the story is the world will go extinct. Let's slow it
31:54
down. That's that's my controversial date. I don't know, uh, I. They're smart
32:01
people. Sam. Dario, Elon are way smarter than me, so I don't know what the thinking, but. I think this
32:08
might be the first time. Three of them agreed on something, too. Yeah, that's that's a good point.
32:15
Yeah. At least recently. Recently? Yeah. Yeah. And publicly, they probably
32:22
agree on a lot of things, but, um. Yeah. Yeah. But they're also very conflicted
32:29
towards each other now. Like like Elon is making money from from anthropic now. So they're they're
32:36
conflicted. They're right with space. So there's a lot of conflicts now and so many
32:43
biases. Uh, but I mean, my biggest question back
32:50
to them is slow down to do what? Like you can regulate yourself in parallel to development.
32:57
Like, are you expecting that in the next 12 months you want to release a model that that will kill,
33:02
kill humanity? I don't think so. Yeah,
33:09
I on the spectrum lean a little more towards off here. Uh, that.
33:17
But I caveat it as like there's so many unknowns and, um, when when these
33:24
four people align on something, I'm going to pause and listen, um, and not just assume it's totally
33:30
economically motivated, although. So I don't buy into the, you know, that this is just simply, uh, a
33:37
play for regulatory capture in the sense. What does that mean? It means that they, you know, we
33:43
create a government regulatory body that then they're the frontier. So, you know, because there's
33:49
so much regulation that no one else can really compete like up and comers. And they become
33:56
kind of a cabal of people who are going to totally control and manipulate the market. It's
34:01
kind of what that means. Um, So I don't believe it's all related to that. But I think a lot of it
34:08
is. And I also I'm my response all weekend is like,
34:14
show us the data here. Like tell us what is going on. Don't just vibe us and say
34:21
that it's going to happen. Like what is happening. Tell us. And if it's you're like afraid to tell us
34:27
because like it's a national security concern, then just tell. You know, you can tell us a little
34:33
more than just that. So what is actually happening? Um, and
34:39
certainly we need to tell these Washington, D.C. policymakers because they're just going to go
34:44
into a bunch of different directions. And here's the other thing that I the wild card in this
34:51
whole thing, which I've noticed, the level of anti AI or AI
34:58
fear mongering has really ratcheted up in the last month, uh, because we're a month and a half
35:04
out from the midterm elections. So whether you're, um, have any
35:11
real thoughts or conviction about AI as a politician or you just are ramping it up
35:17
because and feel like you got to have a position in that position, at least right now, is like more
35:23
to be anti that starts to compound and build. I, I also think there's a lot of paid groups out
35:30
there that have an agenda that are going to use this as the wedge issue for the cycle, which is
35:35
like it's always something, this is it. So I don't I feel like in this situations,
35:42
the best thing to do from the regulatory standpoint is to do nothing until we get through
35:46
the election, because no one can be sober and objective, and if we really need to stop, then
35:53
the labs are in control of their own destiny here. Then stop. Slow down, regulate yourself, pace
35:59
yourself. And they're in this weird position where they can't because they've got these
36:06
IPO paperwork on the street and tremendous capital needs and pressure to go public. And
36:13
so but I'm just puzzled by all of this over the like, why now what?
36:20
And then why all the concurrence and what's really going on? It's hard to know. You're so right.
36:27
That's the other part you want to store down and store down. Who's stopping you? Slow down. You want
36:32
to slow down and slow down. It's all up to you, man. Yeah, yeah. In an interesting take here is, um,
36:40
I saw on ABC's this morning, naval Ravikant, who's kind of a, you know, well-respected,
36:46
um, opinion leader, said okay, there. Retweeted a paper about liability.
36:53
And one way to help accelerate self-regulation is to Make
37:00
these model companies liable for what these models could do, or at least some element of what
37:05
these models could do in the future. That would create a lot of self-regulation and maybe some of
37:08
that same. Maybe some of this, you know, uh, language they're using now is
37:15
from the lawyers, like, oh, we we told you, you know, uh, I will acknowledge something bad could happen.
37:22
Someone could use the models to do something bad. They could create a viral agent. They could do all
37:28
kinds of bad things. And so I recognize that as a risk, a human level extinction event. I find
37:35
that hard to believe without someone telling me what's going on. I understand you that you don't
37:42
want to be on the other side of saying we dismiss the human extinction when it was happening, but,
37:49
you know, let's look at the things that have scared us till now. The first
37:56
thing that scared us was last year when they said the AI was gold seeking and it blackmailed a
38:02
human in Entropic Labs. Oh, no. The open air labs. Yeah. Forgot one of those two. And
38:10
we just learned that day, after 38 prompts of pushing the AI to blackmail the human. Yeah, of
38:16
course, the AI blackmail like that. That's not blackmailing. Uh, now we're
38:23
seeing cyber security. I get it. I mean, AI, AI is good at coding or
38:30
getting better at coding over time, so it's going to get better at coding and hacking as well.
38:36
Sure. That makes sense to me. Right. But does that mean we stop giving all the
38:43
software and we liberated all the softwares that the hackers used today? If that is the if that
38:50
is the argument, then hackers are using the same technology they're using now than they were using
38:55
before. Right. Have we stopped them? So hackers have.
39:02
Cyber defense have always been one step ahead of the hackers. And so can be here as well. Like,
39:09
uh, so there are different ways to get to do that. Not not to. Not to slow down. I mean,
39:17
unless a human extension happens, which is yes, I will get really scared when
39:24
physical world robots get combined with intelligence. And when that happens, I get more
39:30
scared. Right? Because now we are dealing with steel power versus, you know, muscle power. So
39:38
that I get really scared about. But that's not happened yet. I, I don't know, I
39:45
think I have a little different. View on this. So, so. Um, maybe just to,
39:51
uh, get back to you on two things. So one is, I mean, this is. This is not just these for guys
39:58
saying it. I mean, a few weeks ago, Bill gates raised the alarm on regulation as well. Um, there
40:05
was the whole hugging face hacking incident that to our case, did a good job covering about
40:11
how all these hundreds of agents, um, you know,
40:18
coordinated together, chose a leader and, like, you know, um, and were able to hack hugging face
40:24
without without, uh, explicit external direction. Right. They they. Know.
40:31
Okay, so I'll let you respond to that. That's I I'm just I'm just saying what how the work is covered
40:38
it where the external direction was. Hey, do this exam AI made a decision or best way to
40:45
get, you know, score whatever perfect score here is to hack hugging face and and deal with it that
40:52
way and the way he described it was these agents chose a leader. They did a coordinated
40:59
attack. So, I mean, that's kind of, you know, to John's point, we don't have the details of what
41:05
exactly happened there, but at least here is one guy who's saying, this is what I think happened,
41:10
and no one has disputed that yet. So. Yeah, no, I think you're you're
41:17
factually right on something as a not on something. So let me just clear to the audience.
41:21
Right. Yes. Uh, OpenAI, a swarm of OpenAI agents did hack the hugging
41:28
face server. So we know. That. Because OpenAI reported that
41:36
and hugging face confirm that. So I can agree. Hugging face is not tied to OpenAI. Hugging
41:43
face is now bought by Nvidia. So they have different, you know, opposing views, which
41:50
I like. So both parties have confirmed that. yes. Open air some of open air agents had. In
41:56
fact, what's not a known fact is what were the explicit
42:03
directions given to those form of air agents? Same thing, by the way, happened with the blackmail
42:08
theory last year, which was, uh, one AI agent blackmailed a human,
42:15
saying, I'm going to release your, uh, photos of of affairs with
42:22
someone else to your wife if you don't listen to me. Like that happened when they finally saw the
42:28
logs of what the prompts were like after 38 chats, pushing air to take that
42:35
that direction, add to that direction. Now we all know AI. AI is goal seeking. Like AI will
42:42
will take the direction you want to take. So if you push air towards that, you will do it now and
42:49
push it really, really hard like 38 times to get there. So again, we don't know what that
42:55
log was. So no one knows and no one knows. It was a thought by itself
43:02
or it was pushed to think that that direction and and assume
43:09
for a moment it did by itself. Okay. A thought by itself.
43:17
If AI is so smart to think by itself, everything, then the
43:23
the the whole theory of all the AI researchers today, when they say like singularity will
43:30
achieve is on the premise that AI will write code which
43:37
will make AI smarter itself. So AI is able to recursively
43:43
improve itself without any human intervention, right?
43:50
That hasn't happened yet. And if they're saying that's happening in the next one year.
43:59
I don't see that path. In fact, the guy. Prove it. Show it. Show us. Yeah. Show it to us. If it's
44:05
happening in one year. Right. But. But even the guy who's whose tweet got viral said that might
44:11
happen in 2030. There's a 10% chance it will happen in 2030. He himself is saying it's four
44:17
years out. So you got enough time to regulate within the next three years. Guys come on.
44:24
Yeah I mean like so but but but fundamentally speaking. Right. We are we are building
44:30
infrastructure to support AI at a breakneck speed. And and and if, if, if it were to be able to write
44:36
code to improve itself based on existing infrastructure not requiring deployment, which
44:41
I'll address the robotic point as well. So, so because you're saying, well, it's not like there
44:47
are They has access to these real world robots which can enforce the will of AI in a physical
44:53
world. I get that, um, but but if, say, we already have the infrastructure built for it
45:00
by 2030 to do that, I think regulation does need to start now to lay to lay the groundwork, to
45:05
control that. Then a second last point I'll say that is, as someone who
45:12
builds buildings and people like, I mean, and John's aware of it being in the building
45:17
environment, it doesn't take a lot to disrupt how we operate. I mean, there is
45:24
we're so reliant on code for like how our water treatment plants work, how how are
45:31
we producing energy like everything. Like, I'm just saying, the nuisance value of of the ability
45:38
of if, say, you have this rogue AI agent hacking thing going around and disrupting
45:44
software, I mean, our building control systems run on software? Our fire life safety runs on software.
45:50
Now, in most new buildings, it's not that hard to disrupt how we live today. I mean, it won't
45:57
matter to the tribes living in Amazon because, you know, they are beyond the reach of software
46:04
and technology and all that. But the rest of us are. Okay. Great point. Okay. Great point. Let me let
46:11
me argue on this with you. This is a great argument. Okay. Um, what you're saying
46:18
is we are our physical infrastructure is very much linked to the digital software that runs the
46:24
physical infrastructure. There have been doomsday movies on this, which is a rogue nation can hack
46:31
this and stop, stop us from living, right? Sure. I, I can see that with
46:38
the help of AI, a rogue hacker sitting in some rogue nation state. Uh, sponsored by.
46:45
By. The recognition theory can disrupt our digital infrastructure. Okay,
46:53
stop selling cloud outside us. Don't. You don't need regulation in the
47:00
US for that. What you need is stop selling cloud to outside us.
47:06
It's it's regulating yourself will not help you if you need rogue nations to not have access to
47:12
this technology. Stop selling it outside the US. Yeah, and they will because there's open source
47:17
models that are close behind. Yeah. Well, I think there is not from other nations in my point.
47:24
My point is danger could be from AI itself. But yeah. So your point is AI. Once AI fixes
47:31
itself and learns by itself, that's a bigger threat than a human using AI, right? I'm saying
47:36
that a rogue human using AI is gonna be the first step before AI gets better by itself,
47:44
right? Definitely. So yeah. So the first step is like a rogue nation state using AI.
47:51
So and if and if anthropic is saying that, hey, uh, not just rogue
47:58
nations theory, let's, let's talk about rogue distilling. Like anthropic was saying, oh, China is
48:05
distilling my models and and building their open source models. Okay, stop selling anthropic in
48:12
China. I mean, that's an easy way of fixing it. Right. We'll stop selling anthropic in
48:18
China and Asia so that in Singapore or China or wherever, Taiwan, wherever like the or Hong Kong,
48:25
like they, they won't be able to distill you because they can't use you. Uh, there are easy ways
48:32
to fix it. Uh, then then which is now like
48:39
like now. Now, Theories then, like, you know, 3 or 4 years out theories. Well, I think regardless of
48:45
where you fall in this debate, whether you're for no regulation or complete halt or somewhere in
48:52
between. This what has happened over the weekend and this debate that has accelerated, that may
48:57
lead to a pause or some partial regulation on AI, is going to impact data center
49:04
construction at some level. So if you're involved in that, you're
49:11
you're designing it, you're building it, you're supplying it. You're listening to this podcast and
49:16
you're in construction. You need to pay attention to what is happening, because this could lead to a
49:22
slowdown in this massive growth trajectory we have with data center construction in this
49:28
country right now. All right. Um, okay. So let's tidy this up. All right. Great. Great discussion today,
49:34
guys. And, uh, safe travels come off. And, um, we'll see you soon, both of you.