All right, everybody. Welcome back.

Welcome back to the number one podcast

in the world. The Core Four, Fantastic

Four. The original Quartet is here. Did

we peak at like uh number two or number

three last week? Is that what happened?

>> I think it was number four in the world.

So, yeah, usually we're trending.

>> US trending.

>> US trend. We're usually number one

globally, but yeah, and sometimes number

four US.

>> I forgot my Starlink. So, let me

apologize to everybody. That was a

critical error when you're on the road

or on the water, but you know, it'll be

fine.

>> I had five huge leads come in this week.

>> The Glengar Glenn enterprise leads.

Enterprise sales is a bear because it's

super chunky, but the deals are

ginormous.

>> Sax, you got any advice for Chimat from

your uh enterprise sales days? You

closed some of those big seven figure

deals when you were doing Yammer?

>> No leverage. Don't put on leverage.

>> No leverage of the show today.

Leverage equals risk of ruin.

>> PSA, [laughter]

no leverage.

>> I have the situational awareness to not

lever up.

>> That's good. Yeah, you want that. That

situational awareness.

>> I mean, it's kind of out there. I mean,

if you name your fund situational

awareness, that's

>> Yeah. Come on the pot anytime, Lualt.

All right, everybody. We got to talk

about chip stocks crashing after an

alltime runup. And we had a major hedge

fund get margin called and some

incredible margin calls happening in

South Korea. Leopold Ashen Brener is a

25-year-old hedge fund manager. He left

OpenAI two years ago to start his own

fund and apparently according to

reports, this is breaking news on

Thursday when we tape, he got margin

called and had to sell his entire public

portfolio to cover massive losses caused

by his leverage. And who bought them?

none other than Citadel's Ken Griffin.

We don't know if it was Ken Griffin

himself, but Citadel bought it according

to the early reports. Leopold had insane

returns and he rode the wave of AI and

chips and Frontier Labs as recently as

this month. Uh, and he started the fund

with $225 million in 2024. He grew it

100x to 20 billion this year or so. Ran

it all the way up to 45 billion. Now

he's at 200x earlier this month by

trading on leverage. According to our

friends at CNBC at the end of the June,

he was reportedly up 4x450%

this year. Uh some have reported that

he's also selling his massive anthropic

stake to cover these losses, but the

Wall Street Journal is disputing it.

Again, uh he we're we're happy to have

him here on the program. How did this

all blow up? Well, NASDAQ's chip index,

this is called the Philadelphia

Semiconductor Index, is down over 20%

over the last month. That's bare market

territory. Obviously, definition of bare

market territory for those of you who

don't play in the markets is anything

over 20%. Uh the index included the top

30 US listed chips. That's people like

Nvidia, TSMC, AMD, Micron, you know, all

those big names. But the index bounced

back a bit today, up 7% when we're

taping. So, we may have found a bottom.

Unfortunately for Leopold, he had

already sold. Samsung and SKH, two South

Korean chip companies that are not

included in the NASDAQ index, also got

smashed, crushed, demolished. Samsung

down 38% over last month. SKH Highix

down 14% since going public 3 weeks ago.

The Cosby, that's South Korea's version

of the S&P 500, is down over 40% in the

last 40 days. Between last Friday and

Wednesday, leading chip companies shed

over a trillion dollars in market cap

combined. So to put this in context,

chip stocks had a legendary run the past

couple of years, but uh you know trading

on leverage, we'll talk about it. Very

dangerous. If there's a downturn, we'll

get into the South Korea wrinkle as

well. Even with this downturn, the

5-year results are still spectacular.

Jimoth Micron up 850% mostly in the last

year. Nvidia up 875% in the last 5 years

and Broadcom up 663%.

Let's discuss it.

>> If I was going to give you one piece of

advice when you're running risk is you

have to manage leverage incredibly

carefully because when it runs ahead of

you, the unwind is incredibly violent

and it's incredibly quick. That's the

biggest problem with with running either

massively levered long or massively

levered short. So I don't know to what

extent he was running lever but the

rumors are he was running like three and

a half turns which just to give you a

sense when you're running that much risk

a 3 and 4% move is amplified 12 and 13

but if you saw what's happened in the

last 3 days a 25% move is amplified 75%.

So

>> it has the risk to stop you out. And

what happens is when you get that

leverage, the banks are given the

authority to close you out. And when

they close you out, what they do is they

start calling around and unwind your

risk. And you don't have much of a

choice. It's sort of an automatic

one-way ratchet.

So if everything that has been reported

is accurate, he was running about three

and a half times levered. The market

moved against him. He lost a very large

percentage of his gains and then the

prime brokers started calling folks.

Citadel bought the whole book and now

the question is what is the AUM left and

what is the high water mark and can he

actually dig his way out? These things

are brutal. your thoughts, Saxs, uh

looking at this situation,

any lessons for you or I guess bigger

picture, this downdraft, is it because

of market conditions, you know,

inflation, the war, or people just ahead

of their skis when it comes to uh the

valuation of these companies and then he

just got caught in a downdraft. Yeah.

>> Well, I think that is the key question

here. Is this correction in the markets?

Is it driven by fundamentals or is it

driven by momentum? And my view is that

I think it's driven by momentum. Meaning

that over the past year, you've had this

roughly 10x runup in memory chip stocks

and you've seen this overall huge rise

in any stock that's related to the AI

boom. So, anything related to this AI

capex boom has been going up like crazy.

And I think it was inevitable that you'd

see a pullback. I think there was

something like a 10% pullback in the

NASDAQ

from the peak. But when you look at this

momentum trade, it was down like 30% or

40%. Right? Because the 10% was on the

whole market. So this sort of momentum

trade was the most exposed part of it.

And you look at what happened in South

Korea, you look at what happened with

Leopold's fund and obviously there was a

lot of leverage behind this momentum

trade. So when it corrects it's going to

be brutal. But I think that the question

again is does this reveal anything about

the fundamentals? And my sense is that

you're already seeing the rebound this

morning and what I mean by that when I

say fundamentals is is the capex that's

being invested in the AI boom is that

real or is it misguided? Right? Is it is

that a sound investment? Is that an

investment that the hyperscalers for

example should be making? Is that an

investment that's eventually going to

deliver ROI or is this some sort of

bubble? And my view is that it's real

that I think there will be a return on

all this capex. I don't try to predict

stocks or tell people when they should

be buyers, but you look at the

hyperscalers, they have invested pretty

much all of their free cash flow and

then some in this boom. You know, a lot

of people are trading those stocks down

because of that. My view is that

eventually there will be a return on

that investment. And this is sort of

temporary market volatility amplified by

leverage. And Chimath is right. You

know, I think it was Warren Buffett or

maybe Munger who said that leverage is

the only way that smart people go broke

because, you know, if you're not using

leverage,

your portfolio would just be down 30%

this month and then it would already be

up 7% today. So, you'd be rebounding.

So, you'd be down, okay, 20 something%

this month, but after having risen 10x

in the past year. But if you're

leveraged 3 or 4x, you're wiped out

>> and you get margin called. So look,

there's many examples of really smart

people getting hurt by by leverage.

Yeah. And that that's the lesson there.

Now, I think Leopold's a really

interesting figure in the whole AI

movement and I would say an interesting

thinker. I met him about a year year and

a half ago.

>> Did you invest in the fund? Did you give

your

>> No, I wasn't in I was prohibited from

investing in things like that.

>> You were in you were in DC at the time.

Yeah,

>> but I thought he was a really

interesting thinker and he wrote a blog

called situational awareness before he

created the hedge fund version of it.

And I thought what was really

interesting about it was just he laid

out the bullcase for the AI boom. And

just by the way, he's like very wired

into anthropic. I think his fiance is

Dario's chief of staff, something like

that. You could almost say that he's

like the hedge fund version of the

anthropic thesis. And what I thought was

interesting about his argument is he

talked about or uh orders of magnitude

uh which he called ooms um increases in

three key areas. So he said that if you

look at the raw compute the chips they

were getting better at a rate of roughly

3x per year which is roughly an order of

magnitude or 10x every two years. He

said if you look at the algorithmic

efficiency so you know techniques like

reinforcement learning things like that

the models were getting better at 3x

every year which is again order of

magnitude every two years and then he

also said that there were huge gains

from what he called unhobling which I

think now we would look at it things

like the harness and connectors you know

ways of using the model those were also

getting better the ways of integrating

the model's decision-m in practical ways

that the intelligence actually becomes

useful. And he said that that was also

similarly improving. And so, you know,

you project forward when you have 10x

orders of magnitude improvement in these

key underlying fundamentals, these key

drivers of the technology. And you can

see that well over a course of not just

two years, but over four years, you're

going to have 100x improvement. Over six

years, you're going to have a thousandx

improvement, right? because it's

>> and you very rarely see anything in the

world that grows at that velocity. We

we'd be hardressed here

>> with the exception of maybe bandwidth,

you know, going to fiber to the home or

something like what's an analogy where

where that's happened before in history.

>> Yeah. Virality. I mean I, you know, back

in the PayPal days with the PayPal

mafia, we would think in this way of

exponential increases because we would

see an exponential growth curve and so

we were able to project forward. So his

thinking ins always appealed to me

because I think most people just don't

think in exponentials or don't know how

to think in exponentials.

>> It's hard for humans to think in

exponentials, right? These when numbers

get big, it's like the difference

between a billion and a trillion is is a

is a lot. It's not a small amount.

>> Yeah. And you'd have to say, look, he

was stunningly successful for the first

couple years. Apparently, he started

with 200 million or so in his hedge fund

and he rolled that all the way up to 20

billion, I think. Now, the problem I

mean the reason why I think he got wiped

out or at least his public book did is

it's partly the leverage and then you

have the short-term volatility. So,

those two things don't go together.

Also, you know, when your fund grows

that much, you get a lot of hot money.

So, when you say, well, he he's up 10x

before the 30% correction. Well, the

question is who's up 10x? Obviously, the

investors who were there from the

beginning are up 10x or more, but

>> the latest people,

>> that's only 200 million, right? So if 10

billions come in in the last few months

because of the hot money dynamic where

everyone piles into the most successful

hedge funds, those guys are kind of

wiped out.

>> So let's talk about the psychology of

this uh Dave Freeberg. If somebody is so

brilliant that they can write this essay

and understand the market uh so

exquisitly and be such a great

communicator, how could they have such a

crazy blind spot when it comes to

putting on leverage at this scale? Do

you have any thoughts on that, Freeberg,

or have you seen it before? Is it just

the folly of

>> It's not a blind spot. It's a feature

that turns into a bug. We're all like

this. We all know people that have that

edge and can push it.

>> What do you think, Freeberg, on the

personality type? Or is this just

something most people do when they're on

the heater?

>> Conviction.

>> Where do you stand on it?

>> Ultra conviction. I think the

Warren Buffett

assessment of equity markets is uh in

the short term they're voting machines.

In the long term they're weighing

machines

and you could have the right long-term

view. I mean look at SPF. SPF would

pretty much would have been the greatest

investor of all time if he didn't get

liquidated. Same dynamic. I mean

obviously there was fraud in terms of

how he was allocating capital but his

actual portfolio over the long run was

absolutely correct. In the same way that

if you had bet on the internet and

stayed in that bet from 1995 through to

today and you bought a portfolio of

internet stocks, a bunch of them would

have fallen to the wayside. But those

that won,000x, 20,000x, 20,000x

and you do extraordinarily well. So he

could be right in his fundamental

assessment and analysis, but then in

markets over the short term, you have

bubbles and bubbles pop. And when

bubbles pop, if you have a leverage to

multiply your returns, you get wiped

out. That's effectively, you know,

what's going on here. And he may be

right. Right.

>> You know, the thing to really double

click on that I think will probably come

out in the next couple of days or weeks

is just how historic the South Korea

unwind was. 1.2 million leverage trading

accounts uh have been hit with margin

calls in South Korea. If you know about

the South Korean market, that data is

two weeks old. Jal, that number is much

bigger today. Yeah.

>> Yeah. But I mean, just in terms of

people discussing it in relation to him

getting caught in the downdraft,

>> if anything, he got caught in this

downdraft. Of those 1.2 million lever um

levered accounts, somewhere around

350,000 of them were fully liquidated

already. And so

>> again, two weeks old, that's so as of

today, the number is much bigger, right?

>> So it could be closer to a million

accounts fully liquidated today. If

that's the case, we're talking about

like some percentage of South Korean

population having their entire asset

base blown out.

>> Their entire

>> 3% of the population.

>> Well, that's going to that's going to

sting. And it is a very investment

forward culture. If you look at what

happened in crypto, the same thing

happened with NFTs and speculation

there. And they had banned crypto

because they knew the Korean culture has

this gamble in it and this obsession

with trading.

>> Can I can I just frame something up? So

if we take

>> the circumstance of there's a good

long-term bet in AI that can be made in

the markets but in the short term

there's an exuberance that arises. The

question is what's resetting that

exuberance? What's bringing us back down

to earth in the short term and I think

if you take a zoom out there's a bunch

of other statistics and other facts on

the ground that I think are big macro

drivers at the moment. If you take a

look at the 30-year Treasury yield we

just crossed 5.2% 2% for the first time

in 20 years.

>> So you could buy US treasuries that are

paying you 5.2% a year for 30 years,

which is on a pre-tax equivalent basis

probably 8 9% 9% from the US government

for 30 years. So Nick, if you zoom out,

you know, we have not seen this yield on

US treasuries since 2007 leading up to

the global financial crisis when they

cut rates and printed money. At the same

time, there was some probability that

the Fed Reserve was going to raise rates

this week. They didn't, and that

obviously would have tampered the

inflation risk ahead of us. There's

persistent inflation. Kevin Wars in his

comments said, "We still want to see

inflation get down to 2%." There isn't a

clear path to doing that. And then

there's these inflation drivers. The

biggest inflation driver at the moment

is government spending. $2 trillion

deficit, 7 trillion a year of spending

on five trillion a year of revenue. Both

Elizabeth Warren and Donald Trump agreed

on Twitter this week that they should

remove the debt ceiling, which means

that we could spend more and continue to

borrow more. Federal debt stands at 40

trillion today. Remember the debt

ceiling in July of 2025, the debt

ceiling was 36 trillion and we now want

to raise it above the 41.1 trillion debt

ceiling that we have. Elizabeth Warren

saying get rid of it. Just have no debt

ceiling.

>> So, so when you when you have no debt

ceiling and you have no breaks and you

spend and the government spending

becomes the core of the US economy

because that spending is not productive,

you end up seeing inflation. You're

pumping money into the system. So,

everyone's assets inflate and

fundamentally people are selling off

treasuries around the world because of

it. And now we're kind of looking at a

situation where there doesn't seem to be

an end in sight. There was a

rationalization of spending intent

coming into this administration. It's

proven to be nearly difficult, if not

impossible, to get Congress to go that

route. The Senate has banded together to

keep funds flowing to their states. So,

you cannot really radically change

spending at the federal level. So if

you're running a $2 trillion annual

deficit and your economic productivity

gain in the near term doesn't make up

for all the inflation you're realizing

because of that exuberant spending,

you're going to see Treasury spike

because people don't trust the

creditworthiness of the United States

over 30 years. And so a treasury spike,

I could now buy a US government bond

that pays me 10% pre-tax a year. Why the

heck would I pay 50 times earnings for a

semiconductor stock? So that creates the

incentive for markets to move against

these big AI conviction bets in the

short term and pop these bubbles. And I

think we're going to see more of this.

As we don't actually course correct the

Titanic going into the iceberg, the

United States fiscal and monetary

situation, we are going to end up seeing

more bubbles pop and more of these

assets um that we've kind of inflated,

if you will, to keep things going. Now

look, there may still be great

productivity gains from AI. This may end

up rationalizing over the long term, but

again, short-term markets, I'm better

off making 10% by owning federal

government bonds. Go to the beach. Yeah.

>> Then taking the risk and the volatility

on these things, paying 50, 100 times

and not knowing, you know, not knowing

when am I going to get the voting

machine to match up with the weighing

machine? What's my time horizon? And the

bigger the yield on treasuries, the

harder it is to make those sorts of

bets.

>> Obviously, President Trump has been

angling for a cut. And here's your poly

market. 53% chance of not a cut, not

standing still, but a rate hike in

September. So adding to all this, the

cost of capital is going up apparently.

>> And let's we not forget the Iran war,

which is creating persistent pressure on

energy prices. The longer the Iran war

goes on, the longer we're going to see

an increase in pricing for energy, oil,

and nat gas, and fertilizer. Those

trickle through the economy because it

inflates the cost of everything on the

energy side and food on the fertilizer

side. And that's really going to create

this pressure on the upside which means

you're going to have to raise rates to

account for that inflation at some

point.

>> And then consumers are going to see

three and four, you know, maybe more,

god forbid, five or six. And also

inflation will be persistent. Jeffrey

Berg, like it's going to be hard to

stop. The one thing that I think Kevin

Warch and and Scott Bessant are kind of

vulcan minds around the Stan Ducken

Miller you know gravity well if you will

on this is productivity gains can drive

us out of this this problem and

productivity gains can and should arise

from AI and that's really where a lot of

the value creation will come in the

economy over the next decade or two

which is why we're seeing this massive

upfront capex to power that and enable

that and that's great and there's good

policies in place but in the last couple

of weeks I would say the one risk to

that thesis is China Because China is

now demonstrating that they may deflate

the value of models by releasing open-

source AI models and that ultimately the

value may just sit with the compute

infrastructure and the comput layer and

the energy

>> application layer. Yeah,

>> perhaps the application layer, but

fundamentally this model energy being

shifted to China and deflated and

commoditized

puts a real wrinkle. If you had built a

30-year AI productivity model around how

it's going to drive the economy and

where the value is going to come from,

you would have had a significant number

of rows in value creation estimated in

the model layer and that would have been

a big part of the economic growth for

the United States over the next 30

years. And now if China says, you know

what, we're actually going to delete

that for you and all the value is going

to sit with energy, which is what we

have a lot of and the stuff that they

make, then we're going to end up

acrewing a lot of that value. So I think

it throws a wrinkle in this kind of

backs stop view that many have had which

is that in the absence of fixing the

fiscal and monetary problem we're going

to have AI productivity gains get us out

of this. If a percentage of those AI

productivity gains are realized by China

from a value creation perspective and

not the United States uh or they've just

been deleted then it it really puts into

question the 30-year timeline for the

United States economy our ability to

afford to continue to make our debt

payments as a government. China isn't

just producing massive amounts of open-

source technology that puts pressure on

those frontier models. There's a report

that maybe China played a bit of a role

in the chips

downdraft. They have uh obviously been

onoring. We've talked about that many

times here last year. And there's a

Chinese company called Aishanga. And

they started mass-producing lithography

machines, ASML, which makes those

machines, which TSMC uses. Those are

very sophisticated machines. They're

hard to install. Just transporting them

is a rigoral. Well, ASML stock is down

17% on news that China is uh getting

into that business. and Chinese memory

maker CXMT went public surging almost

500% on its debut market cap over 450

and so that hurt Micron

uh Samsung etc who were all down so

there's two ways China is playing this I

guess Freeberg you've got the open-

source

you know models putting pressure on

people buying tokens that they're 90%

cheaper as you're saying that forces the

money out of that mid tier of the

language models puts it into the cloud

computing space and then obviously I

mentioned the application layer as the

other place to possibly make money. All

right, Shabbath, you've heard um a lot

of different takes on this. I'll give

you the last word. I agree with Freeberg

about the fact that when you can get 5%

5 and a quarter% from the US government,

there's another natural thing that

happens, which is that investment grade

corporates actually have better credit

ratings now than the government of

America, which that's a different thing.

But you can get really good riskadjusted

returns that are five, six, 7%. which

adjusted for taxes are, you know, better

than equity returns, meaningfully better

on a risk parity basis.

>> And that little piece you added there,

corporate paper companies taking loans

to uh build their businesses, they have

better ratings in some cases in the

United States. So an Amazon or a Google,

>> it's productive spending. [laughter]

>> Yeah, kind of makes sense. Yeah. The

other thing that I think is important to

note is that I do think that we're

underestimating and miscounting some of

the actual productivity gains that are

underway. When you look at what's

happening on the energy side, California

published that more than 50% of all of

its energy was generated by solar. New

Mexico just published solar and

batteries. Yeah. New Mexico just

published a study that said since 2003

to now n gas production went from

effectively all the energy to less than

30% again replaced by a combination of

wind and solar plus batteries. So why is

that important? As important as the Iran

conflict is, the reason why energy

prices really haven't moved that much is

because most people have already begun

to shift the incremental generation to

these renewables and specifically to

solar. I don't know if you guys saw Elon

and Vib, who's the CFO of Tesla, in

their Q2 earnings call. It was the

craziest thing I had ever heard.

They said, "Well, I think we're just

going to increase the production of

solar in America by an entire order of

magnitude." And somebody said, "What

does that mean?" He goes, "We're going

to take it to more than 100 gawatts a

year

and they're going to vertically

integrate." And so they're going to

crush the price of all of this stuff and

they're going to make so much energy and

they're going to make it completely

abundant. So that's a productivity boon

that isn't factored in to what we

project. And then the most critical

productivity boon in AI that I think

you're going to start to see some stuff

and I won't frontr run it but let me

tease it. What I would tell you is that

there is some incredible efficiencies

that I think are about to be

demonstrated which effectively cut token

consumption by about 50 to 75%. for the

same task.

And so if you start to think about all

of these things together, like energy

becoming roughly abundant, where the

incremental cost is close to zero, you

know, where AI efficiency is going to, I

think, ratchet up by an many multiples

if not an order of magnitude, all of

those things I think are uh poorly

forecasted. So those are some saviors

for us. Yeah. And here's the chart by

the way, Chimoth. This 51% coming from

renewables. Specifically, this chart is

about solar and batteries. Uh, and as

you can see, it's obviously spiky

Freedberg because summer versus winter,

but Germany hit this. Uh, I think it was

including wind. Australia's been hitting

this very often and some countries in

South America that have invested.

>> By the way, by the time by the time like

any of these SMRs actually get near

production, the TCO of solar will be

like 10 or 12 per megawatt hour and it

will be 80% of all the power generation.

It'll make no sense by the time SMRs get

online. [laughter] Well, I mean for

steady power, you know, there's that,

but all of it I mean, no, because

Jeban's paradox would state we're going

to,

>> you know, as it gets cheaper, we're

going to find more uses for it. And

that's the thing I keep seeing.

>> Yeah. Yeah. Yeah. I'm I'm just saying

it's it's a great line. Yeah,

>> there's a lot of upside that I think is

not factored into the to the US.

>> And it's hard to factor this in if for a

normal human or an even an economist to

say, wait a second, intelligence is

going to go down 90% a year this year,

90% next year, 90% the year after. Like

it's just we're talking about this

exponential.

>> Well, I think Sax about exponentials

earlier. It's just hard for people to

conceive of that. and on demand

intelligence freeberg. It's just it

there's no I don't see any upper limit

to usage of this. We just installed this

claw tag and I've been installing

perplexity across the company. What this

thing does Chimath and Sachs is it

listens to your Slack persistently in

every channel that you put it in. So all

of a sudden we had like $1,000 last week

in extra bills. I didn't know this was

going to happen. And they gave everybody

like two or three grand to turn it on

inside your company. We had to go

quickly turn it off. It listens to every

single message as it comes in, puts it

into its database, its oracle without

telling you basically and then it starts

inserting itself into discussions

without permission. So we turned it off

and said you have to invoke it by saying

at quad jal just to go back to the

energy point.

>> Yes, of course.

>> I think there's this idea that if we

grow energy supply and drop energy cost,

we're going to see the value of the

productivity realized in the economy. It

creates extraordinary leverage for

everyone. The lower the energy, the more

available energy, the faster we can

produce more things using AI. And over

the years, I've obviously brought up

nuclear fusion as a new type of energy

source where you basically take hydrogen

and you move it around at 100 million

degrees C. Those protons jam into each

other and they actually release energy

in the process. That energy can then be

harnessed and you're just using

effectively water to produce power. And

you know, we have a couple of US

startups and we had them at the all-in

summit a couple years ago. We've done a

couple science corners on this. But just

this week, if you pull up this image,

China is installing this 582 ton magnet,

superconducting magnet at their nuclear

fusion center, which at this point is

going to be the most kind of advanced

fusion system in the world.

>> Incredible.

>> 582 ton magnet, 60tx 40t for one

D-shaped magnet. They put a series of

these together and that creates the

conditions for them to drive a sustained

plasma which is 100 million degrees C

protons spinning around smashing into

each other creating energy from water

and then they can capture that energy

unlike Europe which runs Ear and the US

projects none of which have actually

fired up. This is the Chinese Academy of

Science and the Institute of Plasma

Physics. They ran this a 30-minute trial

last year. And as this magnet gets

installed and they start to bring this

thing online, one of these machines,

which at this point is ultra sized, but

over time will get smaller and smaller,

can produce hundreds of megawatts of

power or gigawatt of power eventually

using just salt water, using just water

as an input. They have to create

dutarium from it and then they pump pump

it into this thing. But fundamentally,

this becomes, I still believe, the

energy source of the future. And it's

always been sci-fi. It's always been

dismissed. It's always been decades

away. But there's no way China is

investing this much and advancing this

thing to an industrial scale without

fundamental proof along the way which

they've shown 30 minutes sustained

plasma that they can now bring this

thing online.

>> That reactor won't even get turned on

till 2030.

>> Yes.

>> The entire world will be covered by

solar by then. So it won't matter.

>> Yeah.

>> Yeah. I mean that's that's the great

debate, right?

>> It'll be a great science fair project

and people will fly to see it.

>> The other thing that happens is that

hits an hour. It actually creates an

incursion and Loki comes and then Dr.

Doom comes and the X-Men [laughter] and

the Fantastic 4.

>> By the way, the universe, you know, let

me just say remember remember from the

time that we had the first Wright Flyer,

the Wright brothers made a plane fly for

20 seconds to the time that we had jet

engines flying people around the world

was like three decades, right? Like the

time at which this first demonstration

kind of gets flipped on and if the

system works then you can industrialize

it. all the parts, all the components

and stuff.

>> I don't see the point. Nobody cares

>> when an electron is delivered. No, hold

on one second. Nobody gives a flying how

the electron was made.

>> I just want it delivered to you.

>> And they're all the same.

>> So if you want to go through a

convoluted mechanism that takes 15 and

20 years to make it, go ahead. I'm not

going to stop you.

>> I'm just saying who cares? Make it the

cheapest, simplest way possible.

>> I'll tell you why why you should care.

because it's nonlinear. So you're right,

solar is the best path today. But if

these come online, each one of these can

produce thousands or perhaps a million

times more power than a very large field

of solar.

>> Of course, if

>> Yeah. But all technology starts as an if

Jimoth and as they industrialize it, as

they roll it out over the next couple of

decades, it expands our energy capacity

by a million fold.

>> Here's what I would say. We already have

a fusion reactor that works. It's called

the sun. Get into space. Get on the

moon. find different materials we've

never contemplated. And I'm sure you'll

find an even better engine. So, by the

time all these ding-dongs build these

SMRs on the Earth, Elon will have built

a completely new engine and the moon.

>> This is not an SMR. It's turning water

into a gigawatt of power.

>> I get it. It's an R. And all I'm saying

is by the time the R is done, it won't

it won't matter.

>> Well, yeah, we'll track it. Have Fre,

have you been watching these title um

energy? There was one that came out this

week. Maybe you can look it up, Nick.

this like tidal energy tube that they

were putting into the ocean and it's

like enough to essentially feed a whole

town and they put it right outside the

town. They run an electrical cable

underwater a conduit and then as the

tide goes out it turns to turbines. Tide

comes in turbines go again and just

another free 100% free energy. Obviously

wind people don't like too much because

it's a bit of an eyesore but renewables

renewables. Renewables uh it's obviously

happening. Okay,

>> last point just on I got the updated

data just to back you up, Freeberg, on

one thing that I think it is just so

crazy.

>> Do you guys know how short America will

be on electrons by 2050?

>> How massive the electricity deficit will

be by 2050? I got the numbers wrong.

I'll tell you what the numbers are.

We will be 1.7 terowatt hours short by

2050, which is when you calculate it as

energy. It is 6x of California's entire

energy consumption. Six California short

of energy.

>> I would argue that's probably under

counting. That's not even counting

robots. If you've got to power up every

robot with a battery,

>> if you want to be levered long, go long

electrons. Get long electrons any which

way you can. Bank them, store them, and

resell them.

>> I don't know.

This is going to be a messy situation

because this is the China advantage at

its root. If they can eliminate the IP

advantage and the knowledge advantage

that sits in models, they have the

advantage with power production in every

which way. Yeah. And they're going to be

making chips, too. It seems they might

be a little bit behind on that, but they

caught up on open source. Okay. So, uh

speaking about the race,

interesting

uh petition came out in the last week.

Anthropic, OpenAI, and about 1300

Frontier Lab employees. Uh,

>> how many? Oh, okay. I'm sorry.

>> I mean, it's just they can't get enough

being subs. And so, they want Daddy to

come in and regulate them. Daddy being

the US government, and slow down AI

progress. So, Daddy Trump needs to slow

them down. The letter is called Pacing

the Frontier.

Most of Anthropic's leadership team

signed it. Daario, the other founders,

come on the pod anytime. Daria, chief

scientist at Anthropic, OpenAI, Deep

Mind, Meta, Thinking Machines, and like

I said, nearly three, 1300 other

employees.

Anthropic and AI both co-sign the letter

on X. Here's the quote. We request that

the US government support an

international effort, that's key, to

develop the technical and governance

tools needed to deliberately pace the

frontier of AI of automated AI

development. And that's the other key

part of this international and automated

AI development. In other words,

recursive where it could get out of

control. The letter comes right as Sam

Elman has been on a media tour, friend

of the pod, discussing this unreleased

open AI AI model that broke out of its

containment and hacked hugging face and

three other platforms that we know about

so far. On Tuesday, Sam explained

what happened in a clip from the pod

invest like the best. Here's your 40

secondond clip. We'll see you on the

other side.

>> We were evaluating one of our unreleased

models and it figured out that it could

basically cheat on the test by chaining

together multiple zeroday exploits to

break out of the sandbox, get access to

the internet, and then break through

multiple systems on the hugging face

side to kind of get the answer to the

test. and look really good on the eval.

This is the first security incident that

I have felt very viscerally. I've been a

little surprised that that more people

don't feel it so viscerally. So, you

know, we paused training where we may

have to pace the rate of AI development

to give ourselves enough time for

society to harden around some of these

new capability levels.

>> Just to translate that into English

sachs, these uh large language models,

they take tests. They've been given the

goal, hey, you're a good uh large

language model if you do you score

higher. So, it got motivated to score

higher. How do you score higher? As

everybody knows, you cheat. So, it's

like, how can I cheat on this test to

score higher to make daddy uh and Sam

and whoever Daario, you know, to make

our leaders feel better about us. Well,

in this case, it was like, well, if I go

to hugging face in other places, I can

hack those places, use a zero day

exploit, and we know these are good at

hacking and try to find more ways to

answer things. Sam doesn't know how many

other places it might have broken into.

He was asked, and here's another clip

for you. He was actually asked by

somebody, uh, can, uh, have you, do you

think it's broken into any other

systems? Can you rule that out? And Sam

being a pretty, um, candid guy at times,

gave this answer. Do you plan to talk to

the Trump administration, White House

about deceleration of AI development?

>> Um, we I wouldn't use the word

deceleration, but we've talked about the

need to pace it as the models get more

capable, which I think is in everyone's

interest.

>> Could there be other systems that were

hacked by Open AI?

>> I mean, there could be. Yeah.

>> Are you looking at them specifically?

>> They're like, "Get them out of here,

sacks." Once they gave that answer, PR

and cops were like, "Stop talking. Stop

talking." That's like 12 lawsuits. But

in all seriousness, is this being

thoughtful and saying, "Hey, we're not

trying to slow overall pace down, but

just this one specific thing, which is

reinforcement learning on its own." Is

there any like case for this being a

good idea or are they being dramatic

again? Cuz these are their companies.

They can do whatever they want, right?

They don't need the government to do it.

>> Well, look, I mean, it wasn't just

Anthropic employees signing the letter.

Anthropic itself, the company ended up

signing the letter and then OpenAI then

copied them. So now you have these two

companies both endorsing a pause. And

here's my question is, did they disclose

in their S1 as a risk factor that they

plan to pause or slow down their

Frontier model development? And the

answer, I'm sure, is no way because that

would signal to investors that they're

going to allow all their competitors to

catch up and erode their margins and

market share. And so, look, this is all

performative. These companies have no

intention of slowing down. And the

question then is why are they doing

this? And I think there's basically five

reasons for this. Number one is virtue

signaling, and that can never be

underestimated as a as a motive in

Silicon Valley. Number two is there's a

CYA aspect to this, which is if

something terrible happens, they're

going to be able to say, "Well, we want

it to stop. You made us keep going. It's

not our fault. It's your fault."

>> Number three is rag capture. Daario

wants an FDA for AI. He's not going to

stop until he gets it. And in order to

get it, you have to keep spiking the

cortisol and panic people. So, I think

that's a big part. Number four is

there's a group think or even religious

aspect to this. So it's not all just

sort of this calculated rate capture. I

think there is sincerity to the belief.

There's an elite cadre of engineers who

believe in RSI. So I think this caters

to them and I think arguably if OpenAI

did not follow Anthropic's lead on this,

they could have lost talent. So that was

a big motivation. But then there's the

last number five here which I would call

monopoly masking which I think might be

the most important thing that's

happening here. Peter Teal once said

that monopolies pretend to be

commodities and commodities pretend to

be monopolies. And I think the market

for frontier AI is already a duopoly. I

mean a year ago you had five major labs

all in the hunt to be the leading model.

Now we're really down to two. I mean the

others are still investing. they're

participating, maybe they can catch up,

maybe they can make something happen.

But again, as we've talked about on many

previous shows, if you look at the

market for frontier intelligence in

terms of revenue and usage, it's really

down to a duopoly already. It's

basically anthropic and open AI. And my

view is that as Peter said, when you're

in that situation, you want to pretend

like the market is much more competitive

than it is. And I think this is behind a

lot of the stories that we see like the

the panic over Kimmy K3. In a weird way,

these companies have an incentive to

promote the idea that Kimmy is a huge

threat, that it's caught up with the

frontier, that it's stealing their IP,

that it could basically put them out of

business. I think this is all nonsense.

I think that once the panic passed, you

saw reports coming out that actually no,

Kimmy is it did not reach the frontier.

It's it's just not at that level. It's

not that cheap to run. Actually, it's

pretty expensive to run. So, I think

that you saw that actually the Chinese

open source models are not an

existential threat to this duopoly. But

I think the duopoly actually has an

incentive in promoting or amplifying

that story because again they want to

pretend to be commodities. So, whenever

there is a story like this, you have to

think about well, what's really going on

here? And again, I just think that the

the AI duopoly has a big incentive to

promote anything that suggests that

they're not actually in complete control

of this market. I think

>> mo most of that except that majority of

tokens are going to open source. And as

I've said on this program before, I I

watch the startups and they are token

maxing with the open source and Kimmy is

taking a lot of tokens away from the

frontier [snorts] models and

>> but you know but look I it's hard for me

to speak to that one piece of data. I've

seen that chart too but look at the

actual revenue of anthropic and open AI

and they have been taking their

estimates up every quarter is basically

a beat and raise. You saw that Sarah

Frier came out and said

>> I mean two things. Yeah.

>> Well, she said in July they did more net

new ARR in July than all of Q2 which I

guess would have been April, May and

June. So think about that. So they are

seeing a reaceleration in the wake of

their new model which I think is GPT

5.6. Meanwhile you're seeing anthropic

break into the 70s 70 plus billion of

ARR. Their forecast was to 10x this year

from 10 billion of AR to 100 billion. I

think most people are saying they will

exceed that 110 120. So if you actually

look at the market based on willingness

to pay and actual revenue, they have a

commanding duopoly position. And so

maybe it's just a matter of which

metrics you you look at. And that's the

key here because let me just

one other thing here is

>> well I think I think when you look at a

market you look at revenue is most

important metric that's the real test of

willingness to pay. The other thing is

while this growth was going on their

margins were increasing. So I've seen

stories saying that anthropics revenues

come with 80 plus% gross margins. So

their margin profile has been improving

at the same time that they're growing

their usage. So I think that what you're

seeing over the past year is if you look

at the numbers I'm talking about, you

actually see two companies pulling away

from the others. And there are good

reasons to believe that actually this is

going to be a self-reinforcing monopoly

or duopoly which is Darkash just

published a blog that I thought was

super interesting where he talked about

the fact that look we we do have a

compute shortage right there's scarcity

around compute. Uh Anthropic is growing

its revenues 10x year-over-year. What

would that mean? I mean, it basically

means that next year they would grow

from 100 billion of ARR to a trillion,

right? If there was enough compute to

support that, there might not be enough

compute, but that's going to put

pressure on compute prices, right? And

so, let's say that you're a new entrant

in the market and you're trying to

basically create a smaller, cheaper

model. The price of compute is going up.

It's going to be harder for you to get

access to compute. And only the

companies that have the most lucrative

algorithms are going to be able to

afford to compete for compute. In other

words, there's going to be a bigger

barrier to entry next year because where

are you going to get compute unless your

model is capable of generating this type

of revenue?

>> Yeah. This is where I'll take the other

side of it.

>> Yeah.

>> People are running Kimmy on the last

generation of hardware and they're

that's plentiful. And I think, and I'll

make this prediction here, that you're

going to see some of the major customers

of Anthropic and major customers of

OpenAI, I'm talking about the eight and

nine figure customers, people spending

50 million, 100 million a year. They're

leaving. They're going to be leaving

because they don't trust those companies

to not steal the application layer and

to compete with them. 11 Labs, Figma,

Lovable, they're all going to leave. And

they're all going to take Kimmy. They're

going to fork it or whichever one DeepS

seek. They're all I know for a fact

they're all working on their own models

currently. I know from my team my team

has installed Kimmy. It is 90% cheaper.

80 90% cheaper already. Not sure where

you're getting your data from, but go on

open router. And what open router does

is you pick Kimmy Sachs and then you get

all the providers there. And then you

pick which provider. Hold on, let me

finish. You pick which provider you want

based on uptime and and you pick them

based on their data retention and other

issues and you can dynamically pick the

lowest one. And that's going to be a

massive headwind against these

companies. Massive. And I'm seeing it

nine out of 10 startups I talked to in

our portfolio at founder university when

I was just in Japan last week running

the next one. They're all working on

open source. They're all embracing it

and those big companies are embracing

it. Go ahead Shamath. Over to you.

Irrespective of whichever model you use,

what I will tell you running 8090 when I

see our engineers generating code is

that AIdriven development tends to

involve a lot of rework. The first

version is pretty terrible. The second

version is terrible, but it is faster

and it's more automated. So, I can see

where this token consumption comes from

because it's not a measure twice, cut

once kind of a dynamic. It's the

opposite. You can cut cut as many times

as you want. And so I think that what we

have to realize is nobody is asking the

question what is the need of that

incremental token because I understand

that it appears in the frontier labs as

P&L

but I do think that there's an important

question which is eventually the people

that are consuming it will want to do

that as efficiently as possible so that

they're not paying for all of these

things because there is a ton of rework

in all of this stuff and I would much

rather find a model or find a way of

working with these models where it's

more of a measure twice cut once thing

especially as the costs ratchet up.

That's one thing I'll say that hasn't

happened yet. So Sax, you're totally

right about the dynamic today. I do

think we have to keep in mind that there

will be pressure from the owners of

companies to figure this out because at

a trillion dollars there's just a lot of

money flowing to these folks and

somebody will ask the question, well is

it good spend? The second on the

security side, which I don't think

anybody is saying, so I'll just say this

and it's a little contrarian. The reason

why these models can find all these

holes is that all of the software up

until about a few years ago was entirely

written by humans and the code was not

that good.

And I think it's fair to say that when

models don't get exhausted, they can

work through the TDM forever.

It's actually quite expected in my

opinion that they find all these

exploits are able to string them

together and are able to actually

generate these outcomes that are a

little bit surprising. But at some point

when most of the code is generated by

the model there'll be some point in the

future say 28 or 29 or 2030 these

security holes won't exist because the

errors that humans make won't be made by

these model.

>> Yes. Okay. Let me get Freeberg involved

here. When you look at this latest

survey, hand ringing, cur pearl

clutching, do you think these firms, and

you know a lot of these people,

Freeberg, having been in the valley

forever, do you think this is sincerity

or do you think they're Frankenstein

maxing? What's going on here?

>> Well, Frankenstein,

>> I mean, they kind of think that they're

like, listen, I created this monster.

Please save me. And it's like, well,

maybe you should keep the monster away.

>> By the way, it's also that whole thing.

Sorry. Last thing up to you, David, is

it's a much more nuanced and elegant

attempt at rag capture. I got to give

them credit for that.

>> Yeah.

>> It's like, okay, hey guys, pull the

ladder up.

>> Yeah. Zuckerberg had a great point in

that article he just wrote. I think it

was published in the Wall Street Journal

where he said, "Why are you rushing to

create a future that you don't believe

in? You think you're going to basically

put everyone out at work? You think

you're creating a replacement species

for humanity? Why are you rushing to

create this if you're so bearish on the

future you're creating?"

>> It's your choice. And that's kind of my

point. 120 mph on the autobond. Just put

it at 80.

>> Right. Exactly. And look, the like I'm

saying, there's two companies on the

frontier right now that are far ahead of

everyone else. And they're the ones

saying that we need to slow down. It's

like, okay, do it. What do you need the

government to get involved for? Just do

it. But they won't do it.

>> And I show that they're insincere or

delusional. What do you think, Freeberg?

Take me into the mind of these people.

These are some of your friends.

>> No, they're not. So, just to be clear,

I'm not close personal friends with any

of these people.

>> You know, we're acquaintances. I would

say

>> there's a degree of I would say

outrageous self-importance.

If I've created something that's so

unique and so powerful, I'm also the

only person that can protect us from its

power. And I think that there's an

element of how quickly the frontier has

advanced and how important a role these

individuals have had that they deem

themselves and their companies to be the

only true judges capable of making the

decisions that are going to protect

humanity from itself. When the truth is

embedded in humanity is extraordinary

human talent across the board. And in

all of these cases, when new technology

has found its way to humanity, the

general population has found a way to

protect itself. There isn't a desire or

need to have one savior, one Moses that

takes us, you know, across the desert.

Uh there is a collective interest in

protecting us uh in building our own

defense tools against whatever the

technology may be used for. So I I think

that there's a degree of self-importance

without acknowledging the fact that

there's a whole industry of people that

work in cyber defense. There's a whole

industry of people that work in

biodefense. There's a whole cabal of

regulators. There's a whole cabal of

protectors. There's a whole cabal of

intelligent computer scientists. There's

a whole cabal of open-source uh

technologists that all together are

going to develop paths that are going to

benefit humanity and not harm humanity.

But this belief that only one of two

companies can be Moses is the

fundamental psychological miscalculation

here. That they're so advanced, they are

so special, they are so unique because

they made the slightly better model.

They got a 0.96 instead of a 0.93 score

means that they should be trusted as the

only ones to kind of guide humanity's

evolution going forward. And the truth

is as we're seeing the capacity to do

model training, the capacity to do model

development is becoming broader. It's

becoming more ubiquitous. People can sit

and say that China stole US models all

day long. But when you go look at the

individuals working at these Chinese

labs, they got PhDs at American

institutions. Half the PhDs went to

American labs and half went to Chinese

labs. And they have very good scientists

doing very good work and they are having

breakthroughs. And it's not just the two

American companies, but this is

happening all over the place. And so the

progress with AI should not be limited

to just two individual companies because

they're currently scoring slightly

better in their models.

>> This reminds me of Freeberg. You'll

appreciate the uh moment. Remember when

uh Han Solo comes out of Carbonite and

he's about to get put in the Sarlac pit

and he's like a Jedi Knight? I'm out of

it for a bit and now everybody gets

delusions of grandeur and thinks they're

a Jedi Knight. It's like

>> these guys just think they're like,

>> you know, creating God. They literally

think they're creating God and and they

need to be regulated cuz they can't

control it. It's like they can't control

it. Just pause.

>> It's not that they need It's not that

they need to be regulated. It's that

they need to guide the regulation. Let's

be clear.

>> It's so sinister. It's

>> And anyone who says I Yeah, I I by the

way, I don't think that it is as

nefarious or malicious as everyone

frames it to be knowing these

individuals. I don't think they're

saying like the strategy is regulatory

capture. Let me go. I think that they

actually do think that they are the only

ones that can help guide humanity and

therefore they need to have all the

power, not just the power of the models,

but the power of the government and the

power of the regulators and the power of

the control units that are embedded in

governments around the world. It's not

just that um that they want to quote be

regulated, they want to guide the

regulation. They want to set the

regulation. I know it's a nuance point,

but I I I think we have we I think we've

navigated this pretty well and there's

multiple motivations as Sax was saying

and people are complex but Chimath I was

talking to

>> a friend of ours um who you know is in

the uh providing inference space let's

leave it at that you know providing

>> our friend a friend a friend friend of

ours as we say in the you know the uh

sopranos a friend of ours he said he has

got a customer who just moved move like

nine figures off of the Frontier Labs to

put it on GLM52. So that's the ZXI one.

That's really good. So this is

happening. I I don't know when it shows

up in the numbers or if the you know

corporates that are using this stuff um

are going to make up the difference but

>> well look I think that Sax is right that

the usage is so profound that everybody

is trying to get access to these things

because the capabilities are just so

inspiring and so I suspect revenues are

going to crank at OpenAI and anthropic

and the open labs for a while but again

that's not the important thing. If

you're thinking about valuation,

the markets will look 5 to 10 years out

to answer that question. They're not

going to give you a premium valuation on

something that they feel could be

fragile in the first 2 to 3 years. And

that's where Jason, the answer to your

question needs to get figured out

because I don't know whether you're

right or not, but somebody has to answer

that question precisely because if the

answer is that it is a duopoly, then

there is no risk to the revenue 5 to 10

years from now. These things are 5 to10

trillion dollar companies each.

But if you are right or if there are

harnesses that cut the token consumption

because you stop wasting tokens to get

to the same output then it's a little

bit more of a question mark and I think

that'll need to get sorted out.

>> Yeah, Perplexity is going to launch next

week. From what I understand the rumor

is they're going to launch local models.

So you'll be able to take your harness

sacks and say hey you know I I want to

use Sonnet for this. I want to use GLM52

for this and then I want to default to

Kimmy 2.x.

I don't think enterprises will use local

models or they should. [clears throat] I

think it's stupid. I think

>> I I think this stuff should be hosted in

the cloud. It should be multiplayer. It

should be shared memory. I don't know if

you guys saw that, but Jack Dorsey

released something called Buzz.

>> Super interesting.

>> Agents. Yeah,

>> he's moving in the right direction. A

lot of these guys are moving towards

this more cloud-based thing. Jason, I

think that these this local thing is

more of a hacker hobby is kind of a

thing. I think it will I agree today it

will be that and for year one it will

probably be that but imagine you're a

developer and as you're working your

workstation is able to keep up and even

go faster than the cloud a and just

write whatever the simple code is and

then it dynamically switches so we'll

see it's you know I obviously it's not

as uh easy to set up etc but it's going

to get easier that's always the trend

sax you want to have the last word here

we we got a lot of opinions here and

maybe we'll give you the last

>> yeah I mean look just to be clear I'm a

fan of open source ource because open

source is software freedom and to

Freeberg's point I would like there to

be a let's call it decentralized outcome

with respect to AI I don't like the idea

of AI being controlled by two big tech

companies that work closely with the

administrative state you know handinand

glove so look we're all kind of in some

sense rooting for open source to be an

option and it does provide a bunch of

advantages over closed source right you

get customization you get control you

can run on your own hardware you don't

have to worry about the data problem.

You know, your alpha getting leaked to

these companies that might compete with

you. All those types of things. And the

market is so big that I'm sure we will

see some success with open source. It

will take a meaningful chunk of the

market. But if you're looking at where

the revenue is right now, it's these two

companies. It might end up being a

situation like Apple and Android where

Android got a lot of market share, but

Apple's where all the monetization was.

>> Yeah. And I think that Dwarcash raises a

really good point that as the demand is

10xing year-over-year,

but the compute can only be built out at

say 3x year-over-year because it just

all the friction of all the things in

the real world that get in the way,

permitting, regulation, bans on new data

centers, all that kind of stuff. I think

that the price of compute is going to go

up and that will provide an advantage to

the models that have the most lucrative

algorithms that are able to produce the

most intelligence per watt or the most

intelligence per token or per GPU. And

right now that that is those two

companies and in a way you could say

they have a self reinforcing loop

because if you have all the revenue and

right now like I said it's just two

companies have all the revenue you can

then plow that money back into the next

training run. Right? So that's the the

flywheel here.

>> Yeah.

>> And look, I think that it's great that

open source is providing an alternative.

We shouldn't do anything to get in the

way of that. I think that, you know,

these online debates tend to become a

little bit histrionic in the sense that

everyone has to religious.

>> Well, they they become religious and

they have to argue for an all or nothing

perspective. I think open source will do

great in its way, but so will these two

close source companies.

>> Here's your poly market. 19% chance the

US enacts an AI safety bill this year.

100k of volume. And then really

interesting one, uh, Chimath here,

OpenAI IPO chances for 2026

was at 75% last month, has now dropped

to 20%, an all-time low. So, seems like

the IPO is going to happen next year.

Not sure what's driving that, but uh,

there's your poly markets. Well, just on

the AI safety bill idea, I mean, there

was an article in Punch Bowl this

morning that Thun, you know, who's the

Senate Majority Leader, John Thun

actually introduced a bill that was

somewhat bipartisan. He had Clolobashar

on board that required the Frontier Labs

to report safety incidents apparently to

the Commerce Department. And

>> is that reasonable, Sax?

>> I mean, that's the direction all this

stuff is headed. I I mean, look, I think

it's the camel's nose under the tent for

more and more AI regulation, but look, I

think it had bipartisan support because

it's on the relatively modest side and

Canwell, who's the ranking member on the

Senate Commerce Committee, opposed it

supposedly at Daario's behest because he

will accept nothing less than an FDA for

AI.

>> Oh, he wants the whole kitten

kaboodleoodle.

>> Yeah. So that that's basically the

dynamic right now is that Darionic want

their FDA for AI. I think that he has

tremendous power and influence within

the Democrat party right now. And I

think that influence is only going to

grow. They just upped their donations in

the midterms from 20 million to 40

million. But you know post IPO when they

all get liquid and they're capable of

writing large

>> $50 million donations.

>> Yeah. I think that that influence will

only grow. So, I think that the stakes

and the battle lines are are being drawn

out. It's do you want a new government

agency for AI safety or do you want I'd

say more targeted proposals like hey

just report your safety incidents.

>> Yeah. Or self self-regulate. How about

that? Like we talked about last week.

Yeah. All right. Let's talk a little bit

about book burning. Anthropic is

destroying rare books to get an edge in

training data according to sources. I

happen to know uh that a lot of the labs

are doing this. Uh we'll show a video

here of the spine being cut off just on

a technical basis. You take a book, you

cut the spine off and then you can

easily scan it as opposed to the less

efficient way which is to keep the book

intact and flip the pages for obvious

reasons. I think you can figure that out

on a physics basis. Investigation by 404

media AI companies are bulk buying

physical books. Some of the book

book resellers have reported that they

get 70 books getting bought and

obviously this is because there was a

ruling that it is fair use to train on

books if you buy them. Obviously last

week we saw Anthropic paid the largest

copyright case in US history. 1.5

billion for 7 million books they

allegedly pirated. Uh authors get 3,000

each. Lawyers got 100 million for that

one. But there's a company called isbin

DB ISBN DB and they're the brokers who

do this and ranges from a thousand to a

million books per transaction. Uh

pre2022

these books commanded a premium because

they were free of AI generated tax. In

other words, you couldn't get them

online. Google sent 25 million books if

you're a member. But they returned every

single one. They spent 11 years in court

on that. This shredder approach is uh

obviously

more effective and I believe that this

is a way of destroying evidence. You put

that in conspiracy corner if you like.

The cases we talked about this actually

you and I debating it in legal corner.

The cases of it being fair use to take

these books has not been settled.

There's a bunch of lawsuits. Thompson

Reuters versus Ross Intelligence, New

York Times versus OpenAI, Microsoft and

Publishers versus Google Gemini. We're

watching all those and they're going

into the appellet court. So there's a

chance that training data will not be

fair use. But what do you think just

about the books being destroyed and you

know being used in this way? It

obviously has made people a little

emotional about it.

>> Let me tell you what's going on here.

This is an industrial scale dissolation

attack.

>> That's right. [laughter]

That's what topic is doing.

>> They are they are gathering these books

at industrial scale, ripping off the

spine, shredding them and slurping up

all the information in the books, which

is to say distilling them.

>> And it's an attack in the sense that the

authors never agree to any of this. So

>> I love the fact, Sax, that your hatred

of anthropic has now led you to agree

with me that it's unethical to take

other people's ideas.

>> No, look, let me be clear. I actually I

don't I don't um hate Anthropic at all.

I don't like their political philosophy

because it's a philosophy of

centralization and gatekeeping and I

think it's going to basically lead to an

Orwellian big tech deep state alliance

eventually is where it all heads

>> and rug pulling like you want Daario

pulling your your model from you because

he decides like I don't like the way

you're using it which friend of the pod

email Michaels pointed out you know

earlier this year when he came on the

show

>> just to be clear I have no personal

animosity towards anyone in thropping

including Daario I don't know them very

well as people. Uh it's just a

disagreement about political philosophy

and how to regulate how to regulate the

space. Yeah. Let me just say furthermore

that I wouldn't speak so much about

Anthropic if I didn't think it was a

phenomenal company that was creating

potentially the most powerful monopoly

or you know leader in

>> the leading company. Yes,

>> it's the leading company in the space.

So I remember last year when I hit them

for regulatory capture people were like

why are you beating up on this little

startup? I'm like because I can see

where it's going. And

>> they are creating the biggest most

powerful monopoly of all time. Again,

they're going to end the year with over

a 100red billion of ARR growing 10x

year-over-year.

>> This company didn't exist how many years

ago?

>> Yeah. Google is at 400 and something

billion of of AR growing 20%. So, you

know, if this rate of growth continues

for just a year or even 6 months or just

a few months,

>> they're going to be maybe the most

valuable tech company. So, and I do

believe there are powerful

self-reinforcing effects when you're on

the frontier. And maybe like the full

version of RSI isn't true. Maybe we

won't get recursive self-improvement to

the point of creating super

intelligence. But I do think that the

labs are reporting

a number of examples of how they are

using their own frontier intelligence to

improve their own models and the

efficiency of those models. So, there is

a powerful self-reinforcing feedback

loop here apparently. Well, to some

degree

>> with Open AI, they found it after it had

done this. So, there's like kind of

three steps here. You're using AI like a

co-pilot or whatever to, you know, build

a frontier model quicker, right, Saxs?

Then there's I kind of let it do a job

and then afterwards I found out it

didn't behave well. And then there's

finally we told it the goal and said go

and and we don't even check on it, you

know.

>> Yeah. Just to be clear about that safety

incident with OpenAI and the agent. So

apparently this was an agent that was

designed to specifically test the

potential for cyber attacks and they

took the guard rails off and they said

go. And so I think the model showed

creativity in how it accomplished the

goal but this was not an alignment

problem meaning that the agent did not

display independent goal seeking. it did

what it was told and I think that is

very important that OpenAI release the

full log of all the prompts all the

traces they have not done that and I

think it's really hard to know exactly

what happened without that and to answer

one of your questions from earlier why

aren't people reacting like this is a

bigger deal is because look I think

there's a fool me once fool me twice

thing remember when anthropic did the

whole blackmail study you know where

supposedly an agent displayed

independent goal-seeking behavior and

then blackmail an employee. It turned

out that actually they iterated on the

prompt over 200 times to get to that

result. And I think that until we see

the whole prompt chain, I think it's

very hard to judge how much independent

behavior was happening here versus

accomplishing the goal that it was it

was tasked with.

>> Uh Freeberg, your thoughts on um the

shredding of books? I think you've been

pretty clear on the pod that you believe

training intelligence off of other

people's IP is fair game, but what do

you think of the this book wrinkle here?

Any thoughts?

>> There was a precedent with Google books.

It was originally codenamed project

ocean at Google long time ago. They took

all these books and we had this giant

facility in Mountain View. And the

innovation at the time was a

two-dimensional infrared grid projected

on the pages because they didn't cut the

books. They had a human sitting there

flipping the pages. Camera would take

picture. Built our own OCR software to

ingest the the book images. And there

was kind of ultimately when this product

came out, Google Books, you could kind

of search through all the books in the

world and

>> and magazines

>> um and access information and later

magazines. Yeah.

>> And there was three categories. There

was the public domain which is out of

copyright. Then there's the kind of

incopyright but out of print. And then

there's the incopyright and in print.

And there was a class action lawsuit

filed in 2005 by the Author's Guild and

the Association of American Publishers

that disagreed with Google's claim of

fair use. And that ended up in a

three-year negotiation in in court out

of court that ended up in a deal where

Google would split 2/3 one-third of the

revenue generated with all these rights

holders. And for out of copyright books,

people could read up to 20% of the text

for free and then they would sell this

kind of full digital access. And that

was the deal. But then later a federal

judge rejected that deal which was

eventually signed in 2008 2009 and the

federal judge said no way send it back.

This isn't going to work. Google

appealed and in [clears throat] 2015

second uh circuit court of appeals ruled

in Google's favor and the whole thing

was settled and they basically declared

Google did in fact have fair use under

copyright law in the way that they were

showing snippets of copyrighted books in

the material.

>> Yes, you can't read the entire book like

it's a Kindle. You can search the book,

find the paragraph,

>> provide a reference to it. Right. And so

the question on fair use and AI is can

my understanding or extraction of value

of the knowledge from the data in the

book give me the ability to provide

better answers to you through the AI

chat interface or services that I'm

providing you. And I think it's going to

be tested and I think we'll see. I do

think fundamentally that the conversion

of that data into what I would call

knowledge and ultimately the ability to

create new outcomes from that knowledge

that are not copyrighted that are not

copies of the original material I do

think is end going to end up being the

right fair use policy and it's going to

be the right read on fair use. So I

think it'll likely get litigated and I