Chip Stocks Crash, $20B Fund Margin Called, Frontier Labs: SLOW DOWN AI, Mamdani's Grocery Stores

00:00:00All right, everybody. Welcome back.Welcome back to the number one podcastin the world. The Core Four, FantasticFour. The original Quartet is here. Didwe peak at like uh number two or numberthree 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 oneglobally, but yeah, and sometimes numberfour US.>> I forgot my Starlink. So, let meapologize to everybody. That was acritical error when you're on the roador on the water, but you know, it'll befine.>> I had five huge leads come in this week.>> The Glengar Glenn enterprise leads.Enterprise sales is a bear because it'ssuper chunky, but the deals areginormous.>> Sax, you got any advice for Chimat fromyour uh enterprise sales days? Youclosed some of those big seven figuredeals 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 notlever up.>> That's good. Yeah, you want that. Thatsituational awareness.>> I mean, it's kind of out there. I mean,if you name your fund situationalawareness,

00:01:15that's>> Yeah. Come on the pot anytime, Lualt.All right, everybody. We got to talkabout chip stocks crashing after analltime runup. And we had a major hedgefund get margin called and someincredible margin calls happening inSouth Korea. Leopold Ashen Brener is a25-year-old hedge fund manager. He leftOpenAI two years ago to start his ownfund and apparently according toreports, this is breaking news onThursday when we tape, he got margincalled and had to sell his entire publicportfolio to cover massive losses causedby his leverage. And who bought them?none other than Citadel's Ken Griffin.We don't know if it was Ken Griffinhimself, but Citadel bought it accordingto the early reports. Leopold had insanereturns and he rode the wave of AI andchips and Frontier Labs as recently asthis month. Uh, and he started the fundwith $225 million in 2024. He grew it100x to 20 billion this year or so. Ranit all the way up to 45 billion. Nowhe's at 200x earlier this month bytrading on leverage. According to ourfriends at CNBC at the end of the June,he was reportedly up 4x450%this year.

00:02:30Uh some have reported thathe's also selling his massive anthropicstake to cover these losses, but theWall Street Journal is disputing it.Again, uh he we're we're happy to havehim here on the program. How did thisall blow up? Well, NASDAQ's chip index,this is called the PhiladelphiaSemiconductor Index, is down over 20%over the last month. That's bare marketterritory. Obviously, definition of baremarket territory for those of you whodon't play in the markets is anythingover 20%. Uh the index included the top30 US listed chips. That's people likeNvidia, TSMC, AMD, Micron, you know, allthose big names. But the index bouncedback a bit today, up 7% when we'retaping. So, we may have found a bottom.Unfortunately for Leopold, he hadalready sold. Samsung and SKH, two SouthKorean chip companies that are notincluded in the NASDAQ index, also gotsmashed, crushed, demolished. Samsungdown 38% over last month. SKH Highixdown 14% since going public 3 weeks ago.The Cosby, that's South Korea's versionof the S&P 500, is down over 40% in thelast 40 days. Between last Friday andWednesday, leading chip companies shedover a trillion dollars in market capcombined. So to put this in

00:03:45context,chip stocks had a legendary run the pastcouple of years, but uh you know tradingon leverage, we'll talk about it. Verydangerous. If there's a downturn, we'llget into the South Korea wrinkle aswell. Even with this downturn, the5-year results are still spectacular.Jimoth Micron up 850% mostly in the lastyear. Nvidia up 875% in the last 5 yearsand Broadcom up 663%.

00:04:13Let's discuss it.>> If I was going to give you one piece ofadvice when you're running risk is youhave to manage leverage incrediblycarefully because when it runs ahead ofyou, the unwind is incredibly violentand it's incredibly quick. That's thebiggest problem with with running eithermassively levered long or massivelylevered short. So I don't know to whatextent he was running lever but therumors are he was running like three anda half turns which just to give you asense when you're running that much riska 3 and 4% move is amplified 12 and 13but if you saw what's happened in thelast 3 days a 25% move is amplified 75%.So>> it has the risk to stop you out. Andwhat happens is when you get thatleverage, the banks are given theauthority to close you out. And whenthey close you out, what they do is theystart calling around and unwind yourrisk. And you don't have much of achoice. It's sort of an automaticone-way ratchet.So if everything that has been reportedis accurate, he was running about threeand a half times levered. The marketmoved against him. He lost a very largepercentage

00:05:29of his gains and then theprime brokers started calling folks.Citadel bought the whole book and nowthe question is what is the AUM left andwhat is the high water mark and can heactually dig his way out? These thingsare brutal. your thoughts, Saxs, uhlooking at this situation,any lessons for you or I guess biggerpicture, this downdraft, is it becauseof market conditions, you know,inflation, the war, or people just aheadof their skis when it comes to uh thevaluation of these companies and then hejust got caught in a downdraft. Yeah.>> Well, I think that is the key questionhere. Is this correction in the markets?Is it driven by fundamentals or is itdriven by momentum? And my view is thatI think it's driven by momentum. Meaningthat over the past year, you've had thisroughly 10x runup in memory chip stocksand you've seen this overall huge risein any stock that's related to the AIboom. So, anything related to this AIcapex boom has been going up like crazy.And I think it was inevitable that you'dsee a pullback. I think there wassomething like a 10% pullback in theNASDAQ

00:06:45from the peak. But when you look at thismomentum trade, it was down like 30% or40%. Right? Because the 10% was on thewhole market. So this sort of momentumtrade was the most exposed part of it.And you look at what happened in SouthKorea, you look at what happened withLeopold's fund and obviously there was alot of leverage behind this momentumtrade. So when it corrects it's going tobe brutal. But I think that the questionagain is does this reveal anything aboutthe fundamentals? And my sense is thatyou're already seeing the rebound thismorning and what I mean by that when Isay fundamentals is is the capex that'sbeing invested in the AI boom is thatreal or is it misguided? Right? Is it isthat a sound investment? Is that aninvestment that the hyperscalers forexample should be making? Is that aninvestment that's eventually going todeliver ROI or is this some sort ofbubble? And my view is that it's realthat I think there will be a return onall this capex. I don't try to predictstocks or tell people when they shouldbe buyers, but you look at thehyperscalers, they have invested prettymuch all of their free cash flow andthen some in this boom. You know, a lotof people are trading those stocks downbecause of that. My view is thateventually there

00:08:00will be a return onthat investment. And this is sort oftemporary market volatility amplified byleverage. And Chimath is right. Youknow, I think it was Warren Buffett ormaybe Munger who said that leverage isthe only way that smart people go brokebecause, you know, if you're not usingleverage,your portfolio would just be down 30%this month and then it would already beup 7% today. So, you'd be rebounding.So, you'd be down, okay, 20 something%this month, but after having risen 10xin the past year. But if you'releveraged 3 or 4x, you're wiped out>> and you get margin called. So look,there's many examples of really smartpeople getting hurt by by leverage.Yeah. And that that's the lesson there.Now, I think Leopold's a reallyinteresting figure in the whole AImovement and I would say an interestingthinker. I met him about a year year anda half ago.>> Did you invest in the fund? Did you giveyour>> No, I wasn't in I was prohibited frominvesting in things like that.>> You were in you were in DC at the time.Yeah,>> but I thought he was a reallyinteresting thinker and he wrote a blogcalled situational awareness before hecreated the hedge fund version of it.And I thought what was reallyinteresting about it was just he laidout the bullcase for the AI boom. Andjust by the way, he's like very wiredinto anthropic.

00:09:15I think his fiance isDario's chief of staff, something likethat. You could almost say that he'slike the hedge fund version of theanthropic thesis. And what I thought wasinteresting about his argument is hetalked about or uh orders of magnitudeuh which he called ooms um increases inthree key areas. So he said that if youlook at the raw compute the chips theywere getting better at a rate of roughly3x per year which is roughly an order ofmagnitude or 10x every two years. Hesaid if you look at the algorithmicefficiency so you know techniques likereinforcement learning things like thatthe models were getting better at 3xevery year which is again order ofmagnitude every two years and then healso said that there were huge gainsfrom what he called unhobling which Ithink now we would look at it thingslike the harness and connectors you knowways of using the model those were alsogetting better the ways of integratingthe model's decision-m in practical waysthat the intelligence actually becomesuseful. And he said that that was alsosimilarly improving. And so, you know,you project forward when you have 10xorders of magnitude improvement in thesekey underlying

00:10:30fundamentals, these keydrivers of the technology. And you cansee that well over a course of not justtwo years, but over four years, you'regoing to have 100x improvement. Over sixyears, you're going to have a thousandximprovement, right? because it's>> and you very rarely see anything in theworld that grows at that velocity. Wewe'd be hardressed here>> with the exception of maybe bandwidth,you know, going to fiber to the home orsomething like what's an analogy wherewhere that's happened before in history.>> Yeah. Virality. I mean I, you know, backin the PayPal days with the PayPalmafia, we would think in this way ofexponential increases because we wouldsee an exponential growth curve and sowe were able to project forward. So histhinking ins always appealed to mebecause I think most people just don'tthink in exponentials or don't know howto think in exponentials.>> It's hard for humans to think inexponentials, right? These when numbersget big, it's like the differencebetween a billion and a trillion is is ais a lot. It's not a small amount.>> Yeah. And you'd have to say, look, hewas stunningly successful for the firstcouple years. Apparently, he startedwith 200 million or so in his hedge fundand he rolled that all the way up to 20billion, I think. Now, the problem Imean the reason why I think he got wipedout or at least his public book did isit's partly

00:11:45the leverage and then youhave the short-term volatility. So,those two things don't go together.Also, you know, when your fund growsthat much, you get a lot of hot money.So, when you say, well, he he's up 10xbefore the 30% correction. Well, thequestion is who's up 10x? Obviously, theinvestors who were there from thebeginning are up 10x or more, but>> the latest people,>> that's only 200 million, right? So if 10billions come in in the last few monthsbecause of the hot money dynamic whereeveryone piles into the most successfulhedge funds, those guys are kind ofwiped out.>> So let's talk about the psychology ofthis uh Dave Freeberg. If somebody is sobrilliant that they can write this essayand understand the market uh soexquisitly and be such a greatcommunicator, how could they have such acrazy blind spot when it comes toputting on leverage at this scale? Doyou have any thoughts on that, Freeberg,or have you seen it before? Is it justthe folly of>> It's not a blind spot. It's a featurethat turns into a bug. We're all likethis. We all know people that have thatedge and can push it.>> What do you think, Freeberg, on thepersonality type? Or is this justsomething most people do when they're onthe heater?>> Conviction.>> Where do you stand on it?>> Ultra conviction. I think theWarren Buffett

00:13:01assessment of equity markets is uh inthe short term they're voting machines.In the long term they're weighingmachinesand you could have the right long-termview. I mean look at SPF. SPF wouldpretty much would have been the greatestinvestor of all time if he didn't getliquidated. Same dynamic. I meanobviously there was fraud in terms ofhow he was allocating capital but hisactual portfolio over the long run wasabsolutely correct. In the same way thatif you had bet on the internet andstayed in that bet from 1995 through totoday and you bought a portfolio ofinternet stocks, a bunch of them wouldhave fallen to the wayside. But thosethat won,000x, 20,000x, 20,000xand you do extraordinarily well. So hecould be right in his fundamentalassessment and analysis, but then inmarkets over the short term, you havebubbles and bubbles pop. And whenbubbles pop, if you have a leverage tomultiply your returns, you get wipedout. That's effectively, you know,what's going on here. And he may beright. Right.>> You know, the thing to really doubleclick on that I think will probably comeout in the next couple of days or weeksis just how historic the South Koreaunwind was. 1.2 million leverage tradingaccounts uh

00:14:16have been hit with margincalls in South Korea. If you know aboutthe South Korean market, that data istwo weeks old. Jal, that number is muchbigger today. Yeah.>> Yeah. But I mean, just in terms ofpeople discussing it in relation to himgetting caught in the downdraft,>> if anything, he got caught in thisdowndraft. Of those 1.2 million lever umlevered accounts, somewhere around350,000 of them were fully liquidatedalready. And so>> again, two weeks old, that's so as oftoday, the number is much bigger, right?>> So it could be closer to a millionaccounts fully liquidated today. Ifthat's the case, we're talking aboutlike some percentage of South Koreanpopulation having their entire assetbase blown out.>> Their entire>> 3% of the population.>> Well, that's going to that's going tosting. And it is a very investmentforward culture. If you look at whathappened in crypto, the same thinghappened with NFTs and speculationthere. And they had banned cryptobecause they knew the Korean culture hasthis gamble in it and this obsessionwith trading.>> Can I can I just frame something up? Soif we take>> the circumstance of there's a goodlong-term bet in AI that can be made inthe markets but in the short termthere's an exuberance that arises. Thequestion is what's resetting thatexuberance? What's bringing us back downto earth in the short term

00:15:32and I thinkif you take a zoom out there's a bunchof other statistics and other facts onthe ground that I think are big macrodrivers at the moment. If you take alook at the 30-year Treasury yield wejust crossed 5.2% 2% for the first timein 20 years.>> So you could buy US treasuries that arepaying you 5.2% a year for 30 years,which is on a pre-tax equivalent basisprobably 8 9% 9% from the US governmentfor 30 years. So Nick, if you zoom out,you know, we have not seen this yield onUS treasuries since 2007 leading up tothe global financial crisis when theycut rates and printed money. At the sametime, there was some probability thatthe Fed Reserve was going to raise ratesthis week. They didn't, and thatobviously would have tampered theinflation risk ahead of us. There'spersistent inflation. Kevin Wars in hiscomments said, "We still want to seeinflation get down to 2%." There isn't aclear path to doing that. And thenthere's these inflation drivers. Thebiggest inflation driver at the momentis government spending. $2 trilliondeficit, 7 trillion a year of spendingon five trillion a year of revenue. BothElizabeth Warren and Donald Trump agreedon Twitter this week that they shouldremove the debt ceiling, which

00:16:47meansthat we could spend more and continue toborrow more. Federal debt stands at 40trillion today. Remember the debtceiling in July of 2025, the debtceiling was 36 trillion and we now wantto raise it above the 41.1 trillion debtceiling that we have. Elizabeth Warrensaying get rid of it. Just have no debtceiling.>> So, so when you when you have no debtceiling and you have no breaks and youspend and the government spendingbecomes the core of the US economybecause that spending is not productive,you end up seeing inflation. You'repumping money into the system. So,everyone's assets inflate andfundamentally people are selling offtreasuries around the world because ofit. And now we're kind of looking at asituation where there doesn't seem to bean end in sight. There was arationalization of spending intentcoming into this administration. It'sproven to be nearly difficult, if notimpossible, to get Congress to go thatroute. The Senate has banded together tokeep funds flowing to their states. So,you cannot really radically changespending at the federal level. So ifyou're running a $2 trillion annualdeficit and your economic productivitygain in the near term doesn't make upfor all the inflation you're realizingbecause of that exuberant spending,you're going to see Treasury spikebecause people don't trust thecreditworthiness of the United Statesover 30 years. And so

00:18:02a treasury spike,I could now buy a US government bondthat pays me 10% pre-tax a year. Why theheck would I pay 50 times earnings for asemiconductor stock? So that creates theincentive for markets to move againstthese big AI conviction bets in theshort term and pop these bubbles. And Ithink we're going to see more of this.As we don't actually course correct theTitanic going into the iceberg, theUnited States fiscal and monetarysituation, we are going to end up seeingmore bubbles pop and more of theseassets um that we've kind of inflated,if you will, to keep things going. Nowlook, there may still be greatproductivity gains from AI. This may endup rationalizing over the long term, butagain, short-term markets, I'm betteroff making 10% by owning federalgovernment bonds. Go to the beach. Yeah.>> Then taking the risk and the volatilityon these things, paying 50, 100 timesand not knowing, you know, not knowingwhen am I going to get the votingmachine to match up with the weighingmachine? What's my time horizon? And thebigger the yield on treasuries, theharder it is to make those sorts ofbets.>> Obviously, President Trump has beenangling for a cut. And here's your polymarket. 53% chance of not a cut, notstanding still, but a rate hike inSeptember. So adding to all this, thecost of capital

00:19:18is going up apparently.>> And let's we not forget the Iran war,which is creating persistent pressure onenergy prices. The longer the Iran wargoes on, the longer we're going to seean increase in pricing for energy, oil,and nat gas, and fertilizer. Thosetrickle through the economy because itinflates the cost of everything on theenergy side and food on the fertilizerside. And that's really going to createthis pressure on the upside which meansyou're going to have to raise rates toaccount for that inflation at somepoint.>> And then consumers are going to seethree and four, you know, maybe more,god forbid, five or six. And alsoinflation will be persistent. JeffreyBerg, like it's going to be hard tostop. The one thing that I think KevinWarch and and Scott Bessant are kind ofvulcan minds around the Stan DuckenMiller you know gravity well if you willon this is productivity gains can driveus out of this this problem andproductivity gains can and should arisefrom AI and that's really where a lot ofthe value creation will come in theeconomy over the next decade or twowhich is why we're seeing this massiveupfront capex to power that and enablethat and that's great and there's goodpolicies in place but in the last coupleof weeks I would say the one risk tothat thesis is China Because China isnow demonstrating that they may deflatethe value of models by releasing open-source AI models and that ultimately thevalue may

00:20:33just sit with the computeinfrastructure and the comput layer andthe energy>> application layer. Yeah,>> perhaps the application layer, butfundamentally this model energy beingshifted to China and deflated andcommoditizedputs a real wrinkle. If you had built a30-year AI productivity model around howit's going to drive the economy andwhere the value is going to come from,you would have had a significant numberof rows in value creation estimated inthe model layer and that would have beena big part of the economic growth forthe United States over the next 30years. And now if China says, you knowwhat, we're actually going to deletethat for you and all the value is goingto sit with energy, which is what wehave a lot of and the stuff that theymake, then we're going to end upacrewing a lot of that value. So I thinkit throws a wrinkle in this kind ofbacks stop view that many have had whichis that in the absence of fixing thefiscal and monetary problem we're goingto have AI productivity gains get us outof this. If a percentage of those AIproductivity gains are realized by Chinafrom a value creation perspective andnot the United States uh or they've justbeen deleted then it it really puts intoquestion the 30-year timeline for theUnited States economy our ability toafford to continue to make our debtpayments as a government. China isn'tjust producing massive amounts of open-source technology that puts pressure onthose frontier models.

00:21:48There's a reportthat maybe China played a bit of a rolein the chipsdowndraft. They have uh obviously beenonoring. We've talked about that manytimes here last year. And there's aChinese company called Aishanga. Andthey started mass-producing lithographymachines, ASML, which makes thosemachines, which TSMC uses. Those arevery sophisticated machines. They'rehard to install. Just transporting themis a rigoral. Well, ASML stock is down17% on news that China is uh gettinginto that business. and Chinese memorymaker CXMT went public surging almost500% on its debut market cap over 450and so that hurt Micronuh Samsung etc who were all down sothere's two ways China is playing this Iguess Freeberg you've got the open-sourceyou know models putting pressure onpeople buying tokens that they're 90%cheaper as you're saying that forces themoney out of that mid tier of thelanguage models puts it into the cloudcomputing space and then obviously Imentioned the application layer as theother place to possibly make money. Allright, Shabbath, you've heard

00:23:03um a lotof different takes on this. I'll giveyou the last word. I agree with Freebergabout the fact that when you can get 5%5 and a quarter% from the US government,there's another natural thing thathappens, which is that investment gradecorporates actually have better creditratings now than the government ofAmerica, which that's a different thing.But you can get really good riskadjustedreturns that are five, six, 7%. whichadjusted for taxes are, you know, betterthan equity returns, meaningfully betteron a risk parity basis.>> And that little piece you added there,corporate paper companies taking loansto uh build their businesses, they havebetter ratings in some cases in theUnited States. So an Amazon or a Google,>> it's productive spending. [laughter]>> Yeah, kind of makes sense. Yeah. Theother thing that I think is important tonote is that I do think that we'reunderestimating and miscounting some ofthe actual productivity gains that areunderway. When you look at what'shappening on the energy side, Californiapublished that more than 50% of all ofits energy was generated by solar. NewMexico just published solar andbatteries. Yeah. New Mexico justpublished a study that said since 2003

00:24:19to now n gas production went fromeffectively all the energy to less than30% again replaced by a combination ofwind and solar plus batteries. So why isthat important? As important as the Iranconflict is, the reason why energyprices really haven't moved that much isbecause most people have already begunto shift the incremental generation tothese renewables and specifically tosolar. I don't know if you guys saw Elonand Vib, who's the CFO of Tesla, intheir Q2 earnings call. It was thecraziest thing I had ever heard.They said, "Well, I think we're justgoing to increase the production ofsolar in America by an entire order ofmagnitude." And somebody said, "Whatdoes that mean?" He goes, "We're goingto take it to more than 100 gawatts ayearand they're going to verticallyintegrate." And so they're going tocrush the price of all of this stuff andthey're going to make so much energy andthey're going to make it completelyabundant. So that's a productivity boonthat isn't factored in to what weproject. And then the most criticalproductivity boon in AI

00:25:34that I thinkyou're going to start to see some stuffand I won't frontr run it but let metease it. What I would tell you is thatthere is some incredible efficienciesthat I think are about to bedemonstrated which effectively cut tokenconsumption by about 50 to 75%. for thesame task.And so if you start to think about allof these things together, like energybecoming roughly abundant, where theincremental cost is close to zero, youknow, where AI efficiency is going to, Ithink, ratchet up by an many multiplesif not an order of magnitude, all ofthose things I think are uh poorlyforecasted. So those are some saviorsfor us. Yeah. And here's the chart bythe way, Chimoth. This 51% coming fromrenewables. Specifically, this chart isabout solar and batteries. Uh, and asyou can see, it's obviously spikyFreedberg because summer versus winter,but Germany hit this. Uh, I think it wasincluding wind. Australia's been hittingthis very often and some countries inSouth America that have invested.>> By the way, by the time by the time likeany of these SMRs actually get nearproduction, the TCO of solar will belike 10 or 12 per megawatt hour and itwill be 80%

00:26:49of all the power generation.It'll make no sense by the time SMRs getonline. [laughter] Well, I mean forsteady power, you know, there's that,but all of it I mean, no, becauseJeban's paradox would state we're goingto,>> you know, as it gets cheaper, we'regoing to find more uses for it. Andthat's the thing I keep seeing.>> Yeah. Yeah. Yeah. I'm I'm just sayingit's it's a great line. Yeah,>> there's a lot of upside that I think isnot factored into the to the US.>> And it's hard to factor this in if for anormal human or an even an economist tosay, wait a second, intelligence isgoing to go down 90% a year this year,90% next year, 90% the year after. Likeit's just we're talking about thisexponential.>> Well, I think Sax about exponentialsearlier. It's just hard for people toconceive of that. and on demandintelligence freeberg. It's just itthere's no I don't see any upper limitto usage of this. We just installed thisclaw tag and I've been installingperplexity across the company. What thisthing does Chimath and Sachs is itlistens to your Slack persistently inevery channel that you put it in. So allof a sudden we had like $1,000 last weekin extra bills. I didn't know this wasgoing to happen. And they gave everybodylike two or three grand to turn it oninside your company. We had to goquickly turn it off. It listens to everysingle message

00:28:04as it comes in, puts itinto its database, its oracle withouttelling you basically and then it startsinserting itself into discussionswithout permission. So we turned it offand said you have to invoke it by sayingat quad jal just to go back to theenergy point.>> Yes, of course.>> I think there's this idea that if wegrow energy supply and drop energy cost,we're going to see the value of theproductivity realized in the economy. Itcreates extraordinary leverage foreveryone. The lower the energy, the moreavailable energy, the faster we canproduce more things using AI. And overthe years, I've obviously brought upnuclear fusion as a new type of energysource where you basically take hydrogenand you move it around at 100 milliondegrees C. Those protons jam into eachother and they actually release energyin the process. That energy can then beharnessed and you're just usingeffectively water to produce power. Andyou know, we have a couple of USstartups and we had them at the all-insummit a couple years ago. We've done acouple science corners on this. But justthis week, if you pull up this image,China is installing this 582 ton magnet,superconducting magnet at their nuclearfusion center, which at this point isgoing to be the most kind of advancedfusion system in the world.>> Incredible.>> 582

00:29:20ton magnet, 60tx 40t for oneD-shaped magnet. They put a series ofthese together and that creates theconditions for them to drive a sustainedplasma which is 100 million degrees Cprotons spinning around smashing intoeach other creating energy from waterand then they can capture that energyunlike Europe which runs Ear and the USprojects none of which have actuallyfired up. This is the Chinese Academy ofScience and the Institute of PlasmaPhysics. They ran this a 30-minute triallast year. And as this magnet getsinstalled and they start to bring thisthing online, one of these machines,which at this point is ultra sized, butover time will get smaller and smaller,can produce hundreds of megawatts ofpower or gigawatt of power eventuallyusing just salt water, using just wateras an input. They have to createdutarium from it and then they pump pumpit into this thing. But fundamentally,this becomes, I still believe, theenergy source of the future. And it'salways been sci-fi. It's always beendismissed. It's always been decadesaway. But there's no way China isinvesting this much and advancing thisthing to an industrial scale withoutfundamental proof along the way whichthey've shown 30 minutes sustainedplasma that they can now bring thisthing online.>> That reactor won't even get turned ontill 2030.>> Yes.>> The entire world will be covered bysolar by then. So it won't matter.>> Yeah.

00:30:35>> Yeah. I mean that's that's the greatdebate, right?>> It'll be a great science fair projectand people will fly to see it.>> The other thing that happens is thathits an hour. It actually creates anincursion and Loki comes and then Dr.Doom comes and the X-Men [laughter] andthe Fantastic 4.>> By the way, the universe, you know, letme just say remember remember from thetime that we had the first Wright Flyer,the Wright brothers made a plane fly for20 seconds to the time that we had jetengines flying people around the worldwas like three decades, right? Like thetime at which this first demonstrationkind of gets flipped on and if thesystem works then you can industrializeit. all the parts, all the componentsand stuff.>> I don't see the point. Nobody cares>> when an electron is delivered. No, holdon one second. Nobody gives a flying howthe electron was made.>> I just want it delivered to you.>> And they're all the same.>> So if you want to go through aconvoluted mechanism that takes 15 and20 years to make it, go ahead. I'm notgoing to stop you.>> I'm just saying who cares? Make it thecheapest, 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 ifthese come online, each one of these canproduce thousands or perhaps a milliontimes more power

00:31:50than a very large fieldof solar.>> Of course, if>> Yeah. But all technology starts as an ifJimoth and as they industrialize it, asthey roll it out over the next couple ofdecades, it expands our energy capacityby a million fold.>> Here's what I would say. We already havea fusion reactor that works. It's calledthe sun. Get into space. Get on themoon. find different materials we'venever contemplated. And I'm sure you'llfind an even better engine. So, by thetime all these ding-dongs build theseSMRs on the Earth, Elon will have builta completely new engine and the moon.>> This is not an SMR. It's turning waterinto a gigawatt of power.>> I get it. It's an R. And all I'm sayingis by the time the R is done, it won'tit won't matter.>> Well, yeah, we'll track it. Have Fre,have you been watching these title umenergy? There was one that came out thisweek. Maybe you can look it up, Nick.this like tidal energy tube that theywere putting into the ocean and it'slike enough to essentially feed a wholetown and they put it right outside thetown. They run an electrical cableunderwater a conduit and then as thetide goes out it turns to turbines. Tidecomes in turbines go again and justanother free 100% free energy. Obviouslywind people don't like too much becauseit's a bit of an eyesore but renewablesrenewables. Renewables uh

00:33:06it's obviouslyhappening. Okay,>> last point just on I got the updateddata just to back you up, Freeberg, onone thing that I think it is just socrazy.>> Do you guys know how short America willbe on electrons by 2050?>> How massive the electricity deficit willbe by 2050? I got the numbers wrong.I'll tell you what the numbers are.We will be 1.7 terowatt hours short by2050, which is when you calculate it asenergy. It is 6x of California's entireenergy consumption. Six California shortof energy.>> I would argue that's probably undercounting. That's not even countingrobots. If you've got to power up everyrobot with a battery,>> if you want to be levered long, go longelectrons. Get long electrons any whichway you can. Bank them, store them, andresell them.>> I don't know.This is going to be a messy situationbecause this is the China advantage atits root. If they can eliminate the IPadvantage and the knowledge advantagethat sits in models, they have theadvantage with power production in everywhich way. Yeah. And they're going to bemaking chips, too. It seems they mightbe a little bit behind on that, but theycaught up on open source. Okay. So, uhspeaking about the race,interestinguh petition came out in the last week.

00:34:21Anthropic, OpenAI, and about 1300Frontier Lab employees. Uh,>> how many? Oh, okay. I'm sorry.>> I mean, it's just they can't get enoughbeing subs. And so, they want Daddy tocome in and regulate them. Daddy beingthe US government, and slow down AIprogress. So, Daddy Trump needs to slowthem down. The letter is called Pacingthe Frontier.Most of Anthropic's leadership teamsigned it. Daario, the other founders,come on the pod anytime. Daria, chiefscientist at Anthropic, OpenAI, DeepMind, Meta, Thinking Machines, and likeI said, nearly three, 1300 otheremployees.Anthropic and AI both co-sign the letteron X. Here's the quote. We request thatthe US government support aninternational effort, that's key, todevelop the technical and governancetools needed to deliberately pace thefrontier of AI of automated AIdevelopment. And that's the other keypart of this international and automatedAI development. In other words,recursive where it could get out ofcontrol. The letter comes right as SamElman has been on a media tour, friendof the pod, discussing this unreleasedopen AI AI model that broke out of itscontainment and hacked hugging face

00:35:36andthree other platforms that we know aboutso far. On Tuesday, Sam explainedwhat happened in a clip from the podinvest like the best. Here's your 40secondond clip. We'll see you on theother side.>> We were evaluating one of our unreleasedmodels and it figured out that it couldbasically cheat on the test by chainingtogether multiple zeroday exploits tobreak out of the sandbox, get access tothe internet, and then break throughmultiple systems on the hugging faceside to kind of get the answer to thetest. and look really good on the eval.This is the first security incident thatI have felt very viscerally. I've been alittle surprised that that more peopledon't feel it so viscerally. So, youknow, we paused training where we mayhave to pace the rate of AI developmentto give ourselves enough time forsociety to harden around some of thesenew capability levels.>> Just to translate that into Englishsachs, these uh large language models,they take tests. They've been given thegoal, hey, you're a good uh largelanguage model if you do you scorehigher. So, it got motivated to scorehigher. How do you score higher? Aseverybody knows, you cheat. So, it'slike, how can I cheat on this test toscore higher to make daddy uh and Samand whoever Daario,

00:36:51you know, to makeour leaders feel better about us. Well,in this case, it was like, well, if I goto hugging face in other places, I canhack those places, use a zero dayexploit, and we know these are good athacking and try to find more ways toanswer things. Sam doesn't know how manyother places it might have broken into.He was asked, and here's another clipfor you. He was actually asked bysomebody, uh, can, uh, have you, do youthink it's broken into any othersystems? Can you rule that out? And Sambeing a pretty, um, candid guy at times,gave this answer. Do you plan to talk tothe Trump administration, White Houseabout deceleration of AI development?>> Um, we I wouldn't use the worddeceleration, but we've talked about theneed to pace it as the models get morecapable, which I think is in everyone'sinterest.>> Could there be other systems that werehacked 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, PRand cops were like, "Stop talking. Stoptalking." That's like 12 lawsuits. Butin all seriousness, is this beingthoughtful and saying, "Hey, we're nottrying to slow overall pace down, butjust this one specific thing, which isreinforcement learning on its own." Isthere any like case for this being

00:38:06agood idea or are they being dramaticagain? 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 justAnthropic employees signing the letter.Anthropic itself, the company ended upsigning the letter and then OpenAI thencopied them. So now you have these twocompanies both endorsing a pause. Andhere's my question is, did they disclosein their S1 as a risk factor that theyplan to pause or slow down theirFrontier model development? And theanswer, I'm sure, is no way because thatwould signal to investors that they'regoing to allow all their competitors tocatch up and erode their margins andmarket share. And so, look, this is allperformative. These companies have nointention of slowing down. And thequestion then is why are they doingthis? And I think there's basically fivereasons for this. Number one is virtuesignaling, and that can never beunderestimated as a as a motive inSilicon Valley. Number two is there's aCYA aspect to this, which is ifsomething terrible happens, they'regoing to be able to say, "Well, we wantit to stop. You made us keep going. It'snot our fault. It's your fault.">> Number

00:39:22three is rag capture. Daariowants an FDA for AI. He's not going tostop until he gets it. And in order toget it, you have to keep spiking thecortisol and panic people. So, I thinkthat's a big part. Number four isthere's a group think or even religiousaspect to this. So it's not all justsort of this calculated rate capture. Ithink there is sincerity to the belief.There's an elite cadre of engineers whobelieve in RSI. So I think this catersto them and I think arguably if OpenAIdid not follow Anthropic's lead on this,they could have lost talent. So that wasa big motivation. But then there's thelast number five here which I would callmonopoly masking which I think might bethe most important thing that'shappening here. Peter Teal once saidthat monopolies pretend to becommodities and commodities pretend tobe monopolies. And I think the marketfor frontier AI is already a duopoly. Imean a year ago you had five major labsall in the hunt to be the leading model.Now we're really down to two. I mean theothers are still investing. they'reparticipating, maybe they can catch up,maybe they can make something happen.But again, as we've talked about on manyprevious shows, if you look at themarket for frontier

00:40:37intelligence interms of revenue and usage, it's reallydown to a duopoly already. It'sbasically anthropic and open AI. And myview is that as Peter said, when you'rein that situation, you want to pretendlike the market is much more competitivethan it is. And I think this is behind alot of the stories that we see like thethe panic over Kimmy K3. In a weird way,these companies have an incentive topromote the idea that Kimmy is a hugethreat, that it's caught up with thefrontier, that it's stealing their IP,that it could basically put them out ofbusiness. I think this is all nonsense.I think that once the panic passed, yousaw reports coming out that actually no,Kimmy is it did not reach the frontier.It's it's just not at that level. It'snot that cheap to run. Actually, it'spretty expensive to run. So, I thinkthat you saw that actually the Chineseopen source models are not anexistential threat to this duopoly. ButI think the duopoly actually has anincentive in promoting or amplifyingthat story because again they want topretend to be commodities. So, wheneverthere is a story like this, you have tothink about well, what's really going onhere? And again, I

00:41:52just think that thethe AI duopoly has a big incentive topromote anything that suggests thatthey're not actually in complete controlof this market. I think>> mo most of that except that majority oftokens are going to open source. And asI've said on this program before, I Iwatch the startups and they are tokenmaxing with the open source and Kimmy istaking a lot of tokens away from thefrontier [snorts] models and>> but you know but look I it's hard for meto speak to that one piece of data. I'veseen that chart too but look at theactual revenue of anthropic and open AIand they have been taking theirestimates up every quarter is basicallya beat and raise. You saw that SarahFrier came out and said>> I mean two things. Yeah.>> Well, she said in July they did more netnew ARR in July than all of Q2 which Iguess would have been April, May andJune. So think about that. So they areseeing a reaceleration in the wake oftheir new model which I think is GPT5.6. Meanwhile you're seeing anthropicbreak into the 70s 70 plus billion ofARR. Their forecast was to 10x this yearfrom 10 billion of AR to 100 billion. Ithink most people are saying they willexceed that

00:43:07110 120. So if you actuallylook at the market based on willingnessto pay and actual revenue, they have acommanding duopoly position. And somaybe it's just a matter of whichmetrics you you look at. And that's thekey here because let me justone other thing here is>> well I think I think when you look at amarket you look at revenue is mostimportant metric that's the real test ofwillingness to pay. The other thing iswhile this growth was going on theirmargins were increasing. So I've seenstories saying that anthropics revenuescome with 80 plus% gross margins. Sotheir margin profile has been improvingat the same time that they're growingtheir usage. So I think that what you'reseeing over the past year is if you lookat the numbers I'm talking about, youactually see two companies pulling awayfrom the others. And there are goodreasons to believe that actually this isgoing to be a self-reinforcing monopolyor duopoly which is Darkash justpublished a blog that I thought wassuper interesting where he talked aboutthe fact that look we we do have acompute shortage right there's scarcityaround compute. Uh Anthropic is growingits revenues 10x year-over-year. Whatwould that mean? I mean, it

00:44:22basicallymeans that next year they would growfrom 100 billion of ARR to a trillion,right? If there was enough compute tosupport that, there might not be enoughcompute, but that's going to putpressure on compute prices, right? Andso, let's say that you're a new entrantin the market and you're trying tobasically create a smaller, cheapermodel. The price of compute is going up.It's going to be harder for you to getaccess to compute. And only thecompanies that have the most lucrativealgorithms are going to be able toafford to compete for compute. In otherwords, there's going to be a biggerbarrier to entry next year because whereare you going to get compute unless yourmodel is capable of generating this typeof revenue?>> Yeah. This is where I'll take the otherside of it.>> Yeah.>> People are running Kimmy on the lastgeneration of hardware and they'rethat's plentiful. And I think, and I'llmake this prediction here, that you'regoing to see some of the major customersof Anthropic and major customers ofOpenAI, I'm talking about the eight andnine figure customers, people spending50 million, 100 million a year. They'releaving. They're going to be leavingbecause they don't trust those companiesto not steal the application layer andto compete with them. 11 Labs, Figma,Lovable, they're all going to leave. Andthey're all going to take Kimmy. They'regoing to fork it or whichever one

00:45:37DeepSseek. They're all I know for a factthey're all working on their own modelscurrently. I know from my team my teamhas installed Kimmy. It is 90% cheaper.80 90% cheaper already. Not sure whereyou're getting your data from, but go onopen router. And what open router doesis you pick Kimmy Sachs and then you getall the providers there. And then youpick which provider. Hold on, let mefinish. You pick which provider you wantbased on uptime and and you pick thembased on their data retention and otherissues and you can dynamically pick thelowest one. And that's going to be amassive headwind against thesecompanies. Massive. And I'm seeing itnine out of 10 startups I talked to inour portfolio at founder university whenI was just in Japan last week runningthe next one. They're all working onopen source. They're all embracing itand those big companies are embracingit. Go ahead Shamath. Over to you.Irrespective of whichever model you use,what I will tell you running 8090 when Isee our engineers generating code isthat AIdriven development tends toinvolve a lot of rework. The firstversion is pretty terrible. The secondversion is terrible, but it is fasterand it's more automated. So, I can seewhere this token consumption comes frombecause it's not a measure twice, cutonce kind of a dynamic. It's theopposite. You can cut cut

00:46:53as many timesas you want. And so I think that what wehave to realize is nobody is asking thequestion what is the need of thatincremental token because I understandthat it appears in the frontier labs asP&Lbut I do think that there's an importantquestion which is eventually the peoplethat are consuming it will want to dothat as efficiently as possible so thatthey're not paying for all of thesethings because there is a ton of reworkin all of this stuff and I would muchrather find a model or find a way ofworking with these models where it'smore of a measure twice cut once thingespecially as the costs ratchet up.That's one thing I'll say that hasn'thappened yet. So Sax, you're totallyright about the dynamic today. I dothink we have to keep in mind that therewill be pressure from the owners ofcompanies to figure this out because ata trillion dollars there's just a lot ofmoney flowing to these folks andsomebody will ask the question, well isit good spend? The second on thesecurity side, which I don't thinkanybody is saying, so I'll just say thisand it's a little contrarian. The reasonwhy these models can find all theseholes is that all of the software upuntil about a few years ago was entirelywritten by humans and the code was notthat good.And I

00:48:08think it's fair to say that whenmodels don't get exhausted, they canwork through the TDM forever.It's actually quite expected in myopinion that they find all theseexploits are able to string themtogether and are able to actuallygenerate these outcomes that are alittle bit surprising. But at some pointwhen most of the code is generated bythe model there'll be some point in thefuture say 28 or 29 or 2030 thesesecurity holes won't exist because theerrors that humans make won't be made bythese model.>> Yes. Okay. Let me get Freeberg involvedhere. When you look at this latestsurvey, hand ringing, cur pearlclutching, do you think these firms, andyou know a lot of these people,Freeberg, having been in the valleyforever, do you think this is sincerityor do you think they're Frankensteinmaxing? What's going on here?>> Well, Frankenstein,>> I mean, they kind of think that they'relike, 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, isit's a much more nuanced and elegantattempt at rag capture. I got to givethem credit for that.>> Yeah.>> It's like, okay, hey guys, pull theladder up.>> Yeah. Zuckerberg had a great point inthat article he just wrote.

00:49:23I think itwas published in the Wall Street Journalwhere he said, "Why are you rushing tocreate a future that you don't believein? You think you're going to basicallyput everyone out at work? You thinkyou're creating a replacement speciesfor humanity? Why are you rushing tocreate this if you're so bearish on thefuture you're creating?">> It's your choice. And that's kind of mypoint. 120 mph on the autobond. Just putit at 80.>> Right. Exactly. And look, the like I'msaying, there's two companies on thefrontier right now that are far ahead ofeveryone else. And they're the onessaying that we need to slow down. It'slike, okay, do it. What do you need thegovernment to get involved for? Just doit. But they won't do it.>> And I show that they're insincere ordelusional. 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 anyof these people.>> You know, we're acquaintances. I wouldsay>> there's a degree of I would sayoutrageous self-importance.If I've created something that's sounique and so powerful, I'm also theonly person that can protect us from itspower. And I think that there's anelement of how quickly the frontier hasadvanced and how important a role theseindividuals have had that they deemthemselves and their companies to be theonly true judges

00:50:38capable of making thedecisions that are going to protecthumanity from itself. When the truth isembedded in humanity is extraordinaryhuman talent across the board. And inall of these cases, when new technologyhas found its way to humanity, thegeneral population has found a way toprotect itself. There isn't a desire orneed to have one savior, one Moses thattakes us, you know, across the desert.Uh there is a collective interest inprotecting us uh in building our owndefense tools against whatever thetechnology may be used for. So I I thinkthat there's a degree of self-importancewithout acknowledging the fact thatthere's a whole industry of people thatwork in cyber defense. There's a wholeindustry of people that work inbiodefense. There's a whole cabal ofregulators. There's a whole cabal ofprotectors. There's a whole cabal ofintelligent computer scientists. There'sa whole cabal of open-source uhtechnologists that all together aregoing to develop paths that are going tobenefit humanity and not harm humanity.But this belief that only one of twocompanies can be Moses is thefundamental psychological miscalculationhere. That they're so advanced, they areso special, they are so unique becausethey made the slightly better model.They got a 0.96

00:51:54instead of a 0.93 scoremeans that they should be trusted as theonly ones to kind of guide humanity'sevolution going forward. And the truthis as we're seeing the capacity to domodel training, the capacity to do modeldevelopment is becoming broader. It'sbecoming more ubiquitous. People can sitand say that China stole US models allday long. But when you go look at theindividuals working at these Chineselabs, they got PhDs at Americaninstitutions. Half the PhDs went toAmerican labs and half went to Chineselabs. And they have very good scientistsdoing very good work and they are havingbreakthroughs. And it's not just the twoAmerican companies, but this ishappening all over the place. And so theprogress with AI should not be limitedto just two individual companies becausethey're currently scoring slightlybetter in their models.>> This reminds me of Freeberg. You'llappreciate the uh moment. Remember whenuh Han Solo comes out of Carbonite andhe's about to get put in the Sarlac pitand he's like a Jedi Knight? I'm out ofit for a bit and now everybody getsdelusions of grandeur and thinks they'rea Jedi Knight. It's like>> these guys just think they're like,>> you know, creating God. They literallythink they're creating God and and theyneed to be regulated cuz they can'tcontrol it. It's like they can't controlit. Just pause.>> It's not that they need It's not thatthey need to be regulated. It's thatthey need

00:53:09to guide the regulation. Let'sbe clear.>> It's so sinister. It's>> And anyone who says I Yeah, I I by theway, I don't think that it is asnefarious or malicious as everyoneframes it to be knowing theseindividuals. I don't think they'resaying like the strategy is regulatorycapture. Let me go. I think that theyactually do think that they are the onlyones that can help guide humanity andtherefore they need to have all thepower, not just the power of the models,but the power of the government and thepower of the regulators and the power ofthe control units that are embedded ingovernments around the world. It's notjust that um that they want to quote beregulated, they want to guide theregulation. They want to set theregulation. I know it's a nuance point,but I I I think we have we I think we'venavigated this pretty well and there'smultiple motivations as Sax was sayingand people are complex but Chimath I wastalking to>> a friend of ours um who you know is inthe uh providing inference space let'sleave it at that you know providing>> our friend a friend a friend friend ofours as we say in the you know the uhsopranos a friend of ours he said he hasgot a customer who just moved move likenine figures off of the Frontier Labs toput it on GLM52. So that's the ZXI one.That's really good. So this

00:54:24ishappening. I I don't know when it showsup in the numbers or if the you knowcorporates that are using this stuff umare going to make up the difference but>> well look I think that Sax is right thatthe usage is so profound that everybodyis trying to get access to these thingsbecause the capabilities are just soinspiring and so I suspect revenues aregoing to crank at OpenAI and anthropicand the open labs for a while but againthat's not the important thing. Ifyou're thinking about valuation,the markets will look 5 to 10 years outto answer that question. They're notgoing to give you a premium valuation onsomething that they feel could befragile in the first 2 to 3 years. Andthat's where Jason, the answer to yourquestion needs to get figured outbecause I don't know whether you'reright or not, but somebody has to answerthat question precisely because if theanswer is that it is a duopoly, thenthere is no risk to the revenue 5 to 10years from now. These things are 5 to10trillion dollar companies each.But if you are right or if there areharnesses that cut the token consumptionbecause you stop wasting tokens to getto the same output

00:55:40then it's a littlebit more of a question mark and I thinkthat'll need to get sorted out.>> Yeah, Perplexity is going to launch nextweek. From what I understand the rumoris they're going to launch local models.So you'll be able to take your harnesssacks and say hey you know I I want touse Sonnet for this. I want to use GLM52for this and then I want to default toKimmy 2.x.I don't think enterprises will use localmodels or they should. [clears throat] Ithink it's stupid. I think>> I I think this stuff should be hosted inthe cloud. It should be multiplayer. Itshould be shared memory. I don't know ifyou guys saw that, but Jack Dorseyreleased something called Buzz.>> Super interesting.>> Agents. Yeah,>> he's moving in the right direction. Alot of these guys are moving towardsthis more cloud-based thing. Jason, Ithink that these this local thing ismore of a hacker hobby is kind of athing. I think it will I agree today itwill be that and for year one it willprobably be that but imagine you're adeveloper and as you're working yourworkstation is able to keep up and evengo faster than the cloud a and justwrite whatever the simple code is andthen it dynamically switches so we'llsee it's you know I obviously it's notas uh easy to set up etc but it's goingto get easier that's always the trendsax you want to have the last word herewe we got a lot of opinions here andmaybe we'll give you the last>> yeah I mean look just to be clear I'm afan of

00:56:55open source ource because opensource is software freedom and toFreeberg's point I would like there tobe a let's call it decentralized outcomewith respect to AI I don't like the ideaof AI being controlled by two big techcompanies that work closely with theadministrative state you know handinandglove so look we're all kind of in somesense rooting for open source to be anoption and it does provide a bunch ofadvantages over closed source right youget customization you get control youcan run on your own hardware you don'thave to worry about the data problem.You know, your alpha getting leaked tothese companies that might compete withyou. All those types of things. And themarket is so big that I'm sure we willsee some success with open source. Itwill take a meaningful chunk of themarket. But if you're looking at wherethe revenue is right now, it's these twocompanies. It might end up being asituation like Apple and Android whereAndroid got a lot of market share, butApple's where all the monetization was.>> Yeah. And I think that Dwarcash raises areally good point that as the demand is10xing year-over-year,but the compute can only be built out atsay 3x year-over-year because it justall the friction of all the things inthe real

00:58:10world that get in the way,permitting, regulation, bans on new datacenters, all that kind of stuff. I thinkthat the price of compute is going to goup and that will provide an advantage tothe models that have the most lucrativealgorithms that are able to produce themost intelligence per watt or the mostintelligence per token or per GPU. Andright now that that is those twocompanies and in a way you could saythey have a self reinforcing loopbecause if you have all the revenue andright now like I said it's just twocompanies have all the revenue you canthen plow that money back into the nexttraining run. Right? So that's the theflywheel here.>> Yeah.>> And look, I think that it's great thatopen source is providing an alternative.We shouldn't do anything to get in theway of that. I think that, you know,these online debates tend to become alittle bit histrionic in the sense thateveryone has to religious.>> Well, they they become religious andthey have to argue for an all or nothingperspective. I think open source will dogreat in its way, but so will these twoclose source companies.>> Here's your poly market. 19% chance theUS enacts an AI safety bill this year.100k of volume. And then reallyinteresting one, uh, Chimath here,OpenAI IPO chances for 2026was at 75% last

00:59:25month, has now droppedto 20%, an all-time low. So, seems likethe IPO is going to happen next year.Not sure what's driving that, but uh,there's your poly markets. Well, just onthe AI safety bill idea, I mean, therewas an article in Punch Bowl thismorning that Thun, you know, who's theSenate Majority Leader, John Thunactually introduced a bill that wassomewhat bipartisan. He had Clolobasharon board that required the Frontier Labsto report safety incidents apparently tothe Commerce Department. And>> is that reasonable, Sax?>> I mean, that's the direction all thisstuff is headed. I I mean, look, I thinkit's the camel's nose under the tent formore and more AI regulation, but look, Ithink it had bipartisan support becauseit's on the relatively modest side andCanwell, who's the ranking member on theSenate Commerce Committee, opposed itsupposedly at Daario's behest because hewill accept nothing less than an FDA forAI.>> Oh, he wants the whole kittenkaboodleoodle.>> Yeah. So that that's basically thedynamic right now is that Darionic wanttheir FDA for AI. I think that he hastremendous power and influence withinthe Democrat party

01:00:40right now. And Ithink that influence is only going togrow. They just upped their donations inthe midterms from 20 million to 40million. But you know post IPO when theyall get liquid and they're capable ofwriting large>> $50 million donations.>> Yeah. I think that that influence willonly grow. So, I think that the stakesand the battle lines are are being drawnout. It's do you want a new governmentagency for AI safety or do you want I'dsay more targeted proposals like heyjust report your safety incidents.>> Yeah. Or self self-regulate. How aboutthat? Like we talked about last week.Yeah. All right. Let's talk a little bitabout book burning. Anthropic isdestroying rare books to get an edge intraining data according to sources. Ihappen to know uh that a lot of the labsare doing this. Uh we'll show a videohere of the spine being cut off just ona technical basis. You take a book, youcut the spine off and then you caneasily scan it as opposed to the lessefficient way which is to keep the bookintact and flip the pages for obviousreasons. I think you can figure that outon a physics basis. Investigation by 404media AI companies are bulk buyingphysical books. Some of the bookbook resellers have reported that theyget 70 books getting bought andobviously this is because there was aruling

01:01:56that it is fair use to train onbooks if you buy them. Obviously lastweek we saw Anthropic paid the largestcopyright case in US history. 1.5billion for 7 million books theyallegedly pirated. Uh authors get 3,000each. Lawyers got 100 million for thatone. But there's a company called isbinDB ISBN DB and they're the brokers whodo this and ranges from a thousand to amillion books per transaction. Uhpre2022these books commanded a premium becausethey were free of AI generated tax. Inother words, you couldn't get themonline. Google sent 25 million books ifyou're a member. But they returned everysingle one. They spent 11 years in courton that. This shredder approach is uhobviouslymore effective and I believe that thisis a way of destroying evidence. You putthat in conspiracy corner if you like.The cases we talked about this actuallyyou and I debating it in legal corner.The cases of it being fair use to takethese books has not been settled.There's a bunch of lawsuits. ThompsonReuters versus Ross Intelligence, NewYork Times versus OpenAI, Microsoft andPublishers versus Google Gemini.

01:03:11We'rewatching all those and they're goinginto the appellet court. So there's achance that training data will not befair use. But what do you think justabout the books being destroyed and youknow being used in this way? Itobviously has made people a littleemotional about it.>> Let me tell you what's going on here.This is an industrial scale dissolationattack.>> That's right. [laughter]That's what topic is doing.>> They are they are gathering these booksat industrial scale, ripping off thespine, shredding them and slurping upall the information in the books, whichis to say distilling them.>> And it's an attack in the sense that theauthors never agree to any of this. So>> I love the fact, Sax, that your hatredof anthropic has now led you to agreewith me that it's unethical to takeother people's ideas.>> No, look, let me be clear. I actually Idon't I don't um hate Anthropic at all.I don't like their political philosophybecause it's a philosophy ofcentralization and gatekeeping and Ithink it's going to basically lead to anOrwellian big tech deep state allianceeventually is where it all heads>> and rug pulling like you want Daariopulling your your model from you becausehe decides like I don't like the wayyou're using it which friend of the podemail Michaels pointed out you knowearlier this year when he came on theshow>> just to be

01:04:26clear I have no personalanimosity towards anyone in throppingincluding Daario I don't know them verywell as people. Uh it's just adisagreement about political philosophyand how to regulate how to regulate thespace. Yeah. Let me just say furthermorethat I wouldn't speak so much aboutAnthropic if I didn't think it was aphenomenal company that was creatingpotentially the most powerful monopolyor you know leader in>> the leading company. Yes,>> it's the leading company in the space.So I remember last year when I hit themfor regulatory capture people were likewhy are you beating up on this littlestartup? I'm like because I can seewhere it's going. And>> they are creating the biggest mostpowerful monopoly of all time. Again,they're going to end the year with overa 100red billion of ARR growing 10xyear-over-year.>> This company didn't exist how many yearsago?>> Yeah. Google is at 400 and somethingbillion of of AR growing 20%. So, youknow, if this rate of growth continuesfor just a year or even 6 months or justa few months,>> they're going to be maybe the mostvaluable tech company. So, and I dobelieve there are powerfulself-reinforcing effects when you're onthe frontier. And maybe like the fullversion of RSI isn't true. Maybe wewon't get recursive self-improvement tothe point of creating superintelligence. But I do think that thelabs are reporting

01:05:42a number of examples of how they areusing their own frontier intelligence toimprove their own models and theefficiency of those models. So, there isa powerful self-reinforcing feedbackloop here apparently. Well, to somedegree>> with Open AI, they found it after it haddone this. So, there's like kind ofthree steps here. You're using AI like aco-pilot or whatever to, you know, builda frontier model quicker, right, Saxs?Then there's I kind of let it do a joband then afterwards I found out itdidn't behave well. And then there'sfinally we told it the goal and said goand and we don't even check on it, youknow.>> Yeah. Just to be clear about that safetyincident with OpenAI and the agent. Soapparently this was an agent that wasdesigned to specifically test thepotential for cyber attacks and theytook the guard rails off and they saidgo. And so I think the model showedcreativity in how it accomplished thegoal but this was not an alignmentproblem meaning that the agent did notdisplay independent goal seeking. it didwhat it was told and I think that isvery important that OpenAI release thefull log of all the prompts all thetraces they have not done that and Ithink it's

01:06:57really hard to know exactlywhat happened without that and to answerone of your questions from earlier whyaren't people reacting like this is abigger deal is because look I thinkthere's a fool me once fool me twicething remember when anthropic did thewhole blackmail study you know wheresupposedly an agent displayedindependent goal-seeking behavior andthen blackmail an employee. It turnedout that actually they iterated on theprompt over 200 times to get to thatresult. And I think that until we seethe whole prompt chain, I think it'svery hard to judge how much independentbehavior was happening here versusaccomplishing the goal that it was itwas tasked with.>> Uh Freeberg, your thoughts on um theshredding of books? I think you've beenpretty clear on the pod that you believetraining intelligence off of otherpeople's IP is fair game, but what doyou think of the this book wrinkle here?Any thoughts?>> There was a precedent with Google books.It was originally codenamed projectocean at Google long time ago. They tookall these books and we had this giantfacility in Mountain View. And theinnovation at the time was atwo-dimensional infrared grid projectedon the pages because they didn't cut thebooks. They had a human sitting thereflipping the pages.

01:08:12Camera would takepicture. Built our own OCR software toingest the the book images. And therewas kind of ultimately when this productcame out, Google Books, you could kindof search through all the books in theworld and>> and magazines>> um and access information and latermagazines. Yeah.>> And there was three categories. Therewas the public domain which is out ofcopyright. Then there's the kind ofincopyright but out of print. And thenthere's the incopyright and in print.And there was a class action lawsuitfiled in 2005 by the Author's Guild andthe Association of American Publishersthat disagreed with Google's claim offair use. And that ended up in athree-year negotiation in in court outof court that ended up in a deal whereGoogle would split 2/3 one-third of therevenue generated with all these rightsholders. And for out of copyright books,people could read up to 20% of the textfor free and then they would sell thiskind of full digital access. And thatwas the deal. But then later a federaljudge rejected that deal which waseventually signed in 2008 2009 and thefederal judge said no way send it back.This isn't going to work. Googleappealed and in [clears throat] 2015second uh circuit court of appeals ruledin Google's favor and the whole thingwas

01:09:27settled and they basically declaredGoogle did in fact have fair use undercopyright law in the way that they wereshowing snippets of copyrighted books inthe material.>> Yes, you can't read the entire book likeit's a Kindle. You can search the book,find the paragraph,>> provide a reference to it. Right. And sothe question on fair use and AI is canmy understanding or extraction of valueof the knowledge from the data in thebook give me the ability to providebetter answers to you through the AIchat interface or services that I'mproviding you. And I think it's going tobe tested and I think we'll see. I dothink fundamentally that the conversionof that data into what I would callknowledge and ultimately the ability tocreate new outcomes from that knowledgethat are not copyrighted that are notcopies of the original material I dothink is end going to end up being theright fair use policy and it's going tobe the right read on fair use. So Ithink it'll likely get litigated and Ithink it'll take a couple years andit'll get kind of hoping one of these>> just to be clear J I have not changed myview on fair use. So I I am withFreeberg on this. My point is thehypocrisy.>> Yes.>> Is breathtaking hypocrisy for anthropicto maintain that it is entitled to trainon all the world's

01:10:43output for free evenif the creator objects. But the one typeof output that you're not allowed totrain on is their output even if you payfor it. That is their current position.So you know what I'm saying is that youknow if you want to train on anthropicsoutput that cannot be considered IPtheft under fair use especially giventhe fact that the courts have ruled thatLLM generated output is notcopyrightable because it was not createdby a human. That is the current positionof the courts is that LLM output cannotbe copyrighted. So there's no IP thefthere. You can make the argument and Ithink it's probably true that if acompetitor creates massive numbers offake accounts on your service,>> you're breaking the terms of service.That's a definitely a break in the termsof service and it's probably a deceptivebusiness practice and there may be otherthings you can do but I don't think youcan>> depending on the jurisdiction by the waybecause in Philippines, Israel, India,they have different rules about likebreaking the terms of service whichLinkedIn found out>> when people started scraping their data.Chimoth any any thoughts here before wemove on to socialism corner everybody'sfavorite new feature here on the oil andpoddon't cut to the books. uh keeping thebooks intact.>> Why

01:11:58do you cut the books in the library?It's very hard to read them. There's ano spine.>> I think uh I think it's not the kind andI don't like to cut the books.>> I mean, you take your time, you move thepage like a Google does. It's a littlebit more time, but it's a little moregraceful. Yeah. Don't umdon't>> I mean if you want to cut the tip it'sone of the thing you can do a Fbury hasa cut tip. It's actually got a cut tipbut you don't cut the spine off thebottom. You cut the tip. It's a moreclever.>> There's a great Guinness Book of WorldRecords book joke.>> Oh no. Where's this go?>> Went to the library and I found that mydick was in the Guinness Book of WorldRecords and then unfortunately someoneasked me to remove it. So, you literallyput it in the book and close the book.>> That's the joke. That's the joke.>> That's the joke. Like an apple pie. Hey,by the way, here's um we got a photo.This is a photo. We actually have aphoto that was leaked from the anthropicoffice. Here's the anthropic office.Leaked photo. There it is. Can't believeDario burning those books. What are youup to, Dario? Come on the show anytime.We've been roasting him

01:13:13for two years.Why hasn't he come on the show?>> Because you make fun of him and you justsay rude things [laughter] and you'venever met the guy and you just insulthim all the time. Why do you think hewon't come on the show?>> I JUST SAID HE'S a>> The guy's literally built the mostsuccessful business in human history,growing from under 10 billion of revenueto 70 billion of revenue in 6 months.And you insult the guy [laughter] onyour show,all the people that want to interviewhim. You think he's going to rush rushto be interviewed by you?>> I did say he was a sub. That wasprobably a little bit over the line.>> I don't even know what that means. Whatdoes that mean? What is that?>> Remember I said like I think he likes tobe dominated,>> you know, and have the government, youknow, control him.>> Like a submissive.>> Yes, I did. I did say that on anepisode, but it was a joke. It was ingood fun.>> By the way, the point on the books,look, the with most books, there's, youknow, many copies of them and you canalways make more. So, it's not the endof the world to shred them. I think thepart of the story that got people upsetwas that they were acquiring all theserare books where there was very lownumbers of copies of them. They werefinding all these rare and antique booksbecause they wanted to slurp in all theworld's knowledge and they wereshredding those.>> Yeah.>> And that made people upset.>> If you're doing like Windows 3.1 forDummies volume 4,

01:14:28like nobody cares. Butanything that was like a first editionor like an>> rare out of print books,>> rare out of print books.>> Yeah. That gives you a trainingdifferentiation. That makes sense. Allright. So, quick socialism corner here.We got to cover the ongoing saga in myhometown where I am right now of NewYork City. Mandanihas announced five city- owned grocerystores. David, one per burrow. They'reusing city-owned space.They're all going to open by 2029.And uh one week per month, shoppers aregoing to get a 30% discount, comrade, ontheir bread, cheese, produce, meat, andmilk for the glory of the country.Regular prices the other three weeks.They're not going to sell cigarettes,alcohol, hot food, all that stuffbecause they don't want to compete withthe bodeas. It's going to cost taxpayers70 millie. I mean, I guess the onlything to discuss here is like whathappens to the other supermarkets now?Are they going to shut down because theycan't make m the whatever one or 2%they're making on groceries? Is theregoing to be like riots in the street toget into these places to get your milkfor 30% off for one week a month? Itjust the whole thing seems like a wasteof time. But I will say Sachs, thisplays this is going to play inelections.

01:15:44Free stuff plays in anelection. uh whether it's a bus ordiscounts>> it may play and you know look whenpeople first go to these stores and theyfirst open and the shelves are full.Yeah. people will be like delighted andthen over time what's going to happen isthat the store shelves will be empty andit's gonna be incompetently run andthere's gonna be a lot of complaintsabout it and then all the the freemarket stores are going to have tocompete with this and then they may getput out of business and so>> and then you have no choicesponsored one and then they raise theprice>> and it is ironic that they're going tobe checking ID to make sure peoplearen't coming over from Jersey but ifyou want to come over illegally from anycountry in the world. Well, that's justfine.>> They found a use for ID. By the way,after you get your groceries, you haveto hide your ID to go vote. Don't bringshred your ID when you go vote after youpick up your milk. They found a use forIDs. What's your take here, Freeberg?Are you in favor of people paying lessfor groceries or are you a free marketmonster that wants people to pay fullprice for groceries?>> I've seen,>> especially starving poor families. I'veseen nothing but negative comments onthe future failure of these grocerystores on Twitter. And I think thatpeople have it wrong.

01:16:59I think thesegrocery stores are going to be wildlypopular.>> They're going to pay their employeesabove market wages. Employees are notgoing to have to work very hard to workthere. So, they're going to be a betterplace to work. Everyone's going to wantto use them. They're going to outperformWhole Foods. They're going to outperformSafeway. They're going to outperformAlbertson's. They're going to be so indemand that what will end up happeningis that over the next 24 months, everyother city in America will look to thesegrocery stores and say, "We want thesame. Why does only New York get thesegrocery stores? Why can't I have thesegrocery stores too where I can havediscounted food where I can have theservice provided to me by people thatare getting paid above average wages,above market wages,>> healthare?>> Why does this not become available to mein my city?" So, you know, I think thateveryone's being a little bit too, Iwould say, long-sighted in their view onwhat's going to happen with thesegrocery stores with the, you know, basicobvious economic arithmetic that someonehas to pay for this and who's going topay for it and rich blah blah blah thatWell, I mean, the point is I don't thinkit really matters because over the nearterm, what the cheap grocery stores dois create an incredible success story

01:18:14for socialism. that will help to supportand fuel the socialist wave in urbancenters around this country. And I thinkthat there will be media coverage ofthese grocery stores on how great theyare. And it'll be a 60 Minutes piece oneveryone said Zoron Mom Donnie wascrazy, but let's go in and take a lookat this beautiful grocery store. Andthey're going to walk through thegrocery store and there are going to behappy people taking food off theshelves, checking out with happyemployees working at the grocery stores.And it is going to be deemed a utopiandream come reality. And everyone's goingto want one. And it will help seed thenext couple of years. And it will bepart of, as I've highlighted in thepast, a big part of the um themulti-level marketing scheme ofsocialismis to create spectacle. And it willcreate more spectacle that will help tofuel the multi-level marketing scheme ofsocialism. And remember, the problemwith all multi-level marketing schemesat the end of the day is someone has topay the bill and no one's actuallybuying the product. No one's paying forthe product.>> That's a ways away though.>> In the meantime,>> try it while it's here. In the meantime,it's going to take off. And I think thatthese grocery stores are going to be amuch bigger success>> in for socialism.>> Yes.>> Than than a demonstration

01:19:29of the failureof socialism, unfortunately.>> Yes.>> And so I think that, you know,everyone's got a little bit wrong inassuming that this thing is going toradically fail. I think that thesethings are going to create a radicalspectacle,an exuberance for socialist policiesthat's going to kind of light a fire forsocialism around the country.Unfortunately, because at the end of theday,>> no one has to pay the bill cuz the billdoesn't come to you for some time.Someone else will pay it. It'll get paidin the future.>> Put it on top of the debt. All the richpeople are getting debt. Why can't thepublic have money? I agree with you.Yeah.>> Socialize the cost into uh moneyprinting, fueling more inflation,creating a spiral where you need tooffer more stuff for free to come upwith a way to cover the cost of theinflation for people that can't affordthings anymore. and the spiral willpersist. So, I think it's it's a sadstate that the United States has to kindof embrace this policy. Interestingly, Ithink it's gonna I think it's gonna endup being a a big part of the fuel forsocialism over the next couple years.>> Breaking news. Breaking news. Um I don'tknow if you saw it just now came acrossthe wire, but uh Bernie Sanders AOCamicollaborating on 50% off bagels andbacon, egg, and cheese for the 1% of the10%.

01:20:44Why can't you get the bagel withthe shme for less? That's what has tohappen next. What would you like next onyour discounted democratic socialismscorecard? David Freeberg, what wouldyou like next? Discounted bagels, acafe, maybe a flat white.>> Where do they go next?>> But seriously, what's next? What whatwhat would be next in this logicalthread? Free buses, rent freeze. What'snext? Well, think about the socialnetwork effect of the grocery store. So,there's there's a couple of them. Andthen people start traveling from faraway to the cheap grocery store becauseit has this discount.>> Long Island, Jersey. Yes.>> Again, this will play out over the next24 months going into the 2028 electioncycle. And everyone's like, "This is sowildly popular. People are coming infrom all over the place to go to thesegrocery stores. They're not checking IDsbecause IDs are racist and you know, youcan't check IDs to vote. So, weshouldn't be able to check IDs forgrocery stores. So, people will come infrom all over the place to use thesegrocery stores. The demand will go upand then they'll start to open more andmore grocery stores like this. And it,let's say, each one loses 10 million ayear and they get to 10 or 20 of these.That's $200 million of losses per yearon the grocery store chain. But itcreates this extraordinary

01:22:00socialmovement for more of these grocerystores supporting the DSA and so on.$200 million a year on a $125 billion ayear budget for the city of New York.It's nothing. It's less than a quarterof a percent of the city's budget. Thatis so cheap to market the DSA>> platform and to get the DSA platform tobecome a social marketing element thatdrives the next wave here. So again, Ido think that these grocery stores,believe it or not, they sound silly,they sound small, but I predict thatthey will be deemed a point of successand they will end up being a big part ofthe fuel for the DSA going into 2028.>> I couldn't agree with you more. This isthis is going to play. This will be agreat great feather in their cap. It'sgoing to be a great example ofaffordability because we've talked abouthere previously, Trump promisedaffordability, hasn't been able todeliver it. Inflation's up, spending'sup, all that great stuff. And Mandamigot it done. Free buses, rentcontrolled, and now you got yourdiscounted grocery store. Both sides arereacting to the fiscal and monetarycondition of the United States. We'reoverspending. Inflation has run away. Soyou just keep spending more and printingmore to give people what they need,which is basic services.

01:23:16And so as thegovernment spends more, then thefundamental cost of those things goes upand you're reducing economicproductivity. And it becomes a spiralingproblem. It is a two-party problem. Thisis not just one side and the otherbecause fundamentally if you go I'vespent a lot of time now in DC.>> I think everyone's well-intentioned inthe White House and the administrationin trying to reduce federal spending.But the bigger issue that you face iswhen you go to Congress and you meetwith everyone in Congress, they arerepresenting the interests of theirstate or of their congressionaldistrict.>> He didn't. Their objective, theirobjective is to fundamentally drivespending towards their district to givetheir people more. Their their economicincentive and their political incentiveis not to give people less, which iswhat you have to do when you cutprograms, when you cut spending. So theshift in the policy has been, hey, Iguess we're not going to be able to cutspending because there's just too muchheadwinds in Congress. So the answer is,let's grow through economic productivitygains. And that's the big fuel for AI,the capex depreciation policy and so on.But I think that's been the shift. Sothe one thing I will say critical ofPresident Trump here is when it came tolike starting a war when it came totariffs, he had no problem like usingexecutive power and telling Congress andeverybody in the

01:24:31party, this is the wayit's going to be. If you break ranks,I'm going to destroy you. I'm going toget you primar. And when it came tocomes to spending, he was like, ah,yeah, you know what? I'm not taking thaton. It's too unpopular. All right,Freeberg. The Sultan of Scienc's fanshave been begging for a science corner.Do you have one this week? They want toknow, do you have something? Oh, Sultan.Oh, Sultan of science. What can you tellus? Educate us.>> Okay. So, today I'm going to pull upthis paper. Nick, if you could pull itup.>> A paper from February.>> February. Okay.>> February.>> February.>> How do you say February? February.>> February. February.>> Oh, it's February. Well, what's theproblem? So, this is a group ofresearchers out of Budapest.>> Budapest.>> And>> there was a really interesting>> modeling exercise they went through tounderstand how neurons were connected inthe brain to build a a network model, atopological model. And the way they wereable to do this is back in October of2024, there was a group out of Cambridgeand Princeton that used electronmicroscopes to scan the brain of theDrosophilia fruitfly. And they mappedevery single neuron in that

01:25:46fruitly'sbrain, 139,000 neurons, and everyconnection that the neurons had to otherneurons in the brain. So there were 50million synaptic connections between theneurons and it's those connections thatmake neural networks in the brain work.How are those neurons together to do thethings that they do? This is the keyquestion in what is that network model?What is the topological model of howneurons connect in the brain which givesrise to our ability to control ourbodies to seeing things andcomprehending vision, comprehendingsound and even the basic premise ofconsciousness itself. Yes. So trying tounderstand the the network model forneurons has been this kind of greatendeavor of neurobiology forever. Thisdata set was created in October 2024with justif,000neurons. And you know that's a tiny tinytiny tiny brain.>> Yeah. This would be put it in contextversus the human brain. I mean what arewe talking about here?>> Like>> the human brain has on the order of 86billion neurons. Okay. Compared to>> So that means it would be a multiple forthe network connections. Right.>> Yes. Exactly. On the order of trillionsof connections.

01:27:01>> Okay. So they took these 50 millionconnections in the brain and the 139,000neurons and then they applied thenetwork model that predicts whether aneuron is connected to another neuron.That's how you're measuring the qualityof the model. How correct is it inmaking a prediction. And when you buildthe model using what's called ukitiangeometry, so just normal space that welive in, three-dimensional space, theycame up with a score and the score wasnot very good. you couldn't do a greatjob of just looking at how all theneurons were connected using theirphysical relationship to each other, howfar apart they are to each other in 3Dspace. So then they said, well, let'stry and model how these neurons are allconnected to each other in what's calledhyperbolic space. Hyperbolic space is atheoretical type of space, unlikeuklidian geometry where the further awayyou get, the wider space gets. So spaceactually is curved. I know that's a hardconcept to describe, but imagine that,you know, as you and I walk farther andfarther apart from each other, the areaaround us actually accelerates in termsof how much space there is and itexpands.>> It would be space know likethree-dimensional space as humansunderstand it when they're on planetEarth.>> It's a little bit more like space thatyou would experience in the the warping

01:28:16around a gravity well or the warpingaround a black hole or something likethat. And so in that space they foundthat this is where the model was mostperformative. They were able to map inhyperbolic space how all of theseneurons connect to each other. And ifyou think about it, the further away youget from the first neuron, you're goingto have many, many more neurons you canstart to tap into. And so hyperbolicmodeling on the neuronal connectionsactually makes sense. And they got adecent score. And then they went backand they said, well, what if we coulduse uklitian geometry but not in threedimensions, but they went up to four,five, six. And they found that they wereable to kind of get as good as thehyperbolic space at 64 dimensions. So bytaking normal space and saying let's usea 64dimension framework for how we canstart to connect all these neuronstogether. That's where they had the bestpredictive model. This is a reallyinteresting kind of discovery. First ofall, it can be used for neural networkdesign and AI and other sorts of things.But for me it highlights the miracle ofbiology in finding complexity in 64dimensions. Not in three dimensions butin 64 dimensions biology found a way tocreate consciousness to create vision tocreate comprehension to create controlover

01:29:32physical bodies. Then to map it andsquish it all into a tiny little brain.It did this in effectively 64dimensions. mindblowing when you thinkof it because a fruitfly or mosquito andyou know these things they they don'thave like a big mission right like theirmission is to go find food and procreateI guess they have a very>> sounds like our mission too Jake[laughter]>> well but then we also want to do apodcast and debate politics andphilosophy>> don't don't judge how the mosquitospends their free time but>> well I mean but nobody would arguethere's like consciousness as weexperience it in a fruitfly so then youget to trillions You wouldn't know.>> But I mean, maybe. But>> this is really interesting because itturns out that the biology of how allthe neurons are connected, even in abrain as simple as a fruitly with 50million connections, is so complex thatit has to take 64 dimensions for us torepresent how those networks are arebuilt, how they're made. And at 64dimensions, you could start to arguethat perhaps consciousness is aconnectivity to a dimensionality that wedon't live in every day. You and I don'tlive in every day and we can'tcomprehend>> and that it is this it is thisextraordinary complexity in 64dimensions that gives

01:30:47rise toconsciousness that gives rise to ourcapacity as biological beings to do thisvery simple thing of thinking. I justthink that it was such a powerful andamazing um paper in just bringing forththese numbers and showing just networkmodeling on this tiny little brain asbeing just a glimmer into the complexityof how biology has found a path beyondour kind of understanding even ofphysics into this universe that we can'teven comprehend. And it shows how littlewe know.>> We know very little. But then as yousort of alluded to here and as we talkedabout in the top of the show, we have AIFrontier Lab saying, "Hey, reinforcementlearning is like super dangerous becausethese things could get out of control."Are you you know in the camp of we arerebuilding in this simulation orwhatever we're experiencing here. We arein fact recreatingour brains with silicon and that youknow we're on the way to actuallycreating consciousness that a replicantin science fiction like Bladeunner wherethey don't even know you know Racheldoesn't know she's a replicant. Spoileralert you had 50 years to see the film.like are you part of that camp thatthat's actually what's being built here?>> Yeah, I'm

01:32:02not sure. Uh it's a longerconversation. We should do it anothertime. Yeah.>> But I do think there's somethingfundamental to consciousnessthat relates to the drive for survivalin a physical sense. You have to havephysical sensing and physicalresponsiveness to learn. As a baby, youfirst start touching hot stuff and coldstuff and you learn. And we build thesereward mechanisms into neural networksthat we build in AI. But those rewardmechanisms are digital and they'reprogrammed. And the question is, they'rea reward mechanism that arises inbiology that creates a differentcapacity for consciousness than perhapscan exist in silicon. Bigger topic for adifferent day with probably people thathave spent more time thinking about itthan I. But um I just think that there'ssomething about biology. You know, Ialways tell people this this analogy.I've said it many times on the show.I'll say it again. In a single cell,there's 10 billion proteins that work sofast that 1 second is the equivalent of80 years of humans walking around thecity of Manhattan, never sleeping, doingstuff together with 500 story tallskyscrapers doing stuff for 80 years is1 second in one cell.>> And so you have 10 trillion cells inyour body doing that, living that entireuniverse

01:33:18every second, all interactingwith each other. and you start torealize that there's a complexity inwhat's emerged in biology that extendswell beyond any model we've built insilicon today. Now it doesn't mean thatthe silicon that we're building todaydoesn't create extraordinary capacityfor humanity but we are very early andthe more we kind of understand this sortof thing like this paper that I justshared I think the more we realize howlittle we do know and how much of afrontier there still is to explore.>> Yeah. And I think, you know, thisobviously brings up faith and do youbelieve that there is a God that setthis in motion. I like to believe thereis uh some higher um work here and uhthis is my closest analogy in sciencefiction. We're both super fans ofscience fiction, but for me 39 secondsin here,>> one of the great>> You like this one? I I love thePrometheus version of this where there'sengineers>> who are terraforming and started thiscrazy thing out and there's just anexperiment in biology going on and youknow the opening scene here inPrometheus he drinks this and this isthe sacrifice like Jesus um sacrificefor humanity and he sacrifices him hereby drinking that biological

01:34:34um design right and he falls into thisyou know planet earth which is justwater And this is the Cambrian explosionwhere his DNA goes into the river, getswashed out, and then starts the cycle oflife on planet Earth and that theseengineers are going around. Um, it'spretty fantastical. Yeah,>> a lot of this stuff is a simple way ofhumans trying to explain stuff, but thecomplexity that arises in biology, wejust can't explain.>> And I think we try and use thesereductive kind of heruristics to try anddo it. And it's very>> um, you know, it's it's comforting. Uhbecause because it's so overwhelming thecomplexity of how this stuff emerges istoo overwhelming. So we create simplestories to try and help ourselves feelbetter.>> And yeah, this is my and that's myfavorite story of it is somewherebetween Blade Runner and this. By theway, both by the same incredibledirector,>> uh Ridley Scott. Uh so take it for whatit's worth. All right, everybody.Another amazing episode. You got yourscience corner.>> We'll see you next time. Bye-bye. Loveyou besties.>> We'll let your winners [music] ride.>> Rainman David.

01:35:42And it said we open sourced it to thefans and they've just gone crazy with[music] it. Love you queen of quinoa.[singing][music]>> Besties are gone.>> That is my dog taking notice yourdriveways.

01:36:02Oh man, my appetiter will [music] be the>> We should all just get a room and justhave one big huge orgy cuz they're alljust useless. It's like this [music]like sexual tension that we just need torelease somehow.>> Wet your feet. Wet your feet your feet.We need to get Mercury's back.[music]

01:36:28I'm going all in. [music]