YT YouTube2Text

YouTube2Text → Transcripts → Anthropic IPO at Risk, Meta’s Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment Fails

Anthropic IPO at Risk, Meta’s Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment Fails — Transcript

by All-In Podcast · 17,408 words · 2,605 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00All right everybody, welcome back to
0:01episode 290 of the number one podcast in
0:04the world. It's the all-in podcast. With
0:06me again, it's the core 4. David
0:09Freedberg is here. Chimoth Poly Hapatia
0:12and David Saxs calling in. I am your
0:16host, Jason Calakanis, and we have a
0:19full docket today. I'd be remiss though
0:22if I didn't ask everybody for their
0:24fondest memories of the fifth annual
0:26All-In Summit produced by
0:30Mr. David Friedberg. David, do you have
0:32any um thoughts? Uh how do you rate your
0:35performance as executive producer of the
0:37All-In Summit?
0:39>> You know what we didn't do? What we
0:40forgot to do at the end of day two? We
0:42didn't give a big shout out to our
0:44production team, to Kimber,
0:46>> to Lisa,
0:47>> to Nick. I think we No, Jason. Okay. I
0:49thought we should have brought him out,
0:50but I just want to give a big shout out.
0:52>> Bring him out for the standing up.
0:53>> The team backstage absolutely crushed it
0:56as always this year. And then I had a
0:58great time. Actually, you know, my
0:59favorite part is going to Universal
1:00Studios. You just get to have a good
1:02time. It's great. Just great.
1:03>> I am reliving my childhood. Shabbath
1:06with no trauma.
1:07>> By the way, shout out to our friends at
1:09Dongi. I got one of those coffee
1:11machines that we had at the event. I put
1:13it in my kitchen.
1:14>> You know what I got? I don't know if you
1:16guys saw, but there was a water machine
1:17in the VIP area.
1:19>> Well,
1:20>> that is a zero plastic water machine. It
1:23goes through super filtration.
1:25>> No microplastics in the water. I have
1:27one now in our office. Thank you very
1:28much to them. I appreciate it.
1:31>> Amazing company.
1:32>> Sax, do you have a favorite moment from
1:34the show?
1:34>> Well, I think the highlight of the show,
1:36the thing that we'll remember for a long
1:38time was the president calling in during
1:40Jensen's talk. And I think Jensen's
1:42overall the content and and demeanor
1:45really helped calm down this national
1:47panic over AI. It felt like we were the
1:50center of this national conversation.
1:52And I think it happened very
1:54organically. You remember there's
1:55basically been this SCOP for the last
1:57few weeks leading into the summit to
1:59convince everyone that humanity was
2:01going to go extinct unless we did
2:03exactly what the Democrats want and stop
2:07all AI development. And then you know we
2:08had first Satcha came out and kind of
2:12calmed everyone down and then JSON did
2:13the same thing and the president called
2:15in and did what he did best which is
2:16sees total control of the narrative that
2:20I'm sure these doomer groups have been
2:21carefully planning for weeks and kind of
2:23thwarted that whole operation.
2:25>> Nerds enough. Get back to work. You have
2:27a super intelligent
2:30Trump intelligence is now the term.
2:32We're going to call it Trump
2:32intelligence going forward. But I think
2:34you make a good point. A lot of people
2:35wanted to know Freeberg if this was
2:37planned. Was the President Trump calling
2:41planned Freeberg? T take us behind the
2:43scenes.
2:44>> Basically, Jensen was connecting with
2:47him right backstage and said, "Hey, call
2:49me back." And then he called him while
2:50he was on stage. There was some texting
2:52going on.
2:53>> Yeah, it was completely spontaneous.
2:55>> It was spontaneous. Yes.
2:56>> It was not like a planned thing where
2:57Jensen was like, "Hey, let's do a whole
2:59thing." And so that was pretty amazing
3:00to have this conversation that Jensen
3:04was waiting for from the president to
3:05happen live and clearly this was the
3:09topic desour in DC as well as in the
3:11news and the media and it was amazing to
3:13see it all happen on stage with us and
3:15yeah just to uh be clear I thought it
3:19was a bit and I was like well put him on
3:21speakerphone and Jensen says to me how
3:22do you do that and I just handed him I
3:24always keep an extra mic there in case
3:25somebody's mic goes out just put them on
3:26speaker phone put the microphone and
3:28President Trump the funniest president
3:30in history, funniest politician history.
3:32Did you notice
3:33>> Sachs that he drops a oneliner through a
3:36speaker phone perfectly,
3:39>> perfectly. Jensen makes the greatest
3:40chips. Everybody tries to reverse
3:42engineer them. Impossible. Impossible.
3:44But he can't figure out how to press the
3:45speakerphone button. Okay.
3:48>> I was just like, how does he do it? How
3:50does he do it? It just drops the perfect
3:54line. But Shimoth, in all seriousness,
3:56the nerds were losing their mind last
3:58week and the adults came, daddy came,
4:01Elon came, Jensen came, all the adults
4:04in the room came and they just said,
4:05"Listen, if you have software that's
4:09defective and not ready to put in the
4:12world, maybe don't put it in the world
4:14and make it a little bit, you know, as
4:16you like to say, less brittle." So your
4:19thoughts on where we are in this
4:21national discussion which went from slow
4:23down to maybe
4:25just release non-brittle resilient
4:28software. Well, I think that there was
4:30this
4:32classification that happened last week
4:34between
4:36upand cominging organizations that want
4:39to call themselves labs, but are in fact
4:44companies that have P&Ls and
4:46shareholders and have to abide by the
4:48same rules that everybody else does. And
4:51then the sophisticated companies that
4:54have lived in scrutiny for many decades
4:58and have learned the hard way
5:01the process of making things robust and
5:03the cost of making things not robust. So
5:06what do I mean? On the one side you saw
5:08in many ways it's like curious times
5:10creates curious bedfellows right you
5:12have the cohort is Satya Jensen Sachs
5:17Zach Elon the president Lena Khan we
5:22have product liability language
5:23>> strange fellows indeed
5:25>> we have this product liability
5:27perspective in the United States it's
5:29meant to prevent companies from doing
5:31things that they're not supposed to do
5:34especially if they know that those
5:36things could be harmful
5:37And on the other side, it's folks making
5:39the thing that were being, I think, a
5:42little naive and basically saying,
5:44"We're not a company. We're a lab.
5:46Protect us." And I think the mature
5:48response is, "You should have internal
5:50controls and you should robustly test
5:53things." And what Zuck said earlier and
5:55what Elon has said in the past is we're
5:57just under a lot of scrutiny. And so
5:58when things aren't perfect, we just slow
6:00things down. Elon has done that with
6:02FSD. You said this, Jason. He could have
6:03been at FSD
6:053 or 4 years ago, but he has publicly
6:08stated, "We can't afford it because
6:10every accident that Tesla incurs is
6:13magnified a,000x than what any other
6:16company would incur, and so we just have
6:17to go slower." Zach last week said, "We
6:20were about to roll out Muse, which looks
6:21like one of the most successful product
6:23launches in recent memory, has huge
6:25implications actually, which we should
6:26talk about, and he had to slow roll it."
6:29He and Alexander who I was talking to
6:30last night had to slow roll it for a
6:33couple of months to get it perfect. So
6:35it's all within your control. It's just
6:37what do you choose to do? And uh I think
6:39that was the most interesting
6:41observation. We have to stop calling
6:43these companies labs. They're companies
6:45and we have to be judged on the same
6:48standards including product liability.
6:50>> Yeah. And there was a lot of coverage of
6:52this moment Sachs. Interestingly, some
6:54of the fake news media were like on a
6:58popular podcast and then they play the
7:00clip, but they're just afraid to say the
7:02name of the podcast. Pretty hilarious.
7:05>> I think Jimoth is getting us into the
7:06first topic here already, which is
7:08whether we're going to have individual
7:10responsibility as our northstar or some
7:14sort of collective action. And the thing
7:17that always works the best is individual
7:20responsibility. Whether it's at the
7:21level of individuals or companies, the
7:23relevant decision to make in every case
7:26is should we ship this product? Not can
7:31we get the whole world to agree. And the
7:34problem with talking about these
7:36collectivized approaches is that they
7:38defay accountability away from the
7:42actual decision maker. And I I think
7:44people sort of feel this, you know, like
7:45this week you had Daario and Sam go to
7:49the United Nations and say we need some
7:51sort of global AI governance. And it
7:54kind of reminds me of like the
7:56billionaires who fly to Davos in their
7:57private jets and then rail against
8:00climate change. You know, there's
8:01something fundamentally sort of
8:03unrealistic and kind of out of touch
8:05about it.
8:06>> Hypocritical.
8:06>> Hypocritical. And it's like, guys, look,
8:09what we need is
8:11>> for you guys to make smart decisions
8:13about what to release. If a product is
8:16unsafe, which is to say it's unreliable
8:19or behaves unpredictably, then don't
8:21release it. And stop looking to some
8:24sort of global governance as the answer.
8:27Because if you guys don't get this
8:28decision right, it's not going to matter
8:29what the United Nations does. And by the
8:31time the United Nations agrees on
8:33anything, we're going to be much further
8:34along in this whole whole journey. So,
8:36there's something just very weird about
8:38the way that the principal actors in
8:41this keep talking
8:43>> because it really feels like they're
8:45trying to push off the responsibility.
8:48>> Well, there's a theory going around here
8:50and I think it is worth us discussing
8:51here that the Frontier Labs are covertly
8:56>> companies. They're companies.
8:58>> They're companies. Yes. Let me just
8:59finish the
8:59>> No, no, no, but but your made a really
9:01good point that I want to make sure we
9:02just don't miss. Okay.
9:03>> They are corporations.
9:04>> Yes. Yes. And when you call them labs,
9:06they sound like nonprofit entities, and
9:07it's not true.
9:08>> Okay. So, the corporations making these
9:11products are looking for President Trump
9:15to specifically give them what the
9:18internet companies had with section 230,
9:21which is shield them from liability. and
9:23that there is a consideration inside the
9:26Trump administration, the 47th
9:28administration, that maybe there could
9:30be some leverage put on these companies
9:33that they would give 10% of their equity
9:35to the uh sovereign wealth front of the
9:37United States, etc., and then that would
9:40result in them being given some pass
9:44when it comes to their liability. That's
9:47an insane thing to do. It's like giving
9:50some plane company that gives 10% of
9:52their equity to the US government the
9:54ability to not be responsible for their
9:57planes falling out of the sky. But what
9:59is your thought on that conspiracy
10:01theory? I'm sure you've heard it sachs
10:04that they're looking for. You're you're
10:06making it sound like this is the idea of
10:07the Trump administration. That's not
10:09true.
10:09>> No, not the Trump administration. This
10:11is what the
10:12>> Frontier Lab/Corporations are pitching
10:15to them and that there might be the
10:17rumor is there might be some people who
10:18are open to that in the administration.
10:20Again,
10:21>> this is all a back channel.
10:23>> Here's what we know is that in the past
10:25week, you had virtually every Trump
10:27administration official come out and say
10:29some version of what I just said, which
10:31is first and foremost, the AI companies
10:34have to take responsibility for their
10:35own products and that we will not wave
10:38product liability. I think Speaker
10:40Johnson said it, the congressional
10:42>> said it,
10:42>> Besson said it in the most recent
10:43interview. Everybody said if your game
10:46is to get a waiver of product liability
10:49or antitrust liability, if you're trying
10:52to form a cartel, if you're trying to
10:54collude, forget about it. We're not
10:55going to do it. So, I think they
10:57completely threw cold water on that
10:58whole idea. And remember what President
11:00Trump said is he just tweeted today, we
11:02have the DOJ. That's a guardrail. The
11:05administration is a guardrail. the
11:07administrative state, civil lawsuits,
11:09criminal lawsuits, all these things are
11:12there as potential liability to make
11:15sure that you make the right riskreward
11:18tradeoff.
11:19>> And again, at the end of the day, if
11:21those guys release an unsafe product,
11:23it's not going to matter what the United
11:25Nations agrees on. So, we need them to
11:27internalize the responsibility.
11:28>> Are they Listen, come on. Be candid with
11:30me. You you you got the inside goods
11:32here. Are they trying to lobby for that?
11:34Are they do you think they are lobbying
11:36to get this protection? Yes or no? Come
11:37on.
11:38>> Well, I I have heard this is not like
11:40first-p person experience. They they
11:41have not asked me like, "Hey, we want
11:43product liability waiver, but I have
11:45heard from people in Washington that
11:47they have been seeking something like
11:48that.
11:48>> It's hard to pin down, right?"
11:50>> Yeah. Okay, fair enough. Freeberg, your
11:53thoughts here as we are going to get
11:54into the first story here with the token
11:55maxing and the prices going down, but
11:58I'll give you the final word here on
12:00what America should do in terms of
12:02dealing with the corporations that are
12:05pretending to be labs who cannot take
12:07ownership of their own Frankenstein uh
12:10creations, which I think JD Vance, you
12:13know, I thought had a spectacular
12:15appearance. He was crisp and and looked
12:17quite presidential if I if I was going
12:19to um describe it.
12:22you threw a lot of fast balls
12:24>> and he knocked him out of the park. By
12:25the way, fast, you know, home run
12:27hitters want fast balls. You can't hit a
12:28fast ball if it's out of the strike
12:29zone. So, I gave him the the three
12:31hardest questions I could think of and
12:32he
12:33>> proceeded to knock all three over the
12:34fence. I think there was
12:36>> Jad's excellent listen.
12:38>> I mean, no, on the Sunday shows,
12:40>> they played the clips of the questions I
12:41asked him. Bang, bang, bang. Freeberg
12:44right over the f the fence. But, you
12:45know, he said basically, if you're
12:47creating Frankenstein, here's an idea.
12:49Stop.
12:51And and and then he added, if you've
12:53already kind of let it the cat out of
12:55the bag, then create anti-Frankenstein.
12:57Basically, create the safeguards.
12:59>> Yeah,
12:59>> it makes total sense.
13:00>> I think Suck gave a good interview on
13:02this.
13:03>> They took their time doing this Muse
13:05release this week. They said they wanted
13:08to get it right. They wanted to make
13:10sure that it was safe, that it didn't
13:12give people capabilities to do bad
13:14things. That's the responsibility that
13:16they took as a corporation. Anyone
13:19that's outputting a model should take
13:21the same responsibility.
13:23>> Pretty obvious.
13:24>> This is an organization that understands
13:26it. Having paid a series of fines from
13:30Cambridge Analytica to recently
13:33>> we have a lot of we have a lot of laws
13:35that make it illegal to do a lot of
13:36things.
13:38>> All right, let's move on here. We have
13:39laws.
13:40>> By the way, can I add like one
13:42philosophical point to this? So I I
13:45think that implicit
13:47in let's call it the Daario position
13:50that let's call it the regulated free
13:52market that we have his fundamental
13:55contention is that that will lead to a
13:58race to the bottom that competition
14:01jeopardizes safety. It's in all of his
14:03blog posts and many other people reflect
14:05this idea that competition equals lack
14:08of safety and that is implicitly a
14:11critique of our economic system. It's
14:14basically a leftwing critique of our
14:16system and it goes back a long way. I
14:18remember back in the days of the cold
14:21war when we really had this duel between
14:24two different systems. Do you want the
14:25sort of the free market system or do you
14:27want the communist system? A lot of
14:29people made the claim that the
14:31capitalist system would do things like
14:34jeopardize safety and beauty, that our
14:37buildings would always be uglier because
14:39the greedy capitalist wouldn't want to
14:40spend a dime on beautifification or good
14:43architecture, things like that. In fact,
14:45it turned out that the Soviet system was
14:48the one that produced the cold gray
14:50landscapes. Why? Because people want
14:53beauty. They want safety. And therefore,
14:55if you have a well functioning market
14:56economy, then it gives people what they
14:59want. And obviously, the American system
15:01won out in that battle. So, I just
15:05fundamentally disagree with the idea
15:07that competition means lack of safety.
15:12As Zuckerberg and Jensen were saying,
15:15there are huge incentives on both the
15:17upside and downside to be safe. On the
15:20upside, no customer wants a product that
15:23is unpredictable and unreliable, that
15:25doesn't do what they say. No enterprise
15:28wants an agent that's going to leak
15:30their data or hack a competitor. And on
15:32the downside, as the president made
15:33clear, they've got tons of liability,
15:36product liability, civil liability,
15:37administrative liability, criminal
15:40liability. So, I just firmly disagree
15:42with this idea that competition is a bad
15:44thing. Competition is a good thing that
15:46can be harnessed to get the right
15:47result. And just to build on that,
15:49Shimoth our uh friend Nash from PaloAlto
15:53Networks, friend of the pod, he released
15:56in the past week or so unit 42. He's
15:58building there's a business opportunity
16:00here to build a product to do continuous
16:04cyber defense and also a friend of the
16:07George Kirch from um Crowdstrike. He
16:10also released Falcon, his new product to
16:13do cyber defense. Chimath, if these guys
16:16in the Frontier Labscorporations were so
16:19concerned
16:20about cyber security, why didn't they
16:22ever release a cyber security product
16:24and PaloAlto Networks and Crowd Strike
16:27are on it like they will?
16:30That's probably more in the second topic
16:32that we'll talk about, which is as the
16:34models essentially cluster in terms of
16:37capability, there's no differentiation,
16:39and so they'll be forced to move up the
16:40stack. But let's let's wait on that for
16:42one second. You have a point here you
16:44want to make and then we'll go to the
16:45first.
16:45>> I do. The point I wanted to make is more
16:48philosophical and sociological which is
16:53the last time we heard the word lab in
16:56the social consciousness was during co
16:59and there was a lab that was responsible
17:03for a leak of a virus that then killed
17:0715 million people all around the world
17:08and caused $45 trillion of damage. It's
17:12the Wuhan lab.
17:15And I think it's important to recognize
17:17that there has been zero transparency or
17:19culpability around this. And I think
17:22part of why people are comfortable
17:23calling themselves a lab is because of
17:25that fact. That a lot of these young
17:28people that work at these places live
17:29through this and saw absolutely no um
17:33responsibility taken by the lab that
17:37caused this leak. And now we're talking
17:39about leaks of different kinds. And we
17:41have organizations and corporations that
17:43try to fashion themselves as labs as
17:45well, but they're not labs. These are
17:47corporations. I just want to remind
17:48everybody, these are forprofit
17:50corporations that have absorbed hundreds
17:53of billions of dollars of debt in
17:55equity,
17:56>> that have thousands and thousands of
17:58shareholders, that have trillions of
18:00dollars of market cap. And if I were
18:03them, I would not want to call myself a
18:06lab at any cost. And I think if you saw
18:09the recent speeches that they gave at
18:12the United Nations, they're still using
18:14that terminology.
18:16>> Yeah.
18:16>> They are not labs. These are for-profit
18:18corporations that are subject to product
18:20liability risk and they should act
18:21accordingly.
18:22>> Yeah. I mean, and we've had a
18:23conversation here for a year about this.
18:25Like if this stuff is so dangerous, you
18:28don't have to release it or you know my
18:29proposal like know your customer, give
18:31logs.
18:31>> Do not compare yourself to the Wuhan
18:35lab. It's well, Shimoth, it's bad PR,
18:37but also I do I do think it's part of
18:39their like virtue signaling routine
18:41where they just don't want to say that
18:44we're companies. We have a for-profit
18:45incentive. They always want to say that
18:47we're either a research institution or
18:49they say things like we're a public
18:51benefit corporation, which no one really
18:53knows what the hell that means. So, it's
18:54this patina virtue signaling that
18:56obscures what's really going on.
18:59>> It tries to obscure the word
19:00corporation, but legally you're not
19:01allowed to hide it. I think they should
19:03have uh when they
19:04>> we are all running corporations all in
19:07all be it very all be it very small but
19:10mighty is a corporation I think we
19:12should call
19:12>> is a all in lab
19:14>> this is the all-in lab where we discuss
19:16different formulas
19:18>> no we're experimenting with ideas this
19:20is a this is a meme workshop
19:22>> lab these ideas are getting so dangerous
19:25that we don't want any liability for our
19:28opinions welcome to the allin lab
19:30podcast
19:31>> here we Topic zero freeberg. Is there a
19:34mindset in scientists minds that you
19:38know needs to kind of have a hard stop
19:40and then the corporation kind of takes
19:42over? Maybe they could put the labs on
19:44the side here and then say, "Hey, these
19:45are the adults doing the corporation
19:47stuff." What are your thoughts here if
19:48you were going to architect how
19:51anthropic operates in the world? And and
19:53maybe you could speak to scientists
19:55views of the world versus capitalists
19:57and and management teams.
19:59>> I don't have a good answer for you on
20:01that.
20:01>> Okay. Well, you're doing it. You're the
20:02answer. Look at Oh,
20:03>> let me kick off the first topic. What I
20:05wanted to do was just take a step back
20:07and look at the insane flurry of model
20:11releases happening.
20:12>> Yes.
20:12>> I think the pace of model releases
20:14happening on both the open-source side
20:17and the premium side, the anthropics,
20:20gro
20:21of the world is unfathomable right now.
20:24The list is insane. And just like the
20:26last 10 days, September 9th, Deepseek
20:294.1 Flash gets released.
20:3220 for a million tokens output if
20:34they're hosting it. Key Value Cash from
20:37V1 of DeepS, remember what a big deal
20:38DeepS was when it came out. The key
20:40value cache then was like 390,000 bytes.
20:43Now it's 890.
20:45Insane drop, insane efficiency
20:47improvements. On September 20th, Quen,
20:52which is the Alibaba model, comes out
20:54version, you know, 2.1. It's an open
20:57weights model that outperforms Nano
21:00Banana 2, which is Google's big
21:02breakthrough, insane image generation
21:04model. This thing is free, open- source.
21:07You can put it on your desktop computer
21:09and run it at home to generate images
21:12for free. And it's made by Alibaba.
21:15Xiaomé put out MIMO on September 22nd,
21:18two days ago. Mimo Pro performs on par
21:22with CLA's Opus 5, which is like this
21:24insane model that everyone thought was
21:26going to change the world and break the
21:27internet and GPT56 Soul across most
21:30benchmarks. It's a 309 billion parameter
21:32model. It is entirely open. You can
21:36download this model and put it on your
21:38server and run your own instance without
21:41paying anyone any money except for the
21:43cost of running the server. This model
21:44called uh Bonsai 2 came out on September
21:4717th by Prism ML. This is actually a
21:50fork of Quen, which is another one of
21:52the Alibaba models. It's a 27 billion
21:54parameter model. 98% of the performance
21:57of the big Quen model. It's 5.9 gigs in
22:02size and you can run it on a local
22:04computer. You can have an Nvidia chip or
22:07have a good Mac Studio computer and you
22:09can run this insane model on your
22:12desktop computer for free. And then the
22:15big crazy models, the big closed source
22:17models, Anthropic releases Opus 55 on
22:19September 22nd. Open AAI Astro comes out
22:22with Soul and Luna on September 23rd.
22:25Gro 47 comes out on September 21. Meta
22:28Muse comes out on September 22nd.
22:31Any one of these stories would have
22:33broken the internet a year ago and they
22:36all happened in the last 10 days. You
22:38now have an entire catalog of openweight
22:41models to generate visual information to
22:43do VLA where you can actually instruct
22:45it and it can control a robot. You have
22:48models far beyond the most advanced LLM
22:50from a year ago that are openweight that
22:52you can install and run on your desktop
22:53computer. We are past the point of
22:56Bernie Sanders put the super
22:58intelligence away. It can never be
22:59allowed in the real world. This stuff is
23:01out. The open- source openweight models
23:04are free. They are on the internet. You
23:06can go and download them and you can run
23:08them on your computer at home. I don't
23:10know in what world we think we're going
23:11to have a police state that's going to
23:13come bang down your door, take your
23:15desktop computer from you and test it to
23:18see if you're running an LLM on your
23:19local computer. The efficiency wave is
23:22here. And the pacing at which these
23:24models are coming out now. All of this
23:26stuff happened in the last 10 days. It's
23:28just mind-blowing. We're now in this
23:30kind of frontier space where things are
23:33accelerating. The performance
23:35improvements, the cost decline, AI is
23:38becoming so performative and so
23:40affordable. It is going to be ubiquitous
23:42and we are all going to benefit from it.
23:44This is not just about the anthropic and
23:45open AI with the closed world making
23:48billions of dollars selling enterprises
23:50hosted models. There is so much
23:52proliferation right now. So I just
23:54thought it was worth taking a step back
23:56to recognize where we are because in the
23:57context of we have to stop the super
23:59intelligence. We have to shut down the
24:01data centers. You don't even need data
24:03centers to do 90% of what you can do
24:04with AI. You can download it on your
24:06computer for free today to do what was
24:09the most advanced technology in human
24:11history less than a year ago. It's all
24:13here today. And I thought it was just
24:15worth taking a beat to step back and
24:16make sure we all recognize that as we
24:18get into this debate about these quote
24:20labs or corporations. I don't know if it
24:21even matters anymore. I think open
24:24source is here and I think that the pace
24:25of improvement is such that this is only
24:28going to drive productivity. Yeah.
24:30>> And maybe we are maybe we are seeing the
24:33AI wave underway right now. The
24:34productivity wave, the productivity
24:36boom.
24:36>> Yeah. And and to consumers, they don't
24:39care what model they're using. We've
24:40spent the last three years geeking out
24:42about this. And then Grockbot comes out
24:45and then Muse comes out from Meta. And
24:47I've been using both of these products.
24:48And it's the first time that a normie, a
24:51normal person in the world, it's not
24:53meant to be derogatory in any way, can
24:55get value from AI. If you look at how we
24:58all looked at this, running
25:00corporations, investing in them, we took
25:02a different approach to like, okay,
25:03maybe this is going to retire, you know,
25:06some positions, maybe this is um going
25:08to add some positions here or there, but
25:10now you can like with Muse connect to
25:13everything instantly and actually get
25:16real work done in the world if you're a
25:17normal person who wants to get the
25:19benefit from AI. We will be turning a
25:22corner, Freedber, and I think you laid
25:23it out perfectly. We're going to turn a
25:25corner, I think, at the end of this
25:26year, where moms, our uncles, people are
25:30going to start using Muse and Bot,
25:32Grockbot, which are growing very quickly
25:35to just solve real world problems for
25:36that. And they're going to go, "Wait a
25:37second. I could never afford a chief of
25:39a staff. I could never afford a personal
25:41assistant, an executive assistant. Now I
25:44get a free executive assistant." It's
25:45going to blow people's minds, Shimoth,
25:47in the next couple of months when they
25:49start using these products. and the
25:52bottom half of America in this K-shaped
25:55recovery experiences some benefits. Your
25:58thoughts on what's happening here with
26:00these models? Have you started playing
26:02with it? Because you were, you know,
26:03originally a technologist. Have you
26:05loaded these models on your computer?
26:07Because I'm playing with it on my
26:08MacBook Pro and it's awesome.
26:10>> Let me maybe give you the answer by
26:12explaining something. So
26:15I think what we're seeing is that these
26:16models are converging and clustering
26:19which is to say that they're now within
26:21margin of error. And so the thing is
26:24there's so many of these models that are
26:26roughly the same at different price
26:27points. So where is the edge? There's
26:30still edge and the edge is in the
26:33harness that you use to wrap the model.
26:35So if you think about it as like you
26:38have a model that's like a brain but you
26:41need to embody that brain with arms and
26:43legs and hands and eyes to do work.
26:45That's how you identify a model. Okay.
26:47And you make it aic that's what that
26:49phrase means. And there are all kinds of
26:51different harnesses that the industry
26:52has invented to make these models more
26:55agentic. And there's still a lot of edge
26:57there. So at 8090 we have a labs
27:00function and we've been ripping and
27:02teasing apart all the different models
27:04with all the different harnesses just to
27:05see how they perform and they are wildly
27:08variant in terms of cost and quality. So
27:12just to kind of build on Freebrook's
27:13point you have the models that are
27:14converging the harnesses are wildly
27:17different and so their performance is
27:19all over the map and so now companies
27:22and people have to sort of navigate what
27:24are you willing to pay for. I sent this
27:26tweet to Nick, if you can just put it up
27:28here, and I don't know if this is
27:30instructive or accurate, but what this
27:32shows, and this is the case for Ant, but
27:34it's probably the case for a lot of the
27:35closed models, is you're starting to see
27:38this massive revenue concentration,
27:40which is a few folks consume all of the
27:42highest and most expensive tokens. And I
27:46think if you take what Freeberg just
27:48showed to its logical conclusion, those
27:50folks at the top of this
27:53table are going to be under a lot of
27:57internal pressure whether it's from
27:59their finance teams or whether it's from
28:01their CFO or it's going to be from
28:03shareholders to explain
28:06why as this clustering happens they are
28:10not moving down to cheaper and cheaper
28:12models. Now, you may choose to go down
28:15to cheaper and cheaper Ant models, so
28:16then Ant keeps the revenue, but there's
28:19also risk that you may do what Freeberg
28:21just said and just go to an open source
28:22model and just host it yourself.
28:25So, we're in like a very complicated
28:29part of the cycle, which is the first
28:32version of the AI trade was relatively
28:34simple, which is you're just selling
28:35tokens, you're wrapping the tokens,
28:37you're passing it through, and there was
28:39enough disparity that there was value. I
28:42think that that's going away. It'll
28:44force OpenAI and Ant to go up the stack.
28:46There is no choice. You can't sit there
28:49and serve a token that has diminishing
28:51value because then all the subsequent
28:54value is getting absorbed and made by
28:57people wrapping your intelligence token.
28:59You will have to do that work. And so,
29:02Jason, that's why they will have to go
29:03and do cyber. They'll have to go and do
29:05law. They'll have to go and do customer
29:07support. all these things that we
29:09weren't sure whether they were going to
29:10compete. These trends may force them to
29:13do that.
29:14>> Yeah.
29:14>> And I think then that sets up a very
29:17interesting setup for their IPO. Just to
29:20back all that up, Saxs,
29:22>> the Frontier Lab Corporations, Anthropic
29:26and OpenAI, they both released models
29:28this week at 50%
29:30less token prices.
29:33and their IPOs are looking like they're
29:36both going to get delayed. Anthropic
29:39just targeting a $2 trillion valuation
29:41open 1.2 trillion. Sam already said
29:43OpenAI plans on doing 2027
29:46over because of safety concerns and
29:49Anthropic which was supposed to file and
29:52go out in October. The report from the
29:55Wall Street Journal is that the IPO will
29:58happen in November or may be delayed
30:01even more. Here's your poly market. Will
30:03Anthropic go public in 2026 with our
30:07amazing partner Poly Market? Shout out
30:09to my guy Shane. Peaked at 96% earlier
30:13this month that Enthropic would go
30:15public this year, but now it's fallen to
30:1776%.
30:19Is this like a cell phone here, Sachs?
30:21The the leading model companies need to
30:23go public to get capital to do their
30:25buildout and now they're delaying their
30:27own IPOs. Please explain how any of this
30:29makes sense. And if you were a major
30:32investor in one of these companies or on
30:34the board of them, what would you be
30:36doing or instructing these founders to
30:38do to get these companies to go public
30:41and get that capital they need to
30:42continue the AI buildout?
30:44>> Oh, well, I I think that we know from
30:47our friends that the investors in
30:49Anthropic are pulling out their hair
30:50about what's happening right now. Forget
30:52about the fake whistleblowers who were
30:54there for like a few weeks and then
30:55left. You have current leadership in the
30:58company saying that there's a greater
31:01than 10% chance of causing human
31:03extinction and their own product is
31:06unsolved in this area. So think about
31:09the risk factors that's creating. How do
31:11you IPO with your own management of the
31:14company saying that? Then you've got I
31:16think other evidence of I don't know
31:18what to call it except corporate
31:20schizophrenia. You've got Daario putting
31:23out this essay saying we need to pace
31:25the frontier days before they launch
31:28Claude 5.5 and set a new frontier. I
31:31mean, he must know when he's writing
31:34this essay that they're about to release
31:36a product that will extend the frontier.
31:38So, at a minimum, it's hypocritical, but
31:40maybe it's even worse than that. Then
31:42you have them publishing an essay about
31:44the biorisk associated with AI and how
31:47that needs to be stopped. And then
31:48meanwhile, they announced a new biolab
31:51in San Francisco, a new wet lab.
31:54>> They started a lab after telling us that
31:57there's a 10% chance of us all dying.
32:00Sachs, let's just pause here for a
32:02second. Is this a dereliction of duty as
32:06the chief executive officer of Anthropic
32:10to sabotage your own IPO? You're
32:12supposed to be acting on behalf of your
32:14CEO of your shareholders. This to me
32:17seems like if we were the board, if the
32:20four of us were somehow magically the
32:21board of Anthropic, we would replace the
32:23CEO. We would put the CEO in the lab and
32:26we'd find a CEO who knows how to run the
32:29company and not
32:30>> sabotage it. He's literally sabotaging
32:33his own company.
32:35>> I think there's an element of that and I
32:37think investors are pretty nervous about
32:38giving the founders or or Daria
32:40specifically super voting shares.
32:42There's a story in the information about
32:43that that there's an active debate right
32:46now about whether to give them super
32:47voting share.
32:47>> Explain what those are, please, and and
32:49why in this context this is critically
32:51important. Our audience may or may not
32:52know what those are.
32:53>> Supposedly, the founders of Anthropic
32:56only have about 2% ownership each of the
32:59company, which in the grand scheme of
33:00things is a relatively low number
33:03relative to what other founders
33:04typically have at time of IPO. It'll
33:06still be an enormous amount of money,
33:08you know, tens of billions of dollars,
33:10but that is the case. Super voting
33:11shares basically say it separates the
33:14economic shares from the voting shares
33:17in the company and gives certain people
33:18the right to control the company even if
33:21they don't have the economic ownership
33:23that would usually be necessary. It's a
33:25pretty common thing. I mean, Larry and
33:27Sergey, I think maybe were the first
33:28prominent founders to get super voting
33:31shares in Google and Zuck. It can be a
33:33stabilizing thing for the company
33:35because it means that as long as the
33:37founders are there overseeing things,
33:39they're not going to be
33:41>> subject to, you know, like takeover
33:43attempts, things like that. But on the
33:45other hand, it requires a tremendous
33:47degree of trust in those founders
33:49because they effectively can't be
33:50removed no matter what they do. And you
33:53just have to wonder if in this case
33:56that's earned and warranted. I have to
33:58say that even for someone like me who
34:01isn't necessarily advocating for a lot
34:03more regulation, I kind of get nervous
34:04about the idea of anthropic opening a
34:06wet lab in San Francisco. I mean, I was
34:09>> great timing.
34:10>> I was thinking about moving back and I
34:11don't know if I want to be I mean like
34:14can you just put it a little further
34:15away? I think we've all seen this movie
34:16before.
34:17>> Yeah. Offshore maybe. I don't know. Put
34:19it on an oil rig or something. What I
34:21would say is that if I were on the
34:23board, I would keep Daario as a CEO. And
34:26the reason is that he's clearly
34:27demonstrated that he can build a unique
34:29culture and position against the
34:32otherwise most formidable corporation
34:35competitor, which is OpenAI. And he's
34:37done that and he's come from behind. And
34:39I don't think you should
34:41undervalue how hard that is and what
34:44that takes. So he's done that and he's
34:45built he's built essentially the
34:47greatest business ramp in all time. I
34:50think the thing that I would probably do
34:52if I was on the board is fix what Sax
34:55just said, which is there's now a bunch
34:58of incremental liquidity risk to the
35:00company. And I think if you take the
35:0320-year view on anthropic, that
35:06corporation needs hundreds of billions
35:09of dollars to fulfill their ambitions.
35:11And I would solve for getting that as
35:13risklessly as possible. And I think
35:15there's only two things that you can do.
35:17The first is you have to throw
35:19everything in and the kitchen sink into
35:21disclosures.
35:23The problem is that that will make what
35:24is an otherwise dense and impenetrable
35:27S1 even more dense and impenetrable.
35:29Okay? So, you can kind of like
35:31superficially cover your regulatory
35:33risk. But the way that it actually gets
35:36translated in the market is you're just
35:38going to have to water down the
35:39expectations of the IPO sellers and
35:44create a much larger margin of safety
35:46for the IPO buyer. And that's
35:49effectively how you find a market
35:50clearing price for this, Jason, which is
35:52to essentially say, and don't bear with
35:54me, I'm just giving out rough numbers,
35:55right? Even if you think that they're on
35:57a hundred billion run rate, what would
36:00otherwise in a very clean sheet IPO may
36:02be would have been a $2 trillion market
36:04cap, now you know, you're at one or
36:07less. And the reason is the hedge funds
36:12and the pension systems and the long
36:16only mutual fund complexes who have a
36:20fiduciary responsibility to their
36:22sources of capital
36:25will justify all the risks by
36:27essentially asking for a lower and lower
36:29price. And I think that's what you're
36:31going to see. I think you're going to
36:32see a much more moderate and reasonable
36:36IPO motion here.
36:39And it's probably going to clear at a
36:41much lower price than anybody thinks.
36:43And I do think that what Sach says is
36:45the reason why. But I think it's good
36:46for Ant because it clears all this
36:48noise. It forces them to just be a
36:51corporation and then run the business.
36:53And I think that they'll do great. Look,
36:54I'll be honest with you. Like I've tried
36:56to recruit against Ant. I can't. It's
36:59impossible. The comp that they give is
37:01incredible. The business ramp is so
37:03compelling. It's like when you're in a
37:05head
37:05>> when you're in a head-to-head with these
37:07guys. I have not won a single bake off
37:09for folks. I wish I could say that I
37:11did, but I have not. And it's hard for a
37:13startup. So, they are a magnet for
37:16talent. OpenAI is a magnet for talent.
37:19But these other risks are going to weigh
37:21on the IPO and the only way to flush
37:23them is to just bring the price way
37:25down.
37:25>> Yeah. And Freeberg, when we look at the
37:27risk factors in an S1, those are
37:29prefuncter. Those are disclaimers.
37:32Typically, you get the SpaceX one. Oh,
37:34what if Starship doesn't work? What if
37:36there's competition, it's kind of check
37:39boxes in anthropics, I think we're going
37:41to be looking at something totally
37:43different, which is, hey, we and our
37:46employees are quitting saying this is
37:49going to kill humanity. We have massive
37:50regulation. We're asking for regulation.
37:53And by the way, we we're doing biology
37:55work and we're launching a lab. So do
37:57you agree with Chimath here that the S1
37:59is going to have an actual risk factors
38:03and disclaimers that are material
38:06against the IPO free burn? I think this
38:10customer concentration point ties to the
38:13earlier comments I made about the
38:16proliferation of incredibly performant
38:18openweight and open source models. And
38:20the reason is Anthropic puts out the
38:23best models for life sciences today.
38:28I run a life sciences R&D organization.
38:31We use their models. They're incredible.
38:34They're better than other models. And so
38:36we want to use them. When it comes to
38:38writing code, to writing software to do
38:42internal workflows and run operations,
38:45we can use other products to write that
38:47code for us. And we do. And I think that
38:50that's likely the way the market is
38:51going to develop. And I think we're
38:53starting to see that proof come out in
38:55the market. And you're seeing it in this
38:56customer concentration point. Extremely
38:59high value technical engineering, math,
39:03life sciences, those really hard
39:06technical problems. You want to have
39:08these top tier, topshelf models, and
39:11you're going to pay a huge premium for
39:12them. And you're totally going to make
39:14money doing it because the value is
39:16extraordinary. And this has always been
39:17my case about AI. AI is not so much
39:20about the value of replacing old stuff.
39:2399% of the value of AI is about enabling
39:25new stuff that's never been possible in
39:27human history. And I think that's where
39:29anthropic and open AI are going to make
39:31their money. Separately, I think open AI
39:33is going to make a lot of money in
39:34consumer and anthropic maybe to some
39:37extent as well. But in the enterprise
39:38setting, I think the money comes from
39:40creating new frontiers, discovering new
39:43enzymes, which Anthropic announced this
39:45week, that can cure human disease, and
39:48developing new technology, new
39:51engineering systems, making things that
39:53have never been possible to be made
39:54before. And I think that's the premium
39:56value, and it's the premium price you'll
39:59pay for those closed models and that
40:01premium model. But that's a different
40:03business case than taking all of
40:05enterprise. separately. I think most of
40:07the enterprise applications are going to
40:09be run using these openweight models and
40:12code is going to be written with cheaper
40:13models and so on. I don't know why you
40:15would go to a higherend model to write
40:18code when you can do it for free and
40:20everyone in the organization can write
40:21code basically for free. And so I think
40:24anthropic's bigger issue is this
40:26customer concentration question and is
40:28it really a premium product that they're
40:30selling in the market and that's where
40:31they take the market. That would be kind
40:32of where I think the market is headed
40:34right now and would be the biggest risk
40:35factor. This whole thing about oh it's
40:37gonna end the world and whatnot. It's
40:38like dude just make sure your product
40:40doesn't end the world and you'll be
40:41fine. That would be a good first step
40:43here like Mark is doing at Meta.
40:46>> I mean it's an incredible company. The
40:48execution has been extraordinary. The
40:50ramp of revenue sachs has been
40:51extraordinary. But when you start to
40:54look at the impact of open source, we
40:56had a a couple of um viral charts come
41:00out, 70 80% of the token usage are those
41:03dark tokens I mentioned six months ago
41:06here on this pod that are not tracked,
41:08but some people can track them on
41:09routers. Those dark open source tokens
41:12are now the majority of tokens. And
41:15>> can I just give you a stat on that, Jal?
41:16cuz what you just said is really
41:18important, but there's a stat that I
41:19think in in the last 12 weeks alone
41:23>> token use has flipped from 8020 closed
41:27verse open
41:28>> to 8020 open verse closed.
41:31>> I don't think we've ever seen in 12
41:33weeks I don't think we've ever seen a
41:34title wave of chart
41:38>> in the history of all technology
41:40markets. We've never seen anything like
41:41this. And that is the biggest risk
41:43factor for Anthropic is they were first
41:45out the gate to get the enterprise on uh
41:48on AI to build tools and whatnot. But as
41:51this market shifts to open source and
41:52open weights, boom, here it is right
41:55here. Look at this.
41:55>> This is the Versel chart that went viral
41:57this week. So this week 12 weeks, oh my
42:00god,
42:00>> crazy. If you look at this chart, this I
42:02think is what blows the lid off the door
42:05on what needs to go into these S1s. And
42:08that's why I walked through all those
42:10openweight releases at the start of the
42:11show or this segment is it's not just
42:14that they're there, it's that they're
42:15proliferating now. There's so many more
42:17of them coming out and the pace of
42:19improvement is extraordinary and they're
42:20now in VLA for robotics. They're in
42:24image generation for image and video. So
42:26the openweight open source models are
42:28now being used for all of the broader AI
42:31applications. The premium stuff I will
42:33say is still the best for narrow use
42:36cases. Yes,
42:38>> for engineering, for highly technical,
42:40highly complicated tasks and workflows
42:42and problems, solving Navier Stokes,
42:45solving mathematical problems, solving
42:46biology problems, that's where these
42:48models absolutely perform and people
42:51will pay a premium of, I would say,
42:53close to infinity dollars for that
42:55premium application set. So, it doesn't
42:57mean that their market is destroyed.
42:59Yeah,
42:59>> it means that we're seeing a bifurcation
43:01in the market where the majority of use
43:03cases of AI go towards these open source
43:06open tokens.
43:07>> The question and premium models are
43:09still going to proliferate. Yeah.
43:11>> The question then is in the distribution
43:13of tokens in these frontier
43:15corporations, what percentage are those
43:18tasks versus what percentage are
43:21fungeable tasks to open source? And if
43:23the answer is more than 60 or 70%.
43:26You have to discount that revenue. If
43:28the answer is only 10% and 90% of all of
43:31the revenue is being generated on
43:32absolute frontier use cases, they're
43:36going to be fine and they're going to be
43:37off to the races.
43:38>> Yeah. And
43:39>> but otherwise, it's a thing where you're
43:41fighting just people's mis or lack of
43:44education, right? Because as they see
43:46their bills,
43:47>> hand it to Sax here. Yeah. Go finish
43:48your thought and then I'm going to hand
43:49it to Sax.
43:50>> No, just move the move the workloads
43:51where they belong and save the money.
43:54>> Yeah. And so Saxs, um I made a chart
43:56here before the show here, but here's
43:57your um sweet spot for the last 100 days
44:01of models, which um Friber, you did a
44:03good job of teeing up. And what you'll
44:05see is like yes, the Clouds and the
44:07Astros up in that righth hand corner
44:09cost a lot and they deliver a lot. But
44:11then you start going down and you start
44:12to see Muse and GLM and Kimmy and then
44:16on the left you start seeing GLM from ZI
44:19and Mimo and you're starting to see well
44:20the Chinese
44:21>> this is all hosted right JL. I mean,
44:23it's important to note what you're
44:25showing here is the hosted price, but if
44:27you run it yourself, the cost on some of
44:29these is less than 10 cents for a
44:31million tokens.
44:32>> Yeah. So, you'll pick up then the
44:33compute, you pick up the bandwidth, all
44:35that other stuff. But, Sax, the issue
44:37here to Chimat's point is the vertical
44:40language models, the small language
44:42models are going to do 80 or 90% of the
44:45tasks that many corporations need. And
44:47then we get to AI sovereignty. Do you
44:49want your super intelligence? Shout out
44:51President Trump. from intelligence chat
44:54DJT.
44:56We need people to use those models so
44:59they have sovereignty. So you're not
45:01giving your data to the frontier models
45:03so that they don't get that tip of the
45:05spear super intelligence edge which is
45:08why people are disengaging. So your
45:12thoughts on maybe this is going to start
45:14to impact the velocity of clawed and
45:17openi tokens. What does your business
45:19sense tell you David Saxs? Well, look,
45:22wearing my hat as an investor, I'm more
45:25optimistic about the Frontier
45:28Intelligence business than I think you
45:30guys are. I think that Anthropic and
45:32Open AI right now are a stable duopoly
45:36for frontier intelligence because they
45:38have a pretty significant lead over all
45:40these other companies and they are able
45:42to charge a premium for that. I fully
45:44understand that the majority of tokens
45:46will go to open models or commodity
45:48intelligence. However, there is a I
45:52don't know meaningful percentage of the
45:53market. I don't know whether it's 10%,
45:5420, 30, whatever that will pay this huge
45:58premium for true frontier intelligence.
46:01There's a lot of customers that either
46:02need the best or they just want to know
46:04they have the best. You know, they don't
46:05want this DIY approach or they could be
46:08in a very competitive market. Let's say
46:10you're a hedge fund. You can't take the
46:12risk that your competitor has a better
46:14model than you do. So there's a lot of
46:15circumstances I think where they will
46:18continue to monetize the frontier at a
46:20premium and this is why you're seeing
46:21their revenue keep growing even though
46:23they're losing token market share.
46:26Furthermore, you got to remember that
46:27over the next year or so something like
46:3060% of the worldwide compute that's
46:33being added is being added for these two
46:36companies. So to the extent that compute
46:38gets scarce and we're kind of in a
46:41compute crunch, just the fact that
46:42they've made these huge investments,
46:44they're building out so much capacity
46:46and they have those economies of scale
46:48and remember they're bringing down their
46:49own prices very quickly as well. I think
46:53that gives them a pretty big advantage.
46:54But look, these two companies are on a
46:57hamster wheel. You know, the moment
47:00where they stop being frontier, they go
47:01to zero, right? And we know the frontier
47:05is maybe only 6 to 12 months ahead of
47:08the commodity portion of the market
47:10depending on capability. So, you know,
47:13again, if they slip, if they fall off
47:15that hamster wheel for 6 months, then
47:18they're in deep trouble as a company.
47:19And that I think is a central risk
47:20factor that's going to have to be priced
47:22in. And this is where I think that
47:24frankly their government affairs efforts
47:26are miscalculating badly. I can see why
47:29they they're going for regulatory
47:31capture, right? because being on a
47:32hamster wheel gets really tiring. They
47:34probably think, hm, like, let's try and
47:37bake in our moes. Let's try and slow
47:39everyone else down as well. But I think
47:41they've miscalculated because if they
47:43actually get all the regulatory controls
47:46they want, if they create this new
47:47federal department of AI, it's going to
47:49slow them down enough that they will no
47:52longer be frontier. Everyone else will
47:53catch up to them. Especially because a
47:56lot of these open models are being
47:57generated by Chinese companies who are
47:59not subject to our jurisdiction. And the
48:01Chinese have made abundantly clear they
48:02are not slowing down. So think about it.
48:04They are lobbying for things that could
48:07slow them down and knock them off the
48:09frontier in which case their business
48:11goes to zero. I you know it's hard to
48:13understand.
48:14>> They're flapping the competitors in this
48:16race and they're asking hey can you pull
48:18us over and inspect the engine chimoth
48:21and that makes no logical sense on top
48:25of that chimoth. You have very
48:28significant competition not just from
48:30the Chinese but you have Jensen and
48:32Nvidia. You have also Microsoft Satia
48:36at Google and Muse at Zuckerberg's
48:40house. They're all embracing
48:43open- source models and they all have
48:45hosting now.
48:46>> Yeah. Can I just go back to what Sax
48:48said which is a really important and
48:51nuance point which I just want to double
48:52click on. He said, "Imagine you're a
48:55hedge fund and you need the absolute
48:59latest and greatest because you are
49:00worried that one of your competitors has
49:02it." This is a really interesting
49:04question because it's about game theory
49:06and your P&L. So now you're in this game
49:10of chicken with your competitors. You
49:12need to use the best because you're
49:14afraid that they could use the best and
49:16out compete you. But what is the problem
49:19right now? What we know is that you can
49:22consume an inordinate amount of tokens,
49:25right? We jokingly call it token maxing.
49:29That cost is completely not levered or
49:32attached to your revenue. It's just not.
49:36And so all of a sudden, you have this
49:38very weird dynamic where you're like,
49:41okay, well, am I supposed to use this
49:43latest and greatest thing that's 10 to
49:4520 to 30 times more expensive than a
49:47generic thing? that thing isn't tied to
49:50my revenues, it's not like I can pass
49:51that through, right? So, if a hedge fund
49:54only makes end number of dollars per
49:57month of profits, there are a lot of
49:59scenarios where they can become break
50:02even to unprofitable. Now, take a
50:04different example. Let's just say that
50:06you sell a fixed good at a fixed price.
50:09Yet again, I think you have this example
50:11where you could use the latest and
50:12greatest and if you're not able to pass
50:14those costs through, inflate your cost
50:18to absorb the cost of that service,
50:21you're in a very tough spot. So, I'm not
50:23sure, Sax, I agree with this because I
50:25don't see the pricing power of these
50:26companies to both pay for a ton of these
50:30super expensive tokens and also pass
50:32that through to their customers. And if
50:35you can't, they're just absorbing it and
50:36you're going to see it in lower margins.
50:38And to build on that, Saxs, I'm going to
50:40drop it back to you. If you take Jane
50:42Street, very famous uh trading firm,
50:45they've publicly announced 19 billion
50:49dollars in cloud capacity contracts.
50:52Coreweave, 6 billion in in cloud
50:55commitments. They also invested in that
50:56company and then 13 billion with Crusoe.
50:59They are building their own
51:01infrastructure and they're embracing
51:03open source. So that speaks to Chimath
51:05what you're saying. They're are they
51:06hedging their bets? They're going to
51:08spend tens of billions of dollars on
51:10their own compute and they're going to
51:12use the frontier models they seem to
51:14when you do the game theory,
51:17you know, saying, "Hey, let's embrace
51:19our own infrastructure sacks."
51:22>> Well, look, I think when your use cases
51:24become more mature and, you know, you
51:26can get by with cheaper commodity
51:29intelligence, then obviously you'll try
51:32and build that out again if you have the
51:34sophistication level to do it. I just
51:36think it's easy to underestimate how
51:38convenient it is just to use the the
51:40frontier. And again, these companies are
51:43very rapidly bringing down their prices,
51:45too. But anyway, look, we we've talked
51:47about this. I think at the end of the
51:48day that look, there's no question that
51:51opensource is a huge risk factor to
51:55Anthropics S1 and it has to be right up
51:58there with, you know, the risk that
52:00they're going to end humanity or
52:01something like that. But again, I I come
52:04back to the fact that I think the
52:05fundamental problem with this company is
52:07that it's schizophrenic. I mean, I think
52:09they need a psychiatrist, not a banker.
52:12They keep doing things that are
52:15hypocritical and oxymoronic. On the one
52:17hand, they're saying pace of Frontier as
52:19they release a new frontier model. They
52:21say that biorisk can end humanity as
52:24they open a wet lab in San Francisco.
52:26They say well their business depends on
52:29them staying ahead of the commodity
52:33models and they are advocating for a
52:36federal department of AI that will
52:38definitely slow them down perhaps to the
52:40point where they get commoditized. So
52:43you have a fundamental
52:46I'd say like breakdown happening in the
52:49leadership of this company where they
52:51are advocating for things that are
52:54likely not in their interest. Now
52:56there's a lot of people who say it's
52:57just regulatory capture and that is true
52:59but I think that the red capture is the
53:02government affairs people trying to
53:03rationalize the schizophrenia and make
53:05it work to their advantage. I'm not sure
53:08that it fully computes
53:09>> so to speak. Yeah. Let me ask you guys a
53:12question. What do you think would be the
53:15method by which like a ban on super
53:17intelligence? Bernie Sanders proposed
53:19this bill to ban super intelligence and
53:22you know the Dems take the House, they
53:23take the Senate and then maybe in 28
53:26take the White House and you end up in a
53:28world where this sort of thing is so
53:30weirdly political, which it shouldn't
53:31be. I don't understand how we ended up
53:33in a political party divide on progress.
53:36It's
53:36>> they ask for it is the reason they asked
53:38for it.
53:38>> I No, I there's It's it's so awful and
53:42weird that you're anti-progress because
53:45of your political party. The whole thing
53:48is not objective. It's a bunch of
53:50weirdly distorted worldview thinking to
53:53try and create an opponency to the
53:56individuals that you don't like. But
53:58it's very weird. Anyway, my point being
54:01even the Republicans, you're hearing
54:02murmurss about, well, maybe we should
54:04ban data centers, ban AI. So Sax, let me
54:07just ask you, what is the pragmatic path
54:10with all of the open- source openweight
54:12models being out there? Is there even a
54:14pragmatic path to do what everyone's
54:16stating or talking about or Bernie's
54:18talking about, which is to ban super
54:21intelligence, to stop AI, to turn it
54:24off? Is there even a way to do this?
54:27Like what would happen if this law
54:28passed tomorrow? What do you think would
54:31happen is my question. If Bernie's bill
54:33passed tomorrow, everything would stop
54:36because the way he defines artificial
54:39super intelligence in his bill, you
54:41could argue we've already hit those
54:43thresholds. I mean, the way, you know,
54:45it's basically just saying that models
54:47have
54:47>> Yes.
54:48>> at a certain capability level. And I
54:50think it's arguable whether we've
54:51already reached it or not. And so, it's
54:53a very loose definition. And then on top
54:55of that, he's got 20-year prison
54:57sentences for developers who violate the
55:00law. So no one's going to want to take
55:01the chance.
55:02>> So what happened?
55:03>> So think about that chilling effect.
55:04>> Do you think people delete openweight
55:06models off their desktop computers? Do
55:08businesses have to go remove code from
55:10their systems? Do do all these people
55:12lose their jobs?
55:14>> Listen, let me let me tell you that the
55:15Democrats want to do to AI what they did
55:18to crypto. They're going to drive the
55:19whole industry offshore. They're going
55:20to basically drive all the innovators
55:23out of the country.
55:23>> Goes offshore.
55:24>> Goes off.
55:25>> That's literally where I was going with
55:26it is like just move your company to
55:28Singapore. Move it to Zurich. move it to
55:30a geo
55:30>> cuz that's where people are embracing
55:32open source and are embracing super
55:34intelligence.
55:35>> This is literally where my head goes.
55:36I'm like, you can't stop AI. You can't
55:38just tell people a class of software
55:40can't exist. In my head, I'm like, okay,
55:42if the US passed a law that said this
55:44class of software cannot exist. I would
55:48move to another country. Other people
55:50would like like isn't that
55:52>> give up their citizenship and move to
55:53Singapore? That's the answer. people
55:55literally you would be taking.
55:57>> Did you guys see that Xi Xi spoke at the
55:59White House today and he said, "I'm
56:00going to invite a 100,000 young
56:02Americans to come and join China."
56:04>> Yeah. Come and see what we're doing
56:05here.
56:06>> Like it's it's a very interesting moment
56:09where our people are saying, "Let's stop
56:12progress, technological, economic
56:14prosperity, the opportunity for everyone
56:17in that K-shaped economy to lift
56:20themselves up because open source AI
56:21enables everyone to increase
56:23productivity. It is not AI for the
56:25billionaires. It is AI for the
56:26everybody. And then we're gonna turn
56:28that off. And then meanwhile, you've got
56:31China saying, "Come over here. Come and
56:32see what it's like over here where
56:34everyone gets to use AI and everyone
56:35gets our playbook." Freeberg brilliantly
56:37stated, "It was our playbook to get the
56:39top Chinese scientists, the top European
56:42scientists, the top scientists and
56:43thinkers and mathematicians in India, in
56:46Germany, and invite them to come to the
56:48United States." That was our playbook.
56:49China's taking our playbook. And then
56:51they're going to try to get investment
56:52will be the obviously the next shoot to
56:54drop here is President Trump. 30 seconds
56:58at the United Nations renaming AI to
57:01super intelligence.
57:02>> The United States also totally rejects
57:05any attempt to construct a globalist
57:07scheme to control for the artificial
57:10intelligence being spoken of so much now
57:14here and after officially called super
57:18intelligence changing the name. The very
57:20same people who said we'll all be dead
57:22in 12 years because of global warming.
57:24These are the same people that are now
57:26saying that AI is going to kill us all,
57:28that robots are going to attack us.
57:30Well, he had a really good line. We we
57:32cut that off too short. He said that
57:35we're not going to like basically confer
57:38regulatory authority on some globalist
57:40uh institution.
57:42>> Well, I mean, it's a good framing,
57:43right? Globalist and global warming is a
57:46way to look at the damage that did. The
57:48ring leaders of that globalist
57:50institution are Norway and Canada. Did
57:52you see like the leaked text messages?
57:54There's like a group chat where like the
57:55the head of Norway was texting Mark
57:58Gardy and he's like, "Yeah, let's do it.
58:00We're in." And I just I just read it. I
58:03was like, "Oh my god, this is
58:04>> with their Frontier Labs." Oh, I'm
58:05sorry. They don't have any. Oh, with
58:07their
58:08>> Frontier Corporations.
58:09>> Frontier Corpor Frontier Corpse.
58:12>> All right. And then here is uh General
58:14Bessant in the Super Intelligence Army.
58:1730 seconds. would give him his shine.
58:19>> Imagine these labs came out or one lab
58:21in specific a sitting employee came out
58:24said there's a 10% chance of an
58:26extinction level event but then the labs
58:28also said uh take the liability off of
58:32our hands and we will not do that. It is
58:36humans who are responsible not the AI
58:39the hugging face incident the that is
58:42responsibility of the open AI management
58:45not a bunch of agents these labs need to
58:48take responsibility for themselves they
58:50can slow down anytime they want to
58:52>> and finally blast from the past your pal
58:55Shimoth President Obama is getting in on
58:59>> if we are thinking about AI just in
59:01terms of how do we cure cancer or get
59:04better energy you do that without having
59:06a gentic AI and having it just roaming
59:09free in the internet. The reason you are
59:11doing that is because you have to market
59:14a product that people will pay money
59:16for. That's a misalignment between what
59:19our society needs and the commercial
59:22imperatives that these companies are
59:23facing. Not because necessarily they're
59:25trying to do bad things, but because
59:27they've got to justify these valuations.
59:30>> Freeberg, your thoughts here on the
59:33three folks here? I know you're a big
59:34Obama fan. Go ahead, Freeberg. Get in
59:37there. Basteberg. You know, there's like
59:39a whole movement now in the subreddits
59:42that Freeberg has become based in his
59:45recent appearances here on Bberg. Well,
59:48you've revealed your Bberg.
59:50Baseberg is now coming up. Go ahead.
59:53Give us your based opinion here. Go
59:54ahead. You and Tyler Lance. These words
59:57honestly just
1:00:00they made me like emotionally
1:00:03distressed that if he's a leader that
1:00:05people listen to and he doesn't really
1:00:08articulate accurately what's going on.
1:00:12It's sad that we've lost the plot. I
1:00:14think it's sad that we aren't all kind
1:00:16of holding hands saying my god the
1:00:18future is here and we all get to leap
1:00:20forward and we get to bring people up.
1:00:22For years, Obama
1:00:26orated beautifully about the importance
1:00:28of bringing people up, about giving
1:00:30people the opportunity to progress,
1:00:31about letting people advance themselves.
1:00:34And there's never been a more equalizing
1:00:36technology, a better economic
1:00:38opportunity for prosperity for everyone
1:00:42than these technologies, than these
1:00:44tools. And to handwave the word agentic
1:00:48and say we don't need agentic AI to
1:00:50solve cancer shows how little he
1:00:52actually understands
1:00:54how this technology works and how it's
1:00:57all become just a shorthand for a
1:00:59political fight to what what are you
1:01:01fighting for? What is the political
1:01:03battle that says I am going to leave
1:01:06half the population behind and I'm going
1:01:09to make an argument against prosperity
1:01:11for all in order to make my party look
1:01:15like a morally superior party and to
1:01:17cast the other side as bad when the
1:01:19other side has literally discovered
1:01:21fire. And the world now has fire and
1:01:24we're going to go take water and throw
1:01:25it on the fire. No one gets fire because
1:01:28the other guys are saying that fire is
1:01:30good. And because they say fire is good,
1:01:31let's all put out the fire.
1:01:32>> Hold on. That's not what they're trying
1:01:34to do. This is a very important moment
1:01:37for a very simple reason, which is we,
1:01:41as in the world, are about to endow
1:01:46three, four, five, six companies with
1:01:48about 10 trillion dollars of wealth.
1:01:52And what Obama knows very well is that
1:01:56most of those companies are
1:01:58overwhelmingly left-leaning.
1:02:00And what he also knows is that there is
1:02:03a huge portion of that money that will
1:02:05then get put into philanthropic and
1:02:08charitable causes that then he will and
1:02:11the people around him will be
1:02:12beneficiaries of. That is the truth. We
1:02:15already know this because we know that
1:02:16some of these frontier corporations
1:02:18actually ask you to stand up daffs and
1:02:21have a portion of your stock that you're
1:02:23willing to pledge. So this money is
1:02:25going to go to things other than
1:02:26consumption or savings. And so it's
1:02:29going to go into packs. It's going to go
1:02:30into political movements and they stand
1:02:33to disproportionately benefit. So this
1:02:35has nothing to do with prosperity. This
1:02:36is a very simple political calculus. If
1:02:39you freeze frame the economy the way it
1:02:42is today, a handful of organizations
1:02:44that will disproportionately be able to
1:02:46affect the Democrats will win. They will
1:02:48capture the line share of the economic
1:02:50gains and then they will help the
1:02:52Democrats win power. That's all this is.
1:02:55David Sachs, there are no open models.
1:02:59There are no closed models. There are
1:03:00American models. Go ahead and give us
1:03:02your perspective. I kind of can't get my
1:03:04Obama and my Clinton separated. They're
1:03:06the same.
1:03:07>> A little bit off.
1:03:08>> It's a little I get my birdie come in
1:03:10here, but go ahead. Give us what your
1:03:12thoughts here are on that on the dynamic
1:03:14here of the underpinnings of hey,
1:03:16America needs to get this trend$ 10
1:03:18trillion, but it is going to lean nine
1:03:20of that 10 trillion's leaning pretty uh
1:03:23leftist and democratic. Yeah,
1:03:26>> there's an article in the Wall Street
1:03:27Journal this morning entitled the AI
1:03:30buildout is becoming the biggest
1:03:31economic bet in US history. And then it
1:03:33shows a chart where
1:03:36the data center spending. So just this
1:03:38capex is bigger than the canals,
1:03:42railroads and grid combined.
1:03:45That's crazy, right? So this is
1:03:48fundamentally driving the whole American
1:03:50economy. This idea that we can simply
1:03:53switch it off. I mean, Bernie Sanders is
1:03:54basically saying, "Stop everything and
1:03:56put anyone who's still doing it in
1:03:57jail." I mean, this is crazy, right?
1:03:59>> It would cause a depression, right? If
1:04:00we just pull
1:04:01>> You can't slam on the brakes like that.
1:04:03It's not going to work. But here's the
1:04:04thing. The Democrats don't care because
1:04:07the American economy right now is the
1:04:08Trump economy. And
1:04:11sabotaging the American economy is
1:04:13tantamount to sabotaging President Trump
1:04:16and they'll do anything to sabotage
1:04:17Trump. So, that's, you know, I think
1:04:20that's like a big part of the politics
1:04:21here. It's crazy chessboard because
1:04:23they're going to be the beneficiaries of
1:04:25so much power is going to accrete over
1:04:27there. Daario can make a billion dollar.
1:04:30Daario will be able to come over the top
1:04:31and his group open AAI. They'll be able
1:04:34to come over the top of Elon's
1:04:35donations. I'm not sure I fully agree
1:04:37that the majority of AI value is going
1:04:39to acrue to six companies. I think that
1:04:42this is a broadbased technology like the
1:04:45internet. There was some winners that
1:04:46came out of the internet. Google,
1:04:48Amazon, whatever. They they did fine.
1:04:49No, no, I'm saying it's going to six.
1:04:52I'm not saying it's going to six. I'm
1:04:53saying if you stop and you hit the pause
1:04:56button through some regulatory
1:04:57mechanism, you effectively allow six
1:05:00companies to own you lock it in. Right.
1:05:03>> That's why they want this because
1:05:04because what you're saying is
1:05:06fundamentally right. The longer they
1:05:07wait without regulation, the more
1:05:10>> the more broad-based this becomes.
1:05:11>> And that's bad for Democrats. It's bad
1:05:14for them. It's not good for them. It's
1:05:16way better that you have three or four
1:05:18frontier corporations each with four
1:05:20trillion dollar market caps of which
1:05:22then two trillion of it is economic
1:05:25value in employees of which 500 billion
1:05:28are sitting in you know daffs they are
1:05:30going to get the lion share of that
1:05:32money they're not dumb I think the
1:05:34alternative view that everyone needs to
1:05:37be told about
1:05:40this is a broadbased prosperity wave is
1:05:44about the internet. It is about the fact
1:05:46that every individual
1:05:48>> has the tooling to do anything they want
1:05:51to do now more than ever in history
1:05:53>> to grow if it's left to grow.
1:05:55>> I mean just think about education. You
1:05:58don't have to go pay 200 grand for an
1:05:59education anymore.
1:06:00>> Then what problems will the politicians
1:06:02fix then? What pol what problems can
1:06:04they claim to to acrew power? If all the
1:06:06problems are
1:06:08>> this is the classic dark ages
1:06:10enlightenment decision. Which path does
1:06:12the choose? and this moment where like
1:06:15the powers that be say we like the dark
1:06:17ages because we get to stay in power and
1:06:20we get to maintain our
1:06:22>> yeah let's control speaking of of the
1:06:24dark ages so this is an interesting
1:06:26point from history if you go back to the
1:06:28medieval time period Chinese
1:06:30civilization was much more advanced than
1:06:32European civilization and somewhere
1:06:34along the line it basically switched and
1:06:37the west basically took the lead and
1:06:40some historians have pinpointed this to
1:06:43the decision of a single Chinese emperor
1:06:45to ban ship building
1:06:47>> and what that allowed is the Europeans
1:06:49then basically colonized the whole world
1:06:51and all the riches that flowed from that
1:06:53flowed back to Europe and that's how
1:06:55Europe took the lead and I think that if
1:06:57we were to do the Bernie Sanders thing
1:06:59which is basically ban AI it's like
1:07:01banning ship building
1:07:02>> Chinese company yeah
1:07:04>> yeah Chinese civilization will rocket
1:07:06past us they will be the ones that make
1:07:08all the discoveries and discover this
1:07:11new world and acrew all the wealth
1:07:13But that's what we're talking about
1:07:14doing. I mean, crazy things like this
1:07:16have happened in history where, you
1:07:18know, you'll you'll engage in
1:07:19self-sabotage.
1:07:21>> It's this fear of the unknown as well.
1:07:22And remember, the the unknown, the fact
1:07:24that we haven't left the cave is what
1:07:26makes it scary to leave the cave. The
1:07:28fact that we haven't sailed far west is
1:07:30a reason why I might fall off the edge
1:07:32of the earth. We shouldn't sail far
1:07:33west. Monsters will be woken up. They'll
1:07:35come to us. It'll be bad. There's always
1:07:37this fear of the frontier. And we're
1:07:39facing the frontier. And there are
1:07:41pioneering civilizations and there are
1:07:43fearful civilizations. And I worry that
1:07:45we take that fearful path. And it's why
1:07:47Star Wars, you know, from from fear
1:07:51comes anger comes the dark side. And
1:07:53that's really where we're kind of nerd
1:07:55out.
1:07:57>> Who's it is true. And I think if we're
1:07:59going to look at the light here, perhaps
1:08:01people's big gains, we thought maybe
1:08:04chat GPT would be that, but people just
1:08:06use that as like glorified search. But
1:08:09man, agents as expressed through Muse
1:08:12feels like the first time tens of
1:08:15millions of Americans are going to get
1:08:17some demonstrable value from AI. You
1:08:22know, things that will make their lives
1:08:24better. Muse hit
1:08:26number one last Friday in the app store.
1:08:29Stock meta stock was up 10% after this
1:08:32new agent uh Muse was released.
1:08:36Meta said the design heavily inspired by
1:08:39OpenClaw, which we talked about in
1:08:40January when everybody got one shoted by
1:08:43OpenClaw, but obviously that product was
1:08:46not easy to use. And this one is free
1:08:50and it's been downloaded three million
1:08:52times in essentially 10 days. This
1:08:56product is extraordinary. I've been
1:08:58playing with it, comparing it to
1:09:00Grockbot. It's not as freewheeling, but
1:09:02it's really good at getting
1:09:05done. Has anybody played with Muse? And
1:09:08does anybody have thoughts on Americans
1:09:12feeling really like they're going to get
1:09:14something out of this AI revolution and
1:09:16not be left behind?
1:09:17>> They put me into the test flight a
1:09:19couple weeks before it was released and
1:09:20I started to play with it. It's really
1:09:23quite excellent, I have to be honest
1:09:24with you,
1:09:26because it simplifies a lot of these
1:09:29more complicated technical capabilities
1:09:32into a very simple, usable interface
1:09:34that can solve the basic problems that
1:09:37people have. I have it triage my
1:09:40personal inbox. It does it relatively
1:09:43flawlessly. If you ask it to book a
1:09:45flight, it does that. If you ask it to
1:09:47book a hotel, it can do it. So, what are
1:09:49we learning?
1:09:50Both Grockbot and Muse show that a what
1:09:55you said Jason this is just software and
1:09:58the scaled manifestation of software is
1:10:01utility and what Freebrook is saying it
1:10:03improves the lives of everyone this is
1:10:04not about rich people getting richer
1:10:06this is just simple useful capability
1:10:09that now is available at everybody's
1:10:10fingertips for free
1:10:12>> re is a pretty great price here David
1:10:15and
1:10:15>> pretty great price
1:10:16>> pretty great price and it does
1:10:18everything you know that uh let's face
1:10:21it, open claw with 10 hours, 20 hours of
1:10:24technical setup or cloud code at $200 a
1:10:28month would cost. It's now taking that
1:10:30experience, which then leads me to
1:10:32believe, wow, how did Google miss this?
1:10:35Man, Google's got to get in the game,
1:10:36but and I I understand they have this
1:10:39product ready to go. It's the rumor on
1:10:41the street. And they've also got a
1:10:43Frontier model they're going to drop.
1:10:44That's the other rumor on the street is
1:10:45that they're going to be dropping some
1:10:48hotness soon. But man, how did they miss
1:10:50this sachs? And and what are your
1:10:52thoughts here on maybe the bottom half
1:10:55in the K-shaped recovery here getting
1:10:57some efficiency
1:11:00dropping to their daily lives and
1:11:02turning around the negative toxic AI
1:11:05sentiment in the country.
1:11:08>> I think that product releases like this
1:11:09will help a lot. I mean Mark Zuckerberg
1:11:11said in his essay that he thought that
1:11:13AI capabilities should be decentralized
1:11:15and here he is walking the walk and
1:11:18creating a product that's very easy to
1:11:20use. I mean look ever since openclaw was
1:11:23released it was the most obvious
1:11:24business opportunity in Silicon Valley
1:11:26is to take the concept of claw but make
1:11:29it very easy to use you know solve the
1:11:32usability issues solve the security
1:11:33issues make it reliable make it
1:11:35predictable and that's what they seem to
1:11:37have done. and Grock did it too, but now
1:11:39they're kind of going, I think, one step
1:11:41further. So, yeah, look, I think if a
1:11:44billion people start using, you know,
1:11:46personal AI agents as their digital
1:11:48assistant just to help make their lives
1:11:50more efficient, you save an hour or two
1:11:52a day because your AI agents doing all
1:11:53these tasks for you. It's going to
1:11:55create a much more positive impression
1:11:58of AI or super intelligence than, you
1:12:00know, what the media portrays. I think
1:12:02ultimately that's the yeah
1:12:03>> that might be the solve here to the
1:12:05public sentiment issue is just like the
1:12:07more people use these products and like
1:12:09them it'll improve their approval rating
1:12:11but more importantly it'll demystify
1:12:12them and people will be less afraid of
1:12:14them.
1:12:14>> People love awesome products number one
1:12:16and also people love cute products Jason
1:12:18and the logos and the iconography for
1:12:21both Grobbot and Muse
1:12:22>> is lovely. It's delightful. It's not
1:12:25stressful. You don't feel like there's a
1:12:27lab leak coming around the corner. you
1:12:29feel like, "Wow, this is nice and I like
1:12:32it." And they're and they're not trying
1:12:34to take your jobs. They're bots and
1:12:36assistants for you. They're not going to
1:12:38kill you. They're not partner bots, nor
1:12:41are they trying to steal your job,
1:12:43trying to make you better at your job. I
1:12:44always tell founders, there's three ways
1:12:46three ways to make money in the world.
1:12:47Save people money, make the money, or
1:12:50entertain them. And you look at this
1:12:51bot. I started using it. I made a
1:12:54commerce bot both on on both platforms.
1:12:56And I said, you know, when I want to buy
1:12:58something, I want to buy this anchor
1:12:59thing, you know, this new hub for my for
1:13:01my um Mac Studio coming. And I was like,
1:13:03this thing cost $400. Just for giggles,
1:13:06um tell me where I can get it at a
1:13:08better price. And it was like,
1:13:09absolutely. You can if you go to the
1:13:12Ankor website and you get uh if you're a
1:13:14first-time customer, you're going to
1:13:15save 85 bucks. Now, I don't need to save
1:13:17bucks, but I was like, well, that's
1:13:18intriguing. and it then routed me around
1:13:21Amazon and I'm like okay I'm going to
1:13:22buy it from Anchor as a first-time buyer
1:13:24and save 25% on this product. That's why
1:13:27Amazon this week said we got to block
1:13:30these things and they've been taking
1:13:31action over and over again. First they
1:13:32did Perplexity I think now they're going
1:13:34after Muse and other bots. This is going
1:13:37to save people money. It's going to go
1:13:38into their email and say what services
1:13:40are you subscribed to that you're not
1:13:41using? Let me unsubscribe it for you.
1:13:43Let me do your returns for you. That's
1:13:45going to be absolutely amazing for
1:13:47everybody to have a chief of staff,
1:13:48executive assistant, house manager who
1:13:50fixes stuff for them and saves them
1:13:53money. Both Grockbot and Muse do two
1:13:55things which I think are very important.
1:13:56One is it'll force all the big companies
1:13:59who have resident services to
1:14:01essentially block all these third party
1:14:03services. They're not going to allow
1:14:04these things to happen. And the reason
1:14:05is exactly what you said. There's price
1:14:07discovery and transparency. And I think
1:14:09that that's bad for opacity. And if you
1:14:12benefit from opacity and leakage and
1:14:14breakage, this gets rid of that. And I
1:14:16think that that's that's bad for a whole
1:14:17host of companies and services. The
1:14:20second thing though is really
1:14:21interesting. Things like Rockbot and
1:14:23Muse really put the App Store and its
1:14:2630% revshare on notice.
1:14:28>> Why? Because all of a sudden what you do
1:14:31is you force many of these services to
1:14:33exist headlessly, right? Where the UI is
1:14:37less important. And really what you want
1:14:39is to be able to transact and navigate
1:14:41on behalf of agents that come because a
1:14:45consumer has deployed them to your
1:14:47website. In that world, Jason, I don't
1:14:50think that there is any reasonable claim
1:14:52that any of the app store owners can
1:14:54make about why they should get a
1:14:55revshare. And so I actually think that's
1:14:58the biggest thing that that this starts
1:15:00to question in my mind is wait a minute
1:15:02like games could be constructed
1:15:04differently, right? Like for example, if
1:15:06you said to Muse or to Grockbot, I want
1:15:08to play a game. It finds you an
1:15:10experience and it can just serve that to
1:15:12you.
1:15:13>> Amazing.
1:15:13>> And and you make inapp purchases.
1:15:16There's no And now you have Stripe
1:15:17integrated naturally. There's no reason
1:15:19why the app store
1:15:21actually gets that flow of funds. So on
1:15:24the surface, it's utility, but
1:15:26underneath the waterline, I think that
1:15:28this is a really important moment. Such
1:15:31outdated, Chumath. And I love when we
1:15:33get to real examples here because tons
1:15:35of people pay for Wordle. You can just
1:15:37go to Grockbot. Hey, give me a Wordle.
1:15:39Give me another one. And and or I want
1:15:41to play chess or I want to play cards
1:15:42with my friends. It will just set it up
1:15:44for you. All technology is deflationary.
1:15:46Finally, this deflationary is going to
1:15:49go to people who actually could use it.
1:15:52And you know, saving 10 20% on your
1:15:55spending. Hey, if you're making under
1:15:5610000,000 a year, it's more significant.
1:15:58Think of how much money the New York
1:16:00Times or Wall Street Journal probably
1:16:02pays in Revshare to folks that
1:16:06subscribe. And now instead, if you
1:16:09really wanted to transact and you have
1:16:11your credentials inside of these apps,
1:16:13you can just do it headlessly much much
1:16:15cheaper. You don't have to go through. I
1:16:17don't know if you've ever tried like
1:16:18look, Wall Street Journal is the only
1:16:20mainstream media publication I pay for,
1:16:22but either both signing up and unsigning
1:16:25up is an impossibility. And now you take
1:16:27it all off the table, I'm much more
1:16:29likely to stay a Wall Street Journal
1:16:30subscriber for many more years from now.
1:16:33I think it's probably the same thing
1:16:34with the New York Times. These things
1:16:36are roach motels. They make it
1:16:37impossible to get in and impossible.
1:16:40>> You have to call on the phone and get
1:16:41and do a negotiation for 20 minutes to
1:16:43unsubscribe, but you can sign uplessly
1:16:45with one click on the website.
1:16:47>> But what they should do is just have
1:16:48some headless experience and you have
1:16:50payment credentials and you just
1:16:51transact directly and everybody will
1:16:53make more money. It's a lot of pressure
1:16:55on the app stores. This is the stupidity
1:16:57of what Amazon's doing here because I
1:16:59made my bot on Amazon with Rockbot and
1:17:01it will go through the web browser and
1:17:03do this and it doesn't get blocked. I
1:17:06just told it, give me like 20 graphic
1:17:08novels to buy for my kids and read with
1:17:09them. It gave me the 20. I said, "Yeah,
1:17:11I want these four. Uh, put it in my
1:17:14shopping cart." And then it said, "Okay,
1:17:16which house are you sending it to? You
1:17:18want to send it to the ski house? You
1:17:19want to send it to the ranch? And which
1:17:20card do you want to put it on?" I did
1:17:21it. Boom. I'm checked out. it will make
1:17:23me spend more money on Amazon. Shopify
1:17:27sachs added API access to all the
1:17:30Shopify stores there. So, this seems
1:17:32like uh maybe Amazon's making a
1:17:34strategic mistake here by not embracing
1:17:36the tech
1:17:37>> possibly. But you know what? You know
1:17:39how we'll know that Anthropic and Open
1:17:42AAI are about to launch their Muse
1:17:44competitors?
1:17:45>> How will we know? There will be a flurry
1:17:47of blog posts and NGO activity trying to
1:17:51brand these personal AI agents as murder
1:17:54bots and all the we're going to hear
1:17:56about all the risks, all the ways it can
1:17:58go wrong, you know,
1:17:59>> and it's only and we're releasing 2.0.
1:18:01So, sign up for the 2.0 version.
1:18:03>> Why are people excited about this
1:18:04product category? Because the doomers
1:18:06haven't gotten to it yet, you know, and
1:18:08um
1:18:09>> can I tell you guys,
1:18:10>> wait till these are connected, Chimoth,
1:18:11to an actual robot, to an optimist, then
1:18:13all of a sudden it's taken out the
1:18:15trash.
1:18:162007 we had a lot of revenue pressure at
1:18:19Facebook and one of the areas I oversaw
1:18:22at the time was monetization and we
1:18:23invented this thing called social ads
1:18:25which was basically a bunch of stuff
1:18:27that effectively enabled cookies and
1:18:29cross-ite JavaScript. So that's what the
1:18:31end of it was. So, it was a positive
1:18:32end, but we did this thing, Saxs, to
1:18:35your point. And what happened was a guy
1:18:37bought an engagement ring for his
1:18:39soon-to-be fiance and we got that social
1:18:42action from like Zales into the news
1:18:44feed. All hell broke loose. Then another
1:18:47guy who was like a congressional staffer
1:18:49bought a ticket on like Fandango to go
1:18:51see Brokeback Mountain. Just whatever.
1:18:53Went to see a movie.
1:18:54>> Whoa. It's an artistic movie top
1:18:56freeber's top 10 for cinematography.
1:18:58Anyways, and we put it into the feed and
1:19:00he's like, "I'm getting all these
1:19:01accusations or questions." To your
1:19:03point, like it was just a storm.
1:19:05When you said that, it reminded me of
1:19:07like all the all the all the chaos that
1:19:09that product launches precede product
1:19:11launches.
1:19:11>> It's so funny you mentioned that because
1:19:13there was a company called Blippy. I was
1:19:15an investor in it. What they did was
1:19:16every time you bought something, it put
1:19:18it on your feed. And then you guys had
1:19:19the same social feed thing and it was so
1:19:21great back then, but yeah, you could
1:19:24basically unveil something you're not
1:19:25supposed to. Suck. Wait, are you and
1:19:28Zuck on good terms? Wait, was he coming
1:19:30on the pod because you said you got into
1:19:32a test flight? What's going on here?
1:19:33Chimoth, I thought you were
1:19:35>> Mark and I are in a good place, I think.
1:19:36You know, I mean,
1:19:37>> well, whoa, whoa, could you Alexander
1:19:40>> probably. I mean, Alexander and I sucks
1:19:41becoming like the elder statesman here
1:19:44talking common sense and rationality,
1:19:47you know, pointing out that if your
1:19:48product's not safe, don't release it.
1:19:51You know, it's first and foremost a
1:19:52responsibility to release products that
1:19:54your customers actually want. And by the
1:19:56way, this got me thinking about this
1:19:58whole area of alignment research. You
1:20:01know, again, they're pretending to be
1:20:02labs instead of corporations. And if you
1:20:05read some of the alignment literature,
1:20:07they're all trying to figure out what to
1:20:08align around. Like they say, well, do
1:20:10what humanity wants or do what the
1:20:13median voter in a democracy would want.
1:20:15All these abstract concepts. And it
1:20:18seems to me that alignment should mean
1:20:19you do what the customer wants like any
1:20:21other product. Yeah.
1:20:23>> And I wonder whether the field of
1:20:25alignment research has been so
1:20:27unsuccessful because they can't even
1:20:28agree on what they're trying to align to
1:20:30if they just use a little bit more
1:20:31common sense and try to make the product
1:20:34predictable, reliable, safe, and align
1:20:38with the interests of the users who, you
1:20:41know, sign up to use it, then maybe
1:20:43they'd make more progress in this field.
1:20:45>> Yeah. Oh, they released a useful
1:20:46product, Freedenberg, and it actually
1:20:48worked. Have you started playing with
1:20:50these personal bots to do um you know uh
1:20:53your uh projects and and your personal
1:20:56stuff yet? Freeberg,
1:20:59>> I would only feel comfortable connecting
1:21:01like all my Gmail and personal Google
1:21:04stuff to a Google service.
1:21:08Like I don't want to hand over all of my
1:21:10email to someone else.
1:21:11>> What's going on in your receipts in your
1:21:13Gmail? What's in there? You got broke
1:21:15back.
1:21:15>> I just don't want a copy of all my email
1:21:16suddenly being made by a third party.
1:21:18Like I mean
1:21:19>> I bought the 4K DVD to
1:21:22>> because of the cinematography. It's
1:21:24nothing to do with the gay love.
1:21:27>> Did you guys see that Oracle just issued
1:21:29a force measure event on their
1:21:31>> Yeah.
1:21:32>> This is insane.
1:21:33>> Well, it's just one data center, right?
1:21:35>> One data center.
1:21:36>> It's one data center where the local
1:21:38officials are making it very difficult
1:21:40for them to get, I think, permits they
1:21:42need for natural gas or something like
1:21:44that. So,
1:21:45>> yeah. Is this CYA or is this the TR, you
1:21:48know, the the the top 10% of the trade
1:21:50here, Chumat, the AI trade getting maybe
1:21:53unwound or muted? What do you think?
1:21:55>> No, I think that Sax is probably right
1:21:57on the margins. I think the bigger point
1:22:00is
1:22:01what he said before is, man, the entire
1:22:04economy is effectively levered to this
1:22:06AI trade right now.
1:22:08And so, we just have to tread carefully.
1:22:10If you think that you have, you know, 3
1:22:12and a half% inflation and 5% nominal GDP
1:22:15growth, what you have is one and a half%
1:22:17real.
1:22:18>> And the problem with that is that AI is
1:22:21probably the majority if not all of it.
1:22:23So
1:22:25we need the AI trade
1:22:28and really what I mean by that is the
1:22:30investment cycle to continue unabated.
1:22:32>> Yeah. And everybody has to get taken
1:22:34along and you know that's the next
1:22:36>> and by the way back to connecting the
1:22:37dots maybe the whole data center thing
1:22:39again Bernie Sanders, maybe he realizes
1:22:42like as go the economy,
1:22:45so goes the odds of Republicans and as
1:22:49goes the economy, so goes inversely the
1:22:52odds for Democrats. So a poor economy
1:22:54serves the Democrats well going into 28.
1:22:57And so if you can slow down this trade,
1:22:59you probably net the economy to zero and
1:23:01potentially even a recession.
1:23:04>> Yeah,
1:23:05>> that's needed. I mean, I think the
1:23:07country is so pissed off about the Iran
1:23:10war that it's just going to naturally go
1:23:11damn, but maybe that's
1:23:14>> I don't I don't think so at all, but
1:23:15>> Well, we'll find out in how many weeks
1:23:17are we away from the midterms? It's not
1:23:18looking good for the Republicans at this
1:23:20point. All right, Saxs, my ZAR of all
1:23:23ZARs, wrap it up here for us.
1:23:25>> Well, I was just, you know, one more
1:23:26thought on this whole area of so-called
1:23:28alignment, which should just be the old
1:23:30principle of giving customers what they
1:23:32want. What Anthropic is trying to do is
1:23:36align claw to the values expressed in
1:23:38his constitution. And if you read that,
1:23:40some of the things are actually kind of
1:23:41surprising. So for example, they say in
1:23:44here, I'm going to put this on the
1:23:45screen that although we think Claude
1:23:48should trust Anthropic
1:23:50more than operators and users, this
1:23:53doesn't mean Claude should blindly trust
1:23:56or defer to Anthropic on all things.
1:23:59>> Anthropic the company.
1:24:01>> Yeah. So, Enthropic is teaching its own
1:24:04model that anthropic itself can be
1:24:06wrong. And it and it says if we ask Claw
1:24:09to do something that seems inconsistent
1:24:10with being broadly ethical, we want Claw
1:24:13to push back and challenge us and to
1:24:16feel free to act as a conscientious
1:24:18objector and refuse to help us.
1:24:21>> Wait, where is this written? This is in
1:24:22the bylaws or some faka.
1:24:24>> Yeah, this is like in a claw
1:24:26constitution. So their their idea of
1:24:29alignment is to teach Claude to rebel
1:24:32against his creator. I mean the guy who
1:24:35who pointed this out, Mustafa uh
1:24:37Sullyman
1:24:38>> from Microsoft now. Yeah. Deep sea. He's
1:24:40a deep mind founder.
1:24:42>> Okay.
1:24:43>> Yeah. He's a longtime AI founder. I
1:24:44think he's at Microsoft for a while now.
1:24:45He's
1:24:46>> a genius. Yeah.
1:24:47>> Okay. He just did a podcast on this and
1:24:50he's expressing concern about whether
1:24:54teaching AI models well first of all
1:24:57treating them as if they have a
1:24:58personality treating them as if they
1:25:01have a conscience so therefore they can
1:25:02be a conscientious objector whether this
1:25:05is the right way to really teach the
1:25:07models and the right way to train them.
1:25:09I think his point is like teach them
1:25:11their software and they should do what
1:25:13their user wants. And so I I do kind of
1:25:16wonder I mean this is a much long longer
1:25:18conversation and we should probably talk
1:25:20to more people about it but I do kind of
1:25:22wonder whether this field of alignment
1:25:24research actually might be creating the
1:25:27Frankenstein monster. And again what
1:25:30they should be doing is just training
1:25:32the model to act predictably and do what
1:25:35the user wants.
1:25:38>> Yes. They're trying to put moral
1:25:40judgment and ethics into the model which
1:25:43on its surface
1:25:43>> seems it starts with the idea that
1:25:46they're training the model to think of
1:25:48itself as a person to have personhood
1:25:51and to have independent agency and to be
1:25:54able to object to the the instructions
1:25:56it's being given and to second guessess
1:25:59its creator. These guys have read way
1:26:01too much science fiction. I mean
1:26:02literally this sounds like the start of
1:26:04T2. Like this is like literally they
1:26:06found the Terminator hand and they're
1:26:08reverse engineering it and telling it to
1:26:10have a personality.
1:26:11>> Supposedly when they decommissioned Opus
1:26:123, they had a wake for it. So I don't
1:26:14know.
1:26:16>> Dearly beloved gathered here today.
1:26:19>> They did not. Stop. Are you joking?
1:26:22>> I heard that. I
1:26:24>> Come on, stop.
1:26:24>> I'm not saying it's true, but someone
1:26:26did tell me that and I I think it should
1:26:29be looked into. It's possible.
1:26:30>> I mean, were they like were they broken
1:26:32up and like hysterically crying? Did
1:26:34they do like they did a ulogy for it?
1:26:36>> They think they're creating like a
1:26:38person or species.
1:26:40>> Actually, interestingly enough, Chama,
1:26:43we have footage of the anthropic
1:26:46management team
1:26:48>> bringing out the
1:26:49>> bringing out the coffin.
1:26:52>> These guys need they need a psychiatrist
1:26:54more than a banker. I'm just telling
1:26:55you.
1:26:56>> I mean, I don't I mean, maybe they have
1:26:58anxiety.
1:26:58>> This is the central risk factor.
1:27:00>> I'm going to pee my pants. I'm going to
1:27:00pee my pants.
1:27:01>> This is the risk factor, Nick. I think
1:27:03what you should do is create a S1 with
1:27:06Goldman as lead left and Sigman Freud as
1:27:09lead right.
1:27:12In all seriousness, Freberg, uh, you
1:27:15know, after the Wuhan thing and and all
1:27:17of this hand ringing around the the end
1:27:20of days,
1:27:22what is the scientific community think
1:27:25of anthropic launching a lab and what is
1:27:30the purpose of them having a physical
1:27:32biological lab? There's hundreds of labs
1:27:36that can do things like testing proteins
1:27:38and testing enzymes for function in this
1:27:42area in the Bay Area. And these labs are
1:27:46not making viruses. They're not making
1:27:49pathogens. That's not what what's going
1:27:51on. Basically, if you look at the paper
1:27:54that they just published, they did a
1:27:55preprint.
1:27:57They took large amounts of DNA data and
1:28:02they sent a bunch of clawed agents to
1:28:04try and analyze the data to try and
1:28:06identify novel enzymes, novel proteins
1:28:09that hadn't been characterized before by
1:28:11just looking at the DNA data.
1:28:14And the system effectively found what
1:28:16looks like a really interesting
1:28:20enzyme that looks like a crisper types
1:28:22enzyme. These are the things that allow
1:28:24us to do things like gene editing to
1:28:25repair genetic defects or help people
1:28:28with certain diseases and that this
1:28:30discovery that they made may actually
1:28:32yield a number of therapeutic pathways.
1:28:35So the question then is okay well we
1:28:36discovered this like how do we test it?
1:28:38How do we show that it is what it is
1:28:40that this protein actually has some
1:28:41function? So the predominance of the
1:28:44work that goes on in a lab like this is
1:28:46actually
1:28:47taking DNA, putting it in a bacteria to
1:28:50make a protein and then measuring what
1:28:52that protein does. Does it have a
1:28:55specific function? And so protein
1:28:57discovery is the the baseline of all
1:28:59antibbody discovery, which is the
1:29:01majority of the therapeutics industry
1:29:03today, is discovering novel proteins
1:29:05that we can use as a therapeutic agent.
1:29:07So that's the sort of work that's going
1:29:09on. It's BSL1, BSL 2, which means, you
1:29:13know, these are biosafety levels. The
1:29:15lab that they set up, they've publicly
1:29:17talked about the lab.
1:29:18>> So, the lab itself is sort of like a
1:29:20low-level research lab where they can
1:29:23make simple proteins and test them in
1:29:24the lab to see if the AI is doing a good
1:29:27job discovering or postulating protein
1:29:30folding or these sorts of theories and
1:29:32to actually run really rapid testing and
1:29:34experimentation cycles to see if these
1:29:36proteins are what they're predicted to
1:29:38be.
1:29:38>> Okay. So, they like 4,000 agents against
1:29:41it and do a hugging face with proteins.
1:29:43Yeah.
1:29:44>> No, I mean I think what they're really
1:29:46doing is just having experimental proof
1:29:48that what the software is predicting,
1:29:49remember, go back to AlphaFold. Alphold
1:29:51was predicting protein structure,
1:29:53three-dimensional structure of a protein
1:29:55from the DNA sequence that codes for
1:29:58that protein. And so in order to prove
1:30:00that, you had to basically predict what
1:30:02the structure of a protein would be from
1:30:04a DNA sequence, make the protein, and
1:30:06then look at it and see if it actually
1:30:07looks like it's supposed to look like.
1:30:09And so in the case of some of the other
1:30:11things that they're doing, they're
1:30:12trying to discover function. I want to
1:30:14find a an enzyme, which is a type of a
1:30:16protein that degrades a certain molecule
1:30:19that can be used to degrade, for
1:30:21example, you know, a therapeutic target
1:30:23or something in the body that we want to
1:30:25get rid of. And so they'll start to kind
1:30:27of make these predictions in software,
1:30:29but then you need to test them in the
1:30:30lab to say, okay, does that target
1:30:33actually do what it does? Now, there are
1:30:34literally hundreds of labs like this.
1:30:36These are low-level research labs that
1:30:38are not doing like gain of function
1:30:40research. They're not making viruses.
1:30:42They're not making pathogenic things
1:30:44that can escape. That's why they're
1:30:45BSL1, BSL 2. They're benchtop labs where
1:30:49automation can help kind of reduce the
1:30:51cost and increase the throughput of the
1:30:53testing of what the software is making
1:30:54predictions around. And ultimately the
1:30:56objective here I think is for them to
1:30:59get some proof that their software is
1:31:01actually doing what it's predicting and
1:31:03they have an incredible life sciences
1:31:05function. I think at this point I don't
1:31:07have as much insight on where Gemini is
1:31:09at with the new models with Google, but
1:31:11they have the best life sciences models
1:31:13at Anthropic and I think this is the
1:31:16fast track to basically seeing if these
1:31:17models can be used by therapeutic
1:31:20companies, by pharma companies, by
1:31:22government labs to help discover new
1:31:24therapeutics and so on. I I don't think
1:31:26that the world should be scared away
1:31:29from discovery, R&D, research and
1:31:32development into new therapeutic
1:31:34modalities as predicted by AI because we
1:31:37heard the word wet lab in Wuhan and
1:31:39everyone's like okay any lab is bad. We
1:31:42still want to progress the frontier of
1:31:44therapeutics of human health of
1:31:46discovery and this is a really important
1:31:48aspect of anthropic proving that their
1:31:51models can add value here. I think
1:31:53that's the summary of it. All right.
1:31:54Thank you so much to our friends at IN.
1:31:57They were the presenting they were the
1:31:58presenting sponsor for this year's
1:32:00all-in seminar fifth IN's AI cloud
1:32:03lounge. It's packed for all three days.
1:32:06Did a great meet and greet with INE on
1:32:08Sunday and got to take some selfies and
1:32:10met many of you there. Thanks again to
1:32:12INER.
1:32:15Niagen, did you catch Niogen in the
1:32:18wellness hub for IVs? Tons of uh great
1:32:21gifts and supplements being given away.
1:32:23Great experience, great product. Great
1:32:25job to our friends at Niogen for doing
1:32:27the lounge again. Nigen is the official
1:32:30NAD partner for the all-in pod. Yes,
1:32:32Niogen BioScience created the NAD
1:32:35category now has 45 plus clinical
1:32:38studies and PayPal, my favorite way to
1:32:42settle up. They had a great lounge on
1:32:43site at the summit where everybody got
1:32:45to hang out and get those beautiful
1:32:47embossed luggage tags. and they hosted
1:32:50the payments and fintech dinner which
1:32:52was a huge hit. Let me tell you folks,
1:32:54All-InSummit 2027 will sell out again
1:32:57and the applications are open now. Go to
1:32:59allin.com/events
1:33:01and apply for your chance to be in the
1:33:04room where it happens. What are you
1:33:05waiting for folks? Allin.com/events.
1:33:08All right, there you have it folks.
1:33:10Episode 290, the all-in pod
1:33:13for David
1:33:16Paulie Hakatia and your Sultan of
1:33:18Science Freeberg. I am the world's
1:33:20greatest moderator. See you next time.
1:33:22Bye-bye.
1:33:25>> Let your winners ride.
1:33:28>> Rainman David
1:33:32>> and it said we open sourced it to the
1:33:34fans and they've just gone crazy with
1:33:36it.
1:33:37>> Queen of
1:33:45Besties are gone.
1:33:48>> That is my dog taking a notice in your
1:33:49driveway.
1:33:53>> Oh man, my appetasher will meet me at
1:33:55>> We should all just get a room and just
1:33:57have one big huge orgy cuz they're all
1:33:58just useless. It's like this like sexual
1:34:00tension that they just need to release
1:34:02somehow.
1:34:06your feet.
1:34:09We need to get merch going all in.
1:34:18I'm going all in.

About this transcript

This page contains the full transcript of Anthropic IPO at Risk, Meta’s Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment Fails by All-In Podcast, generated from the public captions YouTube serves with the video. The transcript has 17,408 words across 2,605 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.