Everything's a Metaphor

Exactly. I know. I know. Uh we have a lot to talk about though on this episode. Uh we remain in an export controlled situation for Anthropics top models, Fable and Mythos. Meanwhile, OpenAI making a big show of their five point five cyber, showing off some very impressive benchmarks and committing to a bunch of uh ways that they hope that they can leverage this kind of stuff at scale for patching holes in the internet, as it does look like these chained agentic

uh uh hacking techniques or something that uh people are taking very seriously. And then on the horizon, possibly even this week. We see the debut of five point six and a brand new voice model. But let's begin.

You know, we ended up seeing Dario show up at a G six, G seven summit, along with Sam and some other AI people. Uh Trump had said things are directionally positive about Anthropic. The The challenge is this, it's a frontier and I think metaphors got us into trouble. I think using metaphors like this is like nuclear material and that was I think that was a really, really not a good I and Curtis I Dario's built an exceptional company. Cruptial product, great team, but

When you tell the US government, who is in the middle of, by the way, bombing an aspiring nuclear power, that this is a new nuclear power, and then they're like, Hey, we have concerns and you're like, yo, bro, it's got it. Don't worry, that's not a concern.

Particularly when, you know, as Trump had said, hey, one of their investors and competitors had said this, basically alluding to the fact that it was Amazon was the one that went to, you know, the government said, hey, we have an issue and let's let's unpack that a bit. We're not talking about Amazon, the people that deliver you your you know your your latest edition of whatever.

They had concerns. They went to anthropic. Anthropic brushed them off and they didn't feel like they were getting a satisfactory answer. So then they elevated that to the government to say, hey, listen, we have concerns. Other people within Not just administration people. I understand this is not a a a partisan thing. This was Stakeholders in security were concerned about this. And then stories came out. You know, you know, there was

Earlier when Mythos came out, it turned out that more people had access to it than we're supposed to, that basically was being unrestricted access was to other people. Maybe even the Chinese had access to it. Then it came out that there was a government whitelist of a hundred companies and Anthropics said we'll have access to it. Then it turned out there were fifty more companies that had access to it.

And the government's position is we want to know because there are things about companies that we know that you don't know because they're top secret. One of the companies that came out was, you know, allegedly was a s you know a Korean, South Korean telecom company.

That may have had connections to may have had that maybe maybe like maybe infiltrated whatever but they thought was concerns'cause of the Chinese government. Again, maybe true, maybe not. I don't know. I am I'm not sitting there in front of the dashboard, the Palantir dashboard. you know, of the NSA getting all the intel and very few people I know do.

Um and no most m most of the people commenting on Twitter definitely don't have that access to it. So we don't know. But we do know that created this issue of you said this was dangerous and scary and could do these things. But then when we brought concerns to you, you didn't act like these were scary things to deal with the concerns. And that created an issue of trust. A way a lot of the stuff works, a lot of this leads before even formal policy gets.

And that means you have people from your company in DC who are talking on a daily basis. I remember early on when I was at OpenAI, you know, I would go to the White House to go talk about stuff, you know, to partners and stuff on a broad sense. Meanwhile, some of the security team people were talking to the security community and doing this back and forth. Anthropic was been and Anthropic is supposed to be

the experts at this because they've been selling contracts to to CIA and the NSA, apparently, allegedly, you know, for years. So I would say there was a st a lot of uh Not the kind of tight communications that we wanted to have happen.

Came out and said that look, the NSA was really worried about this. And there was questions as to exactly how he described their level of worry of what Mythos could do, but put that to the side for a second. That means that in this period of time leading up to the export control, you had Andy Jassy of Amazon.

The NSA, we have heard from behind the scenes that there was multiple different government agencies that were that these are not politicians. These are these are tech-centric experts that know this tech. They are not babes in the wood on this. They they know what this can do and what this can't do. Uh, and then people that are not adversarial had not been at least adversarial to anthropic in uh Scott Bessent and Howard Lutnick that eventually have this all come to a head.

And it all does just seem to be about trust. And that that's what's sort of come to the come to the fore here. And it does seem like the administration is going out of their way to try to explain it a little bit more because they know that this is something that can be painted on on a more polarizing line, but I don't think it was shocking that they went to the Washington Post with some of this because they wanted the message to be look

The Post isn't gonna do us any favors with with their with their coverage. At least they haven't up till now. If the Post is reporting this, then this is something that that, you know, people should generally feel is essentially vetted. And it all just comes down to for all of their talk about security, security, security, the administration, the NSA, Amazon. All felt that Anthropic wasn't taking security seriously.

They want a successful OpenAI, they like a successful Google. They want to have multiple players. They do not want anybody to run away with this. This isn't very clear. JD Vance has even said as much as this. They don't want anybody to run away with it. And as much as it might play into some fantasy, the idea let one lab run away with it with really good relationships to the government.

Couple things happen. One, administrations change. Two is you create a situation of stagnation and it's not good. So they they they do not want to see. You know, a hobbled anthropic. And I don't think that's in the interest of the American consumer. You know, I don't think that's in the interest of American AI. Okay. Um

But I think that the the every every company right now in AI that every frontier is going through their own sort of learning curves and figuring out how to deal with the changing landscape of stuff. And We're going to, you know, we can we can sit there and say, oh, we'll put this policy in place. And I'm not saying that none of that should be done, but I think things move so fast. And part of the thing that I've been saying for years is some of these things people suggested.

Like I remember years and years ago at OpenAI, it had some some people who are some of the safety people like, Oh, we need to worry about this frontier model. I said, Tell me your greatest fear that you think that once you get above here, that that's a reality. And let me tell you if I can figure out how to do it with a smaller model, limit the capability, and be able to work a way around it. And we're seeing that now.

With people building these harnesses, which is basically building an information system that goes around a model that are able to do really complex stuff. And part of it being people say rightly so that like, hey, if we don't get figured out how to get Fable out there for legitimate security researchers to use.

We're gonna have a Chinese model. We're gonna have something that's not aligned that we don't have eyes into do it. And so I think there's an incentive to figure out as quickly as possible how to make these things widely available. I also think that um Sometimes trying to write the policy can be really hard because it comes at you from so many different angles.

make phone calls to enough people that that can mediate smaller things so smaller things don't turn into bigger things. And it seems like that's that that that might be what's happening now with with the anthropic. Because it it does seem like the further we get away from here is like this was not an exercise in just government bigfooting. an AI lab, even the AI I AI lab that was begging to be Bigfooted, but uh this this really was, you know, just when

You know, there were a persistent line of problems that stakeholders could not get good answers to. The threat was we will export control this, and it was export control. That's the way I understand it right now.

You know, not to be not all are, but I've seen that. Or some other ones like, yeah, they have very terrible practices and anytime that anything they know gets leaked to everybody and everybody knows. We've seen an interesting situation here where you had Amazon, who is heavily invested in Anthropic, and also a partner with OpenAI, who has got a economic interest in Economically this was not a good you know, was not a good play.

for Amazon. It just showed you the severity of the security issue. I think they thought uh there was a deal'cause I think they're afraid of like this thing comes out and somebody finds out another error in MongoDB and shuts down, you know, you know you know, some bucket, you know, in the West Coast which is handling all the credentials and they got a problem.

I think that we are we might be fine trying to figure out who the competent players are and who knows. I think within agencies, you know, some a you know, you'll get a lot of times you get fights between agencies where FBI like we've got the cyber team, we want to do this, and you find out like

Yeah, you guys not to say this is true about them, but a hypothetically native it's like, yeah, you guys have been out of market for a decade and did working for government and doing low level stuff and not skilled to do this, you know, or you have, you know, other areas. So I think we're starting to figure out where these strengths need to be.

Um, I think there's, you know, thinking about like how can we make sure that the government is able to hire really good people to work on these tools and work on this stuff. I think that what's interesting, we're at an interesting point now where you have some people coming out of AI labs. Who are kind of quote what you call post money?

And doing rotations and stuff in the government can be more appealing because it can be going there because these aren't people who are all of a sudden want to go work for another AI company, but really have a vested interest in our country. We've had a couple of people come out and work with like the the the Pentagon has a project project.

Where they they basically enlist like certain, you know, career people, A tech people, whatever to do that. Maybe that's what we're gonna need more of to basically make sure that when you go talk to somebody who's got, you know, a White House lanyard or whatever that

they're out of a lab and they know a lot. You know, there was the, you know, commerce department was gonna there was an appointment. They were gonna do an appointment with somebody straight out of anthropic. And I think that caused issues because you don't want to make it look like one lab is all of a sudden going to put in a whole bunch of their people in there. Yeah. You know? But uh anyhow, I think it was a good wake up call.

Somebody who is very central to op you know uh AI history wrote a thing about like, oh, just to explain to people how uh anthropic is feeling. Step one, there are monsters in the forest. Enough with the monsters. Dario was talking about the treents from Lord of the Rings. Can let's just talk about this. If this is serious stuff, let's just talk about this in serious terms that people can understand without reverting to

Fantasy and science fiction metaphors. Let's, let's, let's please just, you know, put this in a slightly more serious idea. Anyway, that being said. Everybody is wide awake right now. The tree nths are moving faster than they've ever moved before when it comes to cyber capabilities and cyber infrastructure, with the level of intelligence that these models have and the chained reasoning that they are able.

to operate things on. And I think that that is absolutely the the case with Fable. I'm sorry, with with Mythos and Fable getting to Mythos level is is part of what got us here. And then I think it's probably we're gonna talk about five point five and five point six uh from open AI, but I think it's likely true with that as well. That the these are this is just the level that we're at right now when we need to actively incentivize

retrenching some of our infrastructure, our digital infrastructure here, uh, because the capabilities are real. They are, they are very real now.

Everything's a metaphor, it's why metaphors are terrible. I think we we we call things for what they are and I think the metaphors got us into some serious trouble there because we need to use a metaphor like nuclear powers. Um

You know, their open I announced, you know, showed that, you know, GPT five point five cyber, which is their new model for that, and they have a project called Project Daybreak, which they're working with people to sort of re you know, to find a bunch of vulnerabilities and open source software. They've talked about, hey, yeah, we're not, we think that we have a good relationship going forward to continue to release models.

And people have asked like, well, you know, are you know, will this stop everything? And the government said, like, no, this is unique to anthropic, which yeah probably has to do part with the dynamics of relationship. has Anthropic put in place a much clearer chain of command, so to speak, about these things? Because Anthropic kind of everything runs through Dario. That is yeah. That's it's it's Dario decides that open AI, you know

Is obviously uh a very, you know, complicated company in many ways, but also there is a lot of high trust on other people at OpenAI. And I'd say that's what's sort of unique about OpenAI compared to Let's say meta is that it Sam is very good at empowering certain people and saying that this person says go, no go, then that's what we do. He he doesn't tend to

you know, when it's it it seems to be pretty well reasoned out, f you follows through on that. So I'd say that there's also OpenEye has been focusing on has had a different approach towards how they handle model security. So we'll see what happens with 5.6, whatnot. But I think I'd recommend people check out what they did GPT-5 Cyber, 5.5 cyber, to look at that and to see some of the ways in which they look at how they can kind of mitigate these problems.

And there's some technological things too, which maybe we'll see happen in the next week or so, which we could we don't have direct knowledge of, but we have kind of indirect knowing that there's you connect a few dots to things, you go like, oh yeah, really well suited to be able to handle.

Do you think Anthropic throws it right to API? Do they throw it to to to subscription people? Is that the right thing to do? Do you think that any of this affects their decision with that?

And then they got a protest, like within 24 hours protest, they said, okay, we won't do that. We'll just tell you, give you a warning and revert you to there. What else is that involved in? Is there if I'm trying to use it for front, if I'm using fable, not mythos, for frontier cybersecurity, and that looks like I'm doing pen testing.

Am I going to get reliable answers? If I say, look at this entire code base, tell me what the errors are, if there's any vulnerabilities that saw a hacker could do, would it go, ha ha, you're not fooling me, I'm not going to tell you. There's going to be some question about that. You know, are and and is it there like

I don't know how much that was baked into the product, but that was a pretty had a lot of people kind of like the almost people forgot about that. And that's gonna be the question is that if I'm not if I'm doing anything other than designing a a beautiful front end for a React website,

Can I rely on it? So I do think that uh they'd promise people access for it up until like this week or whatever. And of course they didn't get access to it. I do think they have to do they're gonna still do a reset. They've been they've got more compute, they've been they've been catching up.

you know, best they can with the compute space. I I don't know if it goes into the API right away. I think they're gonna want a lot more data from users using it. I could be mistaken. There might be there might be limited rollout to API for certain customers that are relying on it, but I don't know. It's a very I think they've created a very complicated situation.

Meanwhile, across the metaphorical street, OpenAI, you mentioned their cyber announcement. I think it was uh a very Very uh purposefully done to talk about exactly how much they are committed to leveraging their state of the art uh cybersecurity model and every project that they are working on and how seriously that they are taking it.

But the big rumor And this is something that I have heard is that 5.6, their most recent uh or their their what would be their flagship model has been in testing uh out in the public for weeks now. It would stand to reason that this Thursday would be when it is launched. It is rumored to have three different model launches, one 5.6 Pro. This would be the state of the art of the state of the art. I'd be very curious to see what the benchmarks are on it.

There will be 5.6, and then there will be their brand new voice model. This will be their first new voice model since 4.0. And uh this would be essentially a 5.4. uh powered voicemail. Yeah. Those are at least are the rumors.

And people often ask health questions. And one of the reasons why they came up with the switcher and then they came up with the auto was that, you know, people described if you go to a doctor and you find out the doctor's using the free version of Chat GPT to answer your questions. Not a confidence build.

But by making the five point five instant model, they say as well as they say on and uh by the way, open a models you look at actually health benchmarks are like really category leading, like incredibly capable there. It's one of the things that gets measured less because Uh fortunately the average person has far fewer health questions per day than the coder does coding questions. Yeah. But um I think it's gonna be interesting as we try to make these

models like the the voice models more capable because that's often like we've we you and I have had to explain to friends endlessly like, ah, it's so dumb. What do you mean? I I sh here, let me show you. And they show us a voice transcript. They're like, well, you're using the voice model.

And they don't understand, but nobody it's amazing. People don't stop to think, do you think the model that can tell you something in half a millisecond or half a second might not be as smart as the model that tells you it's gotta take 15 seconds to think about it?

And again, it's just, it's, you know, if you stop, you go, oh yeah. So anyhow, that'll be exciting to see the development there, whatever. I also want to just highlight something from the GPT 5.5 cyber, which was um protecting critical infrastructure and syntax systems. This is open eyes web. We're also collaborating closely with government inst government institutions around the world to uplift their defensive security cybersecurity capabilities and protect critical infrastructure.

We have been working closely with the United States government and relevant federal agencies as we prepare for increasingly cyber capable AI models.

Like uh uh'cause I think they are they are very well incentivized to be as communicative as possible in this moment, not only for their own future, but also competitive.

He is a person that like literally is always on his phone, always doing this, whatever. And so he is extremely able to get cap and that was a story we heard. Like that was a frustration from the government was like, Hey, we tried to get hold of Dario, it took an hour and a half or an hour, whatever.

And and I think that um people like that hasn't seemed more reasonable like, well, uh I I can tell you that Sam and a couple other CEOs I know, you're if you're coming if the wide if the numbers coming from Washington D C, you're not gonna wait an hour.

The memory system, things like this, things I had earlier on where I had to restart threads and kind of start over, I don't have to do anywhere nearly as often.

Just a absolute killer for the entire industry. And for folks who are not aware, that's like when the the conversation gets too long, uh, and you know, the context window for what you're asking the model, it can only be so big, uh, to summarize essentially what was said before is the compaction. That never used to be good. It used to suck. You used to have to restart and and and do stuff. And now it's it's just seamless. It just goes forever.

And the the challenge of the time was I had to use like GPT-3 DaVinci for the conversations. Then I was using a model called Curie to sort of create like summaries summaries of the stuff. So you could talk forever.

Terribly slow. Terribly, terribly, terribly slow. And it's one of the reasons why other people played with systems like that. It took a while for people to really implement it to figure out was were we better off with just using a lot of people like, ah, just make context length longer. I'm like, well

You say that, but then you're telling me just use more RAM. And that's not you, and I often I tell people like, I'm gonna tell you what this means from a technical point of view. You're saying use more RAM. And that's the the goal is faster, less RAM, whatever. Anyhow, it's taken years to get to these things are so fast and so good you don't notice it. And I I remember how painful it was with my app, just waiting seconds or 30 seconds for a response.

Now in compaction was really an underexplored area because people just threw it all into the bag of what we'll call it rag. I'm like, well, that's that's like just saying that like library, like data science, like it's just data. I'll just use data. I was like, that's

There's a huge thing there. And I would see companies come out with stuff that like they would think through so poorly, so many different aspects. It's hard. So I think that they really do have a compaction. It's gotten so much better, so good at it. So the question is 5.6. What's going to be miserably different for me? I think speed is really the thing that I'm looking for now. Like like I think that for many of these models that I have to think harder to really stump it and do what?

Obviously, front-end design still like it got better, but it got a certain kind of better. But I would love for better front end, better speed. Speed is a big factor. We know that OpenA has their deal with Cerebrus. which is building has the super fast wafer scale chips that are incredibly capable. And we've seen GPT Spark, which is really fast, not the best model. If we continue if they continue to make the models that are running the Spark models faster, like more capable.

faster than I can snap my fingers. This is this is real time. Like you ask for a thing, it spits the thing back out nearly instantly. And uh If that's part of this, I mean that's that that's a huge thing. The big question is

That these are going to merge into one product. And I wonder whether or not this is the week.

You know, it's like, oh, we'll just throw this interface out there, see how it works. Turned out the interface was all the difference in the world. And I would say that is a lesson, by the way, that I think that OpenI has kind of realized. I don't know if other labs have realized this, because like we we literally saw When we try to tell people, oh, you should switch to codex, yeah, and you're like, well, and it's it it's it's like I it was like me making the argument, like I built a prototype.

uh playground for open I years ago that was let you switch between the just the white canvas and a chat mode, right? Yeah. And then they were and then I saw a show to be like, oh yeah, we're working on this. I'm like, oh cool. So I thought that took them took them like nine months to release it. And I'm like, okay. Well

You know, mine was ready. But anyhow, often it wasn't like wasn't it wasn't apparent to me even though how how big of a difference that would make for a lot of the hardcore users. You know, I could see the benefit of conversational. But I wasn't like, you know, Sam, you've got to do this eyebrow. I'm like, yeah, I think it's better. You know.

Um, I would say that often we really I think the the benefit of OpenI bringing in a lot of people who are really good at interfaces and studying interaction is that they're now looking at the models in a different way than just saying, you know, more, more, more compute.

the public and then then doing the super app when you can also have that team staged. There are these things just like Two huge possible a million things will go wrong when you when you start launching at scale. So you want to have the teams to do it.

Even if you have AI helping you, having, you know, uh, you know, 900 million people using your app or playing around with it is a different matter. And what happens to the all the weird ways the API stuff gets, the edge cases, um, it would surprise me. Yeah.

You know, is this just kind of like flipping the chalkboard? You're mostly web browsing, and then, you know, all of a sudden you can do really heavy duty stuff. Cause I to me, I look at codec. Even in a world where it it's wired a hundred percent for uh uh more mainstream users and is is like only getting you to a coding situation if it

You're really, really asking for something that it's suggesting. Hey, don't worry, let me code something for you, which Codex does a lot of that, you know, just sort of in the background. Uh, but even in that situation, I I want that more as like a central brain for my life. I don't need it to look at, you know. my the things I like to pay attention to online. That's that that's my big my big question is exactly how they fold

these these uh those two products together, Atlas and Codex. Ca Codex and Chat GBT, I I already understand it. I get it. It's Chat GBT. It just looks, it can just do more.

And and and I would even still I find myself going to Codex more and more because the harness in Codex is better. There I'm often now I'm often now just going into Codex and controlling Atlas from Codex. Which you'd be like, that's whack. It's the same model, well, different harness, different, different ways that it goes about doing stuff. And you know, we know that some of the Atlas team members left.

Um that you can do a browser inside Codex. It's not a complete browser. You know, you can't do things like microphone inputs, and there's a lot of other stuff. So

But my my concern is, yeah, it's buried in there. I don't use it because it's not like iMessage, I press a button and the thing is there. That's one of my concerns with the super app is like You might have a lot of great functionality in there, but if I'm three, the lesson we learned at ChatGPT was friction, friction, friction.

And there's great capabilities inside Chat GPT, but I'm three clicks away from getting to use it. And if I want to use it twenty times a day, you know, you're making me click sixty times to do a thing. That's not.

When the G seven summit happened, Donald Trump It's in the center of uh the the the dais there they're talking about artificial intelligence and at his two hands are two top CEOs. Not Dori O'Amade, he was on the other side of the table. It was Sam Altman of OpenAI and Demis Asabis of Gemini. Unfortunately, somewhere between France and Mountain View, Gemini seems to have lost a few of their key brains. Uh one of them goes to anthropic, one of them goes

To open AI. What do you make of the Gemini brain drain? Do you think that it affects where they're going? And are we, you know, are are we do we look at Gemini different today than we did a week ago?

billions, hundreds of billions, who knows how much into this area eventually. That is exciting news because anybody who wants to live a long and healthy life, and I know that that's an area they look it for, where else can they expand to whatever? So I would say that Absolutely net, net positive for people. Like OpenAI has already been doing their initiatives there. It's brought in great people there and then doing that too. To have another player there aside from there is great.

Um Gnome going Gnome was at Google, then he left and he created character AI. And then eventually that got sold back to Google, which became basically Gemini. And so he was like a Gemini lead for running on that. Him going to open AI is interesting because The you know, the rumors are that Demis and Sergei have a different view of the future.

Yes. And we talked about this before. Demis is very much into the idea that you need to, you know, beyond our transformers, beyond these things, you need to be thinking about world models, really, really intelligent systems. From the guy that created the company that made Alpha Go from the guy prior to prior to the excitement around, you know, Chat GPT and GPT three and all that.

It was the only game in town was Deep Mind. And that was the reason OpenEye came to exist was the fear that Deep Mind was going to run away with everything on that. Yeah. So Demis, Demis is a generational thinker. and and about AI, not just sort of where we are right now and very much this. Sergei, another generational thinker. You know, Sergei is a guy that understood the value of information, where they go, very much a quick study. One of the arguments I've heard that why

The reason that Google was able to make a good showing was Sergei was in there making things move, whatever, as much as you know Sundar would like. And he's, you know, I'm sure he's happy to let Sundar take the credit for that, but really kind of came down

He is A believer in the transformer model, which is really what we have seen OpenAI and Anthropic have have moved forward on is in developing that Demis says, look, transformers are nice. Well, I'm not saying that it's bad, but there's an upper limit. There's there's a limitation to it. You need to be thinking bigger with these world models. And so far we have not seen the world model transfer on a, you know, practical layer in the way that we've seen the transformer technology.

And and we know a ton about transformer models. We have hardware designed around them. We have a ton of people who know how to work them. And there's some people, you know, you know, the yawn lacoons who are saying like, oh, there's there are gonna be limited. And he gives you an argument that's out of like in two thousand and nineteen.

Because the way he describes them is not like, oh, it won't, it's gonna have trouble these things. Like, no, we're watching it solve these kind of problems all the day, all day long. Is it the most efficient way? Good I bet not. I bet there's probably much more efficient architectures out there. Might be in entirely different versions of that. What's cool about Transformers is

I can build a tiny transformer model like you do with just a GPT, and I can see really good signs of life and scale it to GPT two and see even more signs of life. World models is a very different thing. You know, opening eye uh chief scientist Jakob Pachotsky. You know, as a guy that part of his reputation was he was very good at figuring out capabilities increases based on amount of compute you throw at stuff. So

Between the Sergei Demis battle is Sergei is kind of pragmatic. Like he learned with Google data centers. I scaled up the data center and then I could do things like Gmail, do this sort of stuff. And then we were able to do this stuff. He understands that really, really well.

Damas is like, yeah, but what if there's this other paradigm and that becomes the question of which which gets you there sooner? You know, is it is it just saying, let's turn all this compute into a a new form of a world model is able to do all of these kinds of things? Or is it just keep scaling compute, keep throwing money at Gemini and keep going forward? Um, I'm an opinionated guy. I don't know the answer to that. I don't. I I I I don't I I think Demis Demis is a super smart guy. My

If you ask me where I was where I'd bet, I'd still keep betting on the Transformer because it just every nine, ten months we get these rapid speed ups in it and whatnot. Um

And like, yeah, it's good, but we're gonna need blank blank. If I went back in a time machine and even told people at OpenI, like, you know, a year and a half from now, it's gonna cost, you know, one tenth this and be a hundred times better, be like, no, and be like, yeah, all in the all these incremental things. So when it comes to if you're, you know, if you're gnome or you're John and you're inside

Deep mind, you're at Google. The problem is every lab is dealing. Even open I had the greatest pitch in the world years ago to researchers. This is the amount of compute you're going to get. Look at this fabulous showroom filled with compute. They still can do that to researchers to an extent because there may be like 7,000 people there, but the research team is still not huge. They didn't.

They didn't crazy scale it, but some of the people they've lost were people that left because they weren't going to get the compute they wanted anymore.

It's not a serious project for them. It's the same problem Microsoft had. There's a reason Microsoft gave OpenAI their entire AI budget was that it was spread so thin between so many different researchers and nobody knew who to bet on. Yeah. And right now, between If you have a different novel idea, guess what? You're not gonna win between demos and or Sergei. No. You know, so

on open source models. Yeah, everybody knew that there was going to be a turn where they were going to utilize these kind of things, but uh boy, things have not been great.

There are people that have spent their entire lives thinking about these things and have a lot of great ideas in their heads and whatnot. And they're researchers. You hear about these crazy salaries for researchers, but that's the differen the differences between, you know,

A group of people put together a paper. It's called Attention Is All You Need and it doesn't do anything at Google. And then a guy like Alec Radford over at Opening Eye, who was at Opening Eye, not anymore, reads it and says, Wow, um This is a very, very, very interesting idea here. Uh what if we if this thing could scale, then you know, that would be really cool. And so that's a thing where you you look for some of these people.

Who are incredibly capable and just sort of figure out, like, you know, what does it mean when they go somewhere? You know, what does it mean? So, anyhow, the point I'm trying to say is that meta. I know in cr there are there are there are super super A plus plus plus plus engineers.

Meta, I think, was still in that mindset of that. They they hired some key people, like, oh, we've got some key people here. We'll throw that there. It's not just that. You have to be able to work with these people. You know, you have to be able to work with above-the-line talent.

You know, you have to be able to nurture that, give them what they need and understand it, not just say, okay, go into a room, produce genius. And I'd say that's Zuckerberg is a super smart guy, and I will not bet against him, but That's a thing that I'm hesitant to say, you know, that'll be challenging.

They've got big brains that know what other big brains are getting them to their vision. They are betting gigantic sums of money on these things coming to pass and they are being rewarded for it. But that is a structure that's rare. You know, that that that doesn't normally happen. And I think it's kind of it's to Sam's credit that Sam's not the one that's doing it. It's to you know, Dario is is is that brain. He's leading that that that operation. Uh

It's it's hard. I think it it it it's not a it's not the same as as some of these other engineering problems where you where the like talent is talent. You you know that talent matters, but a little bit more interchangeable. I don't think that these Frontier Lab uh uh, you know, these the guys, the main players are really all that interchangeable except with each other.

But Sam is not I'd say that the the difference between let's say we'll put Dario and Anthropic in the c you know, out uh in a one sort of area. Difference between Sam, Elon Musk And Zuckerberg Is Sam is okay not being the Sam does not want to be the smartest person in the Sam wants to make sure that he found the smartest people in the room and put them in that room and they know that they're the smartest people. I would say the trouble that Elon had with XAI was Elon has incredibly smart.

But to bring a bunch of smart people in the room and to say, no, you're wrong. We need to do this is hard in that environment. And we saw that. We saw he lost. There were something like eleven co founders. They all left. They all

He hasn't really ran a software company since PayPal and he was organizing Peter Teal, Max Lechkin, a bunch of others. And then revenue wise, the revenue for X, I like I'm on X all the time, but the revenue went down, all these other things went down, and and there's a lot of features they don't support anymore. And then we heard us to do, yeah, I'm like, it's gonna be a challenge because

It's not the same as hardware. It's a different pace than that. And and you're in a team, you're in an environment where the d you if you want to build cars, you can work at Tesla. If you want to build spaceships, SpaceX still wear it at. There are other great companies out there, but you go there.

You want to work in AI. There's a bunch of different places to go work. And that was going to be the challenge he had to deal with. And same with Zuckerberg. Zuckerberg was like he he thought he would just write checks like he was building an MBA team. you know, uh MLB excuse me, M L B major league baseball team, sorry. You know, he thought he was gonna write checks like that.

Jakob, Sam, Greg, they've got a united vision as to where this thing is going to go. Now, where the applications are from that, there's disagreement. Obviously, they went through a big pivot a couple months ago as to what they were gonna. uh keep going, but they are gonna uh w w what they were gonna shudder. But the scaling up of this tech and the idea that it's that it's going to be as robust as it is, that's a singular vision.

They are uh uh uh betting billions and billions and billions of dollars in compute to execute that. If you don't have a united vision from your your big brain, your CEO, and everybody that's executing below it, like I think it's hard. And and that's and that is not the way that Google was built. That was not the way that Apple was built. That was not the way that

Um that that uh uh meta or Facebook was was built. You know, there's are similar lessons, but it's not it's not the same at the scale at it as as it is here. This is kind of a new way to go about things. And it's it's a little bit hard to teach some old dog new tricks.

And they have a new compute, supercomputer coming online. And and sometimes you'll hear like people I talk about like, oh, Elon Musk, he did colossus. He threw a hundred thousand throwing a hundred thousand dollars into GPUs into a system and and using, you know, some gigabit Ethernet to connect it all. ain't the same as knowing how to make a hundred thousand systems. Like that was one of the things they found out with Colossus One was it was just

very inefficiently run. It's fine for doing inference and stuff like this and classes to sort of say mistakes. A lot of people were looking at that like, that's not how you do a bro. Um, and it's not just a matter of scale. And OpenAI has a

uh a very, very massive amount of compute coming online for doing new model training. And we're going to see this acceleration. And I think that's one of the things that we're we're we're seeing people trying to meet demand and continue to training. We're going to start see the acceleration even more so.

For anybody thinking that it's going to get simple, it's it's not. You're going to get to a point. There are metrics that I saw internally basically describing the rate at which you would see, like, you know, how fast it would take to do GPT-4-level models. And it would go from months. Too much faster than that. And and extrapolate that to where you want. And we're going to be able to see the ability for new models, new capabilities that come out at an incredible pace. And

Yeah. Yeah, we gotta get out of here. Unfortunately we won't have time to talk about some of the political uh uh back and forth uh between open AI and anthropic, uh a big, big fight, uh that we can uh, you know, I'd be very curious to see how New York twelve. uh votes. So if you are in Manhattan and that is your district, if you are represented by Jerry Nadler, I'm very curious to see how you decide to select uh

who is there, we can uh we can take a look at all those results. I I I very much wonder whether or not online is real life. Uh You know, or or whether or not those actual constituents care about some of the issues that have been made front and center there, but we'll find out. Yeah. In the meantime, Mr. Andrew Main, where can people find you?

Uh, the movie The Complete Unknown, whole team working on it, trying to bring this to television somewhere near you soon. So that's exciting, but we'll see.