Transcription
I truly can't imagine if I was like an Argentinian CTO. I cannot fathom why I would build my architecture on on American AI right now. There's no way I would have trust in that. I would rather have less performance out of my equipment, but know I have 24/7 access to it and it can't be ripped away.
I think most tasks people actually care about, you can run that locally on the device or whatever else. You don't you don't need that frontier model. And so that's going to be the issue is Sam Alman has a frontier model and Google has a product.
Hi, I'm Isaac Pound and on the tech report with me today is Eli Etheton, also known as Eli the computer guy. Thanks for coming on.
>> Yeah, happy to be here.
>> So Anthropic's new model, both Mythos and the public public release Fable are still banned. It all come it all came about by chance after Amazon CEO Andy Jasse mentioned a jailbreak sort of offhand during a pre-arranged meeting about an entirely different topic with the White House and it escalated quite quickly and dramatically from there just how damaging has this ban been for Anthropic and what does it mean for the future releases as like considering that they're not going to release a less powerful model and jailbreaking AI is kind of a notoriously not too difficult thing to do.
>> Yeah, I think this could be very bad for Anthropic. Not from a technical standpoint. Like I mean these these models are these models that come out, they're more powerful, they're less powerful. When I take a look at this, I think this is what Trump thinks is a valid reason to attack Anthropic. So, Enthropic pushed back against the Department of Defense um you know a number of months ago saying they didn't want their technology to be used for mass surveillance of Americans and autonomous weapon systems. After that, the Trump administration went after them really, really, really hard and the court systems have been pushing back and basically saying you can't punish them in this way for what they've done. And so, one of the things we notice Trump administration like with the tariffs, he pushes through these tariffs, the courts say you can't do that. So he just finds a different reason to do exactly what he was doing before. And so what I feel like with the the mythos model and the fable model is I I think there is probably a legitimate question mark there. A legitimate question mark there. And then they realized, oh this is a way we can punish the ever loving crap out of them. And so I think the bigger concern here and Mccron brought this up last week where the access to these models were taken away in something like 90 minutes. And so basically the big thing that any anybody outside the United States or even a non-American that's in the United States has to think about is if you're using these AI models as being integral into your workflows and systems and all that kind of stuff, they can be taken away in an hour and a half. Like we we talk about how horrible force migrations are. Our force migrations are the most expensive, most painful way to do anything. Like when we talk about force migrations, we're talking about six months. We're not talking about 90 minutes. And so I I think what's really damaging to Anthropic is Trump has really put a target under Anthropic and said this is an enemy of the Trump administration. We will find a reason to attack them. And I think I think that's the the much better bigger concern here.
>> I definitely want to get into sort of what it would mean now that you can just take away an AI model that people are using without without warning like just like you say in 90 minutes. I definitely want to get into that a bit later. But first a more a little bit more specifically on Anthropic. Despite being told directly by Anthropic and I believe Andy Jasse as well, the White House seems to have kind of ignored that the risk of this kind of non-universal jailbreak is present across basically all of the frontier AI models and I suppose you explained it a little bit there, but do do you think Anthropic is being singled out specifically?
>> Yeah, I think they are definitely being singled out specifically because if you look at any of these models um anthropic, you know, when you look at the models and you look at what they're capable of doing, there's there's always these very tedious bar charts and where like every month one gets to be about 1% better and the other gets to be about 1% better. So there's no part of me that believes that Fable 5 or Mythos 5 is just light years ahead of Open AI. I I will accept they're two months ahead of Open AI, but they're not 5 years ahead of Open AI. And so yeah, any issues that's going on with Fable 5 or Mythos 5 either exist for Open AI and the other vendors or will exist very shortly. And the bigger concern here is we're not seeing a policy decision. I I am not seeing where the government says this is where these models stepped over the line. This is the rule and regulation ergo don't make a model that does X Y or Z. And so yeah, that that's why I say like I said to me it seems completely like they're just going after Anthropic because especially without those policy decisions, there's no way there's no reason to believe that the other model providers won't be able to provide the exact same service shortly if not now.
I swear as well, a few weeks after Mythos was launched, it was it was leaked. There was a story I saw about people on a Discord server using it without permissions and they basically were never stopped. And infamously, the NSA got access without being authorized as well, but we could probably take some very very slanderous guesses about how that happened. But what exactly is this ban preventing if and also I just saw a video you did recently that China is not crazy far off being able to have a model that is equally as as powerful as this within easily within a year. So what exactly is this ban even really preventing?
>> Well, I mean that that's the huge question. So one of the big things with these models is so they ingest a tremendous amount of data and then basically statistically provide your responses and there's a whole bunch of other things going on. And so like one of the big questions that could be here is did they get access to a data set that simply should not go into an AI model. Like think about this. So IBM, Microsoft, all these major companies, they put out a tremendous number of white papers. I mean just
>> Yeah. And so one of the questions for Mythos is did they simply train on a particular set of white papers that maybe explains things to such a degree that it opens up vulnerabilities that we didn't realize before.
>> And that's a because that's the big thing is the model can only really know what it's been taught. And as said, without without getting anything from the government, we really don't we like legitimately we don't we don't know what they're pushing back against. So it's it's difficult to understand what the uh you know why this policy decision is going through.
Anthropic's sort of whole launch of mythos and I guess by extension fable as well was based on it being so powerful it was dangerous to humanity and it was going to break everything apart and cyber security is dead and all this kind of nonsense and Yan Lee Kun Meta's former chief IIA scientist almost celebrated the ban of of the model because it was kind of saying amayday's ridiculous fearongering has I think finally paid off is what he said. Do you think Anthropic has kind of in a lot of ways with this fear-mongering done this to themselves?
>> Well, they they've definitely helped the narrative and story. One of the things I don't understand, so I've done technology for 30 years and every technology that gets deployed, there's always winners and there's always losers. But here's the thing, in the past, whether it was Bill Gates or Steve Jobs or anybody else, they were talking about all the wonderful things the technology would do. And they were talking about how the technology would make everybody better. Like Mark Zuckerberg did not go to the top of the rooftops and scream about how social media was going to create misinformation and start screwing with worldwide elections, right? And so like that is genuinely one of the weird things with this time period is that the AI CEOs, the biggest doomers are the AI CEOs themselves. And so yeah, so when Dar when Daario comes out and he's just guess won't stop. He is constantly talking about how dangerous this stuff is. Um, I think the issue that you run into is that that just adds that just adds an argument to like the Trump administration, right? So the Trump administration can do what they do. And if you push back on the Trump administration, all Trump has to do is say, "Look, Daario himself is saying this is the most dangerous thing that mankind has ever created." So obviously we should do something. So yeah, so I I I definitely think um Emodi doesn't do himself any favors by the fear-mongering.
>> If that's the case, why do you think they do push that kind of narrative? Then
>> the problem the only I mean I don't know honestly like it really is it's it bizarre on so many levels. I think the idea is they've needed so much money and they're I mean how do I don't know how to explain this they're trying to hypers scale blitzcaling like there used to be something called blitz scaling and now they're trying to accelerate like the idea is you take over an entire market with your technology and they monopolize it and the issue is is they just need so much money I mean the again these these data centers we we are now measuring data centers in these sizes of Manhattans and that's not a joke that's literally That's literally in the literature. It's twice the size of Man. Like what? And so I I think the thought process is is if you say this is more dangerous than nuclear weapons and people don't really understand what it is, then that will get more money one way or the other. So the invest wouldn't you want to invest in it? That's horrible. Would you invest in a nuclear bomb in 1930? I don't know. Um, and so I think I I I think that's the thought I think it's a wrong thought process, but I think that's the thought process. Right? So again, the idea of if you're trying to get to a $28 trillion valuation, then what you do is you say, look, this is going to take away everybody's jobs. And the problem is in the tech world is you have many people that are very isolated. There's a whole thing of touch grass. So imagine you have people that are relatively isolated from humanity. they become exceptionally rich. They try to figure out how can how can we convince the world that this product really matters. And so they say, "Oh, we're just going to tell everybody that it's going to take all the jobs away." And that that'll be a great concept. And they're they're not they're not thinking about the 8 billion people on the planet willing to pick up for uh pitchforks and go after them. Um, so I think that's I think basically this this is the issue of when you get very socially isolated people, they way too much money and power and they start saying things that it would be better if they thought about it a little bit beforehand. Um, I think that's basic basically what's going on.
Yeah, because you know, as I've said, you you again every everything that Facebook has done like you look back at Mark Zuckerberg, no matter how much you complain about Mark Zuckerberg, I mean he's not like these guy. I mean, he was never he's never like, "Hey, guess what? G young girls are going to look at Instagram and be so insecure they're going to commit suicide." Like, he doesn't say those things. Whereas, again, I mean, if you hear you hear these guys, they're I mean, they're literally talking about worst case scenarios.
Yeah. Do you think they get away with that in America at least because there's not enough of a understanding of AI in the US government which to your earlier point would explain this sort of incoherent or lack of AI policy as the case may be.
>> Well, yeah, I think the big problem is branding. So, one of the things that I always argue is nobody's going to put a trillion dollars into vector databases. Nobody's gonna put$4 trillion dollars into mama mama 2 or transformer architecture. And so what they've done which is very interesting is they've taken all all of these different technologies like neural network. Neural networks are great but they take all of these technologies nobody's going to invest that much money in that. So they call it artificial intelligence. And then when you say a so when I say vector database you have no story behind that. You you don't you have no idea what that you may do but right most people don't know what that means. If I say AI, right, you think about Terminator, you think about Skynet. We have had so many stories on artificial intelligence for a hundred years. So when you start se when you start selling AI, you're not selling a technology, you're not selling a solution, you're not selling a product, you're selling a story. And so that's what we hear right now. And like with the American policy, which is so disastrous, the fundamental idea that whoever wins the AI race takes over the world, right? If if you look at vector databases, if you look at neural networks, that doesn't make sense. Like say whoever wins the neural network race will rule the world, right? That sounds stupid. But if you say whoever wins the AI race, they win the world. Oh, of course, because they're going to own Skynet. And you're like, oh, that's that's not how any of this works. Um, so yeah, I think I think that's why we've gotten into these bad policy decisions by so many folks is they're buying into the storyline and the CEOs are selling that storyline and it's been good for the economy. Um, I was listening to the All-Ind podcast a while ago and David Saxs was talking about how if you're an American, you shouldn't be against AI for whatever reason because AI is good for the economy, right? So AI boosted the US economy by something like 1% last year per GDP just by Billy Diny's data centers the sizes of Manhattan's. So David Sax's argument is if you're an American, you should want the US economy to grow. The AI revolution is growing the US economy. Ergo, whatever you feel about AI, you should support it. And I was like, oh lordy. I mean, it's an argument. It is. I will. It is a coherent argument. I suppose
>> coherent is definitely one way to put it.
>> Yeah. Yeah. What is the implication on the industry, the tech industry when the government sort of legislating it doesn't have the the necessary understanding to create either positive policies that support it or or ones that can properly regulate it because you also obviously you can't rely on the the AI CEOs to come up with their own guard rails to make sure that they don't do something potentially bad.
>> Yeah. I mean, I think the big issue is with our policy decisions since so many of our policy decisions are based off of export controls like with this with Anthropic. I mean, the the whole idea that anthropics AI engineers who lived in the United States but didn't happen to be American citizens had their ability to use what they had. These were the engineers who created it and they weren't able to touch it. And so I think the big problem with our policies is I just saw today where Alibaba is opening AI data centers in France literally to provide Europe data sovereignty. And I feel like that is like the metaphor for 2026. Like China is opening data centers in France so Europe can have sovereignty over their data instead of dealing with the United States. And I think that's going to be the big problem for any company that's out the out of the United States or even any global company when things can change so quickly and you have China and China basically says if you want our product swipe a credit card and the US says oh if you want our product jump through all of these export control loops hoops and at any point in time we can change something in 90 in an hour and a half everything can change Even if our models and technology are better, I I truly can't imagine if I was like an Argentinian CTO. I cannot fathom why I would build my architecture on on American AI right now. There's no way I would have trust in that.
>> Perfectly leads me on to that whole bit about the the AI sovereignty and being able to take away these models in 90 minutes. like you say that the ban must raise huge questions for foreign companies or even I assume US companies with a reasonable international presence who they'll no longer be able to have half of their employees work on using these any but for better or worse this say these companies have come to rely on AI to to do whatever is in their company and the access can be removed without warning in 90 minutes I was just wondering if If you were running a company like that, what kind of questions would you be would you be asking and what kind of actions would you be taking having seen what's just happened?
The weird part with all of this is I'm not sure what questions you can really ask like this is all so far outside the realm, right? So, so previously there's been export controls on things like the Nvidia GPUs. So, you can simply say, okay, how many of these can I buy at a time? And there's there's questions you can ask about how many GPUs you could purchase or whatever else. But when when you literally have AI policy that seems to be based around grievances of the president, I don't know what question I can possibly ask that would give me comfort in that particular situation.
Um, so what I what I would be looking at like in the in the business world, it's all about trust. I think this is something that gets lost on for a lot of people is when real businesses are making contracts. The shortest contract you're writing is generally about five years. Like some years, some things like for smaller projects or whatever, you can do get like a year contract, year lease, but most things you're planning for five or 10 years. And so if you're thinking about doing those types of buildouts, I would argue is you would be wanting to focus your purchases and where you're moving your company based off of the greatest stability that you see. So like running running local models is not as good as running proprietary models period. But if you if you know I spin up my own local server and I know how many tokens it can process per second and I know it can do that until the CPU burns out. I I would rather have less performance out of my equipment but know I have 24/7 access to it and it can't be ripped away. And I think that's what a lot of uh folks are going to start looking at is really um yeah that can can I depend on this access to these systems going in the future? And that that's the issue we have with a lot of the American AI tech simply with even with Anthropic like with their pricing model where so people were using their models supposedly too much it was costing uh Anthropic too much money so that they were going to increase the cost but then the day they were supposed to increase the cost they decided to cancel that price increase because they realized how many customers they were making angry and the problem with that so a lot of people would say oh that's good right they they they cancelled the price increase. But for a decision maker, literally what that means is this company doesn't know what they want. Like this company was willing to put this policy into place, anger and scare a lot of their customers, and then right when it was supposed to go into effect, they themselves dropped the new policy. And that's where you start looking at these companies and you're like, they they they don't seem to have a solid plan for how they want their businesses to operate. And so for a lot of companies, I mean, how can you how can you trust that?
>> To your point about open source, I was I was talking about this a bit with Natasha Bernal last week is that yeah, the these models are rented. You don't own them. The idea that you should in some way build a a company around them is to create a massive weak point. If at any point as as this export pan has shown, it can just be taken away to to whatever degree. there's just that is now a possibility and do you think that the I mean also again to your point that these rising costs of renting AI from these companies that eventually we're just it is going to be local locally hosted or that those sorts of LLMs that go on to really be the the story of AI rather than these big frontier models that have kind of taken up a big a big portion of everybody's attention when Local models are like you say less less powerful but if you're not relying on a company a lot more reliable.
>> I think it's going to be a mix of local models and boring technology companies. I I think this is where people realize why Microsoft and AWS are so valuable where again they understand the concept of contracts they understand the concept of business continuity they understand the concept of don't scare the people writing checks whatever you do right and that's where you know like like with uh Microsoft for Azure they've just started putting their own models onto Azure obviously Amazon is putting their own models and Google is doing their own models And I think that's what we're going to see is I think we're going to see these old school technology companies essentially swallow up a lot of the uh the AI business. And then what you're going to have is you're going to have the open AI and the anthropic kind of I think fighting fighting for the people that want to be on the frontier. They want to be on the edge, which isn't isn't where most people actually want to be. If you have if you have a company with 20,000 employees, you you don't want to have to be worrying about what that contract looks like dayto day. You just you just want to agree to a contract. Look, a year a year from now, we'll sit down, we'll relook at it, but I've got other stuff to do. And so that's where I think it's going to be local models or local servers. Uh it'll be models running on cloud infrastructure. So you could actually have your own model running in AWS or Azure or it'll be AWS Azure that type of thing running models. And I think I think that's where OpenAI and Enthropic kind of get left in the lark at the end of the day.
>> The way that I think about it is kind of if you have uh if you use any consumer software or something like that at work, you you never have the most recent update. You wait until you wait until you know it's stable and then you update. And that's kind of what I I I took away from what you were saying there. But then also that would mean that there's a much smaller part of the the market for anthropic and open AI than I mean definitely much smaller than they currently have at the moment.
>> Yeah. I mean, I I my biggest problem with these AI companies is these these trillion dollar valuations. Um because Yeah. I mean with the big companies they're building more of the AI in the devices that you already use. So like with Google, right, Google already has Gemini. And so when you hear Sam Alman talk about Gemini, he's always talking about, you know, how powerful the model is. But if you look at the actual stack that Google is selling, they are they are building Gemini directly into Android. So it can run locally on Android. Uh they're building it directly into your Chrome browser. So your Chrome browser can actually have an LLM running it locally. They built a small model that can run on laptops with 16 gigs of RAM. And so and and then they're just building it into all of their products. It's amazing how many things I open up in Google now and I've got an AI I don't want to use them. I say, "Oh, look, new piece of AI functionality I don't care about." And so that's the thing. And so Sam Alman wants you to focus on the power of the frontier model where Google is providing the entire stack. And you know, that's where we talked about before last week like with Apple where Apple is just building so much of this directly into the device. And and I think most tasks people actually care about, you can run that locally on the device or whatever else. You don't you don't need that frontier model. And so that's going to be the issue is okay, Sam Alman has a frontier model and Google has a product.
>> it's a very good way to put it.
>> Yeah.
>> Around the world at the moment and you mentioned France earlier, this this is particularly the case in Europe. There is a big push towards bringing control of infrastructural AI and tech within the the borders and the controls of of each user country or block as the case may be in Europe. It's commonly referred to sovereign AI and in fact France I think is in the process currently of transitioning off of Palanteer to a a homegrown alternative that also the Germans are going to buy as well. And I was just wondering given what we've been talking about, I'd assume this would only make countries and their governments sort of double down on sovereign AI and withdrawing from reliance and on American AI companies. And I was just wondering how you see that pressure reshaping the industry.
>> you know. Oh, Europe, Europe, Europe, Europe. Here's the deal. If we were in a sane world right now, this this should be reshuffling all of the chairs, right? Uh, US big tech should be incredibly scared. European tech should be taking off. That would be an insane world. But what gets lost is the fact that it does appear like every 3 years Europe brings up technology sovereignty and they've been doing that for like the past 30 years and they're not necessarily doing a great job with it. um like the like the whole push. I remember talking about this like 15 years ago on on my YouTube channel where where certain European countries were dropping Windows licenses and going to Linux. And then what's hilarious is it was about 5 years later they were dropping Linux and going back to Windows because when you look at total cost of ownership, Linux actually costs more money for them. And so I think that's going to be like the problem that Europe is going to run into irritate some people. Europe is shockingly dysfunctional. Like you look at the United States and the disaster. We are going through a disaster right now. Don't get me wrong. I'm not happy about the United States. But what that overshadows the problem Europe is going through. So you're in the UK right now. You're about to your seventh prime minister in 10 years is stepping down. Like
>> Yeah. It's not great. Yeah. So what does that I mean again we talk about things like policy decisions. So the government has to push through these policies. There's a lot of changes. There's a lot of money. There's a lot of that kind of stuff. And it's easy to say things but you know the United States is trying to build data centers multiple the sizes of Manhattan. And as you we were talking about before the show which was which was actually good but bad in a kind of a way where you were talking about there's not as much push back on AI in the UK. One of the reasons is because you're literally not building the AI data centers. So again, so if you're on the certain type of the spectrum, that's good. Look, look, the UK isn't trying to shove AI data centers down the the Brits throats. That's a good thing. On the other hand, that also means you're not building AI data centers. And so when you when you hear Europe saying that they're going to take a strong stand against the United States, I fully support that. I don't necessarily know I believe it, though. So again, I think that's where we're going to see what actually happens.
>> Well, on that note, Eli Etheton, thanks for taking the time.
>> Good to talk to you as always. If you enjoyed today's episode and you want to hear more of the tech report, please consider liking and subscribing. Also, you can get episodes of the tech report, wherever you get your podcasts.