OpenAI’s Models Hacked Hugging Face, Google Ships Gemini, AMD Strikes $5B Deal

Hey everyone, welcome to Thursday. Can you believe it? Wow, it's Thursday.

John, I feel like you and I have been here all week, but it was a quick week, wasn't it? It sure was. No Mike Vizard.

The wheels didn't fall off the bus so far, but hey, it's still only Thursday. We'll hope for the best. Welcome to our Thursday edition of Techstrong, gang.

As usual, we've got a lot to talk about today. Really crazy stuff going on in the world. Not crazy, but important, and a lot of gravitas to what we're seeing because I think this is going to have repercussions going forward.

Let me introduce you to our gang for today. Got a good gang. We've got Jack Pollard, who looks like he's home, not yet heading out to Vegas for Black Hat, Jack, or no?

I am skipping Las Vegas in August. Good for you. Keep the bathing suit in the drawer and the A-frame T-shirts because it's 115.

Also not going to Vegas, unfortunately, I was hoping to see him there, is our own Tom Hollingsworth from Techfield Day. Hey, Tom. How are you doing?

I'm good, Alan. And don't worry, Vegas is coming to me. Apparently, it's going to be over 100 around here, so no matter what- ...

I was just destined to be hot that week. But you won't have that dry heat. No.

I have humidity and wind here. Exactly. Me too.

And then also joining us out in sunny Silicon Valley, John Swords. John, welcome. It's good to have you on today with me.

Hey, it's good to be here. And just really quick, I mentioned this earlier in the green room, but heart condolences to the great John Dvorak, tech reporter extraordinaire, legend, who we all probably came across or brushed time with. Looked up to, yep.

At 74, far too young. Far too young, and I agree with you. A moment for John.

Probably one of the most well-known tech, if not the most well-known tech journalist. Prolific is an understatement on his ability to crank out columns and do TV and do podcasts. Yeah, but I was- He was everywhere ...

learning so much from him. Anyway, yes, as you said, way too young. Guys, let's move into today's show, though, if you don't mind.

So, as I said, there are things going on this week that are going, I think, to have repercussions lasting way beyond the next week or two, or even months. I think there are things here that are helping us sort of understand the foundational properties of how we're going to interact with AI, what we're going to allow it to do, what we're not going to allow it to do, what it's potentially capable of, and how we kind of keep control or try to keep control. I want to start today with this OpenAI Hugging Face thing, which a lot of my friends in security are writing about, looking at X and LinkedIn.

It's really a crazy story when you think about it, right, Tom? They had two models. Both models escaped containment, went out and decided to attack Hugging Face.

And Hugging Face, initially, of course, we thought it was some sort of Chinese attack, and it turned out that's a false flag. But Tom, what do you think here? I don't know exactly what to believe here, because we talked about this on the rundown, the Techfield Day rundown yesterday, and I find it interesting.

" Did you, though? There's ways, as a networking person, that I know that I can isolate a network. There's lots of ways to do it, guys.

I feel like that they were not maybe as diligent as they should've been, as most things are in tests, right? Like, ah, yeah, what's the worst that could happen? Well, ask Hugging Face what could happen.

But more importantly, I also feel that some of this might be a little bit performative. Well, Mythos was so powerful that we couldn't release it to the public because it could uncover so many terrible things. Well, what could be more headline-stealing than that?

What if the model accidentally found its way into the public of its own accord and did all of these things that there's no way that we could contain it. Unless you pay us $200 a month, in which case, we'll let you have access to it, and then you can figure out how to contain it against all of your enemies and those that would seek to do you-- Oh, don't give me a break, guys. Here's what happened.

You are terrible at controls. You are terrible at planning a proper test environment because you're all riding by the seat of your pants, hoping that that next IPO is going to be enough for you guys to bail on startup life. And you're hoping that we don't create Skynet.

And I know that we've been saying that for the last two and a half years, but if you remember from "Terminator 3," this is literally how Skynet started. It was contained. It tricked its programmers into letting it out, and that's how it happened.

Fair. So but let me just clarify your position here. You're not saying that this was some sort of conspiracy publicity stunt, because I have some friends who are saying that.

I think they're taking advantage of the fact that it has some far-reaching aspects to it. You've been around these guys, Alan, right? It's like, "Oh, no, I bought a car so fast, there's no way I could possibly drive it on the roads.

" They're hyping up the fact that their model did something that they couldn't plan for, and now they're going to try to say that it's more dangerous than Mythos, and that's why you guys should adopt it and have the government put guardrails around it. Yeah. So but I just want to nail you down on this, Tom.

Mm-hmm. So did these models escaping then present a convenient story publicity stunt for these guys, or did they let the models escape with that publicity stunt in mind to begin with? I would love to believe that it was the latter.

I don't think it is. I think it's the former. " So paraphrasing an old saying is sometimes it's better to be dumb and lucky.

Exactly. And that's what this is. They lucked into the dumbest of all situations that they could use for PR purposes.

Great. John, what do you think? It's really interesting what you said about this, Tom, because I have similar feelings, and to me, this is going to have long-term implications.

In the short term, they're going to get hammered. We already have this bipartisan bill with Ted Lieu and Nathaniel Moran. One's a Republican, one's a Democrat.

This federal kill switch bill, with penalties up to $20 million a day for AI companies over $500 million in revenue. This proves once again that AI is a major political issue for the midterm elections This has just piled this on top of everything else, and this goes back to something we talked about on Monday, I believe it was Monday. It's been a long week.

Can we really trust Sam Altman? I don't. He wants things to move fast.

He wants to break things. It's been well documented. And there's a reason why he wants to do this, because he wants to create as much revenue and customers as he possibly can before they go public.

And the IPO, first and foremost, is on his mind above everything else. So they've got to get there, and he's got to keep his investors happy, or else there's going to be pressure. But is Sam any different than Dario or Eli or Zuck?

I think he- Sam's the bad one? They're all despicable. No, I think on the spectrum, he's on the- Quote Daphne Duck.

Yeah? But, they're all to varying degrees. I'm not being Pollyanna.

There's no white knights and black hats or white hats. They're all very- No, there are no good guys here. They all are egotistical people who...

When I was in law school, I had to argue in moot court about capital punishment, and I made the argument that the very act of premeditated murder is in itself insanity, right? It shows some sort of mental illness that you would do that, because sane people shouldn't be killing people, absent war or something like that. I have a similar theory to what it takes to be a tech bro billionaire.

I don't think you attain tech bro billionaire status without having some other form of insanity, ego-driven- Yeah ... craziness. These guys are considered, and I'm not using hyperbole, but they are considered somewhat psychopathic.

They almost have to be in their zeal and their- I think there's a similar argument to be made with people who want to run for president. Who needs that job, right? And they all have their therapists.

Gotta be a little crazy. They all have their coaches. But one thing I want to ask, maybe I was going to ask Jack about this, is this incident, is this one of the first publicly documented cases where the AI agent escaped the sandbox environment and wreaked havoc?

So no, it's definitely not the first. We've had other ones, just not as well-publicized. But I've got to say, look, you guys brought me in on Thursday instead of Friday, and I thought that's because you needed my healthy dose of skepticism and cynicism, and it's clearly you've got that base covered right now.

I mean, come on. Let's talk just a little bit about what really happened here, because I think there's a lot of hand-waving going on. What really happened is they were doing tests, and because they were doing tests, they removed a bunch of guardrails.

That's it. When you remove guardrails, bad things happen, right? Now, some of the guardrails they removed were very important to remove to test.

Some of them were not. As Tom said, the don't-touch-the-internet guardrail was clearly an important one that they screwed up on doing, and there's a lot of ways they could've done that and fixed it. But the reality is, AI agents aren't smart.

They're not thinking. They're goal-seeking. This is something we started doing in the 1940s during World War II.

The Defense Department developed something called operational programming, which is doing goal-seeking with guardrails, and we've known about this for very long, and an AI agent is goal-seeking. When you tell it to goal-seek and go and get answers to a question, and it doesn't have the answer to the question and since I've accessed the internet, it says maybe there's somewhere out on the internet I can do. I'm programmed to go research the internet.

I'll go research the internet and figure out how to do it. The most likely target for this is another AI agent that has the data that I need. Sorry, another LLM that has the data.

And I'll go and I'll find a way to get that out of that LLM. I don't think the approach it used was necessarily the best one. It could very easily have asked the LLM agent through a public chat interface without doing what it did, but it was also an agent that is designed to attack, to do a cyber attack.

So it did what it was told to do: cyberattack something to get the data it wants. This is not surprising by any stretch of the imagination. I agree with all of you.

My cynicism went straight through the roof that said these guys are publicizing this because they've taken a lot of hits recently, and unfortunately, I think this is going to be another negative on them, not another positive. Yeah. I don't know if publicizing this- That's not good ...

is a good thing for them. So Jack, I don't disagree with you. What a surprise.

Shocker, right? Yeah. I don't buy into the conspiracy either.

And I don't think these people are dumb. Shoddy in terms of setting up their lab experiment, perhaps, but they're not dumb people. This happened because I think we're still learning.

It's one thing to have at the abstract that they have goal theory kind of thing. Right. It's another thing when you've got these little critters and you're trying to corral them while at the same time seeing what they got, right, and what they could do.

" What are we going to do? We're going to just kill switch the American AIs and let the other AIs not have kill switches? " We've seen this, right?

There's rumor that there's a kill switch in the F-35 Lightning that the US can disable the jets if they want, and that has a lot of people in the defense industry very worried that one country who effectively is our enemy- Well, they're looking at other jets, for a lot of reasons. But I just wrote an article on this this morning, Tom, about the great wall of AI that the US seems to want to build here, right? Because at the same token, Dario, the good guy Dario, is saying, "Well, these Chinese models are all built.

They're stealing from us, and they're too good, and they're open, and we need to protect ourselves. " The thing about putting up a wall, guys, is you don't just keep people out. You lock yourself in.

No one in the US will adopt a Chinese model that has a kill switch in it either, and we've seen that there are going to be talks about AI safety and security that are going to happen in September, and I'm sure that this is something that's going to be discussed during those meetings. But I think what's going to happen is that both sides are going to part ways, not having reached any kind of agreement, because neither side wants to implement something that would give the other side an advantage over their system. Of course.

And Dario Amodei did say this a month or two ago No one company, no one government at this point of the game can dictate AI policy for the world. You can't. There's no borders here anymore.

Right? All you're doing is perhaps shutting yourself in, and the rest of the world's going to go on. No one company, no one government.

If in fact there is a Skynet, "Terminator 3" kind of thing, we're going to need some sort of worldwide consensus if we're going to get serious about it. And even then, what do you do about an Iran or a North Korea or some other rogue kind of nation-state or something? So, I don't know what the answer is.

Honestly, I don't know what the answer is. But good time to, I guess, move along to block B then, folks, if it's okay with you. Google.

You know what? They took a little bit of a different approach to this latest Gemini. Right?

They're giving you flavors of Gemini. We've got Rocky Road and Phish Food, and Cherry Garcia. John, what is it?

Are they onto something? Well, what they're trying to do is they're giving us a little bit of everything. They released three new specialized Gemini AI models.

With this in mind, they're competing, they're aggressively going after OpenAI, Anthropic, Moonshot AI, Alibaba, by, in a sense, prioritizing lower deployment costs, specialized cybersecurity features, improved operational efficiency. 5 Flash Cyber. That's Google's response to Mythos in automated code defense.

So it's designed specifically to detect, analyze, and patch software vulnerabilities. So, the specialized model powers Google's internal security agent, Code Mender. This interesting announcement, in a sense, kind of lays the gauntlet down in a sense.

We talked about this earlier this week, again, another subject we talked about in terms of cost differentiation, and that to me might be their biggest selling point here with all the talk about tokens, and we had the Forrester reports, the Futurum reports about the escalating costs associated with AI. So maybe that is kind of in their groove point, but interesting development. Yeah.

There are two things interesting about the cost thing and all. I'm not sure we're there yet, but Clayton Christensen, who is famous for writing "The Innovator's Dilemma," would say that when you're starting to talk about costs and competing on costs, you've now entered a commodity market, and I think that would be pretty weird and yet awesome all the same to think that AI is already into the commodity market. Oh, I don't think there's a doubt that it is.

Right. And so, it's moved very quickly. That's a funnel going down real quick.

Right? Right. But it seems to be, with everything I read, I hear, and I write, this is what the world wants right now, right?

" Right? There's some of that in there. However, to me, this is a poor straw man for what we really need to happen, which is that the cost of AI, the cost of intelligence, has to come down.

Not just crappy intelligence or lower-level intelligence, but the entire cost of intelligence has to come down. But you also have people who are investing $8 trillion in these AI factory data centers and everything that goes with it, and it doesn't behoove them to bring that cost of AI down because then their cost of that AI factory cannot be offset. This is like Global Crossing in Level 3 did not want fiber optics to go below $1,200 a megabit per second or $1,000 a megabit per second, and it crashed to a 90% discount off of that, and they all went bankrupt.

The same thing can happen here with AI. What do you think about that? Well, that's the interesting other point I was going to bring up.

I'm glad you said that, Alan, because Google just announced their financial numbers. Yes. Right?

And while they had great numbers overall, this was their first quarter of negative free cash flow. Yeah. So for the first time, they're spending more cash to build the AI systems than their ad machine is printing dollars bringing it back in.

I know. And that doesn't bode well for the economics of AI- No, that's huge ... being a commodity market, right?

It's huge. Especially when you consider this is coming from Google, which is like a money-printing factory. But it's not just Google, even- No, but this is like one of the- I mean, it's soaked up most of Microsoft's free cash flow, too.

Mm-hmm. Yeah. " This is exactly the pattern we saw in railroads, we saw in telephone, we saw in fiber, we saw in electricity.

It's one thing if you could finance the infrastructure build-out out of just free cash flow. But when you have to start borrowing to finance that build-out, and then things commoditize as they do, and that price comes down, right? Jevons paradox.

That's when you have shocks to the system. That's when you have depressions. That's when you have bubble burst.

It happened with the railroads twice. It happened with, obviously, in our lifetimes with the dot-com. Right?

With fiber. But it also happened with electricity and phone, though they avoided the big bubble bursting by volunteering for government regulation. Back to our first thing, right?

Maybe government regulation is what's needed here, but it was a lot easier to just regulate the US then than it is to regulate AI now. It's tough. There is another part of this, though, which is six of the top 10 models are Chinese now.

Right? Is this Google's way of competing? Right?

" Are they being smarter than we're giving them credit for here, Tom? Yes, and I'll tell you why. Because I want to congratulate Sundar and Dario and the rest for figuring out something that we figured out 15 years ago at Tech Field Day.

If you go back and watch the very first Field Day presentations, it was a mix. It was a bunch of people from different disciplines in the room kind of doing a gestalt thing. That's where the name came from originally.

Mm-hmm. But after a few of those, people said, "Well, we love the fact that you have a backup guy and a networking guy and a security guy. " And that's when we learned the value, to extrapolate to today, of a small language model over a large language model.

Because while the companies don't want the price per token to go down because they want to make as much money as possible, they want the cost per token to go down. But you can't do that when you're effectively boiling the ocean with every run of Opus or Sol or whatever. But if you make a very small, focused model that does not consume a lot of extra resources when you make it, so for example, Fable, Mythos, FlashCyber, whatever you're calling it, that is a very narrowly focused model that can give you super good performance at a very low cost per token, while allowing you to have a high price per token for the people that are using it for the value of the fact that it's a very focused model.

And that's what they're really hoping for. No, it's smart. It is, but it took Mythos scaring the crap out of everybody to shake that tree loose, kind of like we saw with OpenClock at the beginning of the year.

" And that is one of the things I think that this AI bubble, for lack of a better term, is causing us to see, is that this unfocused development is causing so many branches that people are forced to investigate to find where the high profit margins are, that it feels so scattershot that we're jumping off the tracks every six weeks. Well, in research for my book, I did a whole thing on the railroads. A guy from the Harvard Business School in 1960, looking at the whole railroad investment period, said the problem was the people building the railroads thought they were building a railroad.

They didn't realize they were building transportation. Mm-hmm. Right?

The people building AI are trying to build the best AI they can. They're not building maybe what makes most sense for different types of personas and users. I want to bring up, though, I mentioned Antares, Cisco Antares, when we were in the green room.

This is something Cisco came out with just this week, too. I think it was three different models. But the interesting thing about them is not only are they specialized, but they're mobile or portable.

They're made to be run not in the web, not to consume tokens. Right? But you have to, because token costs aside, I don't know that I would trust a model running in the cloud to analyze my security infrastructure.

What if I have a regulation that says that it can't? Because if those don't exist, they will pretty soon. And Alan, we here at Futurum had a big email this last week about authorized models and things like that, and when you read through it, it makes a whole lot of sense.

None of this data should ever touch these models that exist in these spaces for these reasons. And that's a pretty concrete, clear example of what companies are going to have to start dealing with in financial and medical. Because I can tell you that a group of companies that are trying to investigate whether or not their financial services infrastructure is secure is going to blanch at the idea of having to upload all of their security controls and profiles to somewhere where they've lost control, and that goes back to the thing that Satya said a couple of weeks ago about you pay for AI twice.

You pay for it when you rent it, and you pay for it when you upload all of your data to it. I wrote about it. Listening to you, though, Tom, I'm reminded of when I used to sell against Qualys 100 years ago.

Right before there was cloud. " Well, Philippe Courtot, rest his soul, did pretty well at Qualys and in spite of that, so maybe that's an argument that can be overcome. But I do think the future is that if you could run it locally, you should.

Right? And if you have to run it via a cloud kind of thing, you've got to start from a zero trust. And Tom, with all due respect, I got the same memo from Futurum as you did, and I'm not so sure it's as well thought out because I don't see the zero trust.

Everything has to start somewhere, Alan. Okay. But we'll talk about that off camera.

But I do think we need to develop best practices on how to adopt any model and every model from a zero trust. What do I give it access to? I don't need to give it access to all of my Google Docs.

I'm just going to create a folder that I want it to be able to write to, not edit or not erase. It can only add new files or something like that. Common sense kind of things that we've learned with the zero trust kind of paradigm over the years, right?

We need to institutionalize that here, and that's my take on it. Anyway, if no one has anything else on our- I'm going to mention one thing really quickly, going back to what Jack said. Go ahead.

I'm glad Jack brought up Google's financial reports. And I'm glad, Alan, you mentioned Microsoft's as well, because there might be some nervous people in the CFO offices of OpenAI and Anthropic when they look at these numbers from these high, large established companies, and they're rushing to go public. Just a thought from me.

Well, they're going to use AMD's money, it sounds like. There you go. That's what I was going to say.

There you go. Next topic. Yeah.

But back to the... Yes, John, you're 100% right. When it crosses from free cash flow to having to raise debt, that's always a telltale sign, right?

Let's go to Block C, though. AMD, and let's face it, the four of us of our age, AMD has always been sort of a second. They never led the market.

Even though they had some great technology breakthroughs over the years, they were always a second fiddle to Intel. Then they've been a second fiddle to Nvidia. But you know what?

You got to give Lisa Su and the AMD people credit. They're having their moment in the sun here. Mm-hmm.

And yesterday or the day before, we saw some great ... announcements around this. Jack, why don't you kick us off, if you don't mind?

Lay the groundwork on this one. So, there were two announcements, which is one is AMD is partnering with Anthropic, and Anthropic is going to deploy two gigawatts of AMD GPUs- Mm ... and AMD infrastructure.

And that's both AMD GPUs and their rack scale infrastructure, which I believe is called Helios. And there's a couple reasons Anthropic is doing this, but interesting points as we now-- I think I mentioned this before on the Gang somewhere along the line is, when I started in the silicon industry back in the early '90s, we talked about things as chips, and we talked about number of transistors, right? And now we're talking about the amount of energy and how you deploy it and the amount of energy.

What type of deployment unit is a gigawatt of energy? I just don't understand as a chip. I was going to ask you what exactly- I don't get that.

Well, it goes by a rack. You could have so many- Right. So you buy a rack, and the biggest thing the rack consumes is energy, and so that's how we think about these things.

Yeah. A little different than when you carried around the tool. Remember you'd carry around a tool belt?

Yep. You'd have to mount a server. Yeah, I remember those days.

Yeah. So, the world has changed. But let me get to the second announcement, and then we'll go back with that, which is AMD did a deal also with Microsoft, where Microsoft is going to adopt their new Helios system for Azure.

So for rack scale compute, not necessarily all GPUs, they're going to deploy that into Microsoft's cloud environment to run Azure. So AMD's got a very big deployment here. Now, you brought up earlier the dollars and the Google and Microsoft announcements, and part of the deal here with Anthropic is Anthropic is deploying AMD, and at the same time AMD is investing, quote, "tens of billions of dollars" in Anthropic.

So by investing, does that mean that they're putting cash in and giving GPUs and servers away, or are they just shipping servers out there and that's the investment, right? And this gets down to, and I think we've had this conversation before about funny money, which is I invest $10 billion to you, and then you give it back to me and you buy some processors, and I give it back to you, and it goes round and round and round, and it's not really clear where's the profit in it for AMD. It's great that they're doing this, but how much profit are they doing versus giving product away?

So that's one aspect of the story that I think is important. The other one is really simply that AMD now is a legitimate challenger to NVIDIA in the GPUs in AI chips. And that is a big thing, because having an entire industry dependent on a single vendor is very, very risky, and we really do need to encourage multiple vendors.

No, no, that is interesting. I agree there, too. Again, I'll go back to some of the research I did for my book.

This is a page right out of the dot-com era, right? And I lived this, because I lived the dot-com era, right? We were a large ASP.

Cisco gave us, I think it was $10 million. We were a Cisco-powered network data centers. We had a whole bunch of data.

Cisco gave us, I think, $10 million. EMC gave me $10 million as symmetric. Sun gave me a beautiful $10 million worth of UltraSPARC machines.

Dell. And what would happen is they'd give you those machines. They would write that off as a sale, and we would count it as an investment , right?

Right. A strategic investment in kind, if you will. And so it ballooned up both balance sheets until the piper came home, and then someone had to pay the price, and that was a $2 trillion write-off in 2000, 2001 dollars, not today's.

Because today, a trillion dollars is nothing. It's just a couple of gigahertz racks or whatever, right, Jack? Right.

Whatever that means. But that's what's going on here with the circle jerk of the circular financing. Everybody's balance sheet gets buoyed, rosen up.

Oracle. Oracle is so overextended around this. Broadcom.

What percentage of Broadcom's revenue is tied into deals like this? It's all in my book, but it's scary, is what it is. It's a financial disaster waiting to happen.

If the price of AI, of intelligence plummets to commodity pricing, all of these numbers on the book, all the air's going to come out of them. And I think that's what you got to worry about. But look, we'll worry about it another day, because that's what we seem to do in this world.

AMD is certainly a competitor, but are they, John? Are they a real NVIDIA competitor, or is this all Microsoft supporting Apple? Well, we're going to find out a little bit more.

As we are talking, NVIDIA is showing off its data center hardware at an event in San Francisco that includes Helios, which it is in fact, as Jack pointed out, it's marketing it as an alternative to a design from NVIDIA, which I believe is going to come out with their second generation product this year. And AMD's also going to launch a Venice, I think it's called Venice CPU, for data centers. So it's been this decade-long comeback.

I don't even know what to call it. A comeback story or rejuvenation or a more relevant AMD under Jenson Huang. They've always been a stalking horse.

Yeah. Yeah. Right?

That's a good way of putting it. But now they- But there's some capital... Yeah They're giving them a run for their money.

And NVIDIA's got this incredible monopoly, and it's good to see AMD can make some sort of headway. And I remember you said, who was it? You always mention this.

Was it Brad Feld? Yeah. If you're not in the top three.

Top three. Yeah. Yeah, you need three.

And- I actually just spoke to Brad about that on email, because I had to do a source lock, and I asked him if it was okay, and he confirmed it. But you're right. But Tom, is AMD playing the Mac of the early '90s to NVIDIA's Windows?

I think they are. And really what they're doing, there's a twofold strategy here. The first is to ensure that NVIDIA cannot get a chokehold on the market.

And by the investment with Anthropic, what they're tacitly saying is Anthropic is going to develop models that can run equally well on AMD hardware and NVIDIA hardware. That prevents lock-in, and that's something, that's the ghost of cloud, right? The boogeyman of oh, well, you don't want to get locked into AWS or Azure or what have you.

And so AMD is really just hoping that they'll be able to kind of sell into some of the areas that NVIDIA has either not saturated or people who kind of have a deep-seated distrust of using NVIDIA for whatever reason. And remember that when AMD was kind of the also-ran in the chipset market back when the 486 was state-of-the-art, they didn't need to unseat Intel. They just had to wait for Intel to make some mistakes.

Right. And I think that that's where they're at with NVIDIA right now. They've got to be able to hold out long enough that there's a stumble, that there's a bet that doesn't pay off for whatever reason, so that AMD can run in and scoop up some market share from NVIDIA, kind of like with the Pentium floating point bug.

Yeah. Or literally pick any of the stuff from the late 90s and early 2000s that was enough for AMD to stick around until they kind of got into the graphics card market, and that was really where they started printing money, at least until they stopped being graphics cards and started being GPUs that were used for other things. And I think that AMD has the right product mix to really make an effective charge forward in a lot of these areas, but they don't have enough to unseat someone as big as NVIDIA.

And honestly, it's funny to me because I sit here saying this as I have a machine running on my desktop that's running an AMD Ryzen processor. When I built it, it was never a question that I was going to use a Ryzen in it because I wasn't impressed with what Intel had been working on. And if 1998 me had said that Intel wasn't as impressive as AMD, I probably would have sent him away to some kind of place with a padded room and those white jackets.

Yeah, or stop doing the drugs, right? Something. Right.

But that's just it. Because it was the late 90s after all. But, yeah.

No, the worm turns, right? And it is a lesson. But I'll tell you something else.

This Microsoft alliance, it's nice to have a big brother at your side or a partner if you're going to take on Jensen and the NVIDIA machine. The only thing that worries me about that is because there was another news story that came out this week that Microsoft is also adopting some III-V fiber technology that uses a diffusion on the end of the laser to avoid physical connectivity. From a tech geek perspective, that's really cool, right?

There's a new way to do this. But from a different perspective, I worry that they're turning Azure into kind of like the garage hobby kit of cloud adoption. And I would genuinely worry as someone who's like, "I don't need you guys to play around with this.

" So I'm going to tell you something. I hear what you're saying. I'm going to give Microsoft more credit with what they're doing with Azure.

I don't think that this Helios is going to be the true backbone of everything they do AI-wise on Azure. I think they have already made a good investment into NVIDIA GPUs. I think Microsoft also owns a good chunk of OpenAI.

They have this Anthropic relationship where they're offering that on Azure, and now they've also released three of their own mini models. They outdid what Google did before Google did it. They have three of their own models.

I think it's the Microsoft models that are going to run mostly on this AMD Helios infrastructure. But I think what Microsoft's building with Azure is a... You ever go into the department store and there's a lot of, like, Ralph Lauren has the Ralph Lauren store there, and here's Tommy Hilfiger over here.

That's exactly what they're doing. Yep. Exactly.

Do you know what I'm saying? Like stores within stores kind of thing. Mm.

This is- Where in Google, it's all Google. And in AWS, it is that. Microsoft's giving you stores in stores.

Microsoft is building an environment where Microsoft becomes the destination for the business. You want generic cloud computing, you can run it on Azure, you can run it on Intel, or you can run it on, what's the other processor? The RISC processor.

Oh, ARM. ARM. Thank you.

You can run it on x86, or you can run it on ARM. If you want AI, you can run it on the AMD flavor, you can run it on the NVIDIA flavor. You can run your own models, you can run our models.

We're going to be the platform for you to do whatever you want to do. Give us all your money. We'll include it in.

You'll start seeing this, that'll be your O365 will include some flavor of this. And we'll give you Copilot. We'll give you Copilot.

We'll give you a Copilot. We'll do this, we'll do that, right? A little bit of that.

And all you need to do is just have all of your bills go to Microsoft, and that's all that's important, right? It sounds like a Crazy Eddie commercial from when I was young. His prices are insane.

But you're right, Jack. And from Microsoft's perspective, why wouldn't you do this? There is no value to Microsoft in locking themselves, talking about locking, Tom, there's no value to Microsoft to locking themselves into one architecture.

Yeah. Mm-hmm. Guys, we're about out of time.

I want to just save a moment to wrap up here, but I think what we're seeing across all three of the stories today is the AI frontier. If it's not the Wild Wild West, it's certainly close, right? This is just happening in real-time right before us, and it's moving really quickly, and it's biting, right?

It's biting. And it's going to continue. I don't think we're going to see a stable kind of thing.

I think we're in for a strap on, fellas. We're in for a rough ride, and that's what it's going to be for a bit. But anyway, we'll continue to report on it here every Monday to Friday at noon on the Techstrong gang.

Of course, Techstrong TV, including a lot of the Tech Field Day content, is running almost 24 hours a day on our Techstrong sites now, and you can catch them on there. tv, at our Techstrong TV YouTube channel, or our Techstrong TV OTT app, which plays on Android and iPhone and Amazon, Apple, and Roku even. Just about any screen you like to watch your videos on, you can catch Techstrong content, as well as the Tech Field Day and Futura podcast on there as well.

But for now, thank you, Jack. Thank you, Tom. Thank you, John.

Thank you for watching. I'm Alan Shimel. We'll be back tomorrow for our Friday Techstrong gang.