[00:00:20] Happy to have you along, my fellow tech enthusiasts. I'm Sam Bushman. Jay Hill is with me. We keep an eye on tech so you don't have to. This broadcast is brought to you by NetworkProvidersInc.com. Check out the website. We've got an offer so that you can learn more about us and see if we're a great fit to help you with all your IT needs. You can also check out the show there, the video version, the audio version, and it won't be very long. We'll put a transcript version up or at least a summary up of the show.
[00:00:45] So check that all out. Welcome, Jay. Hey, Sam. How's it going? It's going very well. There's this massive IPO that we've got to talk about, Jay. I know. It's all in the tech news, right? Anthropic's IPO. They say 2 billion IPO. That are 2 trillion, I'm sorry, IPO. Yeah, trillion with a T. That's a big difference. Isn't that one of the biggest IPOs ever? I think so, yeah. I've never heard of one that big. The closest, I think, was SpaceX coming close to it at maybe 1.5 or something, but this is going to be,
[00:01:16] topping the scales. They're setting it up so that, this is Anthropic. SpaceX's June debut was the largest public offering to date before. Shares are going to be traded on the NASDAQ under the ticker symbol ANT, A-N-T, I guess for Anthropic.
[00:01:34] The company behind Claude, the AI models, took in $16.3 billion in the first half of this year, up from $1.2 billion a year earlier. Also reported, though, there's a problem, $80.7 billion in losses, about $72 billion of which were attributed to accounting changes tied to financial deals.
[00:01:56] And I appreciate them saying that I think the deal is that they just lost money, but here's the idea. Yeah, we lost a ton of money, but we want to be poised as the leader of AI. And if you guys double down and believe in us and make this IPO happen, in other words, fully fund it, $2 trillion, we'll become the leader. I mean, isn't that really the play, Jay? That is the play. I think they're also making some moves politically to try to be the only provider of AI. You and I have talked about this a little bit before.
[00:02:26] I think that some of the legislation and things that they're going for, they kind of want exclusivity a little bit, and I think they're poising themselves to get into that position. And if they are, they will be worth this kind of money, but right now they're just hemorrhaging money. Now, there's a big tank in the tech stocks right now as well. I don't know how long it'll be. Sometimes it tanks, then it comes back strong. I don't know how that looks, but I wonder if that's tied to this, because people are just kind of wondering right now.
[00:02:55] There's a big question out there. AI is incredible. AI does a lot. AI is neat. AI is innovative. AI is the buzzword. AI is the future. Okay, but really, what is AI? And it's kind of like the dot-com bust of yesteryear, Jay. We don't really know what AI is and means and will be. And so one of the questions that I've started to ask the AI bots and ask the corporate world is this. How much is an AI hour? In other words, how much does it cost for an AI hour? That's a good question.
[00:03:25] And the reason that you've got to ask that is because people are offshoring to the Philippines or to India for a lot of different tasks, and those people are working for anywhere from $3 to $12 an hour, way cheaper than what an American employee costs. If you're in many states, you've got to pay at least $15 an hour by minimum wage requirements, et cetera. If you pay less than that, you're almost abusing your employees in America because they can't survive. So, I mean, I get all that. But at the same time, that – and so you've got to compare an AI hour compared to that.
[00:03:53] And then you've got to compare what if I use a Filipino employee at, say, $8 an hour or $10 an hour. And what if I have that employee use AI? Then what am I gaining? Is it more expensive or is it – so they don't know the answers. They don't know the answers. The whole corporate world is trying to figure this out, and nobody knows. I think it's a pay. Because the AI hour for a programmer is different than an AI hour for somebody who might just be asking questions, right? Yeah. Well, two things, though, that factor into that.
[00:04:22] Number one, an AI doesn't take an hour to respond to a question usually. It's just minutes or even seconds. So I think you can get a lot of work done in an AI hour if you really maximize it. Maybe. Because if you're going to sit there and ask questions, you might have to stay there for the hours. And I've got to pay the person an hour and the AI the hour. Right. So what's an AI hour? Is it just what it costs for the AI side?
[00:04:45] Or what does it cost for the person to be there to continue to either make decisions and et cetera, right? I don't think you can count that cost because you're going to have that anyway. Even if you're managing an offshore, you've got the cost of the person that's doing that. And I don't think in doing an apples-to-apples comparison, I don't think you want the cost of the person asking the questions. You just want to know what does it cost for an AI hour. Hold on. What happens if you're a programmer, though, and you have to stay there and babysit that thing and make all the decisions?
[00:05:12] If I have an offshore programmer doing it versus a person in the United States doing it, the costs are drastically different. No, you still have to have somebody who's in charge answering questions and babysitting that other person unless they're just completely autonomous. But an AI could be just as autonomous, really. That's the question. So are we going to now talk about an AI hour if it's running autonomous or an AI hour if it's being babysit and directed on the fly by a human? What is an AI hour?
[00:05:38] And so I bring this up not to debate the terms of it, but really to highlight for everybody this is what the world is really trying to figure out now, and no one has the answers. And I backed into some preliminary data, and I'm convinced that an AI hour isn't as cheap as you think it is. Well, I was about to say, I think that right now an AI hour is actually super cheap, but I think it's being undervalued. And that's why we're seeing companies like Anthropic and ChatGBT, OpenAI, and stuff, losing so much money.
[00:06:03] I think they're actually underbidding it or underselling it and using finance capital and stuff to keep this going because everybody can see the promise that's in there, and they can see the potential that's there. And so a lot of money is going in from other sources that aren't really paying the dollar-per-dollar freight of what the AI costs. So I think it's cheaper now than it should be, but I still think that it's probably cheaper than labor elsewhere. But I don't know. What do you think?
[00:06:33] Well, and that's a debate and discussion. So now we're talking about what you brought up brings up the fundamental question. An AI hour paid for by whom? So if you get an AI hour, Jay, for, say, a dollar, but Anthropic is losing money hand over fist, and it costs them $5 to produce that AI hour they gave you for a dollar, what's the cost of that AI hour? $4, I think, because if they paid $5 for it, and I offset it by $1, I would say $4.
[00:07:01] But it's really $5 because whether I can sell that hour is a different discussion of what it costs, right? Okay. I'll go with that. So because you can't really debate the bring-in-money side of it. Eventually, you can offset it, but that's called revenue as opposed to an expense side of the discussion. Okay. So the reason I'm bringing that up is now here's the next question. So let's say that you ask questions, and let's say that it takes, you know, just for sake of argument, 10 seconds to answer each of your questions.
[00:07:31] 10 to 15 to 30 seconds, whatever. Sounds reasonable. Maybe a minute, two minutes, depending on the hardness of the question. Again, that's why it's very difficult. But you add all that together, and so you take an actual working hour. When it stops and is waiting for you to type, we're not going to count that either. Right. Okay. Okay? Because when you're typing, it's just sitting idle. Yeah, and it's not using any cycles on their stuff, so. And they're not charging tokens either. It's answering somebody else's question, right. They're not tokens either. However, the second it starts to analyze your input, it is.
[00:08:00] So now you take all that questions, put all that together, and say, when it really worked for 60 minutes, it gave you a ton, as you pointed out, right? Yeah. But now how much is that hour in computing? I don't know. I mean, they would know. I would think. But I don't know. I don't know that they know, and that's the problem. They don't know, because now it depends on which engine you're going to use. Yep. Which model. It depends on how much we flush your cache, meaning if you have a long conversation,
[00:08:27] every time they feed that whole thing back to the question again, or whatever it is, and they're working on stopping, you know, reducing some of that. It has to do with token windows and everything else. Right. How many tokens are they wasting to re-skip down to find out where you were, and, you know, flush the cache and do this and that? So how much could an hour could be versus what it is versus... This is a huge, massive debate. And the only reason I spend so much time on it is I want people to understand that we don't know these things yet. And until we know more about these things, everybody jumping to AI with just reckless abandon is crazy.
[00:08:56] Because if they have to back into and come back and say, whoa, hold on. And an AI hour costs $4. Then you've got to ask yourself, if I have an employee for $8 an hour in the Philippines, do I want to give them a $4 bot? Because now when they work, it's going to be $12. If the bot runs and they work and whatever else it's going to be... So you've really got some questions to ask. And you say, does it really speed up work?
[00:09:24] So a lot of programmers are now saying, yeah, in a lot of ways it speeds up work. But in a lot of ways I have to stop it from going down the wrong roads. Oftentimes I have to basically say it coded a ton of stuff. But I don't want that. I have to get rid of a lot of the code that it created because it was so excited to code that I went along with it thinking it made sense. But now that I kind of look at what really is happening, I don't really want all that. It's enthusiastic. And the list goes on and on where, I don't know if you know, but most people, if they write a program and then they make a tiny change,
[00:09:52] they just recompile the change and it doesn't take very long. Well, what these coding bots do is they basically say we've got to run the whole test suite again and we've got to run the recompile again and it takes way longer. And so an employee doing something that takes 10 minutes versus the bot doing it that takes 45 minutes, is it really the savings we think it is? And so people are really starting to debate this. And at the end of the day, they're coming back and saying, yeah, these bots have real value,
[00:10:20] but let's make sure that humans are at the decision process. How many times might these bots, if they're completely automated now, make mistakes or do things that they shouldn't do that you have to now try to recover from? Hey, it ran a report on my QuickBooks. We went to it. We sent it to the bank and now we find out it's wrong. Now we've got to one, redo the reports. Two, call the bank up and go, hold on. We gave you some bogus. And so on. What's that going to cost in terms of time? And you've got to put that against an hour of AI time, right? Don't you? Yeah.
[00:10:50] Because if AI spends five hours making mistakes and you've got to recover those five hours, how long does the recovery time to undo what the unattended or unbabysitted or whatever you want to call it? AI did. There's a big debate about this, Jay, in the whole industry. And that's the reason stocks are tanking. And that's the reason that I don't know that they'll make their IPO of $2 trillion. Well, Anthropic has committed $540 billion to data centers and to chip leases over the next decade.
[00:11:16] Their filing runs more than 90 pages of risk warnings that include advanced AI could pose, and this is a quote, catastrophic or existential risk to humanity. Goldman Sachs and Morgan Stanley are the ones leading the offering. They've got a share count and pricing to be set in the coming weeks. Wall Street has a target range in the top estimate at $2 trillion. And, you know, we're going to see how all that plays out. Do you think they're going to get it all? I think they're going to get somewhere between $1.5 and $2 trillion for their IPO.
[00:11:45] And there's a lot of hubbub about it. I mean, it's all the rage right now, and it's what's making waves. I use a computer right now out in the cloud because I need a super GPU card, a super NVIDIA card. And anyway, I can tell you about it after the pause, but it's very interesting. I'll tell you what I pay for it, and it'll give you kind of a baseline to kind of understand some of this too. I'm Sam Bushman. Jay Hill with me. Network Providers, Inc.
[00:12:14] This is NPI TechGuys. We keep an eye on tech so you don't have to. Cybercrime is exploding. Take Sarah from Sweet Delights, whose world crumbled after having to close with not being able to bounce back. Small businesses are prime targets, but the right strategies can keep yours safe.
[00:12:38] Jay Hill, CEO of Network Providers, has co-authored The Cyber Playbook Simplifies Cybersecurity for Business Owners with Strategies to avoid costly breaches and fines. Build a strong cyber attack response. Secure your business with key protections. Cyber threats aren't slowing down, but you can stay ahead. Protect your business. Ensure its security for tomorrow.
[00:13:03] Get The Cyber Playbook today at networkprovidersinc.com slash cyber dash playbook or call 385-446-5500 now. As you are aware, America is divided over every fault line possible. This is intentionally fostered by those who do not love God, family, or country.
[00:13:31] We believe a peaceful future as a free people absolutely depends on civility. Clarion Call for Civility is looking for funding and volunteers at every level to make our hopes and efforts a reality. Please donate, sign our pledge, and help us in our sacred cause. Please visit callforcivility.com for more details. Call for Civility.com. Back with you live.
[00:14:00] Jay and Sam talking tech. We're talking about this IPO. Anthropic 2 trillion dollars. It's massive, ladies and gentlemen. Whether they'll get the money or not, I don't know. But we're also talking about what it takes for an AI hour. How expensive is it really? And the problem is we don't know. And the problem is the debate is how you categorize what an AI hour is. Is it the bot running for a solid hour all by itself? Is it the bot running with how much ever decision time it takes a human? The problem is how do you gauge a human's time?
[00:14:30] Because if I have to be there to make decisions whenever it wants input from me, I can't go to lunch. Because if I do, it's just going to sit there until I come back. So then is it sitting there a cost? It certainly delays my progress. Anyway, we don't know what that is. But I use a system called RunPod. And it's a computer system, RunPod.io. And you can buy computer time. And you can, it's complicated. I don't want to get into all the details. But you can basically put software on there and have it run tasks. And it's used for audio transcription.
[00:15:00] From audio to text. It's used for, you know, video and audio and all kinds of other tasks. Which are pretty heavy IT tasks, right? Anyway, to rent a computer to do that costs anywhere from 60 or 70 cents an hour to $1.50 to $2 an hour. Depending on how much memory and how much this and that and the other you use and everything else. And so that's kind of a base number. But again, that's just a process. Automated.
[00:15:26] No human interaction or involvement or decision making or anything else. How many tokens does it use when you run Claude or Grok or, you know, ChatGPT. And you have to sit there and make decisions with it. How do we count what an AI hour is? And how do we put a cost to it? And it's a real battle. But in my opinion, until people really get a handle on what that really is, Jay, I don't know that you can decide that you want to move everything to the bots.
[00:15:57] Right. And I think that is a good idea, though. Because the warning 92 pages that they gave, this 92 pages that they gave, that is the real cost. Right? Mm-hmm. Well, that's the real risks, really. But I think that you're right. But risks turning into costs, right? Yeah. But you've got to get this boiled down into, even if it's not perfect, even if there's a lot of variables and you're just kind of taking the averages, you need to get something. It kind of reminds me of NovaBench.
[00:16:24] I use a program called NovaBench to benchmark computers just so I can compare apples to apples. Now, that doesn't mean that if you've got a screaming graphics card in one computer and the other one's got a great chip, you're not going to really get that kind of definement or that resolution granularity. But you are going to get a general understanding. Yeah. But you get a general number. A use of a little quotient. Right. That you can kind of boil things down to a computer. So right now, a good computer is probably about a 4,000 on NovaBench 3, I think is the version we're using. Or maybe it's 4. I can't remember.
[00:16:52] But in their different versions, change that number, which isn't very helpful. But that's a different story altogether. But we need this for AI. We need a way to say, what does an AI hour cost in real dollars and cents? So I've backed into this. And I know my data is just Sam Bushman-ish. It's not scaled. It's not scientific. It's not any of those things. It's just a wag based on all these factors that you've got to consider. You want to know what my number is? Yes. Six bucks. Wow.
[00:17:18] So that's less than half of the minimum wage, you stated, of about 15 on average right now. Yeah. It's about a third of whatever you want to call it, depending on where you are in the country and whatever. And that number, in my mind, takes a little bit of human time to make decisions. Because you're going to have bots that run for a little less than that. But you're going to have programmers paid $100,000 a year making decisions, making that number go up. You're going to have accounting people that have to babysit reports.
[00:17:45] And if it generates something, now I have to have somebody review its work. Do you count the review time reviewing its work? If a bot creates a report, it used to take me 10 hours to do the report. The bot does it in one hour, just say. And then I have to review it and it takes an hour. Do I count it now 10 hours down to two hours? It's hard to say. But here's the one thing about that number. That number. My point isn't to get you to answer that, yes, Sam, you're exactly right.
[00:18:13] The point is to say, hey, all those things need to be considered. And so when I take into all those things that I can best guess, I believe that you end up with about 75% savings in time. Even when you get rid of all the different things. And when you consider it, it'll be about $6 an hour to make these things happen. And that relates to the humans being in charge of it too. So you've got to babysit it. You can't just let all these automated tests happen.
[00:18:42] And you've got to be the decision maker. And you've got to be the one. When it goes wrong, the person's responsible, not the AI bot. And all the recovery time has to be backed into that and everything else. I'm guessing it's about $6 an hour. Well, this is really good news for OpenAI and Anthropic and all these people because that means that it's viable. Right? That means that this is cost savings. Eventually, if that number kind of pans out. I do believe it's viable. It means that it's going to happen. It's not a bubble, hopefully. Right?
[00:19:08] That it's just like, hey, we found out that really it's cheaper to just let somebody else do this work in offshoring or even in America or whatever. That means that they're doing good. And I don't think that it's going to be a problem. I do think that we'll see some bubbles. We'll see some overreaching of it. But overall, it's not like it's just going to be here or there. It's not going to be MySpace. It's not going to be gone in a couple of years. It's going to be.
[00:19:33] And I think personally, in the AI sphere, of course, it's all we almost talk about on any tech show nowadays. But I think that we're just in the very, very beginning of it. I mean, it's still in its infancy. We don't even know how it's going to mature over the next even five or ten years. Yeah. So let's just take six bucks an hour. And let's just say the average person is spending, what, 20 bucks on their chat GPT? Yeah. That's what I do. I have the $20 chat GPT thing. That's the only AI I'm paying for.
[00:20:03] So they're counting on you using approximately three hours or a little bit, three and a half hours of actual computer use per month. And you know what? I know you use way more than that. No, no. I was going to say I use it a ton. But if you're only adding up the time it's chunking and computing, I bet I use less than that for sure. And I think you use more than that still, even though you don't realize it. And that's why they're starting to increase the numbers.
[00:20:28] And that's why guys like me, when I run out of it and then I have to move Claude to be the $100 one. Right? Now Sam is using six times 10 is 60 times 15 is 90. So about 17 hours for me is what they're betting on. 17 hours of literal compute time. I think you use a lot more than I do. Oh, yeah. Because I make it program. I make it run, run, run for hours. And I use way more than 17. And millions of tokens and stuff.
[00:20:57] And I use way more than 17 hours. And they're going to be forced to charge me more. Guaranteed. And so I think you're going to back into about a quarter of an employee cost will eventually be your AI cost. I think that's probably fair. But I think that the bots will get smarter. And this is another part of this equation. And I hate to spend so much time on this. But if you're a business owner, this matters to you. So right now, if it's six bucks an hour and you get a certain amount of value out of it,
[00:21:24] I think in a year, that six bucks an hour is going to give you more value. Right? So how do you account for that changing that bots get smarter and better and more efficient? When I first started coding with Claude, it made a lot of mistakes. Now it doesn't make near the mistakes that it made. Partly because I've trained it. Right. And partly because it's gotten better. How does that factor into the quotient? It needs to be considered too. And so I'm convinced that when all is said and done, you're going to get, at least for now, you're going to get this kind of dialing it in by businesses everywhere.
[00:21:54] And they're going to find out it's about six bucks an hour. And that's why Microsoft is saying we're going to invest in this. And we're going to let tons of employees go, Jay. Yes, they are. A quick thing. I just asked my account, ChatGPT, how much thinking time have you actually spent on inquiries over the last 30 days? And it's not giving me a great answer. It doesn't know. It says, I can give you an estimate, but I don't have access to actual accounting of my cumulative thinking or processing time. So I was hoping to get a good response for you. But what did it tell you?
[00:22:24] Did it give you a guess? It says, for the last 30 days, September 30th, October 30th, I can see we worked through quite a few inquiries. And it's talking about some of them and whatever. Thinking time. No, it's not giving me any numbers in here, really. And that's a problem. So. Yeah. So this is what I mean by it. It talked all around it, but it didn't give any numbers. Yeah. But you know why? It's not giving you an answer, right? No. Why? Because it doesn't know and we don't know. People don't know and the bots don't know yet, Jay.
[00:22:53] I'm very curious now about the number, though, because I'm thinking. Partly. It's a minute. Partly we don't have a good way to track it because what do you account for? Let's say that it thinks down the wrong road for five seconds and then goes, no, that's not what he asked. Hold on a minute. Intelligence says this is what he means now. So there's kind of an intelligence phase, an interpretive phase. Does it count that? This should be really easy because it should just be able to look at the, you know, when it's at the beginning of a prompt, it'll say thought on this for 47 seconds. And then it gives you the answer.
[00:23:20] It should just be able to add all those up over the last 30 days and say, hey, I've done X. You know, I've spent. I know what you're saying. And that's a way that it could be done. But then people are going to argue that. It's the same thing with watching video on the Internet. Every one of these companies has a different way to measure how many people watched what. This is going to get the same. The nuances are you can say that it should be this way. And I don't disagree with you necessarily, but I'm telling you that everybody's going to debate that. They're going to say, well, what happens when it had to switch tasks from my questions to yours, Jay? Right. Now, I understand that.
[00:23:50] Do we count that, too? What do we? So it's complex. But anyway, Microsoft is laying off a bunch of people and they're going to put the money into AI. They are. They're laying off 27 percent of their global workforce. It's an early buyout for people. They're saying, hey, you can leave by December. And it's 60,000 employees globally. This is going to affect. How much do I get to bail? It doesn't go into details of what that kind of severance or that let-go package is. But it's United States, Canada.
[00:24:20] I'm sure it's different per employee. It's some factor of whatever their job is or whatever else. And the more expensive you are as an employee, the bigger the number. It's massive. It's huge. And so they must believe that they can lean into AI and it's going to do a lot of good. So understand the whole point of this is AI is real. There is a payoff in it, I am convinced. And these companies know that. What they don't know is the reality and the quotients. But they're taking a gamble just like the $2 trillion IPO is. I'm sure they're taking questions. And their problems are often時間 می Belle and the Portugal delars are as way盤 on JAKE тогда WHAT LEAD COUNCIL