00:00SPEAKEROkay. I've got a public service announcement here.
00:02SPEAKEROh, a PSA.
00:03SPEAKERA PSA about this DBX.
00:05SPEAKEROh, a DBX PSA. Okay. Hit me.
00:07SPEAKERDBX PSA. Remember the fuzzy garbled, I don't know what you want to call it, you know, my audio issue.
00:13SPEAKERI mean, it wasn't, you sounded not like a, it didn't sound like most audio corruption.
00:18SPEAKERIt sounded real weird.
00:20SPEAKERYeah. There's a tiny little taste of it at the end of, I believe, episode 272.
00:24SPEAKEROh, sorry. 271, two episodes ago, I believe, like at the very tail end.
00:28SPEAKERIf you really want to hear what I'm talking about, but it just sounds like my audio gets kind of
00:31SPEAKERfuzzy, crunchy, rough, something, analog, interference-y sort of.
00:36SPEAKERYeah. It's not like, like there's a sound that a voice meter makes when it conks out or like a,
00:43SPEAKERlike a certain USB audio interface to make, it's not at all that one. That's all I know.
00:48SPEAKEROkay. Well now you know what it's not.
00:50SPEAKERYeah. And it's not like, it's not like a loose connection on an XLR cable
00:53SPEAKERor a headphone jack or something like that.
00:55SPEAKERSounds electrical in nature. I'll say that.
00:57SPEAKERYeah. It's like, I imagine if you brought like a big giant magnet next to a cable that didn't
01:03SPEAKERlike to have big giant magnets next to it, and then also applied some juice to that magnet, maybe.
01:08SPEAKERThere's every possibility it'll happen while I'm in the process of describing it,
01:11SPEAKERwhich would be incredibly appropriate.
01:13SPEAKERIt's not going to happen. I'm sorry.
01:15SPEAKERAnyway, I've been trying to troubleshoot this for a while. It's very intermittent.
01:18SPEAKERWe talked about it on here. I also was talking to Vinny about it and he dug up a couple of
01:22SPEAKERYouTube videos. He's got one of these things too, actually. So does Alex. We all bought it.
01:25SPEAKEROh, you all have them. The DBX-286S. Anyway,
01:29SPEAKERa couple of YouTube videos and then a couple of forum threads on Reddit and a place called
01:33SPEAKERgearspace.com. Okay.
01:35SPEAKERWhich like you can probably guess gear space is a pretty hardcore audio production forum.
01:40SPEAKERYeah. I bet those people have flangers.
01:42SPEAKERI bet they do. I bet they flange all day long.
01:44SPEAKERYeah. But they're constantly flanging.
01:46SPEAKERAnyway, the upshot is this problem I'm having is incredibly widespread on this unit. Incredibly so.
01:52SPEAKERLike maybe dozens of people in that thread talking about it happening. And a couple, like I said,
01:58SPEAKERat least a couple of YouTube videos I found of people having it and trying to repair it.
02:02SPEAKERThere's no like firmware or anything. This isn't a digital thing.
02:05SPEAKERThis is entirely an analog electrical.
02:07SPEAKERHere you go.
02:08SPEAKERIt's just got a power plug and audio ins and outs on it. That's it.
02:11SPEAKERLook. So the good news is you have a Pinesil. You're ready to go.
02:14SPEAKERSo yes, it does in fact come down to soldering issue. People have determined that thing is
02:18SPEAKERjust pretty poorly made. Like the soldering is just bad on it.
02:21SPEAKEROh, that sucks. Did they blame lead free solder? That's always my favorite is like,
02:25SPEAKERHey, millions of companies managed to make a bunch of shit with RoHS lead free solder. No problem.
02:30SPEAKERBut whenever, whenever some small volume places like, yeah, we could,
02:34SPEAKERthe solder is bad because we lead free solder just blows. We can't do anything about it. I'm sorry.
02:39SPEAKERI don't know where the fault lies. I don't know if they just cheaped out on the manufacturing or
02:44SPEAKERwhat, but I can tell you that that one forum thread is years old and there are still people
02:48SPEAKERafter a period of multiple years coming in with the problem. And I just bought this one in November.
02:52SPEAKERSo I don't think they fixed their production lines or whatever's going on here, but the one,
02:56SPEAKERone guy on YouTube fixed the problem by basically reflowing every chip on the board.
03:01SPEAKEROh God. How many chips are on the board? Maybe not all of them. I didn't actually,
03:05SPEAKERI kind of scraped his videos like 30 minutes long of him testing chips on the board.
03:09SPEAKERI didn't actually verify it was every single chip, but he definitely did reflow multiple ones.
03:14SPEAKERApparently fixed the problem just fine. Now it works great. So like, Hey,
03:18SPEAKERHey, what kind of soldering station do you got? Do you want me to come over with this DBX?
03:23SPEAKERThat is more soldering than I am willing to do, but I will tell you the first electronic
03:28SPEAKERflea market is just like two months away. That's true. And you could get an oscilloscope
03:33SPEAKERdown there. Like you can be right. You can have a full on like E workstation that you don't know how
03:38SPEAKERto do any of this stuff with, but it looks real cool. It makes you look like you know what you're
03:41SPEAKERdoing much like several other projects around here. I'll buy all the stuff and then sit there
03:45SPEAKERand look at it for five years. Yeah. That sounds great. Yeah. The guy, the guy, I think it was one
03:49SPEAKERof the YouTube videos. A guy said that you could get like a decent little reflow station off of
03:54SPEAKERvolley express for like 80 bucks, but yeah. So you can get a reflow, reflow stations aren't
04:00SPEAKERparticularly expensive. You can get even that you can get even cheaper than that. That are fine for
04:05SPEAKERsmall work. Twinkle Twinkie talked about this when we were talking about I can't remember,
04:11SPEAKERmaybe when we were, maybe when I was talking about mouse switches a hundred years ago.
04:16SPEAKERBut yeah, they, they, what it is, is it's like a, it's like a, a higher gun that gets hot enough
04:21SPEAKERto melt the solder and suck it away from the, from the thing from the, from the joint. So you can
04:27SPEAKERclean off a joint entirely, and then you have to re solder all this stuff. And you can probably,
04:32SPEAKERlike, I assume that it's all edge, edge mount stuff. So you can probably drag solder it.
04:37SPEAKERSo this guy actually didn't even re solder stuff. He used what the existing solder on the board,
04:41SPEAKERhe just kind of just heated it up. He just heated it and kind of like reseated the chip,
04:45SPEAKERcentered it over the pads and just kind of let it reset. Give it a good wiggle. And yeah. Cause
04:50SPEAKERhe went through with a pick and like determined, like he found a couple of pins that were pretty
04:53SPEAKERnoticeably loose and not really making contact with the board very well. So let me ask you a
04:58SPEAKERmore relevant question is, are you still inside the return window for this device? That's the
05:03SPEAKERthing. If I, if I were, it would just go back, but I'm not. And one of the, the the other YouTube
05:08SPEAKERguy who posted about it had sent his back for RMA repair. Yeah. And let me guess, didn't get fixed.
05:15SPEAKERIt worked for a while and then started going bad again. Weird. Yeah. So that's not a great option
05:20SPEAKEReither. I don't think I might just buy something else. I mean, I think, I think like you could
05:25SPEAKERshare the love on this one and just pass it on to somebody else on the used market and you know,
05:30SPEAKERmake a better choice. Yeah. Yeah. Anyway, I just want to put the PSA out there. It did it for just
05:36SPEAKERa second. Wait, are you serious? Yeah. Okay. That's staying in then because that could not
05:41SPEAKERbe more poetic striking back at me. It's trying to take revenge for talking mess about it. We
05:45SPEAKERedit so many of these out every month. That's yes. It's well, I mean, sort of, I mean, it's,
05:50SPEAKERit's, it's not that bad lately. I mean, I always wonder how often it's happening when I'm not
05:54SPEAKERtalking. I don't like how it might be striking silently when I just happened to not be speaking.
05:59SPEAKERWhoa. If there's, so you're saying if there's, if there's some sort of noise in the signal,
06:05SPEAKERwhen there's no signal, does it even make a sound? These are questions I'm not prepared to answer,
06:11SPEAKERbut I will say if you're looking at a 286 S based on my talking about it, maybe hold off.
06:16SPEAKER386 much better. Anyway, 32 bit, you know, you get a floating point processor if you want
06:24SPEAKERbetter unit all around or whole gigabytes of memory. Yeah. It's the future of computing.
06:29SPEAKER[Music]
06:57SPEAKERWelcome to Brad and Will Made in Tech Pod. I'm Will.
07:00SPEAKERI'm Brad. Brad, I, I did something unwise for lunch today. We usually record like breakfast
07:06SPEAKERtime and today we're eating, we're recording it just after lunch. Yes. And I went to the Asian
07:11SPEAKERmarket and bought ramen the other day because I like to go and get the, you know, like your fancy
07:17SPEAKERlunchtime ramen. And I saw this stuff. I was like, Oh, it's ramen. You can, you know, you can chuck
07:23SPEAKERit in a, you, you cook it and then you save a little bit of the noodle water and you pour some
07:28SPEAKERsauce in there and kind of stir fry the noodles until it thickens up into a sauce. That looks nice
07:32SPEAKERand that looks fancy. And we said it was spicy, but like spicy ramen, not that spicy. It was real
07:38SPEAKERspicy, Brad. So if I have to get up, run screaming from the room at any time, just be prepared. Like
07:43SPEAKERfor, I started to ask for what, but I think I'll leave that for just the listeners to imagine.
07:49SPEAKERSo there's an occasional gurgle happening over here, but I think everything's fine.
07:53SPEAKERYeah. I should have known. I should have known cause the little chicken on the package looked
07:58SPEAKERlike he was very spicy. I think it's, I think it's YouTube meme ramen, but I'm not sure. Anyway,
08:02SPEAKERit was delicious, but hot. I've gotten to the point where I try not to eat before anything on
08:05SPEAKERa microphone. Well, you have some circumstances. I've got some extenuating vocal issues, but still
08:11SPEAKERI'll say I just, sometimes it just makes you logi, you know, as well. That's true. That's true. Big,
08:16SPEAKERbig bowl oatmeal first thing in the morning. And I'm like, Oh man, I should have put an egg on that
08:21SPEAKERand had a little less of the oats. Yeah. This week, speaking of a little less of something,
08:28SPEAKERwe got, I don't know, I was going to call it a twofer, but it's a potential twofer because
08:32SPEAKERI think as is our way, the first topic, the primary topic ended up being more to it than
08:39SPEAKERwe thought that there was going to be. So let's call it a onefer with a potential bonus at the
08:44SPEAKERend. Yes, but neither of us are subject matter experts. So maybe we'll see how deep into it
08:49SPEAKERwe get it. Yeah. So, so a couple of weeks ago, this Chinese LLM rolled out called deep sea model.
08:59SPEAKERPeople call it, call it the AI. You gotta peel back the onion here. Yeah. I'm peeling onions
09:03SPEAKERhere. The it's, it's an AI model from China. And it's actually it, it, it, it's, it kind of threw
09:14SPEAKERthe AI world into a little bit of a tizzy for like four days. And it's because this, this was
09:20SPEAKERto be clear, this was a, this was the day that the Nvidia stock cratered. If people are, yeah,
09:24SPEAKERso it was the 29th, I think two weeks ago in the AI news, but do you remember the, like everybody
09:30SPEAKERrunning around on a Monday morning with their hair on fire going like, Oh God, the world is ending.
09:34SPEAKERSo yeah, they, they took a big hit and they lost a trillion dollars in value or something ridiculous.
09:40SPEAKERIt actually, that, that dip hit the entire, like every, pretty much everybody involved with AI in
09:46SPEAKERthe, in the U S that has publicly traded stock. And the reason it's interesting, the reasons that
09:53SPEAKERthat happened are interesting. Now, before we go, go, I put this at the end, but I'm going to say
09:58SPEAKERit here. I don't think this means that the AI bubble is bursting yet, but it is like, it's,
10:03SPEAKERit's definitely an indicative of the problems with the current, you know, generative AI market.
10:10SPEAKERAnd as we always say, when we do these stories about generative AI, which I still think is kind
10:15SPEAKERof hooey, there's a lot of, there's a lot of good work happening, even in the generative AI space.
10:22SPEAKERLike I talked to programmers all the time. We were like, yeah, I really love like the code formatting
10:26SPEAKERand code structure and like using generative AI to build unit tests and stuff like that is totally
10:33SPEAKERvalid and makes human written code better and more reliable. That's fine. Yeah. Like, I mean,
10:38SPEAKERI, I certainly don't consider myself a programmer. I have programmed, you know, but like for the
10:42SPEAKERlittle bit of code I do write, like I would never, I would never want it to generate the code for me
10:46SPEAKERmyself because writing it is the fun part to me for one thing. But the one thing I could see maybe
10:52SPEAKERat some point using it for is if I have a bug or like there's something not working the way I
10:57SPEAKERexpect it to or something like having it look at what I wrote and see if it can tell me or put me
11:01SPEAKERon the right track of you know, chasing down the problem. Like I could see something like that.
11:06SPEAKERWell, yeah, the building test is the, is the perfect case study. So, so yeah, I mean,
11:12SPEAKERthere was a thing that was going around the other day about pharma pharmaceutical companies saying
11:16SPEAKERthat the success rate on AI derived drug trials is like 20% higher than it is for traditional methods.
11:24SPEAKERSo they're seeing 80 to 90% success rather than 60 to 70% success, which is like, that's good.
11:30SPEAKERIt's like good use. Nice work, everybody. The generative AI stuff much like, well, we'll get,
11:38SPEAKERwe'll get to it. I think generative media is, is I started to say like without value, like I still,
11:46SPEAKERI still feel that way. Like the only reason I stopped myself is because that's really absolute
11:50SPEAKERstatement and like maybe somebody could come up with a use case for it that might convince me
11:53SPEAKERotherwise, but like generating like, like video and imagery and music with it is not something
12:00SPEAKERthat I find particularly useful, but, but more for, for more of this, like, I don't know, like,
12:06SPEAKERwhat would you, what would you call these more text-based approach and also also like sort of
12:10SPEAKERtask-based approaches, you know, like the kind of remember the example a few months ago of like,
12:15SPEAKERtake a photo of your kid's calendar at school and have it generate an ICS file for you to dump in
12:19SPEAKERyour calendar. You know, like that kind of like task-based use for this sort of thing is actually
12:24SPEAKERlike somewhat valuable if it works. So that's all stuff that I've seen people do like pie in the
12:30SPEAKERsky demos of, and it never actually works in practice. So anyway, the point is this DeepSeek
12:37SPEAKERthing rolled out and people lost their minds for a minute. And we wanted to talk about it because
12:41SPEAKERit's interesting why, why that, why that happened. The kind of the high level answer is that they
12:50SPEAKERused a relatively small number of GPUs compared to what OpenAI and Anthropic and the other big AI
12:56SPEAKERcompanies in the US are using. They're less capable GPUs, which meant that they, their cost
13:03SPEAKERto train this model was a fraction of what a cost of, of like the OpenAI GPT-02, I think, cost.
13:13SPEAKERYeah, like 6% of it basically.
13:16SPEAKERWell, that's for GPT-04. 02 is probably more expensive than that.
13:19SPEAKEROh, sure. Sure.
13:20SPEAKERSo, so yeah. So to give, to give an idea of the numbers here, DeepSeek claimed they used 20,
13:27SPEAKER2048 Nvidia H800s, which is a kind of stripped down embargo, export allowed version of the H100
13:36SPEAKERand about $6 million in compute time to train R1.
13:40SPEAKERYeah, I looked it up. The Nvidia H800 is still about just under $34,000 per card.
13:47SPEAKERYeah, they're expensive.
13:48SPEAKERAnd so I think, I think, I think 34K by 2048 is somewhere around 6 million. So
13:54SPEAKERthat's just cost of hardware then, right?
13:57SPEAKERWell, and probably electricity, the interesting-
14:00SPEAKERMaybe electricity's rolled into that.
14:01SPEAKERI, I would assume so, but I don't know, like data center time, I would assume is part of that whole
14:05SPEAKERcost, but I don't, I don't know. Sam Altman, who's the CEO of OpenAI said that the cost to train GPT-04,
14:13SPEAKERwhich is kind of old at this point was around a hundred or just over a hundred million dollars,
14:18SPEAKERI think is the exact quote.
14:19SPEAKERYeah, was GPT-04 came out like two years ago or something like that?
14:22SPEAKERYear, year and a half or something. Yeah.
14:23SPEAKERDoes that sound right? Before we get too much farther, do you want to talk about the
14:27SPEAKERexport restrictions on Nvidia hardware to China? Like I kind of knew that was a thing,
14:32SPEAKERbut like, you know, by what standard are they restricted? Do you know?
14:35SPEAKERSo they, I went and read the orders while we were, while we were setting up to do this today. And I
14:40SPEAKERam not able to tell from the orders as written. The orders that were written in the 2022 and the 2022
14:47SPEAKERfiling don't have a lot of technical details. They list model, model numbers, like, like generate,
14:54SPEAKERlike they say that ADA chips couldn't be shipped. And they said that what was the one in between
15:00SPEAKERADA and Blackwell? There wasn't one.
15:03SPEAKERWhat's the one before ADA then I guess. Ampere? Are you talking in consumer or?
15:08SPEAKERYeah, it's a hopper was hopper is in between. Yeah. Grace Hopper.
15:12SPEAKERHopper is just a data center thing though, right?
15:14SPEAKERJust a data center chip. Yes. So that's what the H-100 and H-800 are based on.
15:19SPEAKERIt seems like they stripped down the memory bus some or something to make these,
15:22SPEAKERto make these export capable. Also some of these were shipped to China before
15:28SPEAKERthe export restrictions came into play. So.
15:32SPEAKERI'll drop links for some of the articles we read to inform this episode into the show notes,
15:37SPEAKERbut there was a life science article I read about this stuff and they, and they did say
15:40SPEAKERthat Nvidia designed the H-800 partially to explicitly to bypass the export restrictions.
15:46SPEAKERYeah. Right. But they didn't say by what metrics it's doing. So I don't know if that's just a raw
15:50SPEAKERperformance thing or like you said, maybe the, it kind of seems like a classic example of like
15:54SPEAKERgovernment tackling a very technical subject that it's not super well informed on and not
15:58SPEAKERdocumenting technical parameters very well. Yeah. And I'm not a lawyer. I don't know what
16:03SPEAKERthe rule, how you usually write these things. My guess is that's a really difficult job and
16:07SPEAKERit's why you're constantly adding to the, to the requirements. Sure. But the, so that,
16:13SPEAKERthat $6 million figure is disputed because there are a lot of like, like there's a big business in
16:19SPEAKERsmuggling these GPU's into China. Apparently people either set up shell corporations in
16:25SPEAKERcountries that aren't embargoed and then it moved them across the border or like, I assume they're
16:32SPEAKERnot, you know, they're not like carrying 55 GPU's in their carry on suitcase through customs, but
16:38SPEAKERyou know, who knows. The, the TLDR on that, on the training side is that everybody freaked out
16:47SPEAKERbecause the assumption from OpenAI and Anthropic and the other US companies, including Microsoft
16:52SPEAKERand Oracle and all the people that are selling compute, cloud compute to train these big models,
16:57SPEAKERis that like there wasn't efficiency to be gained here. That, that the way to make these things
17:03SPEAKERbetter was just by using brute force scaling, which means bigger data centers, more GPU's,
17:08SPEAKERmore money training, compute training models. And the expectation is that maybe we don't need
17:17SPEAKERto invest in all these data centers and power plants to run the data centers. And we can just,
17:22SPEAKERyou know, do better math. Right. And then make some other tweaks that we'll get into
17:26SPEAKERto increase efficiency. I mean, you know, it seems like whatever else you think about this technology,
17:30SPEAKERthis is still a good development because even if, even if the functioning of the technology
17:34SPEAKERis suspect and the uses that it's going to be put to are also suspect as a result, like at least
17:39SPEAKERwe're not going to melt the planets. We're going to melt it more slowly.
17:46SPEAKERIt means it'll get better as we melt it is the thing. Yeah. Fair. Yeah. Yes. Because they're
17:50SPEAKERnot going to stop building the data centers. They're just going to train even better, faster
17:54SPEAKERalgorithms that lie more effectively to us. Anyway. The other big thing about this R1
18:01SPEAKERmodel from DeepSeek is that it does what the open AIO models do, which is it uses chain of thought.
18:08SPEAKERChain of thought is a term of art in the AI space that basically means the model is smart enough to
18:16SPEAKERbreak down complicated problems into more manageable, smaller problems. And one of the
18:21SPEAKERexamples that they use for this is solving quadratic equations. So if you think about
18:25SPEAKERsolving quadratic equations, it's like you have a square, a variable squared plus or minus variable,
18:34SPEAKERthe same variable multiplied, then you have to like cancel out some of that variable and you
18:38SPEAKERget to a number eventually. And the upshot is it'll show its work along the way. So it says,
18:47SPEAKERokay, first I do this, then I do this, then I do this. And the answer is this.
18:51SPEAKERI saw, I saw this exact methodology applied to the strawberry problem.
18:57SPEAKERWhat's the strawberry problem? The, how many Rs are there in strawberry?
19:01SPEAKERBecause previous AI models have been very bad at answering that question. It turns out,
19:05SPEAKERI don't know if you've seen. Yeah, no, I'm aware.
19:08SPEAKERBut I think this was, I think it was just on Reddit. Somebody had put the question to the
19:12SPEAKERDeepSeek R1 model and it eventually got it right, but it exposes its chain of thought.
19:17SPEAKERQuote unquote, you know, like kind of in like a human readable way, you could read through
19:21SPEAKERall the steps it went through to try to count the Rs. And like, to be fair, it took a shocking
19:26SPEAKERnumber of steps to figure out that there were three Rs in strawberry, but it did get at least
19:30SPEAKERget to the right answer, which is an improvement, but it was kind of interesting to like read it out
19:34SPEAKERstep-by-step and see how it went through letter by letter and stopped itself halfway when it
19:39SPEAKERwas about to answer. I think if I remember it, I'd have to pull the text up. I think it was about to
19:42SPEAKERgive the correct answer of three at one point and then stop itself and second guessed and sort of
19:47SPEAKERwent back and performed a whole new methodology. But I guess you want that cause that's error
19:52SPEAKERcorrecting, right? Like to check your own answers. That's one of the things that the newer models do
19:57SPEAKERis they have multiple things going on at once, or it'll apply multiple models to the same problem.
20:02SPEAKERAnd then like use a different, yet another model to decide which of the kind of competing models
20:10SPEAKERare giving the right answer, right? So the other thing that's interesting about DeepSeek
20:15SPEAKERis that the company that made it owns its own data centers. So they, it spun out of a hedge fund,
20:22SPEAKERwhich uses machine learning to buy and to manage. I think it's like six, it's not, it's, it is,
20:28SPEAKERit is a boatload of money, but not a lot by hedge fund standards. Apparently it's like $6 billion.
20:33SPEAKERThat's practically a boutique.
20:35SPEAKERYeah, it kind of is practically a boutique, but the point is they're paying, they were paying
20:42SPEAKERa lot of really smart people exclusively out of China to come in and do AI and machine learning
20:49SPEAKERmath to make money with the hedge fund. And then they were like, Oh, we have all these GPU sitting
20:53SPEAKERhere. Let's see what else we can do. And, and like, that's when they started doing, that's when
20:58SPEAKERthey started working on this stuff. So they're not, it's interesting because they're not buying
21:03SPEAKERcloud compute time, like OpenAI buys cloud compute time originally exclusively with Microsoft and
21:08SPEAKERrecently they were, they expanded to include Oracle and some other stuff as well. But, but,
21:14SPEAKERrunning on bare metal cluster versus cloud compute time, I think probably,
21:20SPEAKERI think there's probably conversations happening in the Valley now about that because
21:25SPEAKERit's a thing that we found, like game developers find the same thing too, right? Like
21:29SPEAKERyour game server, this is something I know more about than AI shit. So when you're running game
21:35SPEAKERservers, if you, you typically buy some chunk of bare metal servers to provide your base level,
21:42SPEAKERcause you can run way more instances of a game server on a bare metal machine,
21:46SPEAKERthen you can run it on virtual machines running on the cloud just cause the overhead is way better.
21:51SPEAKERMy assumption is the same thing happens here. Sure. Let me tell you somebody who has a server
21:57SPEAKERsitting here in the corner of the room while everybody else is just paying for Dropbox accounts
22:02SPEAKERand VMs in some data center halfway across the country. Sometimes there's just value to having
22:07SPEAKERyour own hardware available. Can't run a Valheim server on a Dropbox account anyway. So it's open
22:16SPEAKERalso is the other thing. I mean, that, that is the most like explicitly and objectively
22:23SPEAKERnoble part of this whole thing is it's not noble. It's just, it's just them sticking the knives to
22:28SPEAKERopen AI and then topic. Oh yeah. But it's still, you know, it's nice to see somebody doing this,
22:32SPEAKERsomebody, somebody with this, this level of resources brought to bear on this problem,
22:35SPEAKERjust putting it out there. It's not open, open. It's not like open source. It's just like they're
22:40SPEAKERproviding the models for downloads. So you can run them on your own hardware. Not, not,
22:44SPEAKERnot the front end or not front end. It's not the right word. Not, not the
22:49SPEAKERphase of it. The, the, the phase of it that resulted in its creation remains closed. Let's
22:55SPEAKERsay like the, the, all the code and the work they did to make the model is not something they opened,
22:59SPEAKERbut the model itself is out there. That's it. The training data, the code to train aren't available.
23:05SPEAKERSo you can't replicate the work, but you can download the model and distill it to run on like
23:10SPEAKERthey have models that are distilled to run on local GPU's for people that have like 40 90s,
23:14SPEAKER40 80s, whatever. Yeah. You've got that 50 90. Have you messed with it yet?
23:18SPEAKERI have not downloaded it yet because, well, I've been busy doing other things.
23:24SPEAKERI've thought about it, but I don't have a use for it. Like I, you know, I don't really want to use
23:27SPEAKERit as for practical purposes. Like I would mess with it with it for half an hour just to see how
23:32SPEAKERit worked. And then I would be like, okay, I'm still probably never going to trust anything you
23:36SPEAKERtell me. So that's kind of where I end up. But look, it's just like that startup. It's
23:41SPEAKERgoing to publish 8,600 books next year. All AI generated people are going to eat them up.
23:47SPEAKERDude, I it's, it's been a while, but for a while, my Kindle the, the unlock screen ad,
23:54SPEAKERyou know, like I've got the Kindle with ads. And when you go to unlock it, there's always a book
23:57SPEAKERad on there. I am positive that every book that I was being served on those ads for like a good
24:02SPEAKERsix months stretch had to all be completely generated. And not just the text, not just the
24:06SPEAKERtext, even like the cover art, like I was getting a bunch of kids books, every single one of those
24:10SPEAKERkids books, even in the like kind of low res black and white little screen context, I could still tell
24:17SPEAKERlike, there's no way a human generated this art. Well, so this is, this is the thing that's weird,
24:22SPEAKERright? Is the upshot on what they're doing is that their cost for accessing their models in the cloud
24:29SPEAKERis a 30th or a 50th, depending on which model you're looking for, you're comparing to as the
24:35SPEAKERequivalent open AI machines. Okay. So let me step back real quick and ask. So it was, it was much
24:40SPEAKERmore efficient to train and create this model, but are you saying it's also equally or similar,
24:44SPEAKERsimilarly more efficient to run it as an end user? Well, so, so yeah, when you're running against the
24:50SPEAKERbig, so, okay. So there's little models that you can run locally, right. But if you want to run
24:54SPEAKERagainst the real model, you're going to have to run that on a bigger machine than most people are
24:58SPEAKERgoing to have in their, in their house. You're not going to have an H100 in your office. Yeah,
25:02SPEAKERof course. But what I mean is, is in terms of like for like output, you know, like to get the same
25:06SPEAKERquality and scale of output as you would get out of 01 or whatever. Yeah. Or 02. Yeah. Like the,
25:12SPEAKERwhatever the current chat GPT model is or for similar results, are you saying it's also more
25:17SPEAKERefficient at that scale? I don't know if it's more efficient, but they're charging 150th what
25:22SPEAKERchat GPT pays charges for 02 access. Okay. Then either it's much more efficient or this is a
25:27SPEAKERgigantic loss leader. Yeah. So there you know, when you're, when you're doing AI stuff at scale,
25:35SPEAKERyou buy AI, you buy tokens, you buy, you know, generated tokens and those have a fixed cost. And
25:42SPEAKERlike they're, they're the price for the deep seek tokens is 150th. You know, you can,
25:47SPEAKERyou can put $2 into your API account on deep sync and deep seek, and it lasts a really long time.
25:56SPEAKERThe, the ups like my assumption is that this is widen the door for people who are interested
26:02SPEAKERin building businesses off of creating a whole lot of low value content. And I think that the
26:07SPEAKERbook spam is one of those models, but I think it's probably the least likely to succeed. I think the
26:13SPEAKERmore likely is that we'll see people that we're going to see people feeding whatever publicly
26:18SPEAKERavailable information they have about you, either from your cookie trail or your social stuff or
26:24SPEAKERwhatever, whatever they can associate an advertiser or an advertisee, like basically the cost is
26:29SPEAKERgetting low enough that we're going to start seeing custom advertising generated by AI.
26:34SPEAKERIf we aren't already, that's, that's and or spam. Like that's, that's where we're at. Cause if
26:40SPEAKERyou're looking at like your book thing, how many, how many Amazon self-published AI generated books
26:45SPEAKERdo they have to sell to customers to make that, that, you know, make their cost, cost structure
26:53SPEAKERwork. And if you're talking about the cost of advertising to the Kindle, Kindle with ads load
26:57SPEAKERscreen, plus the cost of generating the book, like what's, what's the person like, it's a simple
27:04SPEAKERspreadsheet problem to know that you're going to make, you're going to arbitrage out X amount of
27:08SPEAKERdollars on every book that you generate. And then they just hone in the books that are selling
27:14SPEAKERbetter. And it doesn't matter if people read them, as long as they pay the dollar for them,
27:18SPEAKERor they get them into the Kindle unlimited thing and they get a certain tiny amount of money per,
27:22SPEAKERper open or whatever. Right.
27:24SPEAKERRight. This is, this is a little bit of a sidebar here, although frankly AI will absolutely be used
27:28SPEAKERin pursuit of this, but I, I'm getting the sense that we are getting closer and closer to the
27:33SPEAKERminority report style of personalized in-person advertising. I mean, I think that's a real
27:38SPEAKERpossibility now. I think it's much more likely that it's used for spam though. I think there's
27:42SPEAKERa creepiness to the minority report. The minority report thing you're talking about for people who
27:45SPEAKERdon't remember the movie is Tom cruise walks into banana Republic or the gap or something. And it,
27:51SPEAKERit does an eye print on him. Yeah. And it says, Oh, Hey, mr. I mean, he's, he's actually gotten
27:56SPEAKERsomebody else's eyeballs at this point. So it's not him, but wait, does he have literal
28:00SPEAKERbiological or organic eyes eyeballs in his eyes? Yeah. They literally popped. He had,
28:05SPEAKERhe had some, like the way he could move off the grid was to have his eyeballs popped out and
28:10SPEAKERput some aftermarket eyeballs in from some dead guy. I need to see you in minority report again.
28:15SPEAKERMovie's okay. I haven't seen it since the theater. So maybe it doesn't hold up super well. I don't
28:19SPEAKERknow. But yeah, in that case they were doing retinal scan at range, but I need to read some
28:24SPEAKERmore about this, but I saw a couple of headlines past recently about technology that I assume has
28:29SPEAKERprobably existed in some form for a while that they could use your cell phone pinging the cell
28:34SPEAKERnetwork to triangulate your location at the resolution of like individual aisles in a grocery
28:40SPEAKERstore now. And so they can literally map your movements through the grocery store and see where
28:44SPEAKERyou spend the most time browsing and the type of the type of this isn't even advertising. The type
28:50SPEAKERof customizability that was, that they were proposing here was like adjusting prices up
28:54SPEAKERand down per customer, like literally offering you high, you know, massaging prices as you walk
29:00SPEAKERaround the store based on your location and what they know about your shopping habits and your
29:04SPEAKERpreferences and stuff. This is why when Kroger started threatening to put screens price tags on
29:11SPEAKERand the screens instead of physical price tags, people raised hell, right? The assumption is that
29:21SPEAKERmy assumption is that as the cost on this stuff goes down, it dramatically increases the amount
29:25SPEAKERof spam. Already, if you're looking for like what would have been a stock, you know, a photographer
29:31SPEAKERtaken or an illustrator drawn stock image five, 10 years ago, you can't find those anymore because
29:36SPEAKERthere's so much AI crap, just crapping up those results. You just have to go to memes and look for
29:42SPEAKERnonsense because there's no, there's no, it's almost impossible to avoid unless you go specifically
29:47SPEAKERto a site that doesn't allow AI art. And like that's, when you look at reviews, like high value
29:55SPEAKERreview keywords, like I think the example that somebody gave was air filters. There was a site
30:01SPEAKERthat had done really good air filter reviews for like 10 years. And now vendors are selling
30:07SPEAKERreview generating bots to places that don't care about their review integrity, like Marie Claire.
30:14SPEAKERIt was a lot of like home magazines and stuff like that. And then they'll have hidden pages that are
30:20SPEAKERlinked invisibly and never exposed to the mainline real life consumer that just generates a bunch of
30:29SPEAKERair filter reviews off of things other people have said and Amazon reviews and whatever that
30:33SPEAKERmay or may not be good that generates affiliate revenue for those sites. Cause their search
30:37SPEAKERtraffic volume is good. Snake is eating his own tail. It's it's it's things are bad.
30:44SPEAKERI think I said this on discord on one of the discords you and I are on not,
30:48SPEAKERnot very long ago, the cyberpunk authors did not go far enough.
30:52SPEAKERNo. Like, like they, they nailed it. Like let's be real. Like they called it. They
30:57SPEAKERsaw where things were going. They didn't see how far it was going to go.
31:01SPEAKERWe're yeah. Like the idea. Look, the financial aids and snowcrasher are up there, but we, we,
31:07SPEAKERwe still need to read neuromancer. We promised we said, we're going to read neuromancer at some
31:10SPEAKERpoint. We should do that. It's, it's going to seem quaint by comparison. Oh, I believe it.
31:15SPEAKERSo, okay. So they did it efficiently. The other thing that they did that was interesting
31:19SPEAKERis they did this really efficiently by they, they, they, they didn't use, usually you start
31:27SPEAKERwith known good data and human reinforced training, right? So you, you start and you
31:34SPEAKERdo the initial kind of cold start training with human feedback or human trained data that is,
31:40SPEAKERthat is labeled and marked and known good. And it takes a really long time to collect
31:44SPEAKERthat data at it's expensive and it's time consuming. They didn't do that. They just
31:52SPEAKERstarted training it and then did the reinforcement after the initial training pass and then built
31:57SPEAKERanother model on top of the first one with good, with a, a very small, as I understand it by
32:04SPEAKERtraining standards, amount of known good data for the cold start on the second,
32:09SPEAKERthe second rev of this model. Now I wonder, is that a situation where the model is able to take
32:15SPEAKERthe much broader lessons that it's learned from that big bulk of data and compare against the,
32:21SPEAKERthe kind of more known good results it's getting from the smaller set and like retroactively learn
32:25SPEAKERmore about what it got right and wrong on the main phase of the training from this smaller,
32:30SPEAKERthis kind of sub phase. Does that make sense? Well, so remember it's not, it's not doing,
32:35SPEAKERit's not learning, right? None of this is actually learning. Well, no, I don't mean in like a kind
32:39SPEAKERof like a sapient human sense, you know, but like it is building a model, you know, it is, it's,
32:43SPEAKERit's establishing parameters that it thinks are correct. Where, where think just describes the
32:48SPEAKERprocess of this thing operating in a calculating way. Like, you know, I, I, I've, I'm, I'm very
32:53SPEAKERclear. This is not actually intelligence. No, no, I know. I just, I just, I just like to be careful.
32:57SPEAKERIt's, it's finding adjacencies. It's finding things that are adjacent. Like it, it,
33:01SPEAKERit's finding relationships between things that it understands. And it's, it's often developing that
33:06SPEAKERunderstanding by like abstracting out the information, an additional layer so that it
33:12SPEAKERbecomes so that like letters and words become numbers. And it knows that this, this number is
33:17SPEAKERusually adjacent to this number. Right. And then it reverse translates those back into, into words.
33:24SPEAKERSo, so I don't know about the answer to your like, yeah, probably that sounds right, but
33:31SPEAKERa lot of the AI stuff is so deep. I can't believe I'm going to get roasted for complaining by this.
33:36SPEAKERThere's so much jargon and all this AI stuff that it's, it's difficult to parse a lot of times,
33:42SPEAKEReven for somebody with a pretty good understanding of what's going on. The, the, the base thing
33:47SPEAKERthat's happening still with all of this is that they're mimicking the behavior of neurons in a
33:52SPEAKERneural net, which, which is a thing that has been able to be mathematically quantified to the point
33:59SPEAKERthat you can run algorithms on GPUs at massive volume. Yeah. See, I did not know that. Like,
34:05SPEAKERto kind of, I guess this is kind of to your point about calling this artificial intelligence or
34:09SPEAKERdescribing what it's doing as thinking or whatever. Like I didn't realize that like
34:12SPEAKERtopographically it is similar to. It's mimicking.
34:16SPEAKERA layout of neurons in a brain, you know, like I didn't realize the correspondence was actually
34:20SPEAKERthat direct. I mean, it may not be that direct, but you get what I'm saying.
34:22SPEAKERI was just saying, that's an important, it's not in a brain. It's in a, it's in an artificial neural
34:26SPEAKERnet. So in a brain, the connections are all three dimensional and they can loop back in on each other
34:30SPEAKERin really weird ways. Okay. In, in the artificial, as far as I know, they're still mimicking 2d neural
34:37SPEAKERnets that only loop back onto each other in a planar way, not in a three dimensional way.
34:43SPEAKERSure. Okay. And yeah, anyway, I did my, I did my senior dissertation on this in college about
34:49SPEAKER25 years ago, 30 years ago. And it was, it was, I don't remember any of the important things at
34:56SPEAKERthis point, it turns out. And also it was mostly theoretical and it was like, at the time people
35:01SPEAKERwere lining up like dozens or hundreds of neurons on plates and then feeding signals into them to
35:07SPEAKERsee what happened. Yeah. So anyway, I mean, I wonder, that actually makes me wonder, I mean,
35:12SPEAKERI'm sure this is well established at this point, but is, is neural net potentially why people
35:16SPEAKERstarted calling this artificial intelligence? Do you think? No, they started calling it artificial
35:20SPEAKERintelligence because it makes people, people understand what that means. And they don't
35:23SPEAKERunderstand what a neural net is. Machine learning does not have the same ring.
35:28SPEAKERWell, yeah. I mean, and machine learning is what it is, right? So the other thing,
35:33SPEAKERthe other thing that they did is they they're using, because they're running on bare hardware,
35:36SPEAKERI assume they were able to run what's, it's not exactly, but kind of the equivalent of assembly
35:42SPEAKERcode for these GPUs. They're running Nvidia's parallel thread execution, they call PTX instead
35:50SPEAKERof a CUDA. So it's, it's one step removed and less, less abstracted, I guess. My understanding
35:56SPEAKERof that stuff, I mean, this is basically what you just said, but from reading people talking
35:59SPEAKERabout this on Twitter, who seem to know what they're talking about, like, it's kind of like,
36:03SPEAKERyou can kind of look at CUDA as equivalent to C or C++ and PTX is assembly language, like you said,
36:08SPEAKERbut it's like CUDA is pretty close to the metal, but not as close as doing hand optimizations
36:13SPEAKERof this nature. And like those hand optimizations are like using 8-bit stuff where 8-bit math makes
36:22SPEAKERsense and using 32-bit when you need the extra precision stuff like that. Well, what I read about
36:26SPEAKERthat stuff is that the previous models trained by OpenAI and companies like that are pretty
36:31SPEAKERpurely 32-bit or predominantly so. That sounds expensive. I think that's for the waiting between
36:37SPEAKERneurons, if I understood. Ask me anything about this technology, but I, I believe what I read was
36:43SPEAKERDeepSeek moved to 8-bit precision and that we're talking about precision on these types of numbers.
36:49SPEAKERIt's basically just how big the numbers can be, right? Or in the context of a decimal number or
36:55SPEAKERhow many decimal places you can have rather than how big the number can be. Look, it's all just
36:59SPEAKERsimple formulas, Brad. I just pasted one into chat for you so you can see it's just work that out.
37:04SPEAKERAnd that's, that's the secret of AI. It's just, just that simple. I'm sorry. I'm really looking
37:11SPEAKERhere. There are next to no numerals in this equation. A lot of pies. I see, I see a numeral,
37:17SPEAKERlet's see, I see three different instances of a numeral one. Yeah. And I'm looking, let's see,
37:23SPEAKERthere's a superscript one and that's about it. Every single other, every, every bit,
37:28SPEAKERevery other element in this equation besides that is something other than a standard number.
37:33SPEAKERIt's a lot. Yeah. Anyway, anyway, the weights, I believe the weights between neurons is if I
37:39SPEAKERunderstood correctly is where they reduce the precision down to 8-bit. That makes sense.
37:43SPEAKERApparently that, apparently that does, as you might imagine, it does decrease the accuracy of
37:48SPEAKERthe model slightly, but the apparently the trade-off in extra speed they gained and extra
37:53SPEAKERsavings on training compute requirements were more than worth it. Well, and then,
37:57SPEAKERso there's one other theory about how they did this so efficiently, which was open AI says that
38:02SPEAKERthey have evidence from Microsoft that someone was extracting large amounts of data from the model
38:08SPEAKERas if they were doing a process called distillation to basically steal the gist of the,
38:15SPEAKERI think the 01 or 02 model. I'm not sure which one. And they were, I'm going to go tell you,
38:22SPEAKERfairly upset about this, which is pretty rich to me, given that they trained all of their models on
38:28SPEAKERthe entirety of the open web that they scraped a few years ago. Yeah. Yeah. And I'm in the
38:32SPEAKERdistinction here and I'm not defending open AI. This is a distinction from a technical perspective,
38:36SPEAKERnot an ethical one, but the distinction here is that they did take a mass quantity of raw data
38:41SPEAKERthat yes, again, they did lift from people without permission, but they did do some amount of
38:46SPEAKERprocessing on it to make it more usable, I guess, for deep seeks purposes. Stealing, stealing is
38:52SPEAKERfine. As long as you don't steal from us, Brad, that's the takeaway I have here. Yeah. So yeah,
38:58SPEAKERwe're not really in a position to evaluate how capable this stuff is. The general consensus among
39:04SPEAKERthe AI experts that I follow and trust is that it seems about as good as Chad GPT 02. The cost
39:11SPEAKERstructure is dramatically better. It's wildly popular as a result right out the gate. It was
39:16SPEAKERthe number one app on iOS the day that this announcement hit, free app. Is that, I assume
39:21SPEAKERthat that app is just an interface to a cloud instance, right? Yeah. It's an interface for a
39:26SPEAKERcloud instance. Yeah. I assume no matter how much neural hardware is on an iPhone, I assume it
39:30SPEAKERcouldn't even run the smallest version of this. I think that I would say you could probably distill
39:35SPEAKERit down to the point that you get something that you're not going to want to run on it, but yeah.
39:38SPEAKERYeah. That would make sense. It's available on PC. I know you said you hadn't messed with it.
39:43SPEAKEROllama offers it. Are you familiar with Ollama? Yeah. Ollama is a tool that you use to,
39:51SPEAKERbut the Ollama thing is a cloud front end. I think you can, you can also download and run stuff locally
39:55SPEAKERwith it. They let you pick models. Oh, that was, that was going around widely. Like the day this
40:01SPEAKERreleased or this news came out, was that like, Hey, it's already built in here. You can just
40:04SPEAKERdownload this and run it. But I started to grab it a, like I said, not just not a lot of interests,
40:09SPEAKERbut B there's kind of an epidemic. I feel like in a lot of open source software these days where
40:14SPEAKERthere's just a very conspicuous lack of a top line statement of, Hey, this is what this is.
40:20SPEAKEROh yeah. There's a lot of that. Like there are a lot of projects that I run into these days that
40:23SPEAKERjust like could really use two sentences at the top that just like explicitly say, Hey, this is a,
40:28SPEAKERthis is an interface to this. You know, this is a thing that will let you use this underlying
40:32SPEAKERtechnology or this does that. You know what I mean? Like, yeah. Like sometimes it's not,
40:36SPEAKERit's not clear where in the ecosystem or the stack of things sits that you're about to download and
40:41SPEAKERrun, which makes it harder to use something if you don't know exactly what its purpose is. Well,
40:44SPEAKERit's funny because the last time I downloaded, downloaded Ollama, I'm pretty sure it just
40:48SPEAKERinterfaced with the web versions of the, of the models. I think it does both be clear. But yeah,
40:54SPEAKERthat makes sense. You can run DeepSeek and some other stuff locally. Oh man. I'm going to turn
40:59SPEAKERsome fans on. If you go to Ollama.com, wait, are you running it right now? I'm installing it right
41:04SPEAKERnow. Wait, which version, the windows one? The windows one. Okay. Yeah. If you go to there,
41:10SPEAKERif you go to Ollama.com, it says, you know, run llama 3.3, which llama is Facebook's model,
41:15SPEAKERis that right? DeepSeek R1, Mistral, Gemma 2, and other models locally. That sounds fun. Yeah.
41:23SPEAKERYeah. I don't know. So one of the other things that happened as a result of this is that there
41:28SPEAKERwas a, there was a multiple bills submitted to Congress in different houses about the varying
41:36SPEAKERlevels of illegality that they want to make DeepSeek. There's a bipartisan bill from
41:42SPEAKERrepresentatives, Josh Gottheimer and Darren LaHood in the house to ban its use on government
41:47SPEAKERdevices, which seems wise because there is a fair amount of concern that especially using the cloud
41:52SPEAKERversion is going to leak sensitive data back to China, which is probably against the best
41:58SPEAKERinterest of the United States for a whole multitude of reasons. Josh Hawley in the Senate,
42:03SPEAKERwho's a kind of noted reactionary, introduced a law that would ally enormous civil and criminal
42:08SPEAKERpenalties for installing DeepSeek for anybody or any other Chinese AI technology. If you're in the
42:13SPEAKERU S seems pretty broad. Yeah. It would seem like it was, that would be a problem in the longterm,
42:18SPEAKERbut anyway don't know what's going to happen with either of those. It seems like the the status of
42:26SPEAKERnew legislation in the United States legislature is fraught right now for reasons to the extent
42:34SPEAKERthat we still have a legislature. Yeah. And like I said upfront, I don't think this means the AI
42:40SPEAKERbubble is bursting. I think people are reevaluating their assumptions. It definitely removed a lot of
42:48SPEAKERvalue from the stock market. I think the next question is what's going to happen when open AI
42:52SPEAKERruns out of money and has to go out with their hat in their hand and who's going to keep pumping
42:59SPEAKERmoney into their system if they don't show more responsible spending. I have had multiple
43:10SPEAKERconversations with venture folks, not at the level that open AI would be talking to people, but
43:15SPEAKERwhen you say, "Hey, I want to run metal servers." They look at you like you're a crazy person
43:21SPEAKERand the result, because if you're building, if you made a bunch of money on apps 10 years ago,
43:28SPEAKER15 years ago, that business was basically AWS, the invention of Amazon web services,
43:37SPEAKERwhich is a cloud-based way to virtualize computer hardware.
43:40SPEAKERIt's the way. Well, I don't mean the in the sense that it's the only one, but as far as I know,
43:46SPEAKERit's still far and away the biggest one. Less so these days than before.
43:52SPEAKERI know Microsoft's Azure is I think still number two. It's quite big too, of course,
43:56SPEAKERbut AWS was the first and by far the biggest. Yeah, AWS invented this whole concept, right?
44:02SPEAKERThat's what I'm... Yeah, that's what I meant.
44:08SPEAKERJust to be clear, they didn't invent the concept of virtualizing machines. What they invented was
44:11SPEAKERthe idea of selling bits and pieces of virtualized machines as individual services. You could buy
44:17SPEAKERstorage and compute and deep long-term storage and databases and all of these other things
44:24SPEAKERseparately. Yeah, and making that easy to automate and spin up piecemeal as it's required and not
44:31SPEAKERhave a bunch of servers sitting there idle when you don't need them.
44:34SPEAKERYeah. There's a reason that sometime around 2012, you stopped having the thing where something
44:41SPEAKERgoing massively viral would take down the website. It's because everybody started putting these
44:45SPEAKERservers in front of your web servers that would just spin up more servers on Amazon if you needed
44:50SPEAKERthem for 10 minutes. Then it would spin them back down when the traffic surge went away.
44:58SPEAKERIt wouldn't surprise me given the Valley's love of cloud compute if nobody ever said, "Hey,
45:04SPEAKERwhat if we built our own data centers?" Well, then we'd own the hardware and that would be a
45:08SPEAKERproblem because then we'd have a bunch of old hardware every five years that we'd have to
45:11SPEAKERfigure out what to do with or upgrade the capital cost on the data centers. Facebook has a shitload
45:17SPEAKERof data centers. Twitter and Google have a shitload of data centers. Apple has a lot of
45:22SPEAKERdata centers. It's surprising to me that the OpenAI and Anthropic are buying access to these computers
45:29SPEAKERrather than just building the data centers themselves. Anyway, I think there's going to
45:34SPEAKERbe conversations about that. Ethan, do you think that speaks to their business trajectory?
45:39SPEAKERI think, look, if they wanted building data centers as Google found out the hard way for
45:45SPEAKERthe first 10 years of the 2000s is difficult. Their Google solution at the time was to just
45:56SPEAKERmake the computers that they were putting in there as cheap as possible and run Linux on them in such
46:00SPEAKERa way that it would just work even if the machines were janky and started failing. Then track the
46:06SPEAKERfailures and make sure you replicate enough that the failures don't take anything down.
46:10SPEAKERUnfortunately, with the AI stuff, they can't do that because the only way they can get the GPUs
46:14SPEAKERis to buy $50,000 GPUs from NVIDIA. Those come in $350,000 boxes that then you put 800 of those GPUs
46:24SPEAKERin a rack or whatever. You have to have enormous cooling and enormous power. In order to build a
46:30SPEAKERGPU data center at this point, you have to basically design the data center from the ground up
46:34SPEAKERto accommodate the power and cooling requirements. You can definitely see what a Herculean task it
46:39SPEAKERwould be. I guess what I meant by business trajectory is does this speak to what kind of a company open
46:44SPEAKERAI wants to be or what their end goal is? Because I would think they would be tripping over themselves
46:50SPEAKERto own their own data centers because they wanted to be the provider of this technology to every
46:55SPEAKERother company on the planet in the same way that AWS is for basic cloud compute, right?
47:02SPEAKERPeople often talk about capital efficiency and startup world. That often means you don't
47:08SPEAKERbite off more things than you can chew. If they're focused on building models, they might say, "Look,
47:14SPEAKERwe're going to spend more money on data training models right now because we're buying that
47:18SPEAKERresource, but it's cheaper than us building up a 3,000 employee division that's just building data
47:25SPEAKERcenters around the world." Yeah. To be fair, my example was really putting the cart before
47:29SPEAKERthe horse because they need to prove sustainable demand to what they have before they go building
47:36SPEAKERthe infrastructure to provide it. I don't think there's no evidence that there's going to be
47:40SPEAKERlong-term sustainable demand for this technology yet, right? Yeah. That's the thing. Even the
47:46SPEAKERplaces where they're charging large dollar tickets for lots of tokens, they're still not breaking
47:51SPEAKEReven on the cost of generating those tokens for open AI, at least in the last announcement that
47:57SPEAKERthey had about that stuff. That's ultimately going to be what pops the bubble is that
48:05SPEAKERMicrosoft has co-pilot, and they recently made co-pilot free for a whole swath of customers
48:11SPEAKERthat used to have to subscribe to an extra tier of their licensing provider to get co-pilot access.
48:20SPEAKERIf people aren't using co-pilot, there's no viable use for co-pilot, then what's the point of
48:28SPEAKERgenerating all these tokens, right? Anyway, that's it. That's the deep-seek. I don't know. It's like
48:36SPEAKERa quick look for this weird thing that happened. Sure. From some people who have played the game
48:41SPEAKERfor 20 minutes. Yeah. We've been paying attention to the AI stuff enough to know what's going on
48:50SPEAKERwith it. The hard part for me about talking about AI stuff is that the places that cover it and are
48:57SPEAKERdoing a good job tend to go from high level overview to terrifying formulas in three paragraphs.
49:06SPEAKERA lot of it is difficult to... I say the basic math, but it's not basic math. If you don't
49:14SPEAKERunderstand the math that's underlying the entire technology, you quickly can get out of your depth.
49:19SPEAKERThere are a handful of people doing a really good job talking about this. Ed Zitron is
49:23SPEAKERfantastic. The Hackaday folks have good explainers about it regularly. I think 404 Media had some
49:30SPEAKERgood stuff that I read leading up to this. We'll put links for the stuff that we talked about
49:36SPEAKERin the notes though. I guess that's probably it. Yeah, that's it. Did we even say what the other
49:41SPEAKERtopic was going to be? It was tariffs. Bonus topic. The bonus topic is going to be tariffs,
49:45SPEAKERbut actually maybe it's not a topic anymore because they seem to have backpedaled on just
49:48SPEAKERabout everything. Well, I guess the China tariffs are still in place, right? Everything got pushed
49:53SPEAKERback except for the China tariffs right now. And I don't know what's up with Taiwan. I think he
49:59SPEAKERthreatened the Taiwan terrorists, but hasn't actually applied them. That was a bunch of
50:02SPEAKERloose talk as far as I know, whereas all the others have been signed. The Taiwan ones are
50:08SPEAKERby far the most insane to me, both the targeting and the scale of them. Taiwan is... Will they make
50:15SPEAKERthe chips? I mean, but also just geopolitically, I think they're still an ally of ours, right? But
50:19SPEAKERthat doesn't mean shit anymore. But the other thing is up to a hundred percent tariffs is what
50:24SPEAKERwas being quoted. Well, okay. So when we talk about tariffs, I like to go reeconomists. I went
50:30SPEAKERand read some economists and the big thing that I saw is that labor and it's cheaper to build fabs
50:38SPEAKERin Taiwan. It's a little bit less stable because of the earthquake zone that they're in and the
50:43SPEAKERwhole mainland China not liking Taiwan situation. But the labor for chip engineers in Taiwan is
50:53SPEAKERroughly... Well, in the US, roughly costs 3.5 times what an equivalent engineer costs in Taiwan.
51:00SPEAKERSo the cost of building the fabs, the cost of staffing the fabs, the cost of running the fabs
51:05SPEAKERbecause of EPA and power and water in the places that we build fabs in the US is a lot more to them.
51:13SPEAKEROne of the economists... When we were reading about the tariff stuff, the general consensus is
51:18SPEAKERthis kind of broad sweeping tariff without any kind of thoughtfulness about the rollout,
51:22SPEAKERit's impossible to know what the impact is going to be in the longterm.
51:27SPEAKERYeah. To the point that I think even if they hadn't backpedaled on a lot of these policies or at
51:30SPEAKERleast delayed them for a month or whatever it is, trying to cover this with any kind of coherency
51:35SPEAKERis a fool's errand anyway, because it will assuredly all be out of date by the time this
51:39SPEAKERepisode runs. The nice thing about the tariffs for the Nvidia folks and all the other AI companies
51:47SPEAKERthat lost a bunch of money is the tariff stuff rolled out the next day and completely erased
51:51SPEAKERthose losses. So you didn't even notice that there was a dip the day before. One of the economists I
51:58SPEAKERread said it might just be cheaper to pay the tariffs and keep buying stuff from Taiwan because
52:03SPEAKERthe pipeline on new fabs, which we talked about a few weeks ago, is four years minimum.
52:10SPEAKERAnd the cost of building the fabs has almost destroyed Intel. They've spent so much money
52:17SPEAKERspinning up US manufacturing over the last five years that it's weakened their entire business as
52:24SPEAKERa whole while they wait for these enormous longterm investments to come to fruition.
52:32SPEAKERIf you don't want to put your business at risk, you might be better off just paying the tariffs.
52:37SPEAKERAlthough Samsung and TSMC and a bunch of other places are building fabs here.
52:43SPEAKERThere were all kinds of other subtopics to address here, like the US postal service
52:47SPEAKERwas blocking all shipments from China and Hong Kong for what turned out to be something like
52:51SPEAKER12 hours. So that was going to cause a massive disruption in the kind of like small boutique
52:57SPEAKERelectronics field from AliExpress and stuff like that. But that is no longer the case.
53:02SPEAKERThe handheld people were flipping out on my subreddits. Your INEOS out there were seeing
53:09SPEAKERtheir business model challenged pretty quickly. Look, I'm a staunch proponent of the Treaty of
53:14SPEAKERBern, which has regulated international mail since 1874. Without that, we'd have to have individual
53:21SPEAKERtreaties with all of the other countries in the world, which is just unworkable.
53:26SPEAKERThe Ulysses S. Grant was a big handheld gaming guy.
53:31SPEAKERHe loved getting mail. He loved getting international mail.
53:33SPEAKERIs that 1870? My late 19th century presidents are a bit of a jumble.
53:38SPEAKER1874. Yeah, that would be in that. It's somewhere in there. Gremlin Cleveland, one of those guys.
53:46SPEAKERSo yeah, the big thing on the small packages is that there's an exception for things that
53:50SPEAKERcost less than $800 shipped from China. So you don't have to pay any kind of duty or tariffs
53:54SPEAKERor anything on this coming in from the US. Is that the de minimus?
53:58SPEAKERThat's the de minimus, yeah. Or policy.
54:01SPEAKERYeah. Whenever there's a Latin name on it, you know, it's serious.
54:05SPEAKERWell, when that went into effect, it wasn't really a problem. And then like Shine and Timu
54:12SPEAKERand AliExpress and all those places have started using that as a way to bypass paying tariffs on
54:18SPEAKERtheir entire business, duties on their entire business. So anyway, we'll see how that goes.
54:24SPEAKERMaybe we'll revisit this topic more comprehensively when it actually settles down and if it ever
54:29SPEAKERactually comes to pass. Yeah, the upshot from everything that I was able to read for electronic
54:35SPEAKERstuff is that no one knows what the impact of a 25% flat tariff on China and no one really,
54:45SPEAKERreally, no one knows what 100% impact on Taiwan, except for that it's probably going to be more,
54:51SPEAKERthe cost increases will probably be more than the bill of goods times the tariffs plus the
54:58SPEAKERindividual line items on the bill of goods times the tariffs that are affected by each of the items
55:02SPEAKERon the bill of goods added up. It's going to be more than that.
55:05SPEAKERMore than that still. I can see that.
55:08SPEAKERBecause there's like middlemen in the way that take extra money and stuff like that. Anyway.
55:12SPEAKERFun times.
55:13SPEAKERYeah. Elections matter. That's the TLDR. This is the time where we thank our patrons, Brad.
55:20SPEAKERIt is. Thank you, patrons.
55:21SPEAKERThank you, patrons. We really appreciate you. As always, Brad and Will Made of TechPod is a
55:25SPEAKERlistener supported show. That means we wouldn't be here without you, the listener. For five bucks a
55:30SPEAKERmonth, you can get access to the fabulous TechPod Discord, which is full of nerds who talk about
55:36SPEAKERthings like duty on handheld gaming devices and keyboards and audio interfaces and Brad's janky
55:48SPEAKERcompressor slash limiter. Sorry, Brad. I think I talked about the spicy ramen in there this morning.
55:54SPEAKERYou probably did. I'm so mad about this thing, man.
55:56SPEAKERI'm sorry. That sucks because you did a fair amount of research on it, too.
56:01SPEAKERI mean, this thing was spoken quite highly of. To me, though, it's like a design flaw is one
56:04SPEAKERthing, but just cheaping out on manufacturing, like having it be crappy because you just didn't
56:10SPEAKERwant to pay. I mean, you tell me. It's just extra per unit, even $5 extra per unit or whatever it
56:16SPEAKERwould have cost to, you know, make the tooling such that this thing would come out of the factory
56:22SPEAKERready to go consistently. Like, come on. If you want to talk about cheaping out on
56:26SPEAKERmanufacturing, I can tell you about the A.O. Smith Water Heater Company, Brad.
56:30SPEAKEROh, oh. I had a, don't ever buy water heaters at Lowe's. This is the TLDR.
56:38SPEAKERThey sell a variety of brands like Maytag and Whirlpool and stuff like that.
56:42SPEAKERApparently ours was, despite having filled out the warranty cards, we'd get notifications of
56:46SPEAKERrecalls. Ours was under recall. And I found this out because I went out to the garage the other
56:51SPEAKERday and it was making sounds like a motorcycle that was backfiring.
56:54SPEAKEROh my.
56:55SPEAKERAnd then when I looked at it, there was fire coming out of the hole where the gas goes in,
56:59SPEAKERwhich is not what you want.
57:00SPEAKERHow long ago did the recall go into effect?
57:02SPEAKERLike six years.
57:05SPEAKERSo I called them and they were like, "Oh yeah, it's under recall. We'll send you apart and then
57:11SPEAKERpay to fix it." And I had already called the water heater person who deals with our water heater,
57:17SPEAKERusually. They came out and I was like, "How dangerous is this?" He's like, "I wouldn't have
57:20SPEAKERone of these in my house." I was like, "Well, of course you wouldn't, because you're trying to sell
57:22SPEAKERme a hot water heater." So I got a quote and I let him go back to the car and I started looking
57:28SPEAKERup what had happened to the water heaters that caused a water heater recall for every water heater
57:33SPEAKERmade by this company for X period of years. And then I called them back and was like, "Hey,
57:37SPEAKERcan you come put the other water heater in?" So I got a new water heater yesterday afternoon.
57:40SPEAKERAnd that's the other thing that your subscription dollars support is my new water heater.
57:45SPEAKEROh, not free of charge, to be clear.
57:47SPEAKERNo, I had to pay $2,500, $2,400. It was fucking expensive.
57:51SPEAKERI'm sorry. I may have zoned out for a second. I thought you were saying that the company with
57:54SPEAKERthe faulty water heater was replacing them.
57:56SPEAKERYes, they were going to, but it's also 10 years old and there's about a 10 year lifespan on water
58:01SPEAKERheaters usually. So I was like, "Let's just fucking do it rather than have the water heater go out
58:06SPEAKERagain in two months." And also, I don't know that I trust the company that was blowing up people's
58:11SPEAKERhouses with their shitty designs. Anyway, so yeah, we appreciate every patron, everybody who listens,
58:19SPEAKERbut especially our patrons. And you can find out how to support the podcast by going to
58:23SPEAKERpatreon.com/techbod. Again, that's patreon.com/techbod. And for five bucks a month,
58:28SPEAKERyou get access to the Discord, you get access to the patron exclusive episode where we talk
58:32SPEAKERabout stuff and things. Last month, people really... It was very divisive. Some people,
58:38SPEAKERnot a fan. Some people really enjoyed just sitting around and talking about stuff.
58:43SPEAKERYeah.
58:44SPEAKERSo let us know where you ended up, I guess.
58:47SPEAKERDid you say that? I'm sitting here second guessing my part in what you just described,
58:50SPEAKERwhich was substantial.
58:51SPEAKERI mean, look, usually I'm the one who rambles, so it's only fair that you do it occasionally.
58:55SPEAKERIt was a lot of a haphazard stream of consciousness about UEFI and secure boots
59:00SPEAKERand drive encryption, various related...
59:02SPEAKERBest to leave the mystery on the exact contents of the rambling.
59:06SPEAKERYeah. So hopefully maybe that'll cohere into a full episode at some point. I think I'm going
59:10SPEAKERto finally put refined on this weekend. In fact, I put my spare SSD in the machine yesterday, so...
59:17SPEAKERIt's all happening.
59:17SPEAKERIt's all down to the software end now. I might install a Linux or two this weekend.
59:22SPEAKERI think the next step is to... I think the next... The other thing we need to do is the...
59:28SPEAKERWhat are they called? The non-carrier carriers? The phone carriers?
59:31SPEAKEROh, the MVNOs.
59:32SPEAKERMVNOs. Yeah. I want to know more about MVNOs.
59:35SPEAKERMobile virtual network operator, I believe.
59:36SPEAKERThat's it. Yeah. I love them.
59:37SPEAKERI switched.
59:38SPEAKERNo. I know.
59:39SPEAKERI switched. I'm on a new carrier now. I'm on one of those.
59:42SPEAKERYeah. Well, I think we should break... When I looked at the offers on those,
59:46SPEAKERthey are all over the place.
59:48SPEAKERThere's quite a bit going on there.
59:50SPEAKERIt's wild. Yeah.
59:51SPEAKERCan I tell you something?
59:52SPEAKERYeah.
59:53SPEAKERI'm still in the free trial period with one I picked.
59:55SPEAKERYeah.
59:56SPEAKERIt might work out. I haven't picked a plan yet.
59:59SPEAKERIt may work out that my phone bill is going to be $10 a month.
1:00:04SPEAKERThe thing I... Honestly, the thing I want to know more than anything else is what it is that
1:00:08SPEAKERwe're losing. Because they're so dramatically cheaper than the Verizon and H&Ts of the world.
1:00:13SPEAKERI have suspicions. I don't know if they can account for a production of that magnitude, but...
1:00:17SPEAKERYeah.
1:00:18SPEAKERMaybe I don't want to burn all the material right now, though.
1:00:20SPEAKERYeah. No. But if people have thoughts about MVNOs, send them in too, please.
1:00:25SPEAKERAgain, the address is... This is the rambliest outro that we've ever done.
1:00:28SPEAKERPatreon.com/TechBod. As always, we appreciate every support. We especially appreciate our
1:00:34SPEAKERexecutive producer tier patrons, including Andrew Slosky, Jordan Lippitt, Bunny Felonies,
1:00:40SPEAKERTwinkle Twinkie, David Allen, James Kamik, and Pantheon, makers of the HS3 high-speed 3D printer.
1:00:46SPEAKERThank you all so, so much.
1:00:47SPEAKERThank you.
1:00:48SPEAKERAnd that'll do it for us this week, Brad.
1:00:50SPEAKERGood luck with your compressor.
1:00:52SPEAKERYeah. Yeah. Yeah. And we will see you all next week.