Can Your Laptop Handle DeepSeek, or Do You Need A Supercomputer?

0:00Have you ever wondered if that laptop sitting on your desk could power these open source models you're always hearing about? Let's talk about that.

0:21All right. Welcome, humans, to another electrifying episode of the Neuron Podcast. I'm Corey Knowles, editor of the Neuron, and here with me is the Neuron's writer, Grant Harvey. How's it going, Grant? Good, good. How are you doing, Corey? Oh, I'm doing good today. Excited to have this conversation because I think there's a little bit of confusion around open source when it comes to people from a non-technical background trying to understand what this means to them. Yeah, definitely. So, with that said, today we're going to dive into DeepSeek, the Chinese AI that broke the internet earlier this year.

1:01Yeah. Have you ever tried DeepSeek, Corey? I have I have I've used their web portal yeah yeah yeah I've done that as well I've never actually run it on my machine which we'll get into as to why um uh but I have uh messed around with the portal a little bit and there's different versions you know kind of like chat gpt uh there's there's you know there's the regular version of it and then there's the thinking version which I think is what has a lot of people excited is that r1 is what it's called it's like chat gpt's 01 I will say I've read some crazy stuff out of its thoughts. Well, I think you showed me one where it was like it could think perfectly in the style of Donald Trump, which I thought was pretty funny.

1:44It did. And there was another one, you know, I read the other day and I can't remember what it said right now, but it was about it was trying to think and trying to work through a problem and it couldn't do it. And then it's like, OK, well, maybe I'm going to just take a different approach here and stop. And it goes through this whole, it was kind of comical, the approach it took to it. From the front end, it's good. I mean, it's not, I haven't, like, stopped using ChatGPT to use it more, but it's fine. Yeah. You mentioned that it broke the internet. It also broke some laws, too, which was an interesting aspect of this that we got to talk about.

2:24because it's not like this is just another chatbot that you can just compare against ChatTPT. I mean, that's kind of complicated. It's got some interesting allegations against it, some alleged military ties to China's military. They may be stealing chips, breaking US laws, and maybe kind of a security nightmare. So I don't know if this is something you want to just like spin up for funsies. You got to know what you're doing. There's kind of nothing we can do about it at all is the other interesting aspect here. But I guess first let's kind of set the stage here with who DeepSeek is. If I was going to put it in one quick sentence, I think what I would say is that it's a Chinese AI company releasing fully open source models that rival GPT-4 but also came out to rival O1.

3:17Yeah. Yeah, definitely. yeah there's uh it's definitely china's version of open ai um except uh they're giving away the secret sauce for free so you know a one obviously was this big guarded secret by open ai they didn't want to tell it they didn't even want to show people its thoughts they were just telling you it's thinking uh when when deep seek came out it was like r1 i'm talking about it was the complete opposite it was like if i were just going to show you the exact train of thought and show you everything oh and you can download it and run it for free how fun that's that's the statement that i think throws so many people off i what i will say is that uh what i remember is it was january 20th 2025 uh you know um i i have some some investments nothing major little little stuff mostly but uh but i i do own a little nvidia just for full disclosure here to everybody and I woke up I woke up that morning and I saw it and my heart dropped into my stomach and I could not, so I immediately got online, started figuring out what was going on and it was somewhere overnight everyone really freaked out about this it was like, I joke, that's kind of the day Wall Street noticed Yeah, definitely, I mean I had people who don't really follow AI talking to me about like, have you heard about deep seeing?

4:41it's like have i heard about deep seek like yes talking about it for months yeah but yeah there was a there was a just to clarify to people who missed it on january 20th there was a big market crash kind of in reaction to deep seek um kind of catching on and i think there was some big stories in cnbc bloomberg about it and and that was when like you said wall street kind of caught on with what was going on yeah it was like we we knew it was that's the thing is like we already knew it existed but but that was the day that wall street had their oh my god ai moment and uh it was uh it was quite a day as i recall yeah i think nvidia crashed 17 percent in one day they lost about 600 billion dollars um biggest single day drop in u.s history i believe oh and also jensen long great guy but personally lost at least 21 billion that day uh so uh it's hard to even fathom yeah man what was it about a about a trillion dollars wiped out or something like that it was it was a lot and it was it was fortunately it was short-lived though yeah that's right it kind of recovered right away everybody had really good talking points i don't remember exactly what they were but essentially they were like well the whole reason everybody freaked out which we'll get into in a minute is because uh deep seek's method of running the ai they can run it a lot more efficiently so it's a lot more efficient when it runs it doesn't require as much power uh doesn't require as many chips allegedly so or or the the rumor was that it didn't require as many trips to train and i think that's where there was a little bit of confusion um yeah i think we even ran a headline that said like you know there's a rumor that it only cost them five million to train this thing that's actually not true research model is is pennies in the bucket yeah and that actually wasn't true that was it was the last version of it that was the version that we saw that was the yeah that was the one that that cost five million and then they still had all these chips that they you know whether they got them legally whether they smuggled them you know we've heard reports that some people are like smuggling chips into other countries with chips and briefcases, like crazy stuff to try and get around export laws.

7:04But yeah, the whole thing was that it can do it more efficiently and then that would crush NVIDIA's market cap and Microsoft and all these other companies. And they're like, hey, more efficient AI, that just means more AI. And the market recovered pretty quickly. Yeah, essentially it killed the moat. The moat you always hear about, all of this about the expense of uh of training the the time it takes it kind of reshaped how the market was viewed for an awful lot of people in first for a short bit i would say it didn't it didn't stay crazy but it was uh it's pretty wild experience to watch uh yeah uh yeah and basically this all happened because deep seek is this tiny it's i mean technically it's not even a an AI company in the traditional sense.

7:55It was like a, I think it was like, I don't know the exact term, but it was an investing company, essentially. Yeah, they were a hedge fund. That's right. I wanted to say hedge fund, but I wasn't sure, yeah. Yeah, yeah, started by a hedge fund. And it's kind of crazy how that all came together. So I've got a question. So I'm going to get to the meat of the discussion here real quick. Let's do it. So these models are free. They're open source. can i just run deep seek on my laptop so technically yes but also technically no and let me explain both sides of that so let's just uh get a couple definitions out of the way so open source right that means in software that means that the code is freely available anyone can spin it up and run it themselves yeah um in ai it kind of modify yes yes that's key um you can fork it is what it's called where you take the code and then you edit it um uh and you know obviously open source and ai is a little bit different they give away what's called the weights and so the weights uh as you know cory it's it's essentially the the what's the best way to put it the setting yeah it's the settings it determines how the ai responds it determines based on this input what is the output that it creates it's not so one-to-one you know because probabilistic it's not deterministic which means like it's it's kind of predicting uh instead of having a uh exact answer every time uh but but they release the weights and usually they don't release the actual underlying code and how how they trained it or anything like that yeah and Essentially, those weights are how they dial up or down expected behaviors.

9:44Like if a model is behaving in a way that for one reason or another is undesirable, maybe it's out of alignment, maybe it's something else, you know, they can adjust the temperature. They can adjust a variety of levels on the back end to, I'd say, dial it in, basically, is what they're doing. Just like, almost like you would a guitar amp. Yeah. Yeah, definitely. and uh because these these machines work like uh human brains do uh they have like a lot of neurons which they call parameters and this thing has like 670 billion parameters um and it requires a ton like 1.5 terabytes of this thing called vram which is virtual ram it's like memory and it's a beast right if you were going to run this whole thing you would need a ton of nvidia chips you would need like probably thousands hundreds of thousands of dollars in in chips and computing software so i can just run to best buy is what you're saying right you would have to run to like 10 best buys and buy everything like i don't think like this is not if you were going to run the full version of this thing uh you're not you're not going to be able to to do it on your home machine so it's kind of like trying to fit a data center into your macbook exactly um and interestingly So even if you were trying to self-host this on AWS, that could cost you$70 ,000 to$118 ,000 per month.

11:09So it gets up there because these things are huge. In case anyone doesn't know, AWS is where it's a hosting platform that hosts a lot of apps that you use every day and the websites you visit and the software you use. so much of it lives on AWS or Azure or other options that are out there as well.

11:34So I feel like this makes the argument that this is like free and open source, which sounds all fun-loving and hippy-dippy. But the truth is, this isn't a thing. You're just going to go download and run on your laptop and have this super secret AI that no one can tap into, is what we're saying, right? uh yeah that's that's unlikely for us people with with deep seek now uh i'll give a caveat there which is um so that's the big version of deep seek right that's like the research model yeah v3 so there's been three versions of deep seek that's come out before that's why you and i had heard about it before it crashed the market uh and then there's r1 which is the thinking version which is kind of uh uh some like an iteration of of v3 essentially like kind of the one that was kind of released around the same time as a one exactly yeah and that's the one that caused all the panic because it was like thinking and it was kind of comparable to a one in terms of its quality now it wasn't actually at a one's level but it was enough that everybody got scared it was close enough that people got real uncomfortable yes exactly exactly and so uh what what you can do with this big v3 version is you can do this thing called distillation which is you make a smaller version of it with less parameters um that you can run on a slower machine so uh you know there's distilled these distilled versions of deep seek that are more reasonable but you would still need a high-end nvidia chip which is called a gpu to run or or uh you know from another brand but as a graphics processing unit uh in order to run it on your computer so that would be like a like what like rtx 4090 or a100 right so those and this isn't gonna feel like you're using a one at that level with a distilled model i'm assuming you're not going to get that level of quality that level necessarily of thought you're not going to get uh the level of everything that a truly built out full model gives you so i mean i'm assuming we're talking about a very pared down version all right so this is this is what deep seek looks like right here this is its web portal so technically if you didn't want to run it on your computer you could use it on like a little web interface like chat gpt um but if you wanted to run it open source you would go to a website like this you would go to hunging face and this here is this is the latest version of deep seek r1 and as you can see here it shows you like how many downloads that it's had it gives you all the details and the performance on it and then if you go to let's go to deep see if you're not a software engineer that's going to be a little intimidating i would say this is very intimidating yeah for some people even if you are yeah yeah yeah but as you can see here here's the distilled versions right so this is a distilled version of other open source models a little too technical to get into but this is 70 billion parameters versus the 600 billion one and i'll give you one more visual this is artificial analysis this benchmarks all the different ai models that are out there and it's one of the sources i like the best in order to compare and i'm going to show you this visual here so this is all of the top ai models by intelligence and it shows you which ones are open weight so which ones are open source like deep seek and which ones are proprietary like chat gbt okay so the blue is open weight yeah exactly so this is deep seek r1 so this is the the newest one that just came out and as you can see the only models that are technically smarter than it uh are all of open ai's top reasoning models thinking models o3 o4 mini and o3 pro and the new Gemini Pro.

15:31And Gemini Pro, yeah. Other than that, this is one of the best ones you can get, which is pretty crazy. Yeah, I mean, that's very comparable. And the idea that you could theoretically run that on your own if you wanted to, if you were in a business situation and had a need and desire to operate your own reasoning model that you didn't want out there in the wild, I would say. It is an option. It's just not cheap and easy. Yeah, because this quality, this 68 version, I say 68 because there's a score that goes from 1 to 100, and the top models are all around 70, so Deep Seek is 68. This version is not what you'll get.

16:23This is not the distilled version, right? The distilled version is not at 68 intelligence. It would probably be a lot lower on this list. It doesn't even show up here, I would imagine. I'm going to say, if it shows up on that list. I'm not even sure it would show up on that list. Yeah, so if you're going to run a smaller version of this on your computer, don't expect it to be like ChatGPT on your computer. Let's put it that way. Now, there are tiny models, though, not necessarily from DeepSeq, where I would say that is possible. Like there are some smaller parameter models from Meta in particular that you might be able to get some performance out of.

16:59But they're not going, still there's a significant difference when you're looking at, you know, these frontier edge research models. What they're doing is light years beyond what a lot of smaller tiny models are doing. But we'll see if that holds true. Drew, there was a little thing yesterday that Sam Altman posted on Twitter that I think is really interesting here about how long do you think the 4 and 0.3 Pro mini model could be ready to deploy on your phone? On your phone, right? Yeah, and he asked 25, 26, 27, or 28. And most people, presumably because he asked the question, click 25. Wow. And I thought that was really interesting.

17:44And, you know, he implied and elaborated on it a little bit that, you know, we do know they've been working on an open source model. We do know that it was delayed because they did something and decided that with just a little more time, they could make it a lot cooler. But we'll see what that is later. Yeah. And just to clarify, Google and Microsoft both have smaller models that you allegedly can run on your phone. So people are making them. They're just not near the level of the frontier and what you could get if you run it on the cloud. Because when you run ChatGPT on the cloud, if people don't know, basically they have all these data centers that are processing it and processing everybody's requests all at once.

18:28And they can do a ton more because they just have so many NVIDIA GPS, basically. Yeah. While we've established that we're not going to be running a big research model on our laptops, at least any time in the foreseeable future, you never know what might happen. So I'll leave that caveat there. Right. You mentioned a little while ago they have a web interface. I understand why a business would have a use for a hosted model and would want the privacy, security that it provides. But as an individual, is there any reason I wouldn't just go use their web interface that's the same models? Yeah, that's a good question.

19:08And they also have an app, too. So, like, Chattop. Oh, that's great. They have an app. I forgot about that. Yeah. So, yeah, I mean, a lot of people download the app. A lot of people use it. It's very popular. But DeepSeek's app and web service have serious security flaws that I think people should know. For instance, it has very weak encryption. It has hard-coded keys and unencrypted data transmission to China because the servers are based in China. So that means all of their databases are potentially exposed. So that means, okay, okay, I get that. So that means it's being your data is all saved on Chinese servers.

19:47So as a result, it's subject to Chinese law. And if the government requests it, companies have to hand that over. With that in mind, can you talk a little, we mentioned military allegations earlier. Can you elaborate on that a little bit? Because I feel like that's where we're kind of headed. Yeah. So this is news that broke about a week ago, if you're watching this now. But U.S. officials say that they have actively figured out that DeepSeek is helping Chinese military and intelligence operations, and that they're actually feeding user data to Chinese military operations. They found it collecting keystroke patterns and device data and sending that directly to Chinese intelligence for their operations.

20:35So it's spyware basically masquerading as a chatbot. Yeah. Okay. That's a fair concern. I know there are some countries, I believe, that have absolutely outlawed it, aren't there? I think it was Italy, maybe Taiwan. Mm-hmm. Mm-hmm. Okay. Yeah, and U.S. agencies too, right? Oh, NASA. Yeah, NASA. NASA had banned employee access or something. Okay. Yeah. Interesting. Interesting. There's also the security testing. Do you want to talk about that? I'm very curious. Yeah. Yeah, that's where it gets really bad. I mean, DeepSeek failed 100 % of its security tests. It could not block a single harmful prompt.

21:25So, you know, last week we talked about safety testing and how all of the AI, you know, big AI companies do that and try to prevent it from doing bad things when people prompt it. Yeah. DeepSea couldn't block a single one. Also, minor fail alert. We absolutely should have discussed this in that episode. Yeah. And I guess it's important to compare that to ChatGPT. And I want to say ChatGPT is somewhere in the mid to upper 80s at blocking harmful prompts. yeah yeah i think it's like 86 or something like that depends on the model i'm sure yeah i'm sure i'm sure it does and i imagine that uh gemini and anthropic are in that same ballpark and uh yeah but it seems to only be here that the safety discussions are ongoing yeah um yeah and as a result uh i think the data that i've seen is that deep seek is like 11 times more likely to be exploited by cyber criminals for that reason wow which makes sense yeah that That does make sense.

22:27And honestly, there's a, not necessarily to get too deep into today, but I think there's an interesting discussion around there and around the risk of open source. You know, we always talk about the need to, you know, be sure that we're not over centralizing the power of AI and do just a handful of companies as often happens with new technology. and while I support that there is an element of open source that while I generally support it does give me chills down my spine yeah definitely I can't remember if we said this last week or not but you've definitely said this to me personally that there's legitimate concerns on both sides of that argument yeah very much so I think that it is in fact true that We don't necessarily want all of the power centralized in the hands of a handful of powerful people.

23:27It is simultaneously true that putting out an open weights model puts it into the hands of undesirable bad actors. And I think Jeffrey Hinton, who won the Nobel Prize, coincidentally, and is actually rather terrified of open source AI. There was a good interview on this a couple of weeks ago that I think we've mentioned in the newsletter. Um, it's, it's really interesting because it creates a situation where two situations are both true and both significant and where, where's the line? Um, yeah. Yeah. Uh, he, he, uh, specifically the quote that you're talking about is he equated giving out the open weights to essentially giving out like how to make nuclear fission to people.

24:17He's just giving and selling it. I think he said it like selling out of Walmart or something. Yeah. Yeah. And, you know, it's not wrong. You know, I mean, in terms of like, it's a lot of money and investment for an individual. But as far as a, you know, let's say a network of bad individuals or, you know, a country with bad intentions, a government, any government, that money is a drop in the bucket when you consider that. Yeah, well, and I also want to provide a counterpoint to the deep seek military allegations, right? Because, you know, if the wherever the information is centralized, that becomes a point of concern.

24:59So, you know, take open AI, for example. People are concerned because I believe there was a guy who's either an active government official or, you know, has ties to the US military on their board now. And people are assuming that, you know, whatever chats you send to chat GPT, now the government can potentially access it. We wrote a whole story about that, if you remember, about how because of the New York Times lawsuit against OpenAI, they're not allowed to delete any chats, even the ones that are supposedly deleted and temporary. Because the government wants the right to subpoena them. Yeah, exactly.

25:37And so just because, sure, maybe the Chinese government is getting your DeepSeq prompts, it doesn't mean that the U.S. government won't be able to eventually get your JGBT prompts. You know, a thing that blew my mind about that was that we had a company that was intentionally not saving user data beyond 30 days. And the government's like, whoa, man, we might want that data. uh yeah that that was a little uh bothersome to me i think yeah i think a lot of people felt that way now obviously like this was the the ruling that this happened it was very specific to the case and open ai was like we're fighting this and and the judge said whoa i'm not trying to enable mass surveillance so everyone needs to calm down um it got out of hand real fast yeah Yeah.

26:34So there's a lot of nuance with that conversation. And crucially, OpenAI had a good blog where they wrote about what data is and isn't to be held at this point. So everyone should look that up if you haven't looked it up yet. Agreed. We can put a link to that below the video. Yeah, I'll include that. But it just goes to show, you know, any any cloud based, you know, where you're sending prompts over the wire does have security risks involved. I mean, just like anything that you post online has a security risk. Yeah. So with all of that in mind. should people just avoid deep seek entirely i mean you could uh but not necessarily you know it depends on how you use it like um i mean i think a safe thing to to bet is unless you're cool with uh china's government potentially seeing your prompts like never use deep seeks direct app or website i think that's a safe safe way to go about it i think that makes sense you You know, but that doesn't necessarily apply to the open source models themselves.

27:39Right. Right. Yeah. Okay. So those are actually safer since you can either self-host them locally, like the really small distilled versions like we were talking about earlier. You can use a tool like Olama to do that. Yeah. Or there's another tool called LM Studio that I find is a lot more user-friendly. Okay. And you can use that to run local models. We can also stick links to both of those down below as well. Yeah. I guess you could also use Amazon Bedrock, Azure. You could use AWS. Base-end, like we talked about. Yeah, there's a variety of directions you could go there. Actually, artificial analysis has a huge list of all of the providers and how much it costs to run it on there.

28:21Those come with risks, too, because you're sending it over the wire. Yeah. But if the key is keeping your data away from Chinese servers, if you care about that, especially if you do sensitive business data or work in government. I think Chat2BD has a government version that you can use. So just be careful and aware, I would say. Makes sense. I agree. I think the way to go is using a US-based provider probably. And the thing is, what makes it attractive is that it's comparable to an O1 type model. It doesn't mean it's better. The truth is you could just go use an American model like you already are, and you're not going to feel much of a difference is the truth of the matter.

29:07Yeah. Yeah, that's true. But sensitive work, I would probably self-host. There's a lot of other models other than DeepSeek that we can get into. Klen is another really popular model from Alibaba in China. Oh, yeah. But that one, people, you know, that's one you can run. there's a lot of versions that you can run locally especially if you have a good a good graphics card um and you know same it one thing we haven't touched on with any of this is that uh the the models do have certain restrictions because they're based in china so you know you can't criticize the government for you can't get the ai to criticize the government for example um things like that we don't have that problem with ours in the u.s though no no yeah grok even Grok is owned by Elon Musk.

29:54It'll criticize Trump and Elon all the time. It does. It does. It's an interesting thing. That's a whole different discussion there. But there is an interesting discussion in there sometime in the future, I think. To cap this off today, let's do our round robin. I've got a question I want to bounce off you. Let's do it. Plain and simple question. Would you personally, Grant Harvey, use DeepSeek for anything of value that you do in your day-to-day job or life? Do you? This is a good question. So I don't remember if I said this earlier, but I have used DeepSeek before. But I used it mostly in a testing environment when it first came out, and I would use it to compare its answers versus ChatGPTs.

30:46gbts as far as for anything serious at this point the model the cloud model i trust the most is clod and the reason for that is because um you know they're they have very specific roles and and and um policies that they share where they you know de-anonymize all prompts you know that's not tied to you after a certain amount and once you delete it i think they delete I think they do delete it within seven days. There's a pretty standard clause between most of them that most of the U.S. companies that they only hold the data for so long basically to go back and find problems and errors and figure out what caused issues, things like that.

31:31You know, I don't know specifically what's kept, but that's cool. For me, I would say I have used it a couple of times, just more trying it out than anything, but I haven't really given it a deep, hardcore test. But I would say you should never, ever use it on a work computer. That's fair. I think it's probably a really good rule. I would also say that I just don't have a use for it. The truth is, until very recently, ChatGPT 4.0 was my model. And it still is for some more basic tasks here and there. But with the cost drop of O3, I'm pretty much just using O3 exclusively throughout the week now.

32:20Okay. Yeah. Even for tasks that don't necessarily need the high power thinking ability? It depends. you know sometimes it happens in a way where you're in the middle of something and you already have a chat going with 03 where you're working and you have small follow up questions that absolutely don't need to be 03 but I also don't want to deal with the headache of going and opening a 4.0 chat giving it the context of what we're talking about for what's a simple request so as long as I'm staying within the limits I usually just keep going I mean if it's If it's a little later in the week and I feel like I've been leaning pretty heavy, I might not.

33:02But as a general rule, I'm pretty much just using O3 for everything now. And the truth is getting the best results I've ever gotten from anything. Wow, that's awesome. That's actually, that's really good to know. Because I don't know if I said this in the last episode, but these days I actually switch between all three of the major models, Clawed, Gemini, and OpenAI a lot. Yeah. And I will compare the results that I get just because I don't want to be I want to make sure that I'm getting the best option or a lot of different options so that I can choose what I like the best. You know, especially if I'm using it for, let's say, like a writing task.

33:43You know, I'm using different ones for research and then I bring them in and I create different outlines and then I can choose which one I like the most and then use that as my jumping off point to edit later. Yeah. Yeah, I think that's, I'd like to get a wide variety of opinions. I like that. And I like the idea of cross-checking them against one another. I like, what I would say is, I guess I was speaking more in terms of what I use during my work day constantly. Yeah. In my personal life, I do tap into other tools. I mean, like I use perplexity daily. I don't use it as much at work. But I mean, I use it for listening to the news while I'm going to get a cup of coffee.

34:23I'll have it because I love that feature where it'll read it to me. I've actually never done that. It's the best. It doesn't work on queries, which, which. Interesting. Aravind Srinivas, please add the read button to the queries too. It's great on Discover articles, but it would be a whole lot greater if I could do that with every entry. I use, I use OpenAI's advanced voice mode every day. I have, I have a variety of uses for that, mostly in my spare time. I use an unknown meta model in my glasses that I don't have on right now. And I use Claude 3.5 to smack talk my buddies in our group chat. Why Claude for the smack talk?

35:08I don't know. It does a good job with, while I would absolutely choose 03 or honestly 4.5. Honestly, I'd choose 4.5 over any riding model. I've finally gotten to where I've spent some time with it, and it's impressive as a writer. But the reason being because I feel like Claude tells a creative story well. If what you're doing is creative writing, I just kind of like Claude 3.5 for that. I still do. And I say that knowing that there are better models. There are better Claude models. but there's something that 3.5 had that I still don't quite feel in the newer models and uh and I I can't quantify it put a name on it uh but there's just something about it that I liked and uh that's what I go back and forth to use like when when when it's late at night and a buddy texts you some nonsense a lot of times instead of you know garnering it with a worthy reply I'll take that chat I'll go throw it into 3.5 and tell it to give me a obnoxious, ridiculous response.

36:17And then I'll go to copy that into my text chat. And, uh, and I've done that. Do you disclose this to your friends? Like, do they know that you're using AI? Figure it out after a while. Uh, we've had some great conversations. It borderline landed into a, a wild role-playing session once. It was, it was the funniest thing. That's funny. Uh, but it was, uh, it's just, you know, it's, it's, it's easy. It's fun. It, uh, when I don't feel like I have the brainpower left, because it's the end of the day and dang it, I'm tired. Yeah, yeah, I get that. I totally get that. You know, this is a topic for another day, but we'll have to talk about the whole fair use thing that happened with Claude and how Claude Anthropics, the owner of Claude, they apparently, like, stole 7 million books, made an illegal database of them to train Claude, so...

37:05You're right, I can read books. You just can't go download them off Napster. Yeah, exactly. Exactly. So we'll have to talk about that. But maybe that's why it's so creative. I think a lot of creative writing. I think it's a good way. It's read a lot of pirated creative writing. That's a good call. Yeah. The last thing I'll say just on the topic of open source is that I would 100 % use a local open source model when one is good enough that you can use it reliably on your computer. Like I think like sending stuff over the cloud all the time, that's not ideal, right? You need something that you can use tangible on your computer, maybe even built into your computer that can be your personal assistant where you are.