Speaker 1: Alright, welcome back. I am going to be talking about DeepSeek v3 today. And I want to answer the question, is this my new go-to LLM? Well, let me just cut to the chase. I spent about 30 hours coding with DeepSeek v3 as my primary LLM, and it is probably the best AI coding assistant I've ever used. Over three days, I put it through its paces, cleaning up code, building APIs, and even playing around with a side project for a chess game where I was going to have LLMs play chess against each other. And I'm going to tell you, I'm really impressed, but I do want to break it down a little bit to give you a little bit more context. So first, what did I use it for? Probably 60% of my use was on the coding side. The type of technologies I use and languages, I use Python. Everything I do is typically serverless. I do have some EC2s that I work with, CDK deployment, UJS, PostgreSQL, TypeScript. Those are the primary ones. I do have some other things that I did send through it as well. I used it a lot for code cleanup. So I was taking a feature that someone on my team had built as kind of a prototype and turning it into production. And to do that, you end up with a lot of cruff in there. If anybody that's ever coded something that's been like a prototype, you know how much cruff is in there. And that is something that I used it for. I would give it a file and I'd be like, all right, it's doing all the functionality I want. I want to remove everything that's not needed. I did use it for some writing stuff, a little bit less on the writing side than I typically would, but typically things around marketing and e-commerce. And I also got some ideas for like wording and stuff around the website. I also used it quite a bit for architecture. So I was planning a bunch of new features and I came up with a few diagrams and things for the data models I was considering. And it actually had some really great ideas that it recommended some that I would say were even slightly better than what I had come up with. And I did tweak some of my models because of it. So I'll give you a few specific examples of things that actually happened. Code cleanup. I gave it a, it was just over a thousand lines of code. And a lot of it was like placeholder, you know, like when you're building a UI, a lot of times you'll do like placeholder format for data to inject into it. Wanted to remove all that bunch of functions that need to be removed. And within one attempt, it gave me back a file with about 415 lines. There was one mistake. And that mistake was it had taken away two tabs that really weren't functional, but I still wanted them there. So it correctly picked them up that they weren't being used. And I didn't tell it ahead of time not to remove it. So I don't know if I want to count that as a mistake or not. But then I said, Hey, put those tabs back. Next time it was perfect. Claude, I gave it the same exact prompt. It gave me a summary of the things to do, but didn't give me the code. I asked it to do the code. There were several errors in it that didn't work anymore. After four attempts, I still had another error, but it was close enough at that point where I was like, okay, definitely deep SQL on this one. From a consistency standpoint, this is just a general statement here. Claude seems to be highly trained with React. So anytime I do anything front-end related, it wants to work in React. And that makes sense because a lot of technology, a lot of people use React. It's probably the number one one, but if I didn't actually go looked it up, I don't know that for certain, but I use Vue.js in this particular project and it never once in the three days switched it to React. And I can say Claude does this to me consistently where I'll be the context of build up. It'll give me React code. And I'm like, okay, Claude, that's React. Give me it back to me in Vue. The new APIs. So basically I had three or four different APIs I needed to develop in a particular controller. I gave DeepSeq an example of how to do authentication, an example in that same controller of like the data fields, the data model and everything that's in there, but then even everything, but enough for them to go through and build what I needed. And first try covered everything, even got authentication. Claude pretty consistently would make up fields and they'd miss the authentication piece quite often. And then I did a side project. I didn't test this in Claude, but within five minutes, I had a working prototype of chess, a visual chess board with LLMs playing each other all running locally, which by the way, LLMs are really bad at playing chess. So I got to rethink that project a little bit because I don't think it's worth it. So a few things that I want to try more. Deep thinking. I did play with this a bit. I've never felt like the reasoning models were great at coding. For example, I don't really enjoy using O1 or O1 preview coding. And I tried this a little bit with deep thinking and I just didn't feel like it worked, you know, as well as just using the regular regular prompt. Search. I did actually use this quite a bit, but my number one AI search, it's usually perplexity. So I want to see over time if this actually replaces perplexity for me. I think it could, it might. But I don't know if it's like enormously better than perplexity. I don't know if like there's a clear standout in search right now. And then the actual API itself. So I didn't actually spend a whole lot of time working with the DeepSeek API. I did a little bit, but I'd like to spend more. I'd also like some of my clients that I run locally to incorporate it. So I can actually run it in my normal environment. So rather than have to go to the web browser. Now, context limits. I never hit a context limit with DeepSeek. I did hear this in Reddit from a lot of people complaining about context limits. So I do want to say that that is something that will happen, but I did not actually hit it, but I'm also pretty good at starting new chat. So I get to a certain point, I started a new chat and then I go to the next one. Now, Claude, I do want to say Claude, I very, very consistently hit some sort of limits. This, when I use Claude prior to these three days, three times a day, I'd hit, I'd hit a limit. I'd hit one, you know, right before lunch. And then I'd hit one in the afternoon. And then if I'm working late at night, you can see an example here. I hit one that basically locked me out to 1am, which really stinks when that happens, especially when you're in the middle of a project. And I know this might change with DeepSeek because the volumes probably nowhere near what Claude is putting off with it. Plus they don't even have the paid version in the web currently for DeepSeek. I know all that's going to change. I can't run all that for free forever, but I will say I did miss Claude projects. This is just an example. I use projects a ton. So I'll work in a project and maybe for a day, you can see some of these, I actually did a couple of the same projects in a day. So what I'll do is I'll make a bunch of changes. I get it to a good point. And then I go and I create a new project. And then apparently I created two V8s, which is not good. But anyway, I create a new project with the files that I want in it. And I also have this for like a particular parts of my application. So I will do like, I need to do some front end code. So I give it some front end details around like the feature that I'm working on. And I use projects consistently. I will say, I mean, it was fine just uploading or copying it into the context of DeepSeek, but I do enjoy the, just the flow of the way Claude works there. And I will say DeepSeek will remain my number one primary LLM probably for the foreseeable future. I will continue using like the Gemini Flash and the, and Claude as my secondary when DeepSeek kind of falls short. And then until I see like some sort of major like issue with limits or pricing or something, I don't see a reason not to use DeepSeek. The other thing I'd maybe add onto this a little bit is that I know DeepSeek is a Chinese company and there's a lot of concern around sending data to a Chinese company, but I would recommend never just putting anything in an LLM that's in the cloud ever that you don't care if the AI company has. And, you know, so don't put anything that's proprietary. Don't put any like API keys or anything. I don't have that much of a concern to this because I'm not going to put anything that's highly sensitive. I mean, to me, code is code, UI is UI. There's nothing like that they're going to get out of what I built that they're not going to already have. But with that said, there is that risk and that concern. So you need to be very careful about which, which ones you use. On the other side, I will say DeepSeek v3 is fully open source as well. So you can run that yourself, but you need one heck of a machine to actually pull that off. And I just don't have the capabilities to run that, nor do I believe the majority of people do. Anyway, I will stop there. What do you guys think? Like, how is your using of DeepSeek gone? Is there anything that you think it does worse? I only covered like a small subset of it and, you know, maybe over the next month, my opinion will change and I'll go back to being clawed as my primary. But I will, I am curious to see, has everyone else's experience been the same as mine? Anyway, let me know in the comments below. Until next time, peace out. Thank you guys for watching. Subscribe and like if you want to see more content like this.