We really need something like opencode to succeed!

0:00 All right, I'm going to run open code, but to do that, I want to actually do this with a local model. So, you can see my open code.json, which is in the root of this Flappy Bird folder that I've created. Now, if you don't know, Flappy Bird was this game that came out a long time ago. Very basic game, but it actually kind of went viral. Uh, so we're going to see if Devstral Small with Open Code can actually recreate it. My guess is it probably can't uh just based on my usage with Devstrol, but we'll see. Maybe it'll surprise us because it is a relatively good coding model. So what we're going to do is we're going to go in and open open code. You can see here that I've already have that configured to be devast small. But if it wasn't, I would go ahead and hit models. And you can see the list of the

0:45 models that I actually have already authenticated. I'm just going to exit out of that. I generated a prompt just asking OpenAI to give me a prompt for it. So we're going to go ahead and paste that in. I don't know how good this is going to work, but I do want to just kind of show you how open code works with local models in particular. So, it's it's loading actually really quickly here. Uh, so you can see here that, you know, first choke to token generated, it's streaming it out. It's working incredibly well. Let's take a look at the GPU usage here. GPU is hitting, you know, 87 86% currently. Uh, this GPU right here, my AMD one, which is built onto my processor, is actually from the recording software I have running because I'm not I don't have that running on my RTX 5090 currently. So, we're going to go ahead and let this

1:31 run. We're going to see what that can actually generate. And while that's happening, what I want to do is pull this over here to the side. And what we're going to do is we're going to talk a little bit about Open Code. I'll have this over here just so we can kind of monitor what's happening here. So, Open Code is really a cloud code competitor. And to me, it's very important that long-term we have an open- source solution that allows us to use local models. That's my dream. That's where I want us to go. But a lot of you do know that I'm a big fan of Cloud Code today. I also love the cloud models. I have actually used Open Code quite a bit with Cloud 4, Gemini 2.5 Pro, and Devstril. I actually spent an entire day using Devstral for life light tasks for me and

2:17 it worked really really well. So I can see over here on the right that it actually did finish creating the Flappy Bird clone, but I'm guessing the prompt wasn't clear enough about actually creating the file. So I'm go ahead and ask it to do that. Again, I just copied this from OpenAI. I didn't spend a lot of time making it. So yeah, now it's going to make that for me. So we'll let that happen. So tips and tricks. So, one of the first things that actually threw me off with open code is adding files to the context. And if you're familiar with claw code, you know, you can do the at symbol and then start typing in a file. Gemini CLI, same way. You can do the at symbol. Very used to that. And I still find myself doing this constantly with open code. The only way I figured out I could do this with open code was to not have it

3:04 be at the beginning of the line, the sentence I was typing, but say something like use the following forward slash, then start typing support or whatever the name is of the file, then I get the autocomplete to come up. I hope they do actually add in the at over time because I think that would make a huge huge difference. Claude 4 is amazing with open code. It scores very well on my eval. It's actually really enjoyable to work with and it typically is pretty efficient costwise. I think I burned a few dollars, you know, using it off and on for an entire day. I think it maybe $67 is what I used. Again, that adds up when you do it for an entire month, which we'll talk about pricing here in a minute, but it's still, you know, very efficient compared to Gemini 2.5 Pro.

3:51 Make sure that you're using SST Open Code, which is this particular GitHub repository. Do not use the Open Code AI one. The SST Open Code one has more stars. It is more recently updated. Every time I go in here, I typically see at least some commits that have happened, you know, within the day, if not yesterday. So, it's just really, really important to make sure you grab the right repository. If you're using Gemini 2.5 Pro, please watch for infinite loops. I had a few that I killed for tests that should have been like a $120.50 with Claude 4 was $8.49 because I just let it keep running. And I'll show you kind of what that looks like over here to the left. Let me see if I can find um find that. Yeah, over here. So, you can

4:37 see here it'll just end up hitting this error, old string not found, and then it'll keep looping over and over and over again. Um so, do not use Gemini 2.5 Pro with that. even tried tuning temperature and stuff to the best of my ability and I could not even square that in like I could with root code. Uh 0.7 is typically what I find to be the best there. Now this also is kind of embarrassing but if you're wanting to authenticate another provider I kept wanting to do /models and do it this way. But the right way to do it is actually to launch Open Code O login. Select your provider put in the credential. Boom. You're good to go. I had to actually search the source code to figure this out. I know this is probably documented somewhere. I did not find it until I actually went into the

5:25 actual source code for the uh project itself and then I used augment code to actually ask questions about the repository and that's how I ended up getting that answer. The opencode.json.config very important. Now, this is what's interesting to me about this is I've had some people report that it doesn't work for them. But I've done this in like four different projects now. Five, I guess, if you count this little Flappy Bird clone. Um, but yeah, I just dropped the Open Code config, open code.json config in the root of my folder, put in the the stuff that I want in there, and it's always worked for me. I don't know why it wouldn't work for some people, but I am calling that out that maybe there's some weird things with looking finding this JSON file. Another thing that I constantly do mess up with this

6:11 is I uh run the clear command. That does not actually exist. If I type clear here, does not actually exist in open code. I think it probably will one day in the future, but instead you need to do new session. I don't think it would be that hard to add clear, but maybe you don't even need it. You just got to get used to it. So anyway, let's jump over and see what this Flappy Bird actually ends up looking like. Taking a look at the results here. You can see we have a kind of a Flappy Bird clone, but there's so much to the physics to it. I don't think I could ever get through this first pipe. Oh, I almost did. Oh, can I do it? Anyway, this is actually surprisingly good, right? This happened with a local model. It's nothing special with open code and it followed tool

6:57 calls properly. It didn't have any errors in any of the output on it. In fact, let me just go ahead and show you what the results look like here. You can see here it just worked great. Like it created everything and I probably could change some of the gravity. I bet you there's settings in here I could configure. Um it works awesome. Devstrol has been kind of a joy to work with with Open Code. Again, the context is kind of bored down here because LM Studio when you have streaming doesn't actually return that to you. So that's unfortunate. All right. And to close this out, the one thing I want to talk about is we need an open source solution. We need local models to actually work. And the reason for that is these companies, a lot of them still are not profitable. I probably talk about this a little bit too much. Age of

7:42 free AI is coming to an end. That coding in the future, it's going to be very expensive. It already is pretty expensive, but we're very fortunate to have things like Augment Code with the fixed price of $50 a month. Claude Max with a $100 and $200 a month plans. And then we have Cursor, which has gone through a quite controversial move over the last few days, which I'll talk about here in a second. A lot of these companies are aiming to like get as much market share as they can, but they're not profitable. Here's an example of someone named Jared Sumner who actually shared their usage of Claw Code Max $2,152 at 70% margins, which I'm skeptical that's even what Anthropic is able to pull off with their uh cloud models.

8:29 They are massively costing claw money. This is money that could be going to API usage that they're paying $200 for onetenth of the price. So, I'm a little bit nervous over time that either the usage we get, which is very nebulous, right? We don't really know how much usage we actually get from these. We just kind of get put on timeout and then we get back some usage, which doesn't bother me too much. I actually prefer that than the monthly limits. I actually don't mind being put on timeout for a few hours and then being able to come back. I think that is the right way. I think more companies should do that. But I think the even better way was just be as transparent as you can about what the

9:15 usages actually are. And right now the usage transparency for claw code is like kind of not there. We just know we get more with claw max. We get 5x more at $100 and 20x more with the $200 plan. But cursor cursor is one that I'm still a little bit bitter at even at months later. I paid for a year of cursor and they have changed so much in the last year. It's actually been very frustrating. I've actually put out a few videos kind of negative on cursor and you get mixed thoughts on it. You know, some people are totally fine with the way they're treating everyone, but to me it's about being anti-consumer. Right now, I think Cloud Code or Claude Max is very proconsumer. I think they

10:03 are very very nice with the limits they give us. They only put us on timeout for a few hours at a time. You know, maybe maybe six, seven hours at most I think I've ever been put on timeout for. Where cursor on the other hand actually sells you one thing. This is the way I felt. Sold me one thing, gave me something else. And then now they've completely changed their pricing. So let me show you the new pricing terms based on uh this this summary here. Unlimited usage when using auto mode. Why the heck would anyone want to use auto mode when it can choose randomly claw GPT4.1 Gemini XAI? That seems awful. So you may

10:48 use auto mode one day and it's amazing and then a few hours later it's awful because you have no control over where they are routing the actual request to and it's going to be in their best interest to do it to the cheapest one. So it's going to be going to GPT4.1. It's 100% going to go be going to the ones that are cheaper. Then you get a $20 month credit for Frontier models. So if you want to pick your models, you pick Clawson Sonet 4. you get 225 requests, which I'm a bit skeptical on. What they're telling me is that each request is about what, 10 cents there. Is that right? Uh I think you're probably getting maybe half that realistically if you're putting any sort of context in there. Then you get 550

11:34 Gemini requests, which I also do not believe. Uh cuz Gemini 2.5 Pro in particular ends up more expensive. I wonder if they're talking about flash there. Anyway, I think this is a bit rough in my mind. I see why they're doing it. I understand why they're doing it. This is not what I paid a year for. I did not pay a year of cursor upfront for this type of billing. It's very unfortunate. And going back to Claw Code, they are amazing right now, but that could change. And if they do, I'll be one of the first ones to talk about it. So yes, I hope you all join me in kind of rooting for these open-source projects like Open Code, Root Code, Klein, and all the others that are out there. We need one of them to actually

12:21 succeed extremely well. And we need one of them to start using local models a lot. Like ideally, we can actually do some really good things with local models. But I think that's going to wrap it up for now. If you have any thoughts on anything that I've talked about today, please let me know. If you've used Open Code, I would love to hear your thoughts on it as well. Anyway, until next time everyone.