Meta Muse Code & Muse Spark Course – Build AI Agents, APIs, and Full-Stack Apps

[00:00] Hey, this is Andrew Brown, your favorite [00:02] tech instructor, bringing you another [00:04] technical course and this time it's the [00:06] MetaMuse Code Essentials. So, we are [00:09] going to learn how to work with the new [00:12] Meta models, Muse Spark, as well as [00:14] their new coding harness uh called Metam [00:17] Muse Code. And the reason I think that [00:19] we should look at this is that when you [00:21] have a first class company that has [00:23] their own models that they've trained [00:25] and they have a coding harness and and [00:26] they're entirely vertically integrated [00:28] that they are a possible option that uh [00:31] we could utilize in the EI space. Uh and [00:33] so you've already seen me make multiple [00:35] courses like this. Uh and I'm going to [00:37] continue doing that as more first class [00:40] companies with their own models are [00:41] vertically integrated. Um but you know [00:43] the routine so let's just jump into it. [00:45] Okay. [00:49] Hey folks, it's Andrew. Before we jump [00:51] into working with Muse Models, let's go [00:53] ahead and learn a bit about uh their [00:57] offerings. Okay, so uh the first thing [00:59] is that we want to cover what kind of [01:00] models are available with meta. So one [01:03] that's very well known is llama, but [01:05] because our focus of this course is to [01:07] work with uh you know spark or glimmer, [01:10] we're going to talk about those ones. So [01:11] musep spark, there's basically only one [01:13] version. It's version one. uh there is [01:15] an earlier checkpoint but there's no [01:16] reason to use that earlier checkpoint if [01:18] you can use 1.2 or whichever is the [01:21] latest checkpoint there uh unless you [01:23] have built something already on there [01:25] and you are locked in with its specific [01:28] intelligence. Um what's interesting is [01:30] they have this contributor offering [01:33] where they're going to give you deep [01:34] discounts if you share your data. Um for [01:36] me I like to build a lot of Japanese [01:38] language learning apps. I don't care if [01:39] they have my data. So for for me that [01:41] could be a really good offering based on [01:43] its use case. Uh it works with multiple [01:45] modalities, image, video, I believe [01:47] audio as well, uh obviously text, and it [01:50] will output it. Um and so it truly can [01:53] do uh a significant amount of things. Um [01:56] and surprisingly, it does really well on [01:58] some of those tasks like uh one that's [02:00] really fun is like trying to build a [02:01] website. It does a really really good [02:03] job uh for what it is. Um and so I do we [02:08] do look at that in the course where I [02:09] make a a fun social tech website um [02:12] using Musepark. Um, so the other one is [02:15] Muse Glimmer. Now, I don't end up using [02:16] it in this course uh just because I [02:18] don't have the local compute and I [02:20] probably could have figured a way to [02:21] serve this. I just decided to stick with [02:23] um the managed uh one here, but there's [02:26] a lot of things going on here. So, it's [02:28] a 30 billion parameter model. Uh it can [02:30] run on a single GPU, which is not the [02:32] one that I have. So, maybe if you have [02:33] an RTX 5090, 32 GB, you could do it. I [02:36] don't have that. Um, so, uh, you know, I [02:39] don't have 5K to to spend on that, but [02:41] if I did, that'd be really really [02:43] awesome to run Glimmer uh, here. Uh, [02:46] it's a open model, so open weights. [02:48] Okay. Um, so it's kind of a mix of open [02:51] sour, it's not open source, but it's [02:52] like it's basically, I believe, what [02:54] Llama has as their model. So something [02:57] very similar to that. And it's trained [02:58] on data from more than a hundred [03:00] languages. Again, that matters to me [03:01] because you know me, I like to learn [03:03] Japanese, but um having a lot of [03:05] languages there is really good. What's [03:07] really interesting is that when I looked [03:08] into benchmarks, and this was not [03:10] because I was making this meta course, [03:12] but I was again trying to build my [03:14] Japanese language um app, I have like [03:16] four of them, and I was trying to uh [03:19] train it uh to detect grammar points. [03:22] And what was interesting is that uh meta [03:24] metas-park was really high up there in [03:27] the intelligence level. And I was like, [03:28] oh, cool. So I could actually add it to [03:31] the mix and I'm currently running an [03:32] experiment on that. So uh it's really [03:34] interesting where this model sits on [03:37] benchmarks. Um in terms of pricing and [03:40] rate limits, Meta has a token usage [03:42] pricing model. They do not have a [03:44] subscriptionbased pricing model. So this [03:46] is something um you can build with like [03:50] in this course I'm running off of tokens [03:53] and I'm running on X high for a lot of [03:56] things just for fun. And you don't run [03:58] have to run an X high, but like I could [04:00] barely get through $100, you know, it [04:04] was I think I spent like $50 out of out [04:06] of my huge budget. And so, you know, I [04:08] would say that if you wanted to uh run [04:11] this thing as your primary thing and and [04:13] it was like you want to spend $20 or $30 [04:16] a month, uh you know, if you are uh very [04:20] conservative in the way that you work, [04:21] and I I generally am, uh I feel that [04:24] it'd be very very easy to uh stay under [04:27] a $50 uh mark or if you wanted to uh you [04:30] know, replace your Claude Max at $100 uh [04:34] with this as a developer, you absolutely [04:35] could. Um, but you know, you just have [04:37] to get used to that um that thing. I [04:40] think people are kind of getting used to [04:41] that. So, I'm not sure how long [04:43] subscriptionbased pricing is going to [04:44] last to be honest. Uh, as I mean in the [04:48] future I could see other models just [04:49] deciding that they're going to say, [04:51] "Hey, you have a $100 cap. When you run [04:53] out, it'll be more transparent." So, um, [04:55] there's that there. Okay. And so, uh, [04:59] these pricing models are available [05:00] obviously for the Spark models. not [05:02] going to be for Glimmer, not going to be [05:03] for Llama as those are ones that you run [05:05] locally. Um, and I don't know if they [05:08] actually serve uh Glimmer or Llama up [05:10] anywhere besides, you know, with other [05:13] providers. I don't believe Meta has [05:14] their own specifically to serve uh [05:17] Glimmer. Uh, but that's a non-issue as [05:20] there's a lot of providers that can do [05:21] it for you. Um, in terms of pricing, we [05:24] are seeing uh pricing for input at uh [05:28] 1.25 25 USD for and then uh10 cents per [05:32] million. So it's very uh inexpensive. I [05:34] think that their input I think it's [05:37] their inputs a little bit more than [05:38] other providers but their output is is [05:40] pretty cost effective here. Uh for rate [05:42] limits we have 3,000 for standard [05:44] contributor is much much lower. So [05:47] that'll be some consideration that [05:49] you'll have to have. [05:51] Um and so you know just consider that. [05:53] But yeah just look at the price [05:55] difference between contributor and [05:56] standard. It's insane uh the jump in [05:59] difference. So you know if you can [06:02] handle this uh lower requests per [06:04] minutes and you have that then that will [06:07] work really really well. The one [06:08] question that I would probably have is [06:09] like you know what does it look for [06:11] batch uh batching right or do they even [06:12] have batch capabilities? Give me one [06:14] moment. I don't think I checked that. [06:16] Okay so currently they do not have batch [06:18] pricing. So maybe that will come down [06:19] the road but what would be really [06:21] interesting is that you know if you have [06:22] contributor and you have batching that [06:25] would be insane for me right? So like [06:27] because I can get around this 100 100 [06:29] per request uh limit. Um but that's just [06:32] one thing you have to consider if you [06:33] are going to use contributor. Um oh they [06:36] also have web search grounding and so it [06:38] is a a higher cost here. Um you know so [06:41] per per thousand search queries. So just [06:43] consider uh that if you do want to [06:45] ground information u it's going to be a [06:48] little bit more. Um so that or you can [06:51] just bring the information in uh that [06:53] you need grounded but depends on what [06:54] kind of app you're building. In terms of [06:55] benchmarks, I would describe Meta as a [06:58] Goldilocks model. It's not at the [07:00] highest, uh, but it's not the lowest. [07:02] It's not, it's somewhere in the middle. [07:04] Um, and in here, we can see it right [07:06] there. It's usually in the upper middle, [07:08] uh, area. Uh, so, you know, I would [07:11] think of it as like, [07:13] you know, we'd say like the third best [07:14] model in everything. And so, that's [07:16] really useful when you want something [07:17] that's fast, cost effective, and you can [07:20] broadly use everywhere. Um, so you're [07:22] going to just again see it. It's going [07:23] to be uh in the upper middle of most uh [07:26] leaderboards. And I had a hard time [07:28] finding a leaderboard that I use [07:29] specifically for Japanese learning, but [07:32] um I have it somewhere in my my [07:34] documentation, but I it would have been [07:36] interesting to show that because again [07:37] and again it just keeps showing up there [07:39] in the upper middle. Uh and so to me [07:42] that that it's kind of interesting. Um [07:44] it uses the REST API as per usual. Its [07:46] endpoint is going to be Meta AI version [07:49] one. Uh you can see that it's using a uh [07:52] bearer uh API token. So that is uh [07:55] pretty straightforward there. Uh what's [07:57] interesting is MetaMuse supports [07:59] multiple protocols. So you have the [08:01] responses, the uh chat completion and [08:04] messages. So basically um because it has [08:08] that coverage, it's really really easy [08:10] to drop into existing SDKs. In fact, [08:14] they don't have their own SDK um because [08:16] they are encouraging you to uh utilize [08:19] existing ones and um I mean to be honest [08:22] when I build stuff I use the rest APIs. [08:23] I don't even use lower lower SDK. So for [08:26] me this is perfect for uh for my use [08:28] cases. But um you know that's their [08:30] approach. Um, in terms of what they [08:32] support, we have obviously the ROS [08:34] responses API, the chat completions, we [08:36] have uh search grounding, image [08:37] understanding, video and audio [08:39] understanding, structured output, [08:41] function calling, uh, file upload, [08:43] streaming, temperature and sampling [08:45] parameters, model selection. So, you [08:47] know, basically everything that you [08:49] generally need in a model. Um I said [08:51] that they don't have their own SDKs and [08:53] that is absolutely true but they are [08:55] meta API open AI compatible uh and also [08:58] with anthropics um messages. So whether [09:01] you want to use OpenAI or Anthropic uh [09:04] you just basically change the base URL. [09:06] Um one key difference here and we find [09:08] this out later is that uh for this one [09:10] we do not put the hyphen or for/v1 on [09:12] here. Um these examples are using 1.1 [09:15] because I just pulled them out of uh the [09:17] documentation from meta. Um, but I did [09:20] test both and they they worked exactly [09:22] as expected. Um, so that was really [09:24] easy. But you know when we say open AAI [09:26] compatible, we don't just mean with [09:27] OpenAI, anything that is OpenAI [09:28] compatible, right? So it's broader than [09:31] what is being shown here. So you can [09:32] work with Langchain, Llama Index, [09:34] Versell AI, a bunch of stuff. Um, so [09:38] that's really really nice. Uh, in terms [09:40] of coding harnesses, uh, you can [09:43] basically work work with a bunch. So [09:44] open code, uh, cloud code, codec. So I [09:47] test cloud code u I don't test open code [09:50] and codeex I mean they're very easy [09:51] obviously to get working with and uh [09:54] they have their own muse code. The thing [09:56] I did notice is that it's not like [09:59] a perfect transition especially with [10:00] cloud code is that you can plug it in [10:02] there uh but you'll get some weird [10:04] artifacts um uh just because that [10:07] doesn't mean that it doesn't work as [10:08] expected but the UI will will show you [10:11] some messages that will make you think [10:12] that things aren't working exactly as [10:14] expected but it works totally fine. Um, [10:16] but that's pretty part of the course [10:18] with claude code. Okay, open code's [10:21] going to be a a very direct integration. [10:23] Uh, but again, I don't explore open [10:25] code. I'm really interested in Muse Code [10:26] and Muse Code performed extremely well. [10:30] Um, I think I only had one little UI [10:33] issue and I think it was mostly because [10:35] I was um increasing the font size on my [10:38] uh text interface uh to show to you [10:40] folks. But when I was running sub agents [10:42] um and I had to do those approvals, I I [10:44] just couldn't always see what what was [10:46] happening, right? Um but other than that [10:48] was like very very good experience, [10:50] extremely streamlined um model. The only [10:54] thing I think I would want in it that it [10:55] didn't have at this time was uh a way of [10:57] troubleshooting MCP as that's always [10:59] kind of a pain and cloud code and other [11:01] providers solve that. So that's just one [11:03] little gap they have but pretty good [11:04] solid uh coding harness to begin with. [11:07] Um, the way we are going to uh get our [11:09] keys and work with stuff, we're missing [11:11] the Dion here, folks. Um, is via the [11:14] dev.ma meta.ai playground. And so we can [11:18] go over there and this is where we're [11:19] going to get our keys, our usage, uh, [11:20] everything like that. Um, so it's pretty [11:23] straightforward. Building is a little [11:25] bit different. So you aren't saying, I [11:26] want to load up $20. you're setting a um [11:30] a very specific [11:32] um uh cap of saying like hey uh you know [11:36] charge up to $30 and then notify me [11:40] right um so it's a little bit different [11:43] um but it's nice because if you are used [11:45] to like claude or other providers where [11:47] you have a minimum of $5 you don't have [11:48] to set a minimum here okay they're going [11:50] to only charge you once you go over a [11:52] dollar uh or or there's something like [11:55] some logic like that so pricing is a [11:57] little bit different, but it's totally [11:58] fine. But anyway, I think that's all we [12:00] really need to uh know to jump into um [12:04] Muspark and Musepark code. So, let's go [12:08] ahead and do that. Okay. [12:14] All right, folks. So, um I want to just [12:16] quickly show you um I already have an [12:18] account, but basically, you'll get [12:19] started. You'll connect. You'll get in [12:20] there right away. And then after that, [12:22] you're going to probably want to load up [12:23] some money. So, I do have some credits [12:25] to play around with here. Uh but [12:27] originally I did load up um 30 bucks on [12:29] my credit card here. I've obscured that [12:30] information. Um and so in order to start [12:33] utilizing you attach your credit card [12:35] and I would set a spend limit. So I set [12:37] mine as 30. So when you reach that [12:39] balance then that's uh when we will pay. [12:42] It's a little bit different than other [12:44] providers. Other providers you load up [12:45] money and then you run out. This one is [12:48] uh different. Okay. So just be aware [12:50] that you are setting your limit and it's [12:51] kind of open uh in terms of uh that [12:54] spend and you know just be aware of that [12:57] with your API keys. Um and I believe [13:00] that yeah there's now glimmer. So I [13:02] think even when I started this glimmer [13:04] wasn't out and now it's out. So we have [13:05] metaspark 1.2 and 1.1 and this is the [13:09] open model that you can download. So I I [13:11] assume it's very similar to llama. Uh so [13:13] if you ever want to download one and [13:15] work with it, glimmer is now here. Um, [13:18] but yeah, if you want to follow along [13:20] with this stuff, you do need to load up [13:21] or you do need to attach a credit card [13:23] and set a limit, whatever it is, $2, $1, [13:27] whatever you're comfortable with. Um, [13:29] uh, but you don't, again, you don't have [13:31] to. There's no minimum of like $5, [13:33] right? Um, or $10 like other providers. [13:35] So, if you want to come in here, [13:36] experiment with$1 or $2, you absolutely [13:38] can, which is a great option for you. [13:40] Uh, but I just wanted to point that out [13:42] before we get started. Um, so yeah. [13:50] So, if we want to start working with [13:52] these models, what we can do is go to um [13:55] developer.ma.com. [13:57] And right off the bat, you can see that [13:58] we have Muse code. Now, if I just clear [14:00] this out here and go to the main page, [14:02] we'll just see where this brings us. So, [14:04] it brings us to developers meta.com. But [14:07] if we go here, let's try AI [14:10] and see if we can get to our respected [14:12] area. So, this will bring us back to the [14:14] same place. So, I've already created an [14:16] account. Um, so I'm going to go ahead [14:17] and just get started here. I think I [14:19] just used um, uh, Gmail like clickth [14:23] through to get through here. And [14:24] obviously, there are many ways that we [14:26] can get installed here. But the thing [14:27] that I want to focus on is actually just [14:28] showing you the model usage. Um, I [14:30] originally did put money uh onto this [14:33] account and then I I was able to get [14:34] credits later on. So, I do have some [14:36] credits that expire at some point for [14:38] experimentation. Um, but I'm here to [14:40] show you what you can do. I don't know [14:42] how you can get credits. Um, but again, [14:44] you know, I would just say load up $5 or [14:46] $10 and you can get the uh similar [14:48] experience and they go really really [14:50] far. I don't even think I needed uh [14:51] credits honestly to do what I wanted to [14:53] do. Um, but anyway, so over here we have [14:56] uh 1.2 and 1.1. Um, so the thing is is [14:59] that I've already said this in the [15:01] slides, but um there there's not a huge [15:03] difference between 1.1 and 1.2. 1.2 to [15:05] is just a newer um checkpoint and this [15:09] might be the future and there could be [15:11] something even newer uh here that is [15:13] here you have your usual controls u you [15:16] can change your reasoning level you have [15:19] streaming you have uh grounded search [15:21] grounded search does cost more per token [15:23] to utilize but um it is something that [15:25] you can do and there are other [15:28] capabilities that we have like putting [15:29] in structured JSON schema and additional [15:32] messages so um the playground is a great [15:34] place to start. We will pro [15:36] programmatically use this later on, but [15:38] let's just see if we can uh see what it [15:40] can do. So, um I think again we'll go [15:43] ahead and start working with uh Spark [15:45] 1.2 here. And uh I really like learning [15:49] Japanese, so that's something that I [15:50] spend a lot of time on. So, and it's a [15:52] good benchmark for me. So, I'm just [15:53] going to see if it can produce that at a [15:55] medium level. And we'll see what happens [15:57] here. So, we'll just say uh you know, [16:00] can you give me a uh list of uh grammar [16:06] points I need to learn for the JPT? I'll [16:09] just pull this up here for the JPT. [16:14] Um I guess I'll have to expand this out. [16:16] It's not it's not the friendliest when [16:18] it comes to expanding here, but the JLPT [16:20] JPTn5. [16:23] Okay. And we will see what it comes [16:25] back. One mistake that always happens [16:26] that if you ask for Japanese sometimes [16:28] they give you everything in Japanese but [16:30] generally it goes up in level but it's a [16:32] really good test because um you know [16:34] languages like Japanese are are very [16:35] comprehensive and uh this is where I can [16:37] kind of spot failures u but we will give [16:39] it a moment uh to produce here and here [16:44] we go. So here is the complete uh JLPTN5 [16:47] checklist. The JLP does not publish an [16:49] official list as true and there's 40 [16:50] grammar points. It's a very good summary [16:53] actually right off the bat seeing this. [16:55] Uh we'll see. I think one thing I really [16:57] like about meta is just not like for [17:00] this. It's just how it brings back data. [17:03] Um I would say OpenAI is pretty good. Uh [17:06] um uh what do you call it? Anthropic is [17:10] uh I do not like its formatting. It's [17:11] always too verbose and whatever. So [17:13] whatever they've done to fine-tune this [17:15] here, it's uh it's really really good. [17:16] So you can see here it's giving me uh a [17:19] really uh concrete example of list here. [17:23] So we got through particles. Yep. And [17:26] then we have basic sentence structures. [17:30] Yep. Yep. And so this is really really [17:32] good to the point that I'd probably save [17:33] this for later. I could give them a nice [17:36] good response here. I'm actually really [17:37] happy with that. Um and I probably want [17:40] to save this for later, but I'm just [17:41] noticing here that they don't have a [17:43] copy feature. It's the playground, so [17:45] it's not that big of a deal. Um but that [17:48] is fine. Uh let's go take a look at its [17:50] vision capabilities. Um so uh there's a [17:52] website called NHK news. NHK [17:58] news easy. [18:00] And so this is a website for people [18:02] learning Japanese. And it's supposed to [18:05] be like really easy articles when you're [18:07] learning. And so I'm going to go ahead [18:08] here and enter it. And I just want to [18:11] take a screenshot and see if it will [18:12] transcribe the content. It shouldn't [18:14] have an issue with it. Um, so we will [18:17] grab maybe this one here. [18:20] And so you can see these are very very [18:22] short. And so I'm just going to go ahead [18:24] and grab this. One trick here is it has [18:26] these little characters for so that can [18:28] sometimes mess up um transcription. And [18:31] so it'd be very interesting to see how [18:32] that does. So we'll go ahead and we will [18:35] paste that in. I can't actually paste [18:36] that in directly. So what I will do is [18:38] open up paint. It's not a big deal. [18:41] We'll just do that. And we'll go ahead [18:43] and save this to our desktop. Okay, that [18:45] is now saved. And then I'm going to [18:47] bring it in here. I just stopped the [18:48] video. I don't like to show all my local [18:49] files. And I have now attached the file. [18:52] And so, uh, I can just say transcribe. [18:54] Now, I'm not going to do this, but like [18:56] if it was a book, let's say it was like [18:58] a full chapter of a a published work. [19:00] Some LLMs will refuse to, uh, to do [19:03] that. They'll recognize that they are [19:05] copyright and might not do it. Um, some [19:09] have more free liberty. They might say [19:10] like, "Yeah, we'll we'll transcribe [19:12] this, but you know, we were not going to [19:13] do certain things with it." Um, so I [19:15] don't know what meta is there. I'm not [19:17] going to push the boundary of that, but [19:18] I'm just curious on how the [19:19] transcription worked. Uh, and so, uh, we [19:22] will just take this as a comparison [19:24] here. [19:26] And actually, you know what? I can just [19:27] feed it in here and ask it. [19:30] So, we'll grab this here. I can turn the [19:32] furry gun off here. So, I can go ahead [19:34] and just turn that off. [19:36] And we will go ahead and copy this in [19:37] here. So, how well did the uh [19:41] transcription go [19:43] compared to the original? [19:46] Let's see if it can reason that [19:50] original [19:53] and we'll see how well it did. It's on [19:55] medium effort. We could turn it up here. [19:57] The transcription was 100% accurate to [19:59] your original. Um, every kanji hiragana [20:02] katakana I like. Maybe it knows what's [20:05] going here. The only difference I had [20:06] the Ferono readings in parenthesis that [20:07] were written in small text. I think that [20:09] was the only thing that was a little bit [20:10] um unusual. There are these three [20:12] characters. So what about what about [20:14] these things? Uh no characters are [20:17] misrecognized. Okay. What about these? [20:22] Maybe it's just not showing encoding. [20:25] Those are not the original image. They [20:27] are replacement characters which means [20:28] my output had encoding glitches and [20:30] failed to render correctly. Um, so aside [20:33] from that, the render the results of the [20:34] transcript was 100% correct. And if I [20:37] just read this quickly, [20:40] maybe the first sentence here. [20:45] Mhm. [20:49] And it looks good. So that is uh one [20:52] good example. Another thing that we [20:53] might want to check is to see how it [20:54] builds out websites as that's always [20:56] really fun. Um, so let me just figure [20:59] out a page that we want to build. Um, [21:02] give me a second to think about it. [21:04] Okay. You know, I'd say something I have [21:06] a lot of, um, nostalgia for is PHP Nuke, [21:09] uh, with this like the lefth hand column [21:11] and the righth hand column kind of [21:13] website. So, let's see if it can [21:14] replicate this for fun. Uh, and so I'm [21:16] going to go ahead and grab a screenshot. [21:18] We'll have to open up Paint and save it [21:20] there. Be nice to be able to paste [21:22] directly, but again, it's a playground. [21:23] It's not, you know, it's not supposed to [21:24] be a full featured chat system here. And [21:27] I'll just go ahead and save that. Okay. [21:29] And so I'm going to go all the way back [21:30] up to here and we will uh go here and [21:32] attach that image. Okay. And for fun, [21:35] I'm going to go ahead and change the [21:36] reasoning up to extra high. I'm also [21:37] going to give a system instruction. Uh [21:39] you are a website uh designer uh or like [21:45] a you are a front-end web developer [21:51] with designer skills. [21:54] Um you will take a concept with [21:56] reference image [21:59] and build out a single page website um [22:03] all self-contained [22:06] uh within [22:08] a single page of a website [22:12] within uh a single HTML [22:15] page. Uh so you know [22:19] please don't use a CSS framework. Um [22:26] okay so just you know straight up do it [22:28] here. uh use flexbox [22:31] uh as much as you can [22:35] and keep the markup [22:38] uh [22:39] keep the uh markup and CSS selection uh [22:44] as lean as possible. Okay, so we'll do [22:47] that. I probably give a better [22:48] instruction for that. So, um, you know, [22:51] I really like the era [22:55] of [22:58] Nuke PHP communities, PHP new [23:00] communities [23:04] where they have lots of uh, plugins [23:08] that show on the left and right. Um, and [23:11] they have the three column design. [23:15] I have uploaded. I've provided you an [23:18] image [23:20] of the [23:23] original PHP new template [23:27] for uh reference. Um I want a concept [23:32] for a website. Um [23:36] I'm trying to think like I could do [23:37] Japanese of course, but I think that it [23:39] would be more fun if we did it something [23:40] grounded that we all we all enjoy. Not [23:43] necessarily a gamer site, but like a [23:44] tech site. So, so I want I I I want a [23:47] concept of a comm a a social [23:51] social community, social tech community [23:54] website [23:56] um for tech professionals. [23:59] um you know learning [24:01] you know [24:03] the broad landscape of tech [24:09] as in you know DevOps cloud AI security [24:15] you know uh you know maybe not [24:18] yeah I think that's fine so so hopefully [24:20] that is a good sense here so please [24:22] produce concept as full HTML page okay [24:27] also make it mobile friendly, [24:31] you know, think of some fun plugins to [24:36] to conceptualize and show. Okay, great. [24:40] So, we'll go ahead and we'll see how [24:41] that does [24:43] and we can fiddle with all this stuff, [24:44] but I think that's fine. The big thing I [24:46] think is the X high. We didn't turn on [24:48] uh search grounding, but we can try that [24:49] later on. Um, additional messages. I [24:52] think that's just you stacking messages [24:54] uh from from prior, but we will give [24:56] this a moment here and see what happens. [24:58] Okay. And here it goes off to the races. [25:01] Producing HTML pretty quick I would say. [25:05] Um it says single file retro PHP new [25:08] concept built for modern tech community [25:09] three column flex layout. Now normally [25:12] what you'd probably want to do is say [25:14] like hey can you give me a concept and [25:16] then I'll review it and give you at [25:17] least let it have its thoughts out and [25:20] uh that but I'm really single shotting [25:22] this because I just want to see how it [25:23] performs at base here. And so we have [25:25] it. Obviously, we can't view it [25:27] directly, but we can um copy it and save [25:30] it as a file. [25:32] Okay, so we'll go ahead and copy that [25:34] and just give me just a moment to go [25:35] ahead and save that. So, I have uh saved [25:38] the file locally. Let's take a look at [25:40] uh the website of produce. And wow, does [25:42] that feel that feels so much like uh [25:45] what I am used to. I'd probably change [25:48] it a little bit, make uh make a bit more [25:49] spacing. I think it' make a huge [25:50] difference. But like uh what do we have? [25:53] We have modules. I like how they called [25:54] it modules. That's really fun. We have [25:56] the system load. Um, I guess that could [25:59] tell you what's going on with the [26:01] current machine, which is fine. Who's [26:02] online? That's a pretty common feature. [26:05] Uh, select your stack. I don't know if [26:07] that really provides much value. Again, [26:09] we the single prompt, right? We're not [26:10] asking to do tons of stuff. We have our [26:12] contributors down below. Um, how the [26:14] server is doing. Uh, nuke chat shout [26:17] box. Um, [26:20] so that looks pretty clear. feed doc [26:23] submit post topics top 10 [26:26] and how do we cut Kubernetes cost down [26:27] by whatever without losing sleep [26:30] so it's fun but you know I think I would [26:32] I would like it to take one more pass at [26:34] it and see if it can make funner module [26:35] so uh can you in summary so the the page [26:38] looks great uh could you um [26:43] uh the uh there are multiple [26:48] plugins about the performance of [26:51] the website. I think the plugins I'm [26:55] interested I'm interested [26:59] in con uh concepting are ones that will [27:03] uh create community engagement. [27:07] Can you first uh suggest in summary [27:13] uh single point bullets um ideas that I [27:18] can uh review? Okay, so we'll go ahead [27:20] and do that. Again, I'm on X high [27:21] reasoning right now. We're not I'm not [27:24] really keeping track of my costs to be [27:25] honest. Um, so I'm not really concerned [27:27] about that for what we are doing here [27:29] right now. Um, and so we'll take a look [27:32] here and see what they have. So we have [27:34] uh deploy confessions anominous 140 prod [27:37] stories. Others react with f or or or [27:42] fixed. I don't understand that part, but [27:44] that's okay. Oneclick random 50-minute [27:46] coffee match with someone outside your [27:48] stack. [27:50] That's interesting. So, like a pairing [27:51] thing. Uh, post a snippet community [27:54] votes. [27:55] Code review. Thunderdome. Uh, okay. [27:58] Stump the stack. Weekly impossible [27:59] bounty question. [28:01] Kudos terminal. Send kudos [28:04] shown as a live tracker. [28:08] Sure. Project M. I'm just trying to see [28:10] if like any of these are are good. I [28:12] mean, like I not expecting any AI to [28:14] really do a good job of of choosing [28:16] these, but I'm just curious on how it's [28:17] done. [28:20] Okay. So, matching up projects upcoming [28:22] live ask me anything uh mentor [28:24] questionship. Okay. So, what I'm going [28:25] to do is like can you uh you know [28:27] instead of grounding it to specific [28:29] examples can we can we create a visual [28:33] mind map [28:35] of uh points of community engagement? [28:40] Uh uh [28:43] you know can you make this a [28:47] mermaid diagram? [28:50] Okay. [28:56] Delete chamber polarizing pole with [28:57] threaded with threaded need data set [29:00] labler. Okay. Let's make sure that is. [29:02] So we have a mind map here. I'm just [29:03] curious as in terms of what it's [29:05] produced. You would think like with meta [29:06] because and I'm not sure what their data [29:08] source is, but like you know they're a [29:10] social platform. So I feel that the you [29:12] know in their training they might have [29:14] something there that might be [29:15] interesting. Uh and so I'm going to go [29:16] ahead here and just uh save this as a [29:18] file. One sec. It can be a bit hard to [29:20] uh get Mermaid to display. So I'm just [29:22] going to go I was trying to get a Visual [29:23] Studio code extension, but I'm just [29:25] going to paste it in here so that we can [29:26] see. And let's take a look at what we [29:28] have in terms of data. Um and so in our [29:31] Wow, this is kind of cool. So we have um [29:34] our tech community. So we have learning, [29:36] growth, skill building, daily [29:38] challenges. And so the idea is that I [29:39] would think that the outside would be [29:40] the kind of task that we would have, [29:43] right? And so here we're talking about [29:45] learning and growth and then we do skill [29:47] building. So we have daily challenges or [29:49] curated vault. Uh I don't really know if [29:51] I know what that means, but that's fine. [29:52] I guess maybe like stuff from prior that [29:54] was really good content that has been uh [29:56] definitively made. Okay, that makes [29:58] sense. So we have experimentation, [30:00] exploration, and so I would think that [30:02] um we have participation. Maybe this is [30:04] more of what we're leaning towards like [30:05] ask questions, debate chambers, shout [30:07] boxes. And so you go all the way around [30:09] here and you can kind of see uh things [30:12] materializing. One thing we do not see [30:13] is like individualized um ideas. So this [30:17] is where I would ask it to go go on the [30:19] outside of the branches and say, can you [30:21] give me now plug-in examples from this? [30:23] So we'll go back over to here. I'll say [30:25] that's great. [30:27] that that that uh that mind map is [30:29] great. Um for the out for the outmost [30:34] branches [30:36] um can you now create you know two to [30:39] three examples of uh plugins or [30:42] engagement [30:44] uh plugins or engagements [30:50] modules that are grounded [30:53] with uh with with act with examples. [30:56] Okay. Um, [30:58] and just color them [31:01] uniquely [31:02] the same color so I can distinguish them [31:08] them in the uh the tree here. So, we'll [31:10] go ahead and do that. We'll give it a [31:12] moment to come back here. [31:17] Okay. [31:19] So, we'll just chill out and give it a [31:20] moment. And literally took a second. I [31:23] didn't even have to stop the video for [31:24] that. I'm not sure why I even bother [31:25] doing that, but we'll go ahead and now [31:26] we'll paste this one in. So, I'm curious [31:27] to see what this one looks like. [31:29] And we'll go ahead and paste it in. [31:32] I'm not sure if that works. So, I'm [31:34] going to go ahead and delete that out. [31:35] We will go back over to here. Copy this. [31:39] Go back over to here. Give it a paste. [31:42] And it is not pasting here today. It's [31:43] just refusing to do so. So, I'm just [31:45] going to hit copy a bunch. It can It's [31:48] often my computer the copy paste. It's [31:50] kind of frustrating. [31:52] There we go. I just I kind of I had a I [31:55] think it was this editor here that was [31:56] giving me the problem here. So, it says [31:58] fix the syntax error. Where is it? I [32:01] don't see it. [32:03] So, let's go back over to here my map [32:06] because I kind of couldn't paste it in [32:07] exactly how I wanted to. So, it's giving [32:09] me a bit of trouble. So, I'm going to do [32:10] is I'm going to just try to copy this [32:11] way. [32:14] It might be the the the class definition [32:17] here down below. It might not be able to [32:18] handle that. That might be the reason [32:21] why. Oh, you know what? It's incomplete [32:22] at the bottom, too. I don't even think [32:23] it copied the whole thing. So, we'll go [32:25] ahead. We'll paste that in manually like [32:26] this. No, it's just this editor is kind [32:29] of mucked up. So, what I'm going to do [32:30] is I'm going to go back over to here. [32:32] Whoops. Can I get a new one here? [32:37] Give us a refresh here. And I'm going to [32:39] just I think it's this editor. This [32:41] editor is giving me some trouble. And um [32:42] there's probably a plugin for Visual [32:44] Studio Code. I just can't seem to find [32:45] it right now. [32:48] It's really not pasting. Let me fiddle [32:50] with it. One second. Okay. So, I I think [32:52] it maybe it doesn't like that custom [32:53] color. It can't handle it. So, I'm going [32:55] to try this and see if that fixes it. [32:59] I think this editor just sucks. [33:04] And uh you can see it's a little bit [33:06] hard to find one. [33:11] This is frustrating. The mermaid part, [33:14] not the the model. The model seems to [33:15] have produced uh what I want. [33:18] Okay. One second. All right. So, I just [33:21] sucked it up and I made a new folder and [33:22] placed them in here. And I got a plugin [33:24] that should be able to render it that [33:25] renders it in line when you do this. So, [33:27] let's take a look and see if this works. [33:28] So, this is open as preview. Um, and [33:30] there you go. Okay. So, now we can see [33:32] it. I mean, it was prettier before with [33:33] all the coloring, but this still does uh [33:35] does the job. And so, what I'm looking [33:37] for here is did we get the same thing as [33:41] before? Learning and growth, mentorship. [33:44] Um, [33:46] is this the same as before? [33:50] I'm not sure. It looks a little bit [33:53] different. So, I think I'd have to have [33:54] the other one pulled up as comparison. [33:57] Let me go back to this one here, the [33:59] original one. So, we have um identity, [34:02] belonging, member presence, who's [34:03] online, stack identity. And so, here [34:06] like we see those, right? And I'll go [34:09] back up to here. And so that one was [34:11] recognition [34:13] uh identity and belonging right [34:18] here [34:20] and [34:22] member presence [34:25] on call status dots avatar. So, [34:29] uh, close, but I think I think you, uh, [34:31] ended up removing a layer from our, uh, [34:36] our tree, uh, you know, example we had [34:38] before [34:40] this. [34:45] And you're supposed to like, you know, [34:49] grounded example, [34:51] grounded example. And I mean, some of [34:53] these don't really require grounded [34:54] examples, but we'll go ahead and try [34:56] this. [34:58] Let's see if it can revise based on that [35:00] information. [35:05] We'll give it a moment. Okay, here we [35:07] have deepest leaf set. So, it went a [35:09] little bit deeper and I don't know there [35:10] was maybe limitations on mermaid. I [35:12] don't know the limitations of mermaid [35:13] that well. So, it might have just pushed [35:15] back so that it wasn't super complex. [35:17] Again, I don't really know. Let's go [35:19] ahead and see if this is even easy for [35:20] us to visualize as it'll have much more [35:23] information. And we'll go over to here. [35:26] Wow, look at that. That's a lot. Can I [35:29] zoom in on this? That's my question [35:31] here. Um, not really. But I mean, again, [35:33] this is the limitation of this thing, [35:35] but you know, if I squint, I can kind of [35:36] figure this out. I know you won't be [35:38] able to read it, but I will just go [35:39] through here. So, we have learning, [35:40] growth, curiosity, experimentation, try [35:43] this Slack button, uh, sandbox labs. [35:46] Okay, fair enough. feedback raids, [35:49] feedback cards, participation, [35:53] um, [35:55] dialogue, [35:57] debate chambers versus debate arena. [35:59] Okay. Or debate or dialogues, [36:03] shout box, um, coffee match, [36:07] low Friday entry, quick polls and [36:10] reactions, Instapole widget. just kind [36:13] of looking at this kind of even, you [36:14] know, even if I don't want to use this [36:15] stuff. Like one thing I thought was [36:16] like, well, what if you had a website [36:17] where you could only join every Friday, [36:21] you know, those kind of engagements I [36:23] think are kind of fun. But uh anyway, [36:24] that is just something that we have over [36:26] there on the side. And I think we kind [36:29] of spelled it out there. So I'm going to [36:30] stop the video here and we will explore [36:32] individualized things like JSON schema, [36:34] which we definitely want to check, and [36:36] grounded uh search grounding. Okay. So [36:38] I'll see you back in just a moment. [36:45] Okay, I want to continue on using Muse [36:48] Spark 1.2 here as um there's some other [36:51] capabilities that I would like to uh [36:53] try. So, one thing might be doing uh [36:55] search grounding. So, that might be uh [36:58] useful to do. Um I've been doing a lot [37:01] of car repair as of late. So, uh [37:03] grounding a search in car repair might [37:05] be useful for me. I'm just trying to [37:08] think of an example where uh it matters. [37:10] So, I have a uh Dodge I have a Dodge [37:13] Grand Caravan. I guess a good comparison [37:16] would be to compare it without the [37:17] grounded search and then afterwards. So, [37:19] that's what I'm going to do. I'm going [37:20] to open up a new tab here and we're [37:22] going to create a prompt. Um, so I have [37:25] a Dodge Grand Caravan [37:30] 2017 [37:33] and um I need to replace my [37:39] control arms. uh my control arms. Can [37:42] you find I'm in Canada. So we'll say I'm [37:44] in Canada. [37:46] Can you find a reputable reputable brand [37:51] um that will be [37:54] cost effective but also good? Where what [37:58] website [38:01] am I buying from? [38:04] And I have high extra high reasoning on. [38:06] You don't necessarily have to have extra [38:07] high for that, but I'm going to leave it [38:09] on. And that's my prompt. I'm going to [38:11] go ahead and hit enter. I'm going to go [38:12] over to here, and we're going to uh turn [38:14] on search grounding, and I'm going to [38:17] give it high for fun. And we'll go ahead [38:19] and give it a comparison here. So, [38:21] normally doesn't take that long, but my [38:23] idea here is that because it's not [38:25] grounding the search, it's not going out [38:28] to the web and getting current [38:29] information. And this one actually is. [38:31] And so, it'll be curious to see what we [38:34] get. Um, and I'm just going to give it [38:36] uh, you know, a minute here and I'll be [38:38] back in just a second. Okay, so [38:40] literally 10 seconds later, and the one [38:42] that's not grounded has already [38:43] returned. This one is still working on [38:45] it, which is fine because it's going out [38:46] to the internet here. So, best brand for [38:48] Canada, Mevotech, Supreme as a Canadian [38:52] company is best price ratio here. Don't [38:54] buy cheap white stuff box off of Amazon, [38:56] eBay. Definitely not doing that. Um, [38:58] Supreme, this is what most Canadian [39:00] shops use. is better than the OEM. [39:02] Lifetime warranty. Comes with ball [39:04] joints and bushings. Pre-pressed, etc., [39:06] etc. with a number. Um, town and country [39:10] price 130 to 190. Best heavy duty. Then [39:15] they have Moog. I I always think that [39:17] this is OEM when I see this, but maybe [39:19] it's not. The old problem solver gold [39:20] standard. Excellent quality, but more [39:22] expensive in Canada. Definitely is. I [39:24] never even heard of this one. I'm not [39:25] sure which one I bought. Give me just a [39:27] moment to check. The ones I bought say [39:29] Dely Technologies, but I don't know if [39:30] that's actually the company. So, here, [39:32] yes, for caravan, you always want to buy [39:34] complete arms of ball joints. That's [39:36] what I did. So, best websites, [39:37] partsavvatar.ca, [39:39] which is I didn't know about this. This [39:40] is the one um I went to, which was [39:43] rockauto.com. Auto parts. Uh, everyone [39:46] recommends this. I've never heard of [39:47] this one, actually. Uh, interesting [39:49] part. My wife used to work for an auto [39:51] parts thing like this called the Wrench [39:53] Monkey. I didn't order from them, [39:55] though. [39:57] Sorry folks, not helping you out. But I [39:59] I ordered from this. The reason why is [40:01] that I know that they don't keep a ton [40:03] of this in stock and so I don't want to [40:04] have a headache trying to order from [40:05] them. Um but anyway, so I guess there's [40:09] parts Avatar which I didn't check. The [40:11] hard part here is the cost of shipping, [40:13] right? And so um that was one thing to [40:16] factor in which we'll take a look here. [40:18] Oh, look. It says the wrench monkey. [40:20] Look at that. I didn't notice that [40:21] there. Also top competitors. Um if you [40:24] want the absolute cheapest rock auto [40:27] But the shipping cannon is inflated and [40:29] you wait for seven days. And that's a [40:30] thing that that I never considered. Um, [40:33] and so buy the pair. If one side is [40:36] gone, the other is close behind you. And [40:37] you save that. So I actually ended up [40:38] ordering both here. And so I'm going to [40:40] go back over to this. I can give it my [40:42] postal code. So P 0 P. Oh, this is uh [40:45] we're mixing them up here. So this is [40:47] the original one here. Sorry. So here, [40:50] parts avatar. It didn't even mention [40:52] wrench monkey if you don't mind waiting. [40:54] So, we have that cost there. Um, [40:58] repair in pairs. [41:01] Do you just need the front lower arms? [41:03] There's there's only two on the bottom [41:04] here, right? So, you will need wheel [41:07] alignment after a budget of 140 for [41:09] that. So you know the question I would [41:11] have for it is um you know can you uh [41:15] uh [41:18] can you determine the cost of rock auto [41:22] shipping [41:24] um and versus [41:28] parts was the first one [41:31] parts avatar.ca [41:33] CA because maybe uh you know the parts [41:36] are cheaper but the shipping cost [41:40] doesn't make it worth it. [41:44] Okay. So, I'm going to go ahead and copy [41:46] that and I'm going to go over to this [41:47] one and I'm going to ask the same thing [41:50] if parts auto is out of stock for more [41:51] than 400. [41:54] And sometimes the thing is that even if [41:56] it's doing grounded search, they can't [41:57] always get to the data. Um, so like, uh, [42:01] I know this for certain because there's, [42:04] um, you know, like when you're looking [42:05] at products and stuff or event pages, [42:07] they'll purposely block, um, agents, [42:11] right? And so then you'll need another [42:12] service that will actually that that can [42:14] get around and have like a bunch of [42:16] different ways of scraping the data. Um, [42:19] uh, and I mean, that's not the purpose [42:21] of search grounding, right? Like search [42:22] grounding can't just penetrate through [42:25] everything because they have to be [42:26] generally generally honest. But there [42:28] are, again, I'm not I'm not going to say [42:29] questionable, but there are ones that [42:30] are more comfortable going and grabbing [42:32] the data that that websites don't want [42:35] them to grab. Okay? So, um, that's why [42:37] I'm not saying like tell me exactly what [42:39] part and find it and whatever. But, I [42:40] mean, technically, you know, like we [42:43] could find things off of Amazon.ca, but [42:44] whether it'll let us search rock is a [42:47] different story, right? So, here the [42:50] shipping for the job is $0. So, parts uh [42:52] avatar provides free shipping for all [42:54] orders. [42:56] Um, so here is 300 to 380 cats. You [43:00] automatically qualify under the the [43:01] thing. I think I paid likeund [43:05] uh 20 or 40 per each [43:09] does ship to Canada everywhere, but for [43:11] control arms it's expensive because of [43:12] that. So no surprise bill at your door. [43:16] Um, the bad news for heavy part shipping [43:18] often raises the cheap part of the cost. [43:20] So for your van would be similar per [43:24] arm. [43:26] So here two supreme [43:30] 384 [43:32] 560 to 650. I mean I didn't pay 560 to [43:35] 650 for these but um I do believe that [43:38] it's right that the cost is going to be [43:39] higher here. Um, but if we go back over [43:41] to here, [43:46] um, parts auto [43:50] shipping to Canada, zero. [43:55] And so it's interesting like this one's [43:57] probably more in line. Um, [44:00] but uh, yeah, I don't even I think it [44:03] was like 300 320 bucks I paid. Maybe I'd [44:06] have to go double check. So anyway, [44:08] what's interesting is that I I learned [44:09] is that uh you know, I should have used [44:12] parts parts avatar.ca. [44:14] But uh you know, I wasn't using models [44:16] for this. I was going based on [44:17] recommendations of what everybody was [44:18] saying. So um and I think if I need more [44:21] parts, I'll probably order from parts [44:22] avatar. Sorry, the wrench monkey. I'm [44:24] just not ordering from you. I'm not [44:25] giving you free advertising here today. [44:27] But you can see they was grounding that [44:29] information. And um surprisingly, the [44:31] built-in data was really, really good um [44:33] for that. [44:35] So but uh obviously grounding is uh what [44:38] is what is interesting here but anyway [44:40] we will continue on with uh other things [44:43] here. So just give me uh a moment here [44:47] uh and we'll stop the video and we'll [44:48] look at structured JSON schema which is [44:50] absolutely essential for any model. [44:52] Okay. [44:57] All right folks, we are back and uh we [44:59] had just done grounded search and I [45:00] would like to move on to structured JSON [45:02] schema as that to me is you know one of [45:04] the most important or valuable things [45:06] here. Oh, they have example schemas. [45:09] Oh, okay. So we have summary text. Um [45:12] and so it looks like we are looking uh [45:14] we're using the usual structured meta [45:16] schema JSON schema. [45:19] Yeah. So it looks like we are using JSON [45:21] schema. Okay. So the first thing is that [45:23] I need to think of like a data structure [45:25] that I want. One thing that you should [45:27] consider with JSON schema is that it's [45:29] not just great for getting out schema. [45:31] It's great for getting a result that you [45:32] want, right? Because you can kind of [45:35] constrain the intelligence through that [45:37] if it has to reason through something. [45:39] And so I want to create a grading [45:41] schema. So I want to create a a grading [45:44] schema [45:46] uh a grading output using JSON schema. [45:51] So I want a rubric. It needs to [45:55] grade my Japanese four and four. Uh give [46:00] me and four and five [46:03] challenges. [46:05] Uh let's create a rubric [46:09] and a feedback and [46:13] ideally [46:14] what would we should have done. [46:18] Okay. Uh this is for structured input. [46:25] Okay. And so we'll go ahead and let it [46:26] do that. We're still on extra high. We [46:28] do not need to be on high for this. It's [46:29] ridiculous. We'll turn it down to high. [46:33] It would take a lot of work to to figure [46:35] out the different side. Like I think [46:36] high and whatever. I've been doing [46:38] really well on medium, but we'll leave [46:39] it at high. I never seem to ever want to [46:41] use extra high. It always seems [46:42] ridiculous to me. But we'll give it a [46:44] moment here to get some structured [46:45] output. Okay. And then we will plug that [46:48] into the JSON schema. And it's [46:50] outputting HTML. So stop stop. All I [46:54] wanted [46:55] all I wanted was the JSON schema [46:58] reference JSON schema structure think uh [47:03] you know like I'm not sure why I started [47:05] creating the rubric which is fine [47:07] because there's a very specific name for [47:09] this format e [47:13] so like oh I I'll just give an example [47:16] here you know like this [47:21] you know eg this. Okay, we'll let it try [47:25] again. Okay, and so now we're getting [47:27] output back. Okay, so we have this here. [47:30] Um, rubric, criteria, particles, [47:32] grammar, [47:34] uh, vocabulary, [47:36] verb, conjugation. [47:41] Okay, I mean like yes, that's that's an [47:44] example of it filled in, but we are [47:47] using uh JSON uh what's the thing [47:50] called? [47:51] JSON schema. [48:03] Okay. Unless it gave me it to me above [48:05] here and I just wasn't paying attention [48:06] because we do have more than one [48:07] example. [48:14] There we go. Okay, great. So, now we [48:15] have it. Now, I'm not saying I [48:16] necessarily like the structure, but I [48:18] mean I think that this is just something [48:19] you need to work with. And so we'll go [48:21] ahead and I'm just going to save this in [48:23] my little thing over here. Just call [48:26] this uh structured. So JP [48:30] JP.json JP output JSON [48:35] just so I don't lose it here. Actually, [48:37] this is a bit easier to see now. Um, [48:43] okay. And so what we'll do here, [48:47] we'll save that. And so now we have that [48:48] structure JSON. I'm going to make a new [48:51] chat here. Uh new chat. And I'm going to [48:54] go ahead and add a new system prompt. [48:59] You are a Japanese um teacher. [49:03] You are you are a Japanese grader. Uh [49:07] you uh I am attempting to [49:12] uh the user [49:16] is provided a challenge [49:18] and you will [49:21] evaluate their challenge [49:25] sentence. Okay. So we'll go ahead and [49:26] we'll save that. [49:28] And down below here we'll add additional [49:30] messages. [49:34] And so I'm going to say uh you know [49:36] please give me a Japanese N4 challenge [49:39] sentence [49:41] that I need to trans uh translate. [49:46] I'm going to go over to here. [49:51] Sorry there. Um I'm going to go over to [49:52] here and um [49:59] just a second here. [50:01] Okay. I just need another uh meta tab [50:03] here. So, give me just a second here. [50:05] We'll open a new tab and we are going to [50:09] go over to what? Um, [50:14] what do I want to do here? I want to [50:17] Yeah, the same thing. I I need the same [50:19] thing here. So, we'll go ahead. We'll [50:20] grab this line, paste it here. [50:28] I just want I probably should have [50:29] turned this down as we'll get more out [50:31] output here. And that might might be [50:32] something I don't want. [50:35] Oh, it has the search JSON on here. [50:37] Sorry. One second. One second. One [50:39] second. We'll take that off here. [50:42] And [50:44] new new chat here. [50:48] I just want to clear this out. There we [50:49] go. Save it. [50:54] So, we're just going to help it along [50:55] here. I could have added it after the [50:57] fact here. That would probably would [50:58] have made more sense. [51:01] Sorry, [51:03] no Japanese, [51:06] just a single English sentence that I [51:09] will translate. [51:14] I barely wrote English there, now did I? [51:17] Now, I'm not going to bore you with the [51:18] the translation. I will pause the video [51:20] and do that. [51:23] In order to pass next week's exam, I [51:25] have to study for one hour a day at the [51:27] library. So, that will be our challenge [51:28] here. I'm going to go ahead and paste [51:30] this in here. Um, like that. [51:35] Save messages. Okay. I'm going to make [51:37] sure that schema is still here. It is. [51:38] Good. And so, the idea here is that now [51:41] that is queued up. And so, um, [51:47] that is what I'm going to attempt to, [51:49] uh, translate. So give me just a moment [51:51] here and I will do that. All right, [51:53] let's check out my uh answer. So in [51:55] order to pass next week's exam, I have [51:58] to study for 1 hour a day at the [52:00] library. So we have rishu next week. No [52:04] shinken. So that's exam. [52:08] I don't know how to say to pass. So kate [52:11] is like I think to win like to basically [52:14] like if you're fighting somebody, you're [52:15] winning a match or competition. So, uh, [52:18] I put nudu, which I don't know if I can [52:20] put sudu on that, but I did it anyway. [52:22] And so, I'm I put, uh, tom is like in [52:27] order to or the cause or reason. I don't [52:28] know if I'm supposed to put a no there, [52:30] but I did it. So, we have I study for [52:32] one hour a day at the library. So, so [52:35] every single day, the library day at the [52:38] library. [52:41] So, one hour, [52:44] which is to study, one hour to study. [52:49] which I believe means like like I should [52:52] I should study. I have to study but this [52:55] is like well I should study. Um so I'm [52:58] hoping that that is close to correct. [52:59] I'm not sure. I mean I always get this [53:01] stuff wrong but it's okay. We just got [53:03] to try and see the grading. Let's take a [53:06] look at what it produces. So this will [53:07] be fun for me. So go ahead and we'll [53:09] paste this in and we will give this a [53:12] try. [53:16] Okay. And while that's working, I'm just [53:17] going to go ahead and save this as a [53:19] prompt just so I don't have to like hunt [53:20] this stuff down later. All right. So, we [53:22] have something back here. And it took uh [53:24] just a minute to get this result back. [53:26] But I mean, it did give us a lot of [53:28] structure JSON. The more structure JSON, [53:30] the more thinking, the more effort it [53:31] takes. And so, we do have a structure [53:33] back. So, I'm going to copy this for a [53:35] moment out of here. And I want a nice [53:38] page for this to display. So, what I'm [53:39] going to do is create a new chat. I'm [53:41] going to clear out uh this schema here [53:44] as I don't want this anymore. [53:47] We'll save it. And uh I'm going to clear [53:50] this out as well. Save. And I'm going to [53:54] put on X high reasoning for fun. I'm [53:56] going to go ahead and give back the JSON [53:59] structure here to it. So, I'm going to [54:01] copy this and say, can you please uh use [54:05] this JSON to create a single HTML page [54:09] so I can uh see my uh graded report. [54:15] Um, okay. So, I wanted to do that and [54:17] I'll download that file and then we'll [54:18] take a look at it together. Okay. And [54:20] so, now we have a uh a page. We'll go [54:23] ahead and I'm just going to go ahead and [54:24] just give me my results. So, I'm going [54:25] to make a new folder here which will be [54:27] just Japanese grader. [54:29] This is something I've actually built in [54:31] uh really really granular detail in [54:33] another project. Um but it's called [54:34] Hajimari that's kind of just sitting [54:37] there that could be revisited as models [54:39] have gotten a lot better at their [54:40] intelligence and and cost. But I'm go [54:42] all the all the way to the top here and [54:44] copy this link here. And we will go and [54:47] make a new file here index html. And [54:50] we'll go ahead and paste that in here. [54:52] And I would like to uh take a look at [54:54] that now. So [54:57] if we could where is our preview button? [54:58] So, I'll have to open this in a browser. [55:00] One second. Here we go. So, what do we [55:02] have? 70 out of 100. Good. Oh, wow. [55:05] There's actually a lot of stuff here. [55:07] Because I didn't like analyze the [55:08] structure so detailed, I had no idea uh [55:11] if all those points were actually useful [55:13] or not. But let's take a look here. So, [55:14] you understood the overall structure [55:16] well, especially the have to grammar. [55:19] Oh, so nice of it. Good attempt. You [55:21] understood the overall structure. The [55:22] main mistake was the vocabulary for it [55:24] to pass. I knew I didn't know that. So [55:27] here we have uh uh yoku yoku deki masha. [55:32] So fairly um able to oh like able to [55:36] could be it. Yeah. Uh boon no something [55:38] something. I I don't have my thing uh my [55:41] tool to turn this on right now. So I [55:43] can't tell you what that word says [55:44] there. Um so we have reasoning. Oh, you [55:48] know what? This is basically the same [55:49] thing here. I think uh they're doing [55:51] here. Yeah, that's what they're doing. [55:52] So they're explaining in uh both here in [55:56] English and Japanese. Um so we will go [55:59] down below here and take a look. So the [56:01] model answer in order to pass the next [56:03] next week's exam I have to study for one [56:05] hour a day at the library. So [56:10] uh [56:12] oh this is the ideal answer. Okay. So [56:15] what's interesting is that this is [56:18] completely almost correct. The only [56:20] thing I didn't do here is I I don't [56:22] think I did Masen. Uh let's take a look [56:25] at what I did. [56:28] Uh Nanai. [56:30] Yeah. And so I mean this is kind of [56:33] nitpicking here, this one part here, [56:34] because this one is plain and this one [56:36] is more uh I guess polite if you will. [56:39] Um but the key thing that I'm messing up [56:41] here is this right here. So whatever [56:44] this word is here, [56:46] u I'm really surprised I did that well [56:48] to be honest. [56:52] And we have yeah gokaku Goku. So that [56:57] was my only mistake. So down below let's [56:59] take a look and see what kind of [57:00] corrections we have and why it thinks [57:02] that we need those corrections. [57:05] Uh so here [57:08] tadashi oh you have an explanation down [57:11] here. So also the particle tadashi [57:14] uh verb that's the word for verb. blah [57:18] should have been used instead. Oh, go [57:21] to pass. Yes s uh kiu to selfishly do. [57:26] That's right. You know what? I knew that [57:28] this is actually a fixed phrase to [57:29] selfishly do. As soon as I put that knee [57:31] there, I should have known better [57:32] without permission. So I that's [57:35] completely wrong, but I was thinking [57:36] because of this it was the uh like to to [57:39] fight, you know, k or whatever it is ka. [57:42] Um also the particle is ni uh not wo for [57:48] this um and for the verbs in order to [57:51] it's tame not to no tame. Okay, fair [57:54] enough. Like oh if it was like ending [57:56] with this. Okay, that that makes sense. [57:58] Again, folks, I know you don't care much [57:59] about this, but I'm going through it to [58:01] see its quality here. We're talking [58:04] about a duration of time. You don't need [58:05] to use well. You just provide the time. [58:08] That makes sense. Okay. Uh perfect use [58:11] of day. correct position correctly. Use [58:14] the N4 [58:16] uh this detailed rubric. So, vocabulary, [58:19] grammar, grammar, vocabular particles [58:22] and try to remember this next time. Wow, [58:24] this is really really good. Now, I spent [58:27] a lot of time last year trying to [58:29] perfect uh this system and uh I think I [58:33] was using open AI at that time and uh [58:37] sadly enough, this is performing [58:39] extremely well. So, [58:41] I think that um [58:45] I think that this might just replace my [58:46] whole tool. Uh and this is just a simple [58:49] experiment here. So, again, I'm saving [58:51] all this stuff because I might want it [58:53] later. Um but I think that was quite the [58:56] fun uh test there and that was a good [58:58] test for us. Um let me just go pull up [59:02] my example code here and we will move on [59:05] from that. [59:07] So in here we have this. So please give [59:10] me the Japanese [59:13] in order to pass. [59:18] Uh [59:20] and I will go back here. I think I'm [59:21] just missing my system prompt that I [59:22] want to save for later. [59:24] Or did I lose that? I lost it. Oh well, [59:28] that's okay. But anyway, uh hopefully [59:30] that was fun. And we have now checked [59:32] out I think basically everything that I [59:34] wanted to uh check in here. We don't [59:36] even really need to fiddle with these. I [59:37] think I think that those are pretty [59:38] straightforward. I don't need access to [59:40] other models. There is a contributor [59:42] model that will give you a lower cost. I [59:43] you probably access that [59:44] programmatically and you probably have [59:46] to ask for access I would imagine. Um [59:49] but we will now move on to programmatic [59:52] integration. Okay. [59:58] Hey, this is Andrew. In this video, I [59:59] want to show you how to programmatically [60:01] work with uh Meta's API specifically for [60:04] the Muse and models. Um, now we can use [60:08] Muse Code and all these other editors, [60:10] but I just want to work with the code [60:11] directly. And so, the easiest way is [60:13] just to start uh utilizing it via the [60:15] API directly. So, we're going to go [60:17] ahead and generate a new key. This key [60:18] is going to get expired anyway, so I'm [60:20] not really concerned about showing it to [60:22] you here on my screen, but I'm going to [60:24] go just say example. I'm going to go [60:26] ahead and generate myself a new key. I'm [60:28] going to copy that key and I'm going to [60:30] go here, make a new folder. I don't know [60:31] what I want to call this right now. It's [60:32] just be called untitled [60:35] and I'm going to make myself a newv [60:38] file here. I'm going to paste it in and [60:40] I'm going to just set this as meta API [60:42] key. Okay, look, I don't need LMS to do [60:46] this basic stuff. It is super simple. [60:48] We're going to make our way over to chat [60:49] and I believe that they might have code [60:51] examples for us. Uh and so the meta meta [60:55] API is compatible with um different uh [60:58] different uh libraries. So you can use [61:00] the anthropic one, you can use open AI [61:03] uh and so they don't necessarily provide [61:05] a direct one. I mean you can obviously [61:06] just use the rest API directly which is [61:09] one way. But I want to plug this into [61:11] something that um uh we we've never used [61:14] before. So we'll say open AAI uh [61:16] metamuse. We'll just see if we can find [61:18] a quick code example, [61:21] right? to see if we can get that. So I'm [61:23] just looking for that code example, but [61:25] it might be as simple as resetting the [61:26] open AI URL. So here we have an example [61:30] of some code where we are doing that. So [61:33] again, I'm not using anything special [61:34] here. We could have used meta to ask [61:36] them for that information, but right now [61:38] it's okay. So we'll just say um uh [61:42] agent.py. It's not really an agent per [61:44] se, but if we're going to go ahead and [61:46] just do this. So we'll go and grab this, [61:47] paste this in here, and I will add my [61:50] requirements.txt. txt [61:52] requirements.txt [61:58] and we will add open AAI here. [62:01] I always kind of forget um [62:05] uh how to make virtual environments. So [62:07] I'm going to try to cheat with not cheat [62:08] but I'm just going to try to not make it [62:10] a virtual environment here as I never [62:11] can remember off the top of my head. [62:13] We'll say python um install r [62:16] requirements.txt txt as that is the only [62:19] requirement we need here. We'll open a [62:21] terminal and I'm going to go ahead and [62:23] see if that works into our untitled [62:26] directory here and [62:30] uh let's see what did we do wrong. So, [62:32] we'll go up here for a second. [62:38] Oh, I think I pasted my code in. I [62:40] didn't actually paste in this line, [62:41] which is fine. Sometimes that happens. [62:44] And so maybe it's Python 3, which is [62:46] fine. [62:49] And that didn't work. So I'm going to [62:50] just type in [62:53] cond list [62:55] Python, right? We have Python. Python [62:56] hyphen hyphen version. We must have [62:58] this. There's no way we don't have [63:00] Python at all installed. Here we do. So [63:02] we have Python 3. Oh, sorry. It's pip [63:04] install. What am I What am I doing here [63:07] today? We'll say pip 3 install. Maybe [63:08] it's pip. Maybe it's pip 3 install. [63:10] Depends on what you have set up. [63:15] Okay. [63:17] And we will go ahead and install that. [63:18] So now we should have OpenAI installed. [63:20] We have our ENV. Another thing I'm going [63:22] to need is env. I believe that's all [63:24] it's called here. Env. [63:27] And we will do a pip install on that. [63:30] And we just need to import that. [63:33] So my guess here is import.env. [63:38] Just going to look it up really quickly. [63:44] Python env. [63:47] Yeah, it's load env. That's what it is. [63:51] And folks, we can use AI for this, but [63:53] like it's so easy and I've done it so [63:55] many times that I feel that I can just [63:56] do it this old way and save myself a few [63:58] credits. [64:00] Okay. [64:03] And yeah, that will load the file that [64:05] is sitting in that directory. Uh, this [64:07] is model API. So, we're calling this the [64:09] meta API key. At the meta endpoint, [64:13] assuming that is the correct endpoint, [64:14] it says 1.2. You're helpful coding [64:16] assistant, write a a quick hello world [64:18] function. And so, this should be in [64:20] theory the most basic thing that we can [64:23] do. So, we'll just say basic API usage [64:26] or we'll just say OpenAI [64:29] uh [64:31] SDK as an example of this here. Okay. [64:36] So, uh, we'll go ahead and do Python [64:39] 3, uh, [64:42] agent, I think we said agent.py, and [64:44] we'll see if that executes here. [64:47] It'll work or it won't work. Could be an [64:49] endpoint wrong. Again, that was, uh, [64:50] Google's results producing it. We didn't [64:52] look up the actual key here. [64:56] And it also said the role is developer, [64:57] which I'm not even sure if that's [64:58] correct, but uh, we'll either get [65:00] something back or we won't. [65:03] And so, uh, I don't see anything here, [65:04] but we didn't log anything out. Uh, and [65:06] it did take some time. So, that's [65:07] usually a good indicator. So, we're [65:09] going to go ahead here, and I'm just go [65:11] ahead and print out the response. [65:15] Okay, because it probably did work or [65:19] came back with an error, whichever. Uh, [65:21] we didn't specify things like its effort [65:23] or level or things like that. I think [65:24] we'd have to look up the chat [65:26] completions API. So, we do get a result [65:29] back here, which is really nice. And so [65:31] you can see here uh here's a quick one. [65:33] Print hello world. You uh can also make [65:36] it a string. So it's giving back an [65:38] example. So there you go. It's as simple [65:39] as that. Um let's go ahead and show off [65:41] the anthropic one for fun. Anthropic [65:44] SDK. [65:47] And I'm just going to put in here. Uh I [65:50] just wanted to ignore any files. I don't [65:52] have this committed to a repo, but I'm [65:54] just going to go ahead and make sure [65:54] that that is committed like that. So [65:57] that's a non-issue for me. I'm going to [65:59] go ahead and just copy the contents [66:00] here. We're going to go over into this [66:02] one here and paste in the same contents. [66:04] And instead of having this one, we're [66:06] going to get the anthropic SDK because [66:08] I'm told that you can use it, the [66:10] low-level one, anthropic SDK, [66:13] uh, Python [66:15] over here. [66:17] And so this one is going to be [66:20] just anthropic. [66:23] So I'll go over to here and we will [66:27] replace OpenAI with anthropic. We'll go [66:30] into our agent code and we will just [66:33] port this over line by line [66:36] as I do not mind taking my time to do [66:39] things manually. Okay, so we have that [66:41] and [66:44] I think that it's going to use the same [66:45] thing. Uh it's an API key. I think it's [66:48] base URL as well. So we'll go ahead and [66:49] try that out. [66:53] and paste that in there. [66:56] So, this one, it's not chat uh chat, but [66:58] it's uh it looks like messages create [67:00] here. [67:02] And I'm not sure if we set the model [67:04] here, but I'm going to try it. Yeah, I [67:06] think so. Yep, same thing. And here it [67:08] says message content. So, it says bring [67:11] back a message. [67:13] Um [67:16] I mean, that's the one they're showing [67:17] us. I'm not sure if there's a newer one [67:18] here, but um that's okay. Okay, we'll [67:20] just give it a go here. [67:24] Now, the structure might be different. [67:26] It says role user and so we're not say [67:30] developer, you're a helpful coding [67:31] assistant role. So, uh that that might [67:34] be a little bit different. Um so go over [67:38] here and say anthropic message schema [67:42] as theirs is probably a little bit [67:44] different. I don't think they call it [67:45] developer. [67:52] Yes. Can we see an example, please? [67:56] Python as an example. [68:00] So, all I'm doing is I'm just looking [68:02] for an example. Create a message [68:07] content ro user. [68:11] And we're looking for the messages [68:12] structure here. [68:15] Assistant. [68:17] Okay. So, we'll go over here [68:21] and write me a quick thing here. So, I [68:23] think the structure is more or less the [68:24] same. Let's go ahead and see if that [68:25] works. [68:28] And so, we'll just go back. We'll say [68:29] anthropic SDK. [68:32] And we'll do a um what would you do for [68:35] this? Uh Python 3, sorry, Python 3 or [68:40] sorry, pip 3 install-r requirements.txt. [68:44] So that's going to go ahead and install [68:46] the anthropic libraries. Let's go ahead [68:48] and see if this now works. Python 3 [68:50] agent.py. [68:53] Um expected max tokens messages models [68:56] or max token messages models. Um [69:01] where [69:02] on what line? [69:06] Line 14. So on line 14 it's expecting a [69:09] model. We provided that [69:13] messages. is we provided that and max [69:16] tokens. So I'm not sure what max tokens [69:18] we need, but we'll just do the usual [69:19] 1024. That might be too little. I don't [69:21] really have a good sense of token size [69:24] ever, no matter how much I've worked [69:26] with this. Says invalid API key [69:28] unauthorized param authentication error. [69:30] So this is where [69:33] uh we are having a little bit of an [69:34] issue. Let's go ahead and take a look [69:36] here. TheMV is here. So that should be a [69:38] non-issue. [69:42] I'm going to ask the meta model. [69:46] Oops. So, I'm going to go ahead and copy [69:48] my code into playground here. [69:53] Copy. [69:55] Paste. [69:56] And we'll just go here. Uh, [70:00] and I'll just give it the information [70:02] that it has here, the error. [70:08] How do I fix my code? [70:11] Okay, we'll hit I'll put it on high. We [70:13] probably don't need extra high. Medium [70:15] is probably fine, too. [70:18] Bad habit just putting everything on [70:19] extra high. And so, we'll give it a [70:21] moment to think. All right, so let's [70:23] come back with some code here. And um [70:28] I think the structure that's different [70:29] is the messages. I mean, it's saying [70:31] system. I guess that makes sense. It [70:32] would probably be actually system. Um I [70:35] know that this meta API should be [70:36] loading. So, I think that's a non-issue. [70:38] Alternative use the open. Well, I mean, [70:40] we already used the open AI. We don't [70:41] want to do that because we hear that we [70:43] can use the anthropic one. So, that is [70:45] our goal. I'm going to go ahead and So, [70:47] it should be [70:49] an assistant with a trailing space is a [70:51] valid role. Your partition prompt should [70:52] be in messages as assistant. So, [70:54] probably that messages structure was [70:56] wrong and I wasn't paying close enough [70:57] attention. That is kind of my guess. Um, [71:00] but it doesn't really matter. If it just [71:01] works, it's fine here. So, we'll go [71:03] ahead and paste it in and [71:07] yes. Yes. Okay. So, we'll go ahead and [71:09] give this a try. [71:12] And [71:15] version one was not found. [71:20] Okay. So, [71:25] compact anthropic version one was not [71:27] found. [71:29] Okay. So, we're going to go back over to [71:30] here. [71:33] Go back here. We'll just say uh [71:35] anthropic SDK with uh metamuse because [71:38] I'm almost certain that you can utilize [71:40] with it [71:42] with the anthropic SDK allowing you to [71:44] point. Okay, sure. [71:47] But is only with the NodeJS one and not [71:49] the Python one. It says the Python one [71:50] here. Basure replace with the meta's [71:53] actual model API base URL. Okay. Max [71:57] messages ro user token. So it shows an [71:59] example that it can can be utilized [72:02] and they suggest that it isn't here in [72:04] the anthropic code [72:09] right and so it's saying that it can be [72:11] done [72:15] if we go up to here we have the messages [72:18] API [72:20] and it could be the endpoints a little [72:22] bit different so here we have base URL [72:23] this is what it's telling us to do right [72:25] and so here's the example code and so [72:27] here you can is a little bit different. [72:30] So I think that is our problem. Okay. [72:34] Um 1024 you like I just put that in [72:36] there as if I know any better. Uh it's [72:39] just again something hardcoded my [72:40] memory. So that's probably the problem [72:42] we're facing here. So we'll go ahead and [72:43] paste it in there. So they had the code [72:45] examples all along folks. We don't need [72:47] to use AI for everything especially [72:48] getting started. Um but it definitely is [72:51] something that is fun and useful to use. [72:55] And I'm going to use the example they [72:56] have because I'm hoping that it just [72:58] works. Okay. And we'll go ahead here and [73:00] go up. Hit enter. [73:03] And what do we have? [73:07] I think it's working. [73:09] And we get a nice print afterwards, too. [73:12] Message is not defined. [73:17] Oh, because I call it response. That's [73:19] why. Fair enough. Okay. So, we'll try [73:21] that one more time. Now we're not [73:22] setting any of the options like uh [73:24] reasoning and uh you know things like [73:26] that. I'm not worried about that right [73:27] now um as I'm sure we can solve that [73:30] very easily. But here we are getting [73:32] data back and our structure. So you can [73:34] you can see you can use open AI or [73:36] anthropic [73:38] uh for whatever reason meta decided not [73:40] to roll their own. They just want to [73:41] integrate with other ones which is [73:42] totally fine. That is just an approach [73:43] that you can take. Um I mean personally [73:46] I just like using the rest API directly. [73:48] I don't even use low lower level SDKs um [73:52] because it's so easy to use REST API [73:54] directly and they have examples for that [73:56] but you know if you already have one you [73:58] want to drop it in place you can [74:00] absolutely do that there. So I think [74:03] that is pretty good for [74:05] that case. Um [74:09] I'm just curious if there's anything [74:10] else here. I mean obviously we could use [74:12] VL VLM but I think that covers basically [74:15] the first level which is working with [74:17] SDKs and libraries for the low-level [74:19] ones. Um I think I would like to take a [74:22] look at uh using any kind of um [74:27] uh agent SDK because that's where I mean [74:31] I again I've rolled my own but that's [74:33] where it would be very useful to find [74:34] out. So we'll do that in a separate [74:36] video here. But there you go. We have [74:37] the anthropic SDK and we have the open [74:39] AI SDK. [74:45] Hey folks, in this video I want to take [74:46] a look at utilizing the agent SDK [74:48] framework. That's something that's going [74:50] to get you up to speed quicker uh [74:52] building out agents. Um I personally [74:54] roll my own because I I believe [74:55] investing my into building my own [74:57] baseline agents, but you know if you are [74:59] familiar with things like agent SDKs [75:01] from Anthropic or OpenAI, then you can [75:04] apparently utilize them directly. Uh, so [75:07] we could try that. I think Langchain [75:08] would be a good example. I'm sure [75:10] Langchain will work because it works [75:11] basically with everything. Uh, but let's [75:13] first just do Anthropics SDK. Uh, open [75:16] eyes is okay. I just don't care for it. [75:18] So, we'll go ahead. We'll just say [75:19] anthropic agent SDK. [75:21] And we'll go ahead here and make [75:23] ourselves a new requirements.txt. [75:26] I'm going to bring in from our anthropic [75:28] one. Uh, well, I'll bring over the meta [75:30] key, of course, and we will paste that [75:32] in here. And then I will go into our [75:34] requirements.txt txt and copy that over [75:36] as well. Um, the import's going to be [75:40] different. It's probably be anthropic [75:42] SDK. I'm just guessing here. That's [75:44] probably what it is. And we will just cd [75:46] into the correct directory here. So, [75:49] we'll say agent SDK. I'll bump up the [75:51] font as well as I know it can be a [75:53] little bit hard on people's eyes. And [75:55] we'll do a pip 3 install [75:57] rquirements.txt. [75:59] Now, I do believe that they have an [76:01] example for us. [76:03] And um [76:09] I I'm not sure why it's saying that. So [76:10] we'll try that one more time. pip 3 [76:12] install r requirements.txt. [76:15] Oops. [76:21] Okay, sure. I mean, I don't care as long [76:22] as it's there. I might have already [76:23] installed the agent SDK before on this [76:25] machine. Uh but normally I would expect [76:27] that to show up. So we'll go over to [76:28] Python. It says claw agent SDK. So let's [76:31] go take a look here. Maybe the name's a [76:33] little bit different. [76:36] Yeah, it's Cloud Agent SDK and it hasn't [76:39] changed since then, right? [76:41] As they like to sometimes change names. [76:44] What I really dislike, and this is a a [76:45] GitHub complaint, every time I'm trying [76:47] to drag down my bar here. I always get [76:49] this darn button that I just want out of [76:51] here. So, we'll go ahead and we'll fix [76:54] that there. And so, that's Cloud Agent [76:56] SDK. [76:58] Now, Cloud Agent SDK can help you to I [77:01] again, I don't know why this is doing [77:02] this. Just give me a second to fix this [77:04] problem. Okay, chat GBT says it's not a [77:07] a problem here. So, I'll ignore that. [77:08] But, um any I was saying something I [77:11] don't remember what it was. We'll go [77:13] ahead and we will create our PI file. [77:14] And now we're actually creating agent. [77:16] I'll just zoom out so I can have a way [77:17] of dragging this bar up and down. And [77:20] Meta seems to have an example for us. [77:21] So, we will go in here and take a look. [77:24] Um so, I do believe that they have a [77:26] starting point for us. [77:28] And I mean, this looks pretty good. So, [77:30] we'll go ahead and copy it. [77:35] Okay, we're just really validating [77:36] making sure these things actually work, [77:38] right? Because sometimes you do use [77:40] these things and they don't tend to work [77:42] out. But so far, everything's been on on [77:44] point, on key here. And uh I'll just go [77:47] ahead and grab this top here. And we'll [77:50] go to the top here. Paste that in. And [77:51] I'll just take out anthropic as we're [77:53] not using that right now. And OS is [77:55] already defined here. [77:57] Put that at the top like that. And put [77:59] that at the top like that. There we go. [78:01] Okay. And so we have an example. Add [78:03] input validation to sign up handler for [78:05] run tests. And we'll just switch this [78:08] over to 1.2. [78:10] Current working directory is actually [78:11] the directory we're in right now. Um so [78:15] I would probably want it to be here. I'm [78:18] going to go ahead and just make a new [78:19] folder here as an example. So we'll just [78:21] say um [78:23] well we'll make a folder in here. So, [78:24] just say code codebase [78:28] and uh I'm going to go ahead and CD into [78:32] No, I don't need to CD into there, but [78:33] I'm going to go ahead and get the full [78:35] address here. So, just say pwd on my [78:37] machine. [78:38] And I'm going to go ahead and grab this. [78:40] And above, I'm just going to say uh [78:43] path, [78:45] right? Like this. [78:47] And this will be codebase. [78:50] and they'll just this is changing the [78:52] current working directory. your coding [78:54] agent make minimal well- tested changes [78:56] readedit whatever bypass permissions um [79:00] that means that it just it won't ask you [79:01] for permissions which you know I [79:03] understand if you're trying to get [79:03] things done so that's totally fine um [79:06] though you have to be a bit careful with [79:07] that so down below we're mapping them to [79:09] different providers so obviously all we [79:12] have is muse spark so we'll go ahead and [79:14] do that it is not model key we call it [79:17] um [79:18] meta meta API key which is fine we're [79:21] getting readedit whatever your your [79:24] coding agent. Well, minimal tested [79:26] changes. Add input validation to the [79:28] signup handler and run the test. So, um [79:30] we don't have a signup handler here. So, [79:32] I'm going to have to go ahead and just [79:33] make some code. So, I'm going to go here [79:36] and [79:38] uh make a new file. And you know, I like [79:40] my Ruby. So, I'm going to go ahead and [79:42] say game.ruby. [79:45] And [79:48] uh we'll define a very simple game. um [79:52] create tick tac toe [79:56] uh for the uh command line. [80:00] And I'm going to go over to here back [80:01] over to our agent. [80:05] And I'm just going to put my prompt up [80:06] here just so it's a little bit easier [80:07] for me to work with. [80:15] Please code a very simple [80:18] tic-tactoe [80:20] game [80:25] in the game.rb [80:27] file uh using Ruby [80:31] for the command line. [80:34] Okay, so we'll go ahead and do that. And [80:37] it seems like this should be fine. So [80:39] we'll go ahead and see if that works. [80:42] But anyway, there's agent SDKs that are [80:44] like coding agents like what we're using [80:46] right now. And then there's agent SDKs [80:48] for like building actual agents. Um, so [80:51] we might need to do a test with lang [80:52] chain for fun, but uh I'm going to go [80:54] ahead and see if this works. We'll go [80:56] ahead and give that a run. [80:58] And yeah, we'll go ahead and just do [81:01] Python 3 agent.py and we'll chill out [81:03] here and see if it comes back and [81:05] actually codes that game. All right, so [81:07] let's take a look here. Uh it didn't [81:08] take very long and uh I want to see if [81:10] the code is actually here. So we do have [81:12] a code output and so you can see it's [81:14] very simple to work with existing agent [81:16] SDKs. I don't feel that we really used [81:18] an agent SDK as in to make an agent uh [81:21] as this is obviously a coding agent. You [81:23] can use the framework for for that but I [81:26] wouldn't recommend it as this is a [81:27] specialized SDK. Um so we will go and [81:30] make another one and maybe use lang [81:32] chain. Okay. [81:38] Okay. So, u what I want to do in this [81:40] video is take a look at using uh the [81:42] meta API to uh use a framework like Lang [81:45] Chain. So, let's go ahead and look up [81:48] Lang Chain. It's been a while since I've [81:49] used it. I generally don't like using it [81:52] to be honest. It's probably gotten a lot [81:53] better since last time, but I just want [81:55] something that we can plug this into. [81:57] Right off the bat, it's trying to make [81:58] me sign up for something. I don't want [81:59] to do that. So, we'll go over here and [82:02] type in Langchain agents and see if we [82:04] can get to those uh docs with with less [82:07] trouble here. Um, and see if we can get [82:10] installed very very quickly. I want to [82:12] use Python if I'm going to keep using [82:15] something. Uh, do we have to use [82:17] TypeScript? I don't care. We can use it, [82:19] but it would have been nice if we could [82:22] use Python. I don't see the option. So, [82:24] I guess we're using TypeScript here [82:25] today, folks. That's okay. [82:28] Personally, I prefer Ruby, but that is [82:31] whatever. So, we'll just say uh we'll [82:33] just say lang chain uh lang chain SDK [82:38] and we will go over to here and make [82:42] ourselves a [82:44] new file. This is a little bit different [82:46] because we're using TypeScript. So, what [82:48] I'm going to do is go ahead and copy [82:50] this line and I'm going to go back for a [82:52] second and we're going to go ahead and [82:55] put that line in there. And so, that [82:56] should create a package.json JSON here [82:58] in this directory. While we're here, I'm [83:00] just going to add a dot getit ignore as [83:03] I'm going to want to ignore the node [83:05] modules. I'm not committing this to a [83:08] repo, at least not yet. I can later. Um, [83:11] but right now, I'm not doing that. So, [83:13] we're going to go ahead and ignore the [83:15] node modules directory in this directory [83:17] here. Um, and so now we can kind of get [83:20] set up. This should be straightforward. [83:22] Lang provides integrations to hundreds [83:24] of LMS. So the question is like okay can [83:28] you do that but can you do it for meta? [83:30] I mean like it's open AAI compatible so [83:32] we probably can do it through there. I'm [83:34] just curious if they have a first class [83:35] one here. So that's what we're going to [83:37] look for here is the meta one and I'm [83:40] searching. I don't see an option for it. [83:41] All providers are over here. Um [83:46] so I'm looking for meta meta meta meta [83:49] meta meta. [83:53] And we'll just type it here. meta [83:56] have got metadata [83:58] and so that's fine. We'll just have to [83:59] use the OpenAI integration I believe in [84:02] order for us to do that. So we'll go [84:04] over to here and we'll look at the [84:05] OpenAI integration. [84:07] And [84:09] it does say that uh Huh. [84:14] So this is where I'm not totally sure. [84:17] So I mean we could guess, but I'm just [84:19] going to go ahead and see if Meta can [84:20] help us out here. So I'm going to go [84:22] say, you know, I want to use lang chain [84:25] with um metamuse model uh for [84:30] TypeScript. [84:32] Can you give me a uh skeleton start uh [84:35] skeleton file to get started for [84:38] configuration? [84:40] Okay. And so what I'm expecting to do is [84:42] to pick up OpenAI and show that as the [84:44] installation method. I might also have [84:45] to ask it for the um the notes here uh [84:48] like the what to install as in the [84:50] packages. But we'll give it a second [84:52] here and see what it produces for code. [84:55] So we have output here and I'm going to [84:56] go all the way to the top and yeah [84:59] that's what it's doing. It's bringing in [85:01] um OpenAI which is what I expected. Uh [85:03] and we are using MV so that's good. So [85:05] we'll go ahead and while I'm doing this [85:07] I'm going to save the instructions as [85:09] you never know when you want these [85:10] instructions later. [85:12] Now, we could again be using the um [85:16] uh like Muse code, but I just I don't [85:18] feel like doing that right now at this [85:19] stage. [85:22] And so, we're going to need av here, [85:25] which says Muse API key, Muse Mo um Muse [85:29] model. Okay, sure. Fine. Fine. That [85:31] sounds fine to me. [85:34] And we do have [85:36] here that key. [85:39] So, we go ahead and paste that in. [85:43] Uh the model we're going to use is muse [85:45] spark [85:47] 1.2 [85:50] muse base URL. [85:56] Okay. So we'll go ahead and we have all [85:57] those three in there. Now we can go [85:58] ahead and create ourselves uh our [86:01] initial file. I'm going to go ahead and [86:02] just call this app.ts. [86:05] They're like call it muse. I'm not [86:06] calling it muse. I'm calling it app [86:09] or bane. you know, that's what we [86:10] normally call these things. We'll go [86:11] ahead and paste that in. And we have a [86:14] bunch of stuff. I haven't uh [86:17] uh set any of the installation yet. So, [86:19] that might be a little bit what we have [86:21] to do there. So, we'll go all the way to [86:23] our readme here, grab these three lines, [86:26] and get initialize. Though, I think it [86:28] would initially create that file. Maybe [86:29] it wouldn't. [86:31] Usually creates a main, right? [86:34] And so [86:37] it looks like we are good. And we have a [86:39] TS config now. Type scripts in here. [86:43] So that is good. [86:45] Uh but we'll go back over to here. Oh, [86:47] so we have an index and the muse. Okay. [86:49] So that's what it's they're doing. All [86:50] right. Well, we will just do what they [86:52] want to do. [86:54] It's not that big of a deal. We'll [86:56] follow along with what they want. [86:59] And I'm going to go ahead here and make [87:01] an index. [87:04] index ts. Yes, index ts. Interesting. [87:07] They're like include the js file, but I [87:10] guess it gets compiled out. So that's [87:12] probably why. [87:14] And so here it's trying to import those [87:16] there. Package. JSON. Yeah, that's what [87:19] we're going to need to get it started. [87:22] So in here, we will grab in [87:25] this top part. [87:27] So that should get started. And then we [87:31] just need to do npm rundev and see if [87:33] that works. Now it's having some [87:36] resolution issues [87:39] because it cannot find the lang chain [87:42] SDK. Import from cannot find module. Oh, [87:46] we're not in the source directory, [87:47] right? So we'll go ahead here and just [87:49] make a source folder. [87:52] And we'll drag these into their [87:54] appropriate location. That was probably [87:55] the problem. [87:58] And now it's having an issue with the [87:59] import. So does not have an export name [88:02] of chain example. I think it's because [88:03] it has to compile the code, right? So if [88:06] we go back over to our package JSON file [88:09] here, [88:10] we did mpm rum dev, but that doesn't [88:13] compile the code. But it it does do tsx, [88:17] right? [88:18] So [88:21] okay, you know, why do you import [88:26] a JS file if we are using TypeScript? [88:29] And so this might be just a failing on [88:31] that output here. [88:38] Give it a second. All right. So let's [88:41] take a look here and see what it says. [88:45] Um, [88:48] good catch. It's confusing, but it's [88:50] intentional for the ESM. If you use ESM, [88:52] you must use the GS file, which is fine. [88:54] This is correct. So it depends on what's [88:57] configured there. I mean, I'm following [88:58] the instructions they had. And so, if we [89:01] go here, I'm not seeing I see ES next. [89:07] Uh, it said is node next as opposed to [89:09] that. [89:13] This is the default. If you did, then [89:15] just remove the extension. [89:18] Uhhuh. Well, that's not what it actually [89:19] did. Then do this. [89:23] Okay. So, what we'll do is I'm going to [89:24] go ahead and just change that out. [89:28] as that might give me less problems to [89:30] be honest. [89:34] And we'll go back over to our source [89:36] here. [89:38] And [89:40] I would prefer [89:43] to do this [89:46] and then we'll just do muse model on [89:48] this. [89:51] This example is a little bit more [89:52] complex than I would have wanted to be [89:54] honest. Um, we just needed simple chat. [89:57] We didn't need the chain example. [89:59] But if we go back over to our muse.ts. [90:05] Oh, let's just save that. Maybe I never [90:06] saved the file. That could have been the [90:08] problem the entire time. [90:12] And so what I'm going to do is just undo [90:14] this for a second [90:17] because I'm wondering if I just didn't [90:18] save that file and that was my actual [90:20] problem here. [90:27] missing credentials. Please pass the API [90:29] key uh and other stuff. So now it's [90:31] saying there's no key. Um and so we do [90:34] have a key here, but maybe the key is [90:35] supposed to be something else. [90:41] Now what's interesting is that we could [90:43] have used our um I don't know why we [90:46] don't. In fact, I think I might do that [90:48] is that in our previous one, we actually [90:49] have the anthropic uh agent SDK which [90:51] would allow us to uh uh fix it. So, I [90:54] think actually I'm going to go ahead and [90:55] do that right now and just change the [90:56] repo. So, I'm going to go ahead here and [90:58] just switch this out to this. [91:00] And I'm going to just say in here [91:02] instead [91:04] the we say lang chain SDK [91:11] and in here I can just fix the prompt. [91:17] I am trying to get my uh example [91:21] of [91:22] lang chain uh working with meta muse [91:26] API. I have the uh I have av [91:32] with the um meta API key. [91:39] I think that's what we called it. [91:43] in this one here. [91:49] And [91:54] we have too much stuff open. So now I [91:56] can't see where I am. So just close a [91:57] few here. [92:04] Oh my goodness, there's so many. Let's [92:06] just close all these. I can't even see [92:07] what I'm doing here. There we go. Okay, [92:09] so um we'll go back over to our agent. [92:12] So, I'm trying to get my example of [92:14] Lchain working with Metamuse AI and I [92:17] have it with the meta API key, which I [92:18] do believe that I do. Um, but I'm [92:21] getting [92:31] I'm getting this error. [92:35] Okay. And so I need a way to do block uh [92:38] quote blocks. [92:40] So, let me see. Python uh block uh block [92:44] uh strings here. I'm just doing this off [92:47] screen here really quickly. It's [92:49] triples. It's triples [92:53] three like this. And then I can go all [92:56] the way down here and do this. [93:00] Um [93:02] I was using back ticks. It's just uh [93:04] three triple equals. There we go. Okay. [93:07] So we go all the way back to the top [93:08] here. [93:10] There we go. And so now we have that [93:12] prompt and we've changed its location. [93:15] And so I'm going to go back a step here [93:16] and then we'll go into our enthropic [93:18] agent SDK. And I'm just going to ask it [93:21] to fix that for me. Uh so we will do [93:24] Python 3 [93:26] agent agent.py. That's going to go ahead [93:29] and try to fix that other code there [93:33] as uh it's a bit of a headache. So we'll [93:34] give it a second. So after quite the uh [93:37] little wait there, it seemed because it [93:39] would have been doing iterative, it [93:40] apparently has fixed it. So I think what [93:42] it did was it not only um uh fixed it [93:45] and made sure that it worked. It's kind [93:46] of cool to to watch it uh do that, but [93:49] this would be no different than using um [93:52] uh what do you call it? Clawed code with [93:54] it. Um and so if we go back over to [93:56] here, I think it just solved all my [93:57] problems, whatever they were. Uh so [93:59] yeah, that's good. Like I don't know [94:01] exactly what it did, but it Oh, I know [94:03] it told us up here. Okay, so uh [94:05] credentials error your original message [94:07] muse API is missing. So it needs to [94:11] accept either this so we have the wrong [94:13] base URL and it fixed that. So even [94:16] though we originally used right we [94:19] originally used beta uh but within this [94:22] uh coding harness agent SDK it just it [94:25] had more time to loop around and fix its [94:27] problems. Uh so that's good to hear. [94:29] Let's go over and actually look at the [94:31] code now. Um, and I would like to run it [94:34] manually myself first. So, we will go [94:37] into [94:40] lang chain SDK and it's going to be npm [94:44] rundeendev I believe [94:48] and we'll have a little simple chat [94:49] here. We'll watch the results first and [94:51] then we will go look at the code. [94:54] How can I help you today? Okay, that's [94:56] fine. So, now we have chin example. [95:01] I don't know what it did as the chain [95:02] example, but maybe it just asked the [95:04] next question. [95:07] This one's supposed to be streaming. It [95:08] doesn't look like it's streaming, [95:09] though. [95:12] And uh the agent wouldn't know that it's [95:13] not streaming. It would be very hard for [95:15] it to detect the streaming information. [95:17] So, I would imagine that this entire [95:19] time it's supposed to be streaming. [95:21] Okay. And so, it didn't it didn't [95:23] actually stream the last example. But, [95:24] uh that's fine. We'll go over to here. [95:27] Uh I I don't care about trying to fix [95:28] the streaming. I'm just more interested [95:30] to see how this code works. And so I [95:32] don't really need three examples. One is [95:34] good enough for me. And we'll go over to [95:36] Muse here. And let's just take a look at [95:39] the implementation. So we have uh chat [95:42] open AI. We set the Spark model here. [95:46] Here it's defaulting to [95:49] Muse API key, Met API key or open AI [95:52] key, which is fine. And then we have the [95:56] uh the fallback over here 1024. As per [95:58] usual we are going to invoke our model [96:01] add our system message our human [96:03] message. [96:04] And so that is our simple chat. [96:10] Okay. Our chain doesn't look like a [96:12] whole lot uh whole lot of a difference. [96:13] It just looks like we are creating a [96:15] template for messages. [96:17] Okay, which is totally fine. [96:20] And then down below we have our [96:22] streaming example which didn't actually [96:24] stream but I'm sure we could fix that if [96:26] we wanted to. So there we go. So um we [96:31] are able to get meta working with [96:33] different coding harnesses and an agent [96:35] SDK. I'm sure we could utilize anything [96:37] else as long as it's open AI compatible. [96:38] You basically can slot Muse in there. [96:41] But uh now that we finished this we'll [96:42] move on to something else. Okay. [96:49] All right. So, I think it's time for us [96:51] to go ahead and use this with an actual [96:53] coding agent. Now, I do want to uh show [96:55] you how to use um uh their own coding [96:59] agent, but I think because there's [97:00] multiples, I would rather start with a [97:02] another one first and then we'll move [97:04] over to uh uh Muse Code. So, there is [97:07] Open Code, Cloud Code, Codeex, stuff [97:09] like that. We'll go over first to Cloud [97:11] Code. And so, I already have Cloud Code [97:12] installed on my machine. Uh and it's not [97:15] hard to install, but the idea is that we [97:16] need our API key. we're going to have to [97:18] export it or set it somewhere. Um, and [97:21] so it looks like we need to [97:24] configure all of this information. So, [97:26] what I'm going to do, I'm just looking [97:28] at this for a second because [97:31] we do have these set, but these are the [97:34] default keys, and this could be kind of [97:36] annoying for me if I um [97:39] have to load them in there. So, what I [97:42] think I'm going to do is make a new [97:43] start file here. here. So, I'm going to [97:44] go over to here and I'm going to make a [97:46] new file or new folder. And this will be [97:48] uh claude code launch. [97:51] And [97:54] uh where is this? And so we will make a [97:57] new script or actually I might just go [98:00] make this top level here. So I'll make a [98:03] new file called or new folder called [98:05] bin. [98:08] And we'll just say cloud code. [98:14] Okay. [98:17] And the idea here is that we have [98:20] this stuff here. [98:25] And so what I want to do, and I'm [98:28] probably just going to use my regular [98:30] Cloud Code account because I just want [98:31] this to to work here in two seconds. [98:36] And I'm just going to go ahead and just [98:37] type in Claude here real quick. [98:40] I'll just say, you know, I have a I have [98:42] a claude code [98:44] uh file. This is supposed to be a bin [98:47] file, and it I want it to uh load the [98:53] envir [98:55] load. [98:58] I'll just hardcode this. No, I'm gonna [99:00] actually just copy this over. [99:04] Paste. [99:06] So, I have a cloud code file. This is [99:08] supposed to this is supposed to launch [99:12] cloud co cloud code [99:15] configured for meta uh meta muse model. [99:19] Um I have in my meta API key. So, I want [99:25] you to write a bash script [99:29] that will extract [99:32] that will um inject the meta API key [99:36] into the correct place. And um we load [99:41] these envirs [99:48] and we start up cloud code. The only [99:52] thing is that if I do this, I'm not [99:53] going to be able to point it at other [99:55] directories. But um [99:59] I think that we could probably fix this [100:00] by, you know, I want you to sim link [100:06] the script into a global path pathway. [100:13] So I can do uh muse claude [100:17] claude code museclaude to start uh start [100:20] up with the config. Okay. So we'll go [100:23] ahead and uh have that do that there. [100:26] And so I'm just running this on opus 5 [100:28] right now is you know I don't I again we [100:31] could install meta and I'm sure it can [100:33] do this but I'm just trying to get this [100:35] to work really quickly. So [100:38] and I'll just come back here when the [100:39] script is ready. All right. So looks [100:41] like we have a wrapper. So it resolves [100:43] through its own sim link. Uh so it works [100:44] from there. [100:46] Uh so the question is like will it [100:48] contextually work anywhere? So um I [100:52] guess we could just check the the tech [100:54] tech.nuke. So I'm just going to cancel [100:55] out of this for a second. We'll go back [100:57] up a directory and we will go into [100:59] tech.nuke and I'm going to type in [101:02] museclaude and we'll see if it loads the [101:04] model. So it's saying yes, it wants this [101:06] location. Excellent. We'll accept that. [101:08] And we're now seeing it's using Muse [101:10] Spark 1.2. Cool. Um, I'm kind of curious [101:14] on the code. I guess it doesn't really [101:16] matter because it does work, but I'm [101:17] just curious what the code looks like. [101:19] Uh, [101:20] says Claude code. That was the original [101:22] one. Muse code. And so it actually named [101:23] it what it should be, which is fine. [101:25] We'll delete this one. And it works. I [101:28] don't know. It does something. Okay. So, [101:31] we'll go here. And I'm using Muse now [101:33] specifically Claude Code. And [101:37] I'm just trying to think here [101:40] what I would like to do to this file. [101:43] Can you refactor refactor this file into [101:46] several files? [101:48] Um uh so [101:52] uh I'll just say like you know can we [101:55] refactor out uh the CSS and JavaScript [101:59] into a assets JS and assets CSS and can [102:05] we add a a bin uh serve script that will [102:12] start up HTTP [102:15] uh server [102:17] which should be already installed via [102:20] node global. Um, [102:24] okay. And so we'll just ask it to do [102:25] that. Very simple task. I'm not sure [102:27] what level we're using. I guess it's on [102:29] high effort. And so that would probably [102:30] mask because you're masking to those [102:33] specific models underneath. Um, so it [102:35] would probably mask to that. So we have [102:37] bash uh denied by auto mode. Again, I'm [102:40] not sure. I've never seen that before. [102:41] So that's a new one. It says classifier [102:44] unavailable. So again, we're not using [102:45] the direct model. So there could be some [102:47] differences here. Um, but it is working. [102:50] So we'll give it a moment here and see [102:52] how it turns out. [102:54] All right. So it looks like we're back [102:56] with some code here. It has moved things [102:58] into assets. So we have CSS and [103:00] JavaScript. Um, was there any [103:01] JavaScript? There apparently is some [103:03] JavaScript. We'll go back over to our [103:05] index html. [103:07] And so we do have [103:10] some stuff going on here. I mean, it's [103:12] just our our usual code here. Um, which [103:14] is fine. [103:18] I'm just checking how clean the uh the [103:19] markup is. I don't like the fact that [103:21] there's uh inline styling. So um I am [103:24] seeing inline styling [103:28] uh in the HTML. [103:32] There should be none. [103:37] It should be using selectors. [103:40] Okay. So I mean there could be a case [103:42] where there's like a little bit [103:43] variations here. Yeah. Yeah. like the [103:45] width and stuff. I'll leave it alone for [103:46] now. But anyway, my point is is that we [103:48] have loaded um uh cloud code. And so we [103:51] have the process there. And you know, if [103:54] we want to load other ones like codecs, [103:56] we're going to have something similar. [103:58] So we're going to be using a [103:58] config.toml. [104:00] Okay, fine. Um if it's open code, it [104:03] looks like you just install it and use [104:05] it. I'm don't really care for open code [104:07] too much, so that's not really my [104:09] preference. Um, and so I think now that [104:11] we are we've at least touched this [104:12] cloud, anything else besides the Muse [104:14] code, we can go ahead and start you [104:16] working with Musecode. Okay. [104:22] Hey folks, we are back and in this video [104:23] I want us to start using Musecode. So [104:25] we'll go ahead and install it as it's a [104:27] single line install. Um, and I don't [104:30] think it matters where we run this, but [104:31] we'll go ahead and do that. And so it is [104:34] downloading now. This is a very minimal [104:37] um not in a bad way but like it's [104:39] minimal in the sense that uh there's not [104:41] so much configuration nonsense or [104:43] constant features being pushed out like [104:44] cloud code is. Um so that is very good [104:49] that at that regard. I'm just trying to [104:51] point that out that your experience is [104:52] going to be a little bit different. Um [104:54] we'll go ahead and just trust and [104:55] continue here. And it looks like I might [104:58] already be authenticated. If you're not, [105:00] it's very easy to fix. Uh I can see [105:02] right off the bat it's trying to load uh [105:04] bubble wrap which is a way of sandboxing [105:06] your environment has voice input. I have [105:09] never used voice input and I don't have [105:10] the means to do that on this computer so [105:12] I'm not going to be able to show that [105:13] off to you. Um but you know there is [105:17] basically all the usual suspects in [105:19] here. And so we will systematically go [105:21] through some of these features here. Uh [105:23] but I just want to show you that it does [105:24] work. Notice we have bunch of skills in [105:26] here that it's loaded from all over the [105:28] place, which I think is kind of cool [105:29] because um uh one issue that I have is [105:32] that if you use something like uh OpenAI [105:35] or Claude, uh you don't get those skill [105:37] those skills uh across the board. And so [105:40] for whatever reason, Musecode will bring [105:42] them all in, which is what I like. Um [105:45] but anyway, so we'll go over to here and [105:48] I just will do a demonstration of it [105:50] doing something here. And so, uh, I'm [105:53] going to ask it to build out the other [105:54] pages. So, um, can you please build out, [105:58] uh, the, uh, a few more key pages in our [106:04] technique example? Check the navigation, [106:08] uh, for that. Okay. And so, we'll go [106:10] ahead and hit enter, and we'll be back [106:12] in a moment when it is completed. Um, [106:15] and so, if there's any permissions, [106:17] we'll just accept them, but should be [106:19] pretty easy to work through. And just to [106:21] show you, it did break up the tasks and [106:22] it understands what it's supposed to be [106:24] doing. So, it's off to the races, doing [106:25] a good job here. Okay, I just wanted to [106:27] show those uh that task breakdown. And [106:30] here it is asking me to uh run [106:32] particular commands. I would just say [106:33] the only thing that I would watch out [106:35] for for abuse code as it is, at least at [106:37] the time of this creation, it is a newer [106:39] um uh coding harness and so sometimes it [106:43] likes to delete files and recreate them [106:45] and maybe you have things you don't want [106:47] to be destroyed. And so, um, I would [106:51] carefully watch that. Or you can update [106:52] your agents, your agents file to just [106:54] tell it to never, you know, just move [106:56] things. But for the most part, I think [106:58] they might have corrected those things [106:59] as that was in version 1.0 and now we're [107:01] on version 2.3. So, um, I think I I [107:04] mentioned it to somebody and uh, they it [107:07] might have been passed forward there. [107:08] But here, you're just it's constantly [107:10] asking to run these commands. I guess [107:12] it's just inspecting those pages [107:14] manually. Um, so it's outputed them and [107:16] they're just making sure that they [107:17] actually work through. I don't know why [107:20] it doesn't just look at the code here, [107:21] but that's totally fine. As you can see [107:23] that it's curling it and then looking at [107:24] the results. It has the files, right? [107:27] But it's treating it because it has a [107:28] web server. Right? Now, if this was um [107:31] uh, you know, server side compiled or [107:33] something, I can understand that, but it [107:34] doesn't necessarily need to do that. But [107:36] also, there's nothing wrong with it [107:37] working that way. So, I have no concern [107:39] about that. We'll go over to our ports. [107:41] It seems like it's running on 8788 as [107:44] it's running this uh Python script here. [107:47] Um I never told it to do that uh because [107:50] we have a bin script, but I mean like [107:52] it's fine. It's just using what tools it [107:53] has to test. But I think I would have [107:54] preferred it uh to just utilize what we [107:58] have here. But I do again I have Python [108:00] installed so it's a non-issue. So we'll [108:02] go over to here and we'll take a look. [108:03] And so we have this expanded. What did [108:05] it code? It probably will tell us here [108:08] in summary. So, we have the feed, the [108:11] docs, the submit, the topics, the top [108:12] 10, and then it add the navigations. [108:16] Love the look of this. I'm sorry, but I [108:17] just love I love uh PHP Nuke. Um because [108:22] it's it like I'm it's trash, but it's [108:24] not trash for me, right? So, we'll go [108:25] ahead and and give the feed here. And [108:28] so, [108:30] we have our community page [108:33] and our docs and our submit post [108:37] and our topics. I love this. This is [108:40] just so much fun for me. [108:44] And uh let's take a little look at the [108:45] mobile. I never did that earlier, so I'm [108:47] kind of curious. [108:49] Oh, it's perfect. It's perfect. [108:54] Okay. But obviously like a lot of these [108:55] things are not linked up. So, uh we'd [108:57] have to do more more work with it. Um I [108:59] kind of feel because I have all those [109:00] credits, I kind of feel I want to make [109:02] this for real. But um what we'll do for [109:04] now is we'll let this sleep and uh we'll [109:08] start looking at some capabilities of [109:10] what we can do. Now I've covered this [109:12] stuff a lot of times in the past but we [109:14] should give the treatment here and make [109:15] sure we go through all the features or [109:17] is most of the features that we should [109:18] care about when using Muse code. Um but [109:22] yeah that's the start of this. Okay. [109:28] All right folks in this video I want to [109:30] show you uh the agents.md file. Also, [109:32] agents.mmd is just like your claw.md or [109:36] uh you know gemini.md. [109:38] Um and so meta has chosen to go with the [109:42] generic agents.mmd and we can create it [109:44] manually but I figured since um we [109:47] already have kind of a project set up [109:49] here. I'm going to go ahead and do uh [109:51] muse init. And what that will do is [109:52] analyze this repo and we'll create us an [109:55] agent.mmd. And wow, was that fast. So, [109:58] Musecode reads this file as a project [110:00] rules and common commands. No standard [110:04] build or tests detected yet and it has [110:06] the layout. So, I don't know why that [110:08] was so darn fast. I know that for larger [110:10] ones it's it's taking a bit more time. [110:12] Um, but that is totally fine. So, let's [110:15] go in here and see what we could do for [110:18] our uh agents.mmd. [110:20] Um, so it gives it some basic [110:21] information. So, the question is like [110:23] what are we going to benefit from adding [110:25] to our agents.md? Um, I think it doesn't [110:29] understand some of our goals or our our [110:31] information there. Um, so that would [110:33] probably be something that I would want [110:35] it to know is the the context of this [110:37] project. Um, and so we'll go back over [110:40] to here and [110:42] uh, so we don't need that because we [110:44] know that's what it does. [110:47] We don't need this. [110:50] So it didn't really do it didn't really [110:51] do a whole lot for us. Okay. Um, so what [110:54] we'll do is we'll just go I guess go [110:55] ahead and write whatever we want in [110:56] here. So we'll just give it um business [110:59] use case. [111:03] So um TechDoc Nuke is a tech community [111:09] uh that is inspired by the experience [111:16] of um Nuke PHP forms [111:20] found in in the uh 2000s. [111:25] Okay. So the idea is that we right so [111:29] lots of uh plugins [111:32] and uh [111:36] different points of interaction [111:40] with a busy uh three-page layout. [111:45] However, it's a modern website, so it [111:49] works in mobile and will [111:54] bring forward [111:58] other uh [112:00] modernities. Uh, is that a word for like [112:03] I'm think I'm making up words here, but [112:04] is that a word for like modern stuff? [112:08] Modernities. [112:10] Or am I just making things up? [112:14] Oh, no. Oh, it is a word. Wow, look at [112:16] me. I'm fancy. So, um, we have that [112:20] there. Um, [112:23] and I'm just going to go here and just [112:24] say like, uh, target audience, [112:28] tech professionals, [112:32] uh, that work. [112:35] So we have like DevOps, [112:37] security, [112:39] uh AI, [112:42] um cloud computing, [112:45] data, things like that, broad broad [112:47] categories, uh you know, stuff like [112:49] this. Okay. So, [112:59] and I think that is a good start there. [113:03] Um, so now it has context every time it [113:05] starts up. It knows what it's trying to [113:07] work in. One thing that I would like is [113:09] that if it can load um conditionally [113:11] other files is that's a very common [113:13] thing that we would like to do when [113:15] we're working with uh this. So let me [113:18] see if it has that capabilities. And it [113:21] doesn't seem like it can do that. Okay. [113:24] So, um [113:27] I the only thing I'm thinking is like [113:28] there could be things like you you might [113:30] specify um [113:33] you know like the style that we want to [113:35] adhere maybe some like guidelines that I [113:37] I would provide here. So stack [113:39] guidelines, stack guidelines [113:42] or technical goals, right? We want uh [113:46] the uh we want to [113:49] we want the stack [113:52] to be the lowest [113:55] cost possible operating on a single uh [114:00] VM. Okay. [114:04] So, and so that's something that we [114:06] probably would want to explore [114:07] separately here, but I we can come back [114:10] and expand that later, but that's our [114:11] agent. MD. And if we were to uh you know [114:14] do something with this, I'm just trying [114:15] to think of something that I would like [114:16] to do uh that we have not done here. [114:20] Go here and take a look at our site [114:22] here. Um at this stage, [114:30] yeah, I'm not sure. I don't think I want [114:31] to um do anything right now. So again, [114:34] just showing you the agents.mmd and [114:36] that's what it's called. Um but I really [114:38] can't think of anything that I I want to [114:40] add here. So we will stop it here. Okay. [114:46] All right. Let's take a look at where we [114:48] can set our configuration for uh Muse. [114:50] So Muse allows us to have a docig file [114:53] in our home directory. So we'll go ahead [114:55] here and I'm going to cd into my home [114:58] directory here. And I'm going to make a [114:59] new one called config if it doesn't [115:01] already exist. Whoops. That is not a [115:05] proper file name here. I'll just remove [115:08] that one as that makes no sense. But I'm [115:10] going to it looks I already have one [115:12] which is fine. So we'll go in here and [115:13] it looks like we have muse already which [115:14] is great. And so in here we have our [115:16] settings.json. We can cat out our o as [115:20] I'm going to get rid of that access [115:21] token anyway. So it's not a big deal. [115:23] Actually this is a different access [115:24] token so I will have to remove both of [115:26] them. But you can see that we have our [115:28] credential information here. That's [115:29] where it's being stored. And if we uh [115:32] cat out our settings page here, we can [115:35] see some of our settings for uh that uh [115:38] and we can cat out our [115:41] um trust file here. Trust. And so we can [115:44] see what folders are trusted. So I'm [115:46] going to go ahead here and just for fun, [115:47] we're going to update the settings.json [115:50] file here as we should be able to set [115:53] something here. Let's see if there are [115:54] any options that we can configure. And [115:57] what I'm looking for [116:00] here is a simple simple settings. Let me [116:03] find one. Okay. So, here it's suggesting [116:06] um Google suggesting some settings. I [116:08] couldn't find where their settings were. [116:10] And [116:13] I mean, there's nothing for me to do, [116:16] but I guess we could just set these here [116:18] for now. [116:21] Okay. [116:24] And so now it's set to extra high [116:27] reasoning. If we go to this one, it is [116:31] set to something. I don't know. We'll go [116:33] ahead and restart this. [116:36] And from this, you know, I think they [116:38] should have probably put on what level [116:40] of reasoning is at at uh whatever it [116:43] was. So um we'll just take a look at the [116:45] effort. Oh, there we go. Now it's [116:47] telling us it's extra high. Okay. And so [116:50] we'll just stop this again. Start it [116:51] again. Oh, I think it's it would [116:53] normally show. It's just because bubble [116:54] wrap is not showing installed here, [116:56] which we'll do in a separate step. So, [116:58] yeah, you notice as soon as we started [116:59] typing it, now we can see what's loaded [117:01] and what's in here. So, we're on extra [117:03] high. Uh, and so we'll just make sure [117:04] that this actually works. So, maybe we [117:06] always want to come in as medium. [117:08] Okay, we'll just stop this. Start it [117:10] again. [117:13] And I'll just type that. And so, now we [117:15] are on medium. And there you go. [117:20] All [117:23] right, let's just take a look at how we [117:25] can set our reasoning effort. We [117:26] actually just saw in the last video as [117:28] we used it as a demonstration, but I [117:30] just want to show that if you want to [117:32] override this medium here, we can go [117:33] ahead and type in muse and we'll say [117:35] reasoning effort and I'll go x high [117:37] here. And so now uh I'll just type [117:40] forward slash. We can see that it [117:41] started with extra high. If I was to [117:42] start it normally, you can see I'm [117:45] hitting forward slash here that it's now [117:47] at medium. And of course, we already did [117:49] this, but we can go here and say low. [117:51] And so now we have low. Uh, what's [117:53] interesting is that it just changed that [117:55] default there. Oh, does it actually [117:57] change it each time? Effort X high. Oh, [118:02] look at that. That's interesting. You'd [118:05] think that like if you set it that that [118:06] wouldn't overwrite it. Um, and that this [118:09] would only be per session, but I guess [118:11] that's the choice that they made there [118:12] and that's just how it is. Um, so there [118:15] you go. [118:21] All right. So, we just ran the resume [118:23] command. I'll just show that again so we [118:26] can get back to the same spot. So, we'll [118:27] type in claude. Oh, I'm typing in [118:29] claude. We'll type in muse. I'm sorry [118:31] about that. And we will go ahead and [118:33] type in resume and we'll go with last so [118:35] that we don't have to click in again [118:37] here. But what I want to see is that we [118:38] can compact the history here which will [118:41] uh save our our size here. Sometimes [118:43] they they give us an option to see how [118:46] large our current context is. I'm trying [118:48] to see if that command is there. Show a [118:51] recap of the current session. Actually, [118:53] I kind of wouldn't mind doing that [118:54] first. Let's do a recap. [118:57] And let's say we just forgot what was [118:59] going on here. And so we it's really [119:00] long. We don't remember. [119:03] Delivering the technique site with five [119:05] key pages wired to navigation. So that [119:08] was nice. Um, but what I'm looking for [119:11] in here is a way to see if we can see [119:13] how much like how size how large our [119:16] context is. And so I'm going here and it [119:21] probably is status. There we go. And so [119:24] we have extra high uh what we're set to [119:28] our usage tokens, our context which is [119:30] 93% right now, how many turns we've had [119:33] so far. And so that is great. So let's [119:35] go ahead and compact that so that um [119:38] we're going to [119:41] get fewer information. No compact [119:43] compatible compactable run is available. [119:46] Uh well I mean if we go back and resume [119:51] right can we compact that? Nothing to [119:55] summarize. Okay. So I guess it's not [119:57] there's not enough going on there for [119:59] anything to compact. So let me give it a [120:01] large job for it to do. um so that we [120:04] have something to compact. Okay. So, [120:07] what I'm going to do is I'm going to go [120:08] ahead here and make a new doc. And um I [120:12] just want to get the base functionality [120:13] implemented here. And so I'm going to go [120:15] ahead here and just say skeleton uh [120:18] backend app. And so I want the most cost [120:21] effective solution. I think it's going [120:23] to probably be a golang application. So [120:25] skeleton backend. I like doing this when [120:28] I'm making plans. So just say name [120:30] skeleton backend [120:32] and this will be uh description [120:36] um and we'll say to create the initial [120:40] uh to create the initial MVP [120:43] uh background uh MVP backend application [120:48] that will run our uh tech form. Okay. [120:52] And so I'm going to go over to here and [120:55] uh specify so scope or business use [120:58] case. [121:00] We are building a uh social plat social [121:03] community platform driven around uh [121:06] around forms [121:09] around forums like PHP Nuke. [121:14] Uh so we are building [121:18] similar to PHP Nuke. [121:22] Um [121:26] we want this [121:28] we want the implementation [121:31] to be very coste effective. Um and so [121:37] we should choose a a stack [121:41] that is known for being [121:44] uh extremely optimal. [121:53] Okay. And so maybe we don't have to be [121:55] so descriptive here, but like uh tech [121:57] choices [121:59] uh tech stack [122:02] golang [122:04] um [122:10] API only. [122:13] Okay. Uh runs on a single VM delivered [122:18] uh delivering a single [122:21] um [122:23] delivering a single binary. [122:27] Actually, that's not even sure how that [122:28] would work because what would you use [122:29] for that? I don't even know. I'm going [122:31] to go over to here to the meta API [122:33] really quickly. So, we'll say metad API. [122:36] I think I actually already have it open [122:37] as a tab right here. [122:41] Okay. So, you know, I want to use Golang [122:43] to build my uh my form, but I need to [122:48] choose a uh I need but I would think we [122:53] need a web server that will load balance [122:58] multiple [123:00] instances [123:02] like Ruby Puma. [123:04] What are we using for Go? [123:07] because I don't actually don't know [123:10] for that. [123:18] And a few other things here that I would [123:19] like is I would like to have runs on a [123:21] single VM delivering delivering a single [123:26] uh it should um in place deployments. [123:31] So like SSH will SSH in. [123:37] Similar to Capistrano, [123:44] Capistrano was very popular because it [123:46] had very fast deploy. So very fast [123:49] deploys [123:50] uh zero downtime, [123:54] very fast roll backs, [123:57] right? Uh database lives on on the same [124:01] server. [124:03] I don't know if SQLite would be a bad [124:05] choice. [124:07] because it's not super performant, but I [124:11] mean like it's probably fine. [124:16] So, let's go up to here for a second. [124:18] And so, [124:20] and Go doesn't need it. Go routines are [124:21] extremely lightweight. Uh, so they're [124:24] packaged already. Each request is a Go [124:26] routine and gets to your own code [124:28] compiled by a single static site. So, in [124:30] Go, you just do guil go build your web [124:32] server. It uses all CPU cores by [124:34] default. You only run multiple instances [124:36] for high availability or horizontal [124:37] scaling, not to handle concurrency on uh [124:40] one machine like Ruby. Oh, that's good. [124:42] Um, [124:44] and so you run multiple copies of your [124:46] binary. [124:48] Do we even need that? We probably just [124:50] need one, right? [124:58] What is the what is the use of the form [125:00] itself? You don't need a framework like [125:01] that. A router is enough uh for a form. [125:10] Uhuh. [125:12] So, there's that. We might want to make [125:14] a choice in terms of what we might want [125:16] to use. I I know I know Jyn, but I don't [125:18] know echo. And like I said, I mean like [125:20] I know the name of it. So, we'll go here [125:22] and say echo golang. [125:29] Yeah, let's use echo for fun. It says [125:31] it's minimal. [125:33] And so we'll go back over to here and [125:36] we'll say uh [125:39] echo. [125:44] If it's going to be a small community, [125:49] uh let's say [125:51] like really how much can you really have [125:53] active? Let's say like 10,000. [125:55] Even even if it was like let's say [126:00] 10k active at max. Can we do that with [126:05] Esco Light [126:15] because I'm not sure what the limits [126:16] like I I know Postgress can do it but [126:17] I'm not sure what SQLite can do. So I'm [126:20] just trying to push the limits of a [126:22] smaller production site. But I think [126:24] what I will do is I'll just use escolite [126:26] for now. So database lives on the same [126:27] server. So, SQLite Escolite 3 um [126:32] authentication lives on the server. [126:41] So, it says yes 10k is the perfect size. [126:44] It seems to think it for form you don't [126:45] even need postgress at the start and [126:46] personally I that's what I prefer is [126:48] just to have it. So, how handles [126:50] concurrency default SQL light is bad for [126:52] web but wall mode fixes it. Under wall [126:55] mode, you should be able to default any [126:56] production deployment concurrent uh [126:58] never blocks. [127:05] Okay. So, [127:08] yeah, that's true. This what you must do [127:10] for go delete default journal mode [127:16] will give you the database is locked [127:17] instance instantly. So, [127:23] and so I'm going to go here and just say [127:24] like [127:26] make sure [127:29] we need to ensure we can handle 10k [127:33] users uh not not active at all [127:38] all at the same time. [127:40] So like eg, you know, journal mode [127:46] equals wall, [127:48] things like that. And obviously it said [127:50] that up there, so I feel like it could [127:51] figure that out. Um, so we have that [127:54] database lives on the same I'm doing [127:56] this all just to show you compact by the [127:58] way folks. So [128:01] um [128:03] Golang API only in place deployment [128:08] D this is the same server authentication [128:10] lib server. So the the only thing like I [128:13] care about is I want raw SQL raw SQL um [128:17] saved in a SQL directory using a [128:20] template language uh so I can easily [128:24] uh read [128:28] read queries [128:31] use some kind of active record. [128:35] Yeah, use some kind of uh [128:40] active record like structure but uh but [128:43] only for uh saving and uh [128:49] and very basic queries [128:54] find by like find by ID not for uh any [128:59] complex queries [129:01] use SQL instead. [129:05] and skip [129:08] the OM. Okay. Um, [129:11] what other options do I want? And so now [129:14] we're just talking about like feature [129:15] set [129:18] MVP feature set. [129:22] Um, create an account with [129:27] uh display name [129:34] avatar upload, [129:38] email, password. [129:41] Okay. And then um maybe like description [129:46] or profile [129:48] description. Um, we'll just say disk [129:51] here [129:53] and you know like forms, [129:56] threads, posts, [130:00] comments, right? Pretty common stuff. [130:08] So like I should be able to go let's [130:09] take a look at the actual application as [130:11] I've already forgotten what it does [130:15] here. We'll go over to here. So the most [130:19] basic thing is probably [130:25] we have post type, topic type, title, [130:29] tags, [130:32] repo if there is a repo, body markup [130:34] which has markdown supported, toggle [130:35] preview, [130:37] topics. [130:39] So I guess there's no like there's no [130:40] forms per se. I mean generally you would [130:42] have um [130:44] forms within topics [130:48] but here if we just have topics and then [130:50] posts. [130:54] So [130:56] that I think is one issue that we have [130:58] in our design [131:01] that is an issue here. [131:04] Okay. And so what I want to do here is [131:08] is just think about this for a second [131:10] because this isn't fully um realized. [131:16] I mean should you have [131:19] forms here? [131:23] H [131:26] every stack tribe has its own feed vault [131:29] and chat. Says feed vault and chat. pick [131:32] your poison or run random stack [131:34] roulette. [131:36] And so I mean there's some kind of [131:38] interesting things here, but like um you [131:41] know we'll probably have to reverse [131:42] engineer some of these features here. [131:44] And so I'm going to go ahead and save [131:46] that [131:49] front end back uh you know [131:53] and so here we have a problem which is [131:55] like open questions [131:57] open questions uh we have a design [132:03] uh a front-end design [132:05] but we haven't uh extracted out [132:09] the um [132:12] uh minimal structure for creating posts. [132:16] Uh, and we don't have a design for those [132:20] pages. [132:21] So, maybe that's what we should do. [132:23] First, I'm going to go here and just say [132:25] like um in our design uh can we mockup [132:31] for our page, for our website. the uh [132:35] clicking into [132:38] a topic and seeing posts and clicking [132:42] into a post and seeing [132:45] what a post looks like [132:48] because if we if we can't see that it's [132:50] going to be hard to map this backwards [132:51] and so we'll give that a moment. Okay. [132:54] All right. So, we should have a little [132:55] bit more design here. We'll go ahead and [132:57] give it a nice refresh. And I just want [132:59] to see if I can click into any of these [133:02] as I would suspect [133:05] Okay. Are we clicking to one? Like I [133:07] think they all go to the same place. [133:08] They probably should, right? Oh, no. [133:10] They made them all. [133:13] That's a problem. I mean, it's not [133:14] really a problem, but like if if we're [133:16] worried about consumption in terms of [133:18] our our credits, then I guess so. Like [133:20] again, I have a lot of credits here, so [133:22] I'm less concerned. But I guess I'll [133:24] take a look at my usage here uh really [133:26] quick and just see what I have. One [133:28] second. No, says I still have quite a [133:30] bit. So, there's a non-issue here. But [133:32] anyway, um [133:36] it went a little bit overboard, but I [133:38] mean like it's nice to see that. Wow. It [133:40] really it really didn't have to do all [133:43] that. [133:45] Look at that. Wow. [133:47] Okay. [133:49] Uh-huh. [133:53] Can you believe it? It like hooked up [133:54] everything. [133:56] Did it hook up this as well? [133:59] Are you serious? Okay. Well, anyway, [134:03] I mean, it's cool that it works. The [134:04] only thing it didn't do is pageionation, [134:05] which is fine. But, you know, my [134:07] question is like from a design [134:08] standpoint, is this how it actually want [134:09] to work as there's no form level stuff? [134:12] Um, so I'm like imagining I'm in here [134:14] and it's like you want to go take a look [134:15] at Rust. The only problem is that um [134:18] like yeah, topics are interesting, but [134:20] there's no top level feed. [134:23] Oh, I guess it just be the feed here, [134:24] right? [134:27] And this would be across all the [134:28] categories. Okay, [134:32] sure. I guess that's fine. Um, I mean, I [134:37] kind of like traditional forums where [134:39] you have some strange groups, but I [134:42] guess the thing is that if you have [134:44] topics there, you could put anything you [134:46] want in a topic, right? So, there's [134:48] necessarily no reason to group things [134:50] together unless you're thinking, well, [134:52] okay, I want to group all cloud things [134:54] under under stuff. And so I would say [134:57] the only that's missing within our [134:59] category. So you know we we have topics [135:02] but we don't have forms. So you know you [135:05] might have [135:07] cloud as a as a form and under it you [135:12] would have uh [135:15] well you would know would you do that [135:16] because like if you had programming [135:18] uh you know we have Rust and Go and [135:20] other very specific languages but like [135:23] do we even like in this time and day and [135:25] age do we even need to have specialized [135:29] language groups? Does anyone actually [135:31] really care about that or is that more a [135:33] uh a flag within something? Right? [135:35] Because you do have tags, right? And so [135:39] that would make more sense like instead [135:41] of rust and go, you have programming as [135:44] a as a tag. And so there is a bit of a [135:47] challenge here to think about the level [135:50] of of granular detail. Um, but we [135:53] definitely want we definitely want to [135:54] have topics and topics should have their [135:57] own uh tags within them, [136:00] right? And so, yeah, there's nothing in [136:03] that one. We'll go back to this top [136:04] level here. [136:07] And so, yeah, we'll click into this one [136:09] here, [136:14] which is fine. I'm not seeing avatars in [136:16] here at all. So, I would just say a few [136:19] things in here. Um, [136:26] you know, topics should have uh tags [136:29] within them. [136:31] So, instead of like Rust, [136:34] you know, uh we'd have a programming [136:38] topic and you could tag something [136:43] as Rust. [136:46] uh you know and um [136:48] uh I don't see avatars. [136:53] Okay. So, we'll go ahead and we'll give [136:55] that feedback there and let it go [136:56] through that for a moment. But, you [136:58] know, that's one thing I think that was [136:59] is missing. And really think of that [137:01] structure is hard. But again, we're not [137:02] doing this for for real. So, um what I [137:05] will do here is is after this we'll let [137:07] it uh go ahead and try to make that [137:09] structure. All right. So, we have a [137:11] little bit going on here where it has [137:13] produced a bit more. I'm just going to [137:14] go take a look and see what we have. And [137:16] uh now we do have an avatar. It's really [137:20] small. I'm not a big fan of it, but I [137:22] mean it is a an okay start for this as [137:26] we we can't really click on profiles, [137:28] but I guess you kind of get a a little [137:30] sense of it. [137:32] Yeah, you don't really you can't really [137:34] see anything. Okay. Yeah. Well, anyway, [137:36] that's fine. Um I think that we have [137:37] enough to compact and so I think I will [137:40] stop this video here and compact in the [137:41] next one. Okay. [137:47] Okay, so now that we have more stuff [137:49] here, I want to go ahead here and take a [137:51] look at status uh in this current [137:53] conversation that we have. And notice [137:54] that we have 93% left. So, it's really [137:57] hard for me to use up uh that there. But [138:00] we'll go ahead and run compact, which is [138:01] what I wanted to show you earlier. And [138:04] we'll see what we get [138:06] as it should reduce that size. 49 turns. [138:12] uh 68.7 [138:15] uh use so far and now it's compacted. So [138:18] we'll check our status again and yeah [138:21] it's greatly reduced. Now if we want to [138:23] clear that out we can go ahead and type [138:24] in clear. There you go. Now it's been [138:27] cleared. So there we go. There's your [138:28] compact. There's your clear. [138:35] All right. So in this video we're going [138:36] to take a look at the goal feature. Now [138:38] this is something that I'm never exactly [138:39] sure what to uh do. So I just fed it my [138:42] skeleton back in into chatbt and I asked [138:44] it like what I could do for that. So it [138:46] says build initial deployable MVP uh [138:49] supporting the accounts or whatever. So [138:51] it would not put these into the goals. [138:52] So the idea here is that um [138:56] uh I guess it just drives it uh drives [138:58] it and keeps checking to make sure [138:59] that's the goal. But here so for [139:01] document goal get a functioning backend [139:03] constraint scope open questions. [139:07] Um, [139:09] so it's actually a good example of why [139:11] goal could matter. Muse might spend 40 [139:13] minutes designing a sophisticated [139:14] persistent layer. The goal gives it [139:15] enough context because this doesn't [139:17] material advance the initial deployable [139:19] of the MVP. Okay. So maybe it just keeps [139:21] it on track a bit better. So we'll go [139:22] ahead and give that a try here. And I'm [139:25] going to go ahead and do this. [139:28] Okay. So we've set a goal. [139:31] Oh, is it now executing it? [139:38] Oh, I guess it's off to the races. Well, [139:41] um I guess I'll just let it run and if [139:43] there's anything interesting to check, I [139:45] will check into here. But, uh I think [139:47] it's going to go ahead and do that. All [139:49] right. What's really interesting, it's [139:50] starting to queue up a a ton of tasks. [139:52] Um and so, yeah, it really is [139:54] interesting to see this go off and and [139:57] do this. Now, I've seen goal setting for [139:59] many different um uh coding harnesses, [140:02] but I just generally don't use them as I [140:04] I don't really know the fit here. But it [140:06] seems like it is driving it and breaking [140:07] up the task as it goes. As a goal [140:09] management de decomposer, [140:12] I think that they may just maybe not [140:14] describe I'm not talking very clearly, [140:16] but maybe they just do not describe very [140:18] clearly how goals are supposed to work, [140:21] right? So I think what it does is it [140:23] takes a goal, it decomposes it into [140:24] multiple tasks and it drives it across [140:27] that and that might be a more efficient [140:28] way um to run larger tasks. Okay, here [140:32] it has stopped to ask me a question. So [140:34] I will go ahead and uh give it access to [140:37] what it wants so it can continue on. [140:39] Also it's interesting it's showing the [140:40] goal progression here on the right hand [140:42] side. So that's also interesting as [140:44] well. [140:45] There we go. And after our really long [140:47] wait um the MVP backend is completed and [140:50] verified. So we have a gossql light. Um [140:53] I guess the question is like how can I [140:54] run this as I don't necessarily have any [140:56] instructions here run locally. So we [140:58] have that uh information. Um normally [141:02] it' be very useful if we had a means to [141:04] oh we do actually have a bunch of stuff. [141:05] So we do have a make file. Okay. I was [141:07] going to say like well maybe we don't [141:09] have a make file or whatever but let's [141:10] go take a look at the readme and see [141:12] what information we have. Project [141:15] layout. We can see our stack [141:16] information. Um, [141:19] we have all the instructions to start [141:22] it. Yeah, we have everything. We have [141:23] tons and tons of endpoints. Um, so that [141:27] is good. You know, I would assume at [141:30] this point it's not integrated into um [141:33] the front end. So that would be my next [141:37] step. Um, let's take a look at the code [141:40] and see what we have. So we got a main [141:43] go application which is great. [141:46] Yeah, looks really straightforward to [141:48] me. Um, [141:50] so we have our database over here. No [141:52] avatar stored currently. Um, [141:56] and [142:00] we can go into [142:05] uh our back end here. I'm just curious [142:07] like what the models look like. [142:11] Okay. Because I was curious if it would [142:12] use something that is like a an OM. It's [142:15] not exactly doing that. But here we do [142:16] have [142:18] um our statements here. It's not using a [142:21] templating language. So that was one [142:23] specification I had that I wanted. It's [142:25] not doing it. It's using strrus which is [142:26] totally fine. And we have our very [142:28] simple inserts. So that is totally fine [142:30] here. I I don't mind that. Um I guess it [142:33] is a compiled language so it might have [142:35] an issue trying to bring it in that way. [142:36] But uh we do have an SQL life file so [142:39] that is uh like good there. Oh there's [142:41] our queries. Okay. So, we do have it. [142:45] Nice, nice, nice, nice, nice, nice. [142:47] Okay. And we have the question mark [142:48] where it's getting inserted. Um, so I'm [142:50] pretty happy with that so far. I don't [142:52] know if there's any caching layer right [142:53] now or if it's even necessary to do so. [142:55] We never asked it to do that. The big [142:58] thing I want to do is to close the gap [143:00] with the front end. And so, I'm going to [143:02] need another doc here uh for our our [143:05] front end. So, we'll just say um [143:08] backend frontend integration [143:13] MD. And so, I'm going to go backend [143:16] front end integration. [143:18] So, we we have a backend implemented. [143:22] Um [143:24] but uh we need our front end working. [143:29] We should do the following. uh create a [143:33] uh docker compose that with uh with a [143:37] single [143:40] uh docker compose up will [143:45] serve the front end [143:49] uh serve the back end. Okay. Um [143:52] generally you want I would assume like [143:55] like let's use engine X for the uh [144:00] front-end serving. [144:03] Okay. [144:07] Um [144:09] I guess the thing is that we never [144:11] talked about where the uh files would [144:14] go. So, um for avatar uploads, they will [144:20] be locally [144:22] uh stored with the um [144:26] virtual machine, but I mean like they're [144:27] really supposed to be in something like [144:28] S3, like an asset layer, right? So, [144:32] I'm not sure how we would mock that, but [144:34] the the idea is like, you know, for [144:35] uploaded assets, for uploaded avatar [144:38] assets, [144:41] they should be in S3, but for uh local, [144:45] we need some way to mock that behavior, [144:49] however you want to do that. [144:53] Okay, [144:55] so we'll just say that for that. Um [145:00] and so there's that. So [145:05] infrastructure work, [145:10] right? Um I don't know if we have [145:13] deployment solved in any way. So like [145:15] this is just like like local like [145:17] develop development infrastructure work [145:20] the front end [145:23] integration. So, [145:26] you know, we need to uh [145:29] uh port our [145:32] um static website over to React. I I [145:37] mean, I prefer Myithil to be honest. Uh [145:39] that's what I like, but we could port it [145:41] over to React. We don't want to use uh [145:44] existing components from a component [145:47] framework. [145:49] uh we don't want to use [145:52] uh a CSS framework. [145:55] Okay. So [145:59] uh so we want to port [146:03] over to React [146:06] uh we will need to hook up [146:10] uh the MVP endpoints to our UI. We might [146:15] need to [146:17] adjust our UI or our uh backend [146:21] endpoints. We are not hooking up [146:23] everything. [146:28] Okay? you know, like it's just [146:32] what we do want [146:35] registration. [146:38] Uh, reset password. [146:42] Um, [146:44] login, [146:46] create a post, uh, create a topic. I [146:49] guess you can't create a topic. So log [146:51] in and [146:56] seed the initial topic uh topics and so [147:02] uh [147:07] seed initial topics users uh posts [147:13] be able to create a post in a topic. be [147:17] able to create a comment in a post. Be [147:21] able to vote a post. Okay. So, vote up a [147:26] post. All right. So, we have some very [147:28] basic stuff here. [147:31] And I really like that. That goal stuff [147:32] was awesome. So, I'm going to go all the [147:34] way down the ground here. And we're [147:35] going to go ahead and clear that out. [147:37] And I'm going to ask I want a new goal. [147:39] So, I'm going to go over here and I'm I [147:41] know I'm using static PT, but we'll go [147:43] ahead and just like [147:45] so I can say how do I formulate how do I [147:48] make this a goal in for [147:54] um meta muse code because it did a good [147:57] job last time [148:00] and [148:03] no no it's trying to like write me a [148:04] whole thing. No, it's it's it's the goal [148:07] command. [148:15] Mhm. [148:23] Docker compose starts. [148:28] Okay. So, I'm going to go here. [148:31] I mean, like I can't paste that whole [148:33] thing in there. goals are the finish [148:35] line the not the design document use [148:37] goal tracking. So but like how does it [148:39] know [148:41] those uh they can discover or you can [148:43] use so does goal know of my doc [149:04] Okay, there we go. So, described [149:06] described in Sure. [149:09] So, we'll go ahead and do that. So, [149:10] described in docs uh backend [149:16] front end integration [149:20] MD. Okay. So, we will let it go ahead [149:23] and uh do that because last time it went [149:26] really well. So, let's get the front end [149:27] working here. [149:29] Yeah. I would just say the one thing [149:31] that is kind of a pain is that you'll [149:33] get these commands that pop up, but you [149:34] can't exactly uh always see what they [149:36] are, what you're confirming. Um, but [149:38] yeah, I mean it's fine. So, I just have [149:40] to keep hitting enter here and watching [149:41] it carefully, but it's been very good at [149:43] driving the skills. Miniio might be a [149:45] little bit overkill or maybe it's fine, [149:47] but we will just wait here. But it's [149:48] getting pretty close to being completed. [149:51] All right, so there we go. It should be [149:52] done. Let's go take a look here and do a [149:54] docker compose up build. [149:56] And [149:59] uh well I want to make sure I'm in the [150:00] correct folder here. So just a moment [150:01] here. Uh we will go into the php nuke [150:10] technique. Technuke. Here we go. We'll [150:11] do docker compose up. Wait for that to [150:14] go up and we'll give it a moment here [150:19] to build. We'll be back in a moment when [150:21] all the ports are open and we can take a [150:23] look at it. It might actually already be [150:24] running and so there might be a conflict [150:26] here. I'm not sure as I might be running [150:29] some [150:30] stuff already here. [150:34] So, we do have stuff running and so I'm [150:36] not sure if that would be a conflict for [150:38] it, but we will go ahead and build it [150:39] anyway. [150:41] Okay, I can always just stop it here in [150:43] just a moment. So, I'll just wait till [150:44] it's done. [150:46] Okay, so it seems to be running in the [150:47] background here, which is great. Um, [150:49] what port it's running on, I'm not sure [150:51] as we have quite a few here. We have [150:52] 8080 um 3000 9000. It probably would [150:57] help if we took a look at the actual um [151:01] uh Docker Compose file here. So we have [151:03] back end, front end, and Minos. So [151:07] Mino's running on 9,000 and the front [151:09] end is on port 3000. So that's how we're [151:13] going to take a look at this. So we will [151:15] go and have a bit of fun. [151:19] And it loaded just for a second and then [151:21] it vanished. I'm not sure why. See, it's [151:24] flashing. And so that's not something [151:27] you want to see, but I mean like it's [151:30] there. [151:34] And then advantages. So I'm assuming it [151:36] might be a small thing. We do have [151:39] uh looks to be an error. [151:42] So we'll go ahead and feed that back in [151:44] there. That's not particularly a goal, [151:46] but it is a problem. So, we'll go to [151:48] here and just say um the uh the page uh [151:54] flickers [151:57] the UI and then it vanishes [152:00] shows blank uh blank blue background. Um [152:05] I think it's erroring out the front end. [152:09] And I'm going to take a look at the [152:10] network console here. [152:17] I mean there is one error but usually [152:18] it's XHR. So [152:21] uh we'll go here and we get payload [152:23] responses. So um no errors for XHR [152:27] payloads. Um so maybe [152:31] front end issue. [152:35] We'll see if it can resolve that problem [152:38] there. All right. We still seem to have [152:40] a persistent issue here. Um though I [152:43] mean I didn't restart it. So maybe [152:44] that's something I should do. So [152:46] remaining action docker compose build [152:48] without a host in the sand uh sandbox. [152:50] So I'm going to go ahead and just do a [152:52] compose down here and stop the entire [152:55] stack. Um and then I'm going to go over [152:57] to here and we are going to bring that [152:59] right back up. So we will stop that [153:02] there and we will do a docker compose [153:04] build and we'll see if that resolves our [153:06] issue here. [153:09] I'm not sure we'd have to build from [153:10] scratch. [153:16] Uh, and the avatar upload will succeed. [153:18] Files changed. [153:22] Seemed like it did quite a bit. U, but [153:24] we will go back over to here [153:28] and wait for it to come uh come up and [153:29] we'll just confirm. All right, looks [153:31] like it is now up. We'll go over here [153:33] and take a look. Um, and so we are in [153:36] the site. I'm going to go over to topics [153:39] into DevOps pipeline and we can see [153:43] stuff in here with reply comments. Uh [153:46] obviously I should not be able to reply [153:47] right now. We have to log in. So we will [153:50] go and figure out how to log in. Um I [153:54] will go ahead and create myself account. [153:55] So say andrew@examppro.co. [153:58] This shouldn't go to any real email. So [153:59] it should be fine. And we'll say testing [154:03] capital T testing [154:05] capital T testing one two three whoops [154:09] capital T testing 1 2 3 4 5 six and [154:12] register. [154:14] And so now I'm registered. I'm logged [154:16] in. I can go ahead and submit a post. So [154:20] we will go and I mean I don't like the [154:22] UI. The UI needs to work but it's you [154:24] know it's fun that it there's something [154:25] here. And we'll go here and say [154:29] AI, anyone know how to do X? I don't [154:35] know how to do X. [154:38] And we'll just put a code block of [154:42] something here. [154:45] And [154:49] it says here needs two Thunderdome [154:52] approvals to enter Vault. Interesting. [154:54] So, we'll go ahead and publish to the [154:56] feed. And so, if we go to our feed, we [154:59] can see it over here. We can't we can't [155:01] upvote our own. Well, I guess we never [155:03] told it to told it to code that reply to [155:06] thread [155:08] comments on the first post. [155:11] So, reply to that. [155:13] Uh, again, the forms are confusing, but [155:15] the point is is that it does work. It's [155:17] completely running, which is really, [155:19] really awesome. So, I would say that uh [155:22] I like it. It's it's cool. I mean more [155:24] work would have to be done to uh [155:28] make this no no tables are harmed here [155:32] built with flex blocks in pure [155:33] nostalgia. I love that part. But um you [155:35] know more work would have to be done to [155:37] to make it more sense. I might just drop [155:39] it down to two column and I don't know [155:42] other things like that. But this is just [155:43] for fun, right? And uh you know it's [155:46] just done an incredible amount of work. [155:47] I'd have to do a bit of work to audit [155:49] the code. I know I like the backend [155:51] code. It looks really really good. U but [155:53] we will call it quits here and uh yeah [155:57] the goal management works. Okay. [156:03] All right. Let's take a look at how we [156:04] can set the model. Again, this is going [156:06] to be pretty straightforward. We'll do [156:07] muse hyphen model muse spark 1.1. And so [156:12] now we're on 1.1. We could probably go [156:14] here and type in model and then uh hit [156:17] enter. And then we could choose which [156:18] one we want. So 1.2. There's only a [156:20] couple models in here. And it does seem [156:22] to be updating the settings.json [156:24] which we just happen to have open there. [156:26] Pretty straightforward folks. [156:33] All right. So let's say we are getting [156:35] tired of um having to confirm everything [156:37] that uh this coding harness is doing. [156:40] Well um you've probably heard of skip [156:42] dangerously permissions, but for uh this [156:44] one I believe for muse we can do yolo. [156:47] And so that is going to give us the [156:49] equivalent of uh full permissions. And [156:51] so it's not going to ask us anything. Um [156:54] and so you know obviously you want to be [156:56] very careful here especially if you [156:57] don't have a sandbox turned on which um [156:59] I don't think I have bubble wrap [157:00] installed at this time. I will have it [157:02] installed at sometime in this video but [157:04] this is a new uh virtual environment [157:06] underneath so it does not have bubble [157:09] wrap installed right now. And so that's [157:11] just one consideration here. But it's [157:12] not going to um ask me anything which [157:15] will be a lot less uh annoying. Okay. [157:24] Okay. What I want to see here is if we [157:26] can resume a prior conversation. So I'm [157:28] going to go ahead and type in muse here. [157:30] And I'm going to see if we have a [157:31] resume. And we absolutely do. Oh, we [157:33] just do last flag. That's nice. But I'm [157:35] going to hit enter and see what we have. [157:36] And so it seems like the last one we [157:38] have is only that last one there. And [157:39] we'll hit enter. And so we can continue [157:41] that conversation from prior which is [157:44] really nice. Um so there you go. [157:52] Okay, let's go ahead and take a look at [157:54] skills for um Muse here. Now obviously [157:58] every coding harness has skills and you [158:01] know this one is no different. Though I [158:04] do believe that you have to type in [158:05] for/kll and then the name of the skill [158:07] as other ones you would just start [158:09] typing what it is. is. So, if we go take [158:10] a look here and list at our skills, I [158:12] have not imported any skills. Notice [158:14] here that it's importing ones from [158:15] Claude. So, I have a tech pros one, a [158:17] skill creator. Uh, we have built-in [158:20] ones. So, I never obviously created [158:22] these. These come with it. So, we have [158:23] crate skill, which is obviously a very [158:25] useful one, and grill and other things [158:27] like that. Though, I think that if I hit [158:29] escape here, if I type for/grill, oh, it [158:32] does show it. Okay. So, what if I type [158:34] in um techros? Okay, so they do show up. [158:37] Um, so I guess we can do skill in the [158:39] name or whatever, but let's go ahead and [158:41] create a new skill. So we'll say, can [158:43] you create me a [158:46] new skill in this repo? [158:51] Um, that will I I don't really have a [158:54] use for any skills right now for what [158:55] we're doing, but I'm going to have to [158:57] come up with something. Um, so our [158:58] technique application is a social [159:00] platform and uh, you know, maybe we want [159:04] a Yeah. Can you create a new skill that [159:07] will uh create that will create a new uh [159:11] plug-in idea and ground it to the [159:16] community [159:18] MD. [159:20] Okay. And so we'll go ahead [159:23] and we will see how that works. Now it [159:25] should invoke the correct one. The only [159:26] thing is that I do have that uh that [159:28] claude skill. So it could get confused. [159:31] It's checking existing skills and it's [159:33] using the built-in one and that's what [159:34] we wanted to use. So that's perfect. So [159:36] we'll wait for that to complete. I'm [159:37] sure it won't take long and we'll be [159:39] back in just a moment. Okay, so this [159:41] skill is been created. I want to make [159:43] sure it's in the correct directory. Um I [159:45] mean like I prefer it when it's in my [159:46] repository here. And it is here and we [159:48] have our skill. So I'm going to go ahead [159:50] here and we will try our new skill. So [159:51] we'll say plug-in idea. [159:56] And I'm not sure if it's picking it up, [159:58] but I'm going to go ahead and restart [159:59] that. [160:01] There we go. And so to make a new skill, [160:04] you got to restart. That's probably why [160:05] I didn't see one prior before. And so, [160:07] uh, you know, I want a plugin, um, uh, [160:11] that will, [160:13] um, have [160:17] well, I'll say I have a plugin that is [160:19] a, you know, hot or not, uh, but it's [160:22] for a series of photos that is tech [160:28] hardware. [160:30] Right. Okay. So, we'll see what it comes [160:33] up with here. Like, we probably should [160:35] have told it a format that we wanted, [160:37] but I just we'll see what it comes up [160:39] with for fun. [160:42] Okay. And so, [160:45] um now we have our idea. So, we have uh [160:47] swipe hot or not through the created [160:49] hardware, photo sets, home labs, racks, [160:51] GPUs, rigs, identity. So, we have [160:54] community anchor. And so it's going to [160:55] that to where it actually goes. Audience [160:58] problem for developers etc. The home lab [161:00] is identity members already share racks [161:02] wire closets and stuff. Oh, I like that [161:04] idea. So the idea is that people will [161:07] showcase it and then people will publish [161:08] hot or not or like choose hot or not. [161:10] But uh solution core mechanic creator [161:12] posts a rig set of 36 photos. Viewer [161:15] sees one set at a time and votes hot or [161:17] not. Score the decay weekly to keep the [161:20] board fresh. Heat service discovery [161:22] daily hot rig in the feed. user flows, [161:25] lurkers, encounters, hot rigs in the [161:27] right rail of the feed. UI placement. So [161:31] cool. Yeah. And it gives you some [161:32] structure. I mean, it gives me a little [161:33] bit more than I need, but I have it on [161:34] extra high. So, that kind of explains. [161:36] It's not a dating app. Yeah, we we're [161:38] not doing that. No faces, no no stuff [161:39] like that. Lurker to voter conversion. [161:42] Sure. So, yeah, there you go. That [161:45] That's a skill. Cool. [161:53] All right, let's go take a look here and [161:55] see how we can execute uh in headless [161:57] mode. So headless mode allows us to [161:59] basically uh use it uh via a single line [162:02] command. And so here I could say uh you [162:06] know can you tell me what uh lang we are [162:09] using for the back end. [162:13] And the idea here is that we can go [162:15] ahead and implement that. So it did not [162:17] like what we did here. So I'm going to [162:19] go all the way back here. here. I don't [162:20] think I need the the doubles. And now it [162:22] is going ahead and executing against our [162:24] code space. [162:27] Okay. So obviously um headless mode is [162:29] very useful if you are uh if you just [162:33] want to run commands from a bash script, [162:35] you know, so maybe a remote server, [162:37] that's where it's going to be useful. [162:38] And so here we have uh the output. Very [162:41] straightforward. Um we could ask it for [162:45] JSON elements. So, we could go ahead [162:47] here and say uh JSON and see what we get [162:51] back here. So, yeah, it's streaming out [162:53] JSON information. Uh we could also [162:55] specify a prompt file. I'm not going to [162:58] do that, but I'll just show you really [162:59] quickly. It's just prompt [163:01] file and then the name of the file, [163:03] task.txt. Okay, so you get the idea [163:06] there. Pretty straightforward. Um yeah, [163:09] there you go. [163:15] All right, let's take a look at uh the [163:17] ability to have local memory. So, uh [163:20] coding agents, sometimes you want them [163:22] to remember very specific things um that [163:24] it keeps making mistakes on. And so, [163:26] normally the way this works is you just [163:27] tell it, hey, can you remember this [163:28] fact? And it will store it in a [163:30] location. So, um in the documentation, [163:32] it's not really specifying exactly. [163:34] Maybe it's hidden in a tutorial, but [163:36] it's not showing uh how we do that. And [163:38] so, my guess is that we're probably just [163:40] going to ask it to do that. So, I'm [163:42] going to say, "Hey, can you remember uh [163:45] that um I want all CSS to be uh external [163:50] and not in line. [163:53] Okay. And so I'm going to see if it [163:54] actually creates a memory here and where [163:57] it will create that memory as it could [163:58] be creating it in a few places. So, we [164:01] have personal project memory which is [164:03] the default stored on your machine [164:05] outside the repo, private to the scope, [164:06] committed to the repo and and then your [164:09] machinewide notes. Um, so I'd rather it [164:11] store it in here, but if it doesn't, we [164:14] can go check the configuration [164:15] directory. It's not that big of a deal. [164:16] Says save to your project memory. So [164:20] it's suggesting that it's saved. [164:23] Okay. Sorry. Did Did you save it in the [164:25] repo? [164:30] Project memory committed to the repo. [164:31] Yeah, because I'd rather I'd rather have [164:33] it in the repo. And so it will rework [164:35] that right now. [164:39] We'll give it just a moment. So, they [164:41] probably have obviously a skill [164:42] underneath that is uh the thing that is [164:43] doing that. And we'll just give it a [164:46] moment. It took it a while saying no, I [164:49] didn't save it there. Yeah, move it. [164:50] Move it to this this repo. So, you [164:53] shouldn't have to ask it twice, but [164:55] that's okay. Not a big deal. So, um [164:57] it'll take a moment here and it'll move [164:58] it into the correct directory. We're [164:59] expecting to go to uh I guess agents [165:02] memory. So, we will refresh here and see [165:05] what happens. Now we are in Technuke. [165:08] Oh, you know what? It might already be [165:09] in the correct location. [165:13] Uh, well, I mean, it's now in the [165:15] location we want it to be. So, I think [165:17] we didn't expand this directory and we [165:18] were working here. But you can see here [165:20] we have memory.mmd. Uh, there's a lock [165:22] file. I'm not exactly sure what the lock [165:24] file does as I explain it here that [165:26] there is a lock file. Um, [165:31] nope, no information about a lock, but [165:32] that is totally fine. [165:35] Okay. So [165:38] yeah, I mean there's memory and you know [165:40] it will pick that stuff up next time we [165:42] are implementing something. Okay. [165:48] Okay. So we have approval modes here and [165:50] there are three modes that we can set it [165:52] to. On request, untrusted and never [165:54] approval modes. Um you know they protect [165:57] you against it uh the the coding harness [165:59] taking action when it's doing dangerous [166:01] things like removing files and things [166:03] like that. And by default, it's going to [166:05] uh always ask you when it runs pseudo or [166:08] running a remove command uh for your [166:10] approval. You can move it up to [166:12] untrusted, which is a stricter mode. Um, [166:14] and I believe that it's going to be [166:16] asking you a lot more questions, not [166:18] just dangerous ones. So, I think it's [166:19] based on uh, you know, your your rule [166:22] settings or things like that. And then [166:23] never it's never going to stop for [166:24] approval, but everything's going to be [166:26] contained in the sandbox. I would [166:27] imagine to set that you probably would [166:29] want to have sandbox uh enabled which I [166:31] don't have at this point but you know if [166:33] we wanted to switch to untrusted we can [166:34] do that here. Um it's not showing it [166:36] anywhere in the interface but if we go [166:38] here and go to status there's no [166:40] indicator uh where it is because it says [166:42] trusted here but that's the workspace [166:44] and access is approval normal. Um if I [166:47] go here and just launch it regularly. [166:49] Let's see what happens now. We'll check [166:51] status. So I'm not exactly sure where [166:54] that setting is because it's saying [166:55] approval normal. Uh but obviously there [166:58] is some difference. There is another [166:59] flag as well if you want to turn [167:01] approval judge off. Um so uh that is [167:05] another option and that way every every [167:08] decision you are making. Okay. So there [167:10] are those two options. Um but there you [167:12] go. [167:17] All right. So in this video let's figure [167:18] out how to get um sandbox mode turned [167:20] on. It's usually pretty straightforward. [167:22] You'll end up installing um bubble wrap. [167:26] Uh so that will be what we'll need to [167:28] do. Um let's go take a look at the [167:29] instructions and see what they have for [167:31] that information. Um so it says run [167:34] inside the sandbox, but it's not [167:35] specifying how. So I think what we'll do [167:37] is just go install bubble wrap. Um as [167:39] that's usually what we just need. So [167:40] bubble wrap uh install Ubuntu. Different [167:45] systems use different layers. So it's [167:47] going to tell you what it wants. And [167:48] yeah, it's just going to be pseudo app [167:49] bubble wrap. I mean, that's what I was [167:50] thinking it would be, but for fun, we [167:53] looked it up initially. Okay, so let's [167:55] go ahead and just install that. And I [167:57] will put in my password, and that will [167:59] install here. And we will make our way [168:02] back over to here. Um, and close this [168:05] out and restart it. Seems like I have it [168:07] launched in two places or I don't know, [168:10] but I'll just close it out and let's see [168:12] if now it detects it. So now it's not [168:13] saying any issue about sandbox. And so I [168:16] believe that we are now operating within [168:18] a sandbox. I think it's as simple as [168:20] that. Um, so let's go take a look here. [168:23] Shell commands run inside the file [168:24] system network policy. Muse code refuses [168:27] to run a shell command when it can't [168:28] confirm that the sandbox is active. Fair [168:30] enough. And maybe that's why I was [168:32] getting a lot of um uh issues earlier. [168:35] But over here we have control network [168:37] access. So sandbox network where we can [168:39] control whether it's proxy only, it's [168:41] restricted or enable full access. So, [168:44] this is something I noticed that when it [168:45] wants to go out to the internet or [168:46] something, it often will ask you. And [168:49] so, here you can see that it is proxying [168:51] uh by default, but if that really annoys [168:53] you, you can say enable all. But, you [168:54] know, if you start Skynet, that's your [168:56] fault. Okay, because you're letting it [168:58] out of the box. Um, but anyway, I didn't [169:00] even have the sandbox installed, so I [169:02] guess um maybe I started Skynet. Anyway, [169:06] uh there you go. So, sandbox is now [169:07] enabled. Um we'll just go ahead and [169:09] check our status because it probably [169:11] would show up under here. and sandbox is [169:14] normal. Okay, it said normal before even [169:15] when we didn't have it installed, but [169:17] there you go. [169:23] Okay, so we already looked at the YOLO [169:24] command, but I just want to show that [169:25] there are multiple guard rails that we [169:27] could uh do. And so I just want to show [169:30] you the three. So YOLO will disable [169:31] approvals and the sandbox and trust the [169:34] trust the workspace. So that means it'll [169:36] disable everything. You can uh [169:38] conditionally disable uh approvals or [169:40] conditionally disable sandboxes. So, I'm [169:42] curious to see like what it shows when [169:44] we do that. Um, so we'll go ahead and [169:47] just say disable sandbox [169:49] and we'll go status [169:52] here and go up here and just take a look [169:55] here. So, it's set it's showing that [169:56] it's disabled. So, if we probably go [169:58] ahead and set this to uh yolo, it [170:02] probably would disable both of those. We [170:05] go here and they're now both disabled. [170:07] And then for the last one, we will go [170:09] ahead and say disable approval. [170:13] Okay. And then we will try uh status. [170:17] And they're both disabled. Oh, I mean [170:19] that one's disabled, of course. Okay. [170:20] So, you know, pretty straight [170:22] straightforward. There you go. Just [170:24] wanted to show that. Okay. [170:30] All right. So, something that we will [170:32] absolutely want to add is MCP. So, I'm [170:34] just curious to see how we can do this. [170:36] Uh, I just I didn't I thought there [170:38] might have been a subcommand here for [170:39] MCP. So, I'm just going to take a quick [170:41] look here with the help. Um, so maybe [170:44] not that way, but we'll go ahead and [170:47] open it up this way. So, we can see MCP. [170:51] It doesn't list it this way. What I'm [170:54] seeing in the docs over here is uh MCP [170:57] servers via file. So, maybe this is the [170:59] only way that we can set it. So connect [171:00] external tool through the model context [171:03] protocol declare servers and MCP block [171:05] servers. So let's go take a look at MCP [171:08] uh servers [171:10] or directory. [171:12] Let's see if we can find something that [171:14] is interesting. There's obviously a lot [171:15] of marketplaces. [171:17] Um so I'm just trying to find something [171:19] that is fun. [171:23] Most stars most installed. [171:27] Um, [171:29] well, we got a duck.go one. Sure. [171:33] And so, [171:37] we'll go over to here [171:41] and we'll go ahead and add this. And so, [171:42] I'm going to go over here and say, you [171:44] know, I want to have [171:47] access. I want to be able to use [171:51] MCP server with Muse. Can you add [171:57] this one for me? And I will just help it [172:00] along here [172:03] with the MCP server [172:05] information here. And maybe it'll just [172:07] place it in the correct directory. It's [172:08] obviously going into the settings file. [172:10] So, it's like it's not that hard to open [172:12] up, but if it can configure it for me, [172:14] that will take out a bit of my headache [172:16] for me. I kept saying 0.2.1 [172:19] or three. We're on 0.2.1, which is fine. [172:22] I don't think the little version [172:23] difference is going to matter when I'm [172:25] talking about versions here, but anyway, [172:27] we will give it a moment here and see [172:28] what comes back. And here it's asking to [172:30] edit the file. So here it's adding the [172:33] duck.go reference, which looks fine by [172:35] my standards. And we'll go ahead and hit [172:37] enter. Like that's what it would look [172:38] like. I would expect it. Um though I [172:40] don't think we have a command to see [172:42] what MCP servers we have, but we will [172:43] find out here in just a moment. Um so I [172:46] will uh once it's all done here, we'll [172:48] stop it. We'll and we'll open it up and [172:50] we'll see if there's any way for us to [172:51] detect uh that it's loading that MCP [172:54] tool. Okay, [172:57] I'm just going to keep hitting enter [172:58] here. Okay, so let's go ahead and [173:00] restart the MCP server and well, Muse [173:03] Muse. Okay, and we'll see if we can find [173:06] it here. Um, so I'm not sure uh if there [173:10] is any way for us to see it. [173:16] skill site settings. [173:20] That'd be local settings. I mean, it [173:22] would show it here, I guess, if it was [173:23] loaded, but there probably wouldn't be [173:24] any debugging options for this. [173:27] Um, most providers will have it because [173:29] then you don't have to figure out [173:31] whether it's working or not. But, uh, [173:32] the way this is imported is pretty [173:34] straightforward. [173:36] Um, we don't have a slash command, so [173:38] I'm going to go ahead and give it a try. [173:39] So, can you use duckgo? Use the mcp [173:42] duck.go go to uh tell me what the most [173:47] uh popular JRPG of 2026 in the last [173:51] quarter is. And so we want to see if it [173:55] actually triggers [173:57] um that MCPU. So that's what we're [173:59] watching out for here. It's searching [174:01] the web. I I don't know if there's any [174:03] way for me to know if it's using the MCP [174:06] server. [174:08] How would I know? I don't know how I [174:09] would know. [174:14] So the MCP server is installed but [174:16] there's no indicator uh for it. So we'll [174:19] go ahead here and just say did you use [174:21] duck go? [174:31] It's not available in the session. I use [174:32] it available in the muse web search. [174:38] Yeah. So, you know, I, you know, I added [174:40] it to our MCP, uh, set. I had you add it [174:44] to our, you know, it's added to our our [174:48] settings, JSON. How do we get it to [174:50] work? [174:58] You know, and you to be honest, like [174:59] anytime I do MCPS with any provider, [175:02] they're always kind of finicky. Um, but [175:04] we'll give it a second here to see [175:05] what's going on. Let's take a look here. [175:08] So settings JSON is not in the MCP [175:09] location. [175:11] Um no MCP servers are correct. Why it's [175:13] not working the session? MCP servers are [175:15] loaded at the session start adding it [175:18] requires Muse code to take effect. So [175:20] exact changes I would need. Keep the [175:22] file above. Pre-install the duct.go [175:24] manually uh locally to node module. So [175:27] npx doesn't need to cache it. Restart [175:29] muse. Okay, go for it. [175:32] And to be honest, you know, again, even [175:34] with cloud code, I I run into this [175:36] problem where I just I have to fiddle [175:37] with MCPS to get them installed. But as [175:39] long as the coding harness can get me [175:41] there, that's all that matters to me. [175:43] And so we will find that out here [175:46] shortly. Okay. Okay. So, it says that [175:48] it's fixed our issue. So, let's take a [175:50] look here. MCP, what was wrong? You [175:52] added it to I didn't add anything. You [175:53] added it, but placed it into this [175:55] directory. Um, current location is now [175:57] this. Fix to apply. [176:00] Um, [176:02] so couldn't use the MPXY.gov [176:05] MCP server [176:07] because it was read only. Okay, that [176:08] makes sense. Um, so [176:13] we have verification. So I'm going to go [176:14] ahead and I want to stop this and we [176:19] will try this again. What am I typing? [176:21] Muse. Muse. Okay. And actually in [176:23] particular, I'm mostly interested in a [176:26] game called um oh, what's it called? [176:30] Well, let's just go ask again. So using [176:32] duck.go [176:36] MCP, can you please uh tell me the most [176:41] popular [176:43] um [176:45] JRPG in 2026 last quarter? [176:51] Okay. And so here we're going to see if [176:53] it uses it this time. [176:57] Now to me there's no indication here [176:59] that it's Oh no, it checks for MCP [177:00] configuration. Excellent. Okay. MCP JSON [177:03] failed to read the file. Not sure as to [177:06] why. [177:13] Checking the server. [177:18] Okay. Okay. So now it is it is using it. [177:21] And normally it will ask for approval. [177:22] This time it didn't bother which was [177:24] nice. [177:28] So it's trying to use MCP. That's good. [177:30] Excellent. [177:34] Come on. You can do it. [177:46] Okay. So again, we're just seeing if it [177:48] will work. [177:56] Now, it's never asking me for approval. [177:58] I mean, like, I would give it approval [177:59] if it just asked, but I I think it knows [178:01] that it's a lowrisisk thing and that [178:03] it's just going ahead and and trying to [178:04] do that. I mean, the tool of it is MCP, [178:06] so it makes sense here. [178:14] And maybe I just chose a a hard to use [178:17] MCP server. But we saw earlier that it [178:20] was able to work. At least it looked [178:21] like it. [178:32] So go here. [178:35] It is installed but currently failing [178:37] for all JRPG uh queries. Uh, some failed [178:40] to get to the view viewd for the most [178:42] popular. Um, fall back via web search as [178:45] of today. [178:48] Want me to rerun [178:52] once the I don't not sure what VQD [178:55] clears. What is VQD? [178:58] VQD. [179:04] I mean, I meant to say VQD. [179:09] internal search token. So what it is so [179:12] when you visit here this contains that [179:13] information a shortlived hash token all [179:15] subsequent API requests must include it [179:17] without it duct rejects request why it [179:20] failed. Um it does a live check and this [179:23] is the sandbox request failure. So the [179:25] regx finds nothing and throws that it's [179:28] not your query it's the network anti [179:30] antibbot blocking. Oh. [179:33] Oh. So duck duck. So despite us having [179:37] an MCP, maybe Duck Doto [179:41] doesn't want agents uh using it [179:43] directly. [179:50] Exactly. The VQD failed its clue. So I [179:53] would say that we got it to call MCP. [179:55] Now this is a different reason as to [179:57] why. Um, so if you need a reliable [179:59] agent, brave search uh API, Bing or [180:04] I never can say that Tavilli are [180:06] designed for it, but duck go scrapeable [180:09] keep being flaky. Fair enough. So we [180:10] could keep running it and get the [180:12] result. What was the most popular JRPG [180:14] of last quarter? Dragon Quest uh 7 [180:18] Reimagined. Cool. Tim would love that. [180:21] He probably already owns it. Probably [180:23] already played it. Probably would [180:24] recommend it. Um, but anyway, we got MCP [180:27] working. Uh, yeah. Okay. So, there we [180:31] go.