Transcription
Hello everyone. In the comments, people wrote to me about the Kmi model when I was writing about connecting GLM 4.7. And it's indeed an interesting model. I looked at the benchmarks. Uh, Kmi outperforms GLM 4.7 in many ways. I decided to try it. First, I'll tell you about an interesting place where I found it. It's on the build.nvidia.com panel. Here you can go and try it, firstly, in chat, and secondly, you can go, look at the View code, uh, look at the code here. And, essentially, here is your token, meaning you can even write code. Try it. For example, you can insert this code. Let me copy it. You can copy it into a notebook. And let's create a code here, yes. Paste it. And let's run it, run this code. Yes, as you can see, we got a response. Uh, you can see how the model reasons. Well, it's a direct stream. That is, uh, you can see that it says, "Hello." This is in Russian, so, uh, what is this in Russian? I should respond in a friendly manner and so on and so forth. Here it reasons, reasons, and then it replies: "Hello, how can I help you?" You can display all of this more human-understandably. Uh, now I will copy the code. So, ChatGPT suggested I change the code a bit. Let's run it so that it outputs not a stream, but the result. Waiting for execution. So, and the code has executed. It took 36 seconds. We asked a question. Well, more precisely, our prompt was. Hello. Here, yes. And it answered us: "Hello, how can I help you today?" In a more familiar, usual format, yes, we received a response. I don't know how many limits there are here, I don't know how much this token can be used, yes, but the very fact that you can try it is very interesting. So, but I wanted to check, uh, this model, essentially, with our MCP servers, but I don't have a subscription. It costs. Maybe I'll buy it for myself? I'll think about it. It costs. Let's see now. So, uh, pu-pu-puum. Yes, there are some here, uh, you can talk to it. Uh, it writes that the discount depends on its mood, yes, that supposedly you can get a subscription for 0.99. I've already tried the last two. Let's try this one. No, well, it offers, essentially, the first month for 12 bucks. But before buying, I'd like to try. Let's try it right in the browser. I'll also show you how to use MCP Super Assistant, if someone, perhaps, isn't aware. Uh, let's try it with our MCP servers. For this, first, I installed a browser extension called MCP Super Assistant, yes? Here it is. After installing it, the MCP icon appeared here. We need to use it with our local MCPs, which are not on the internet, yes? It needs to be configured. There is an instruction here to configure the proxy. So, we installed the browser extension. I think everyone can handle that. After that, what do we need? Configuration. Yes, we need to create a file, essentially the same as, uh, the same as we have for Cursor or for Visual Studio Code. And after that, we need to configure MCP Super Assistant Proxy. We first need to have Node.js. You can download it here as well, yes. I installed Node.js. Here it is. And then you need to run this command. Let's just run it. So, let's start CMD. And so, let's essentially copy this line. So, and we need to specify the path to config.json. So, our path is slightly different. It's like this. So. So. If anything, this file looks like this. Essentially, it looks the same as config MCP in all. Yes, exactly the same. The only thing is, I added streams here. Stream was HTTP. For some reason, SSE didn't work for me. GPT told me that it's because my MCP, in general, is not the kind of true MCP that MCP Super Assistant needs, but I made a small improvement and will post it, I think, today. So you can download it from my fork, which I showed in the previous video, and it will also work with MCP Super Assistant. So, we start, we start this file. A proxy should start. Essentially, what is this? It's a connection, well, it's a proxy, yes, it connects the MCP servers and this browser extension. Everything, it has started. Proxy. Here's its path. Localhost 3006. Uh, so. And now we go to Kmi, yes, here we will chat. Go to Configure MCP. Go. If you turn it on/off, yes, a tab appears here. Go to settings. So, localhost MCP. Uh-huh. And it also says here that you need to start it like this. Output transfer streamable http. Let's write it like that just in case, so that everything is according to the instructions. So, stop it. So, okay. And here you need to add it like this, yes. And now transport streams http. Let's start it again. It has started. Okay, it's working. Look, it's detected here, uh, so, everything, we'll do save, cancel, yes, and our MCP servers are detected here. Look, Search 1C Documentation and 1C Syntax Helper. So, we have connected our MCP servers. Let's check if they work, yes? Let's go, uh, click insert, click add. So. Now it says it's ready to help us. It has access to the tool. And we will ask how many requisites the customer has. So, for the buyer's order. Let's write it as a user would. Yes, let's see. It accesses MCP, gets the requisites. And look, it got the header requisites. So, everything works perfectly. Wonderful. Now let's try our task. So, now we will ask it to write code. We have, uh, an Excel file. Well, the same prompt, essentially, that we used, yes. The only thing is that we will remove the module, yes? Because we are not writing from the IDE, so we won't write "write code in the file enrich Excel." So. Uh-huh, and let's run it. So, we send it, yes? Okay. Excellent. I found the customer order structure. Now it's looking for contact information. So, now it needs to get information about the syntax for working with Excel. Okay, it got it. We sent it to him. Combjects. It got information about combjects. So, it got that too. In general, it's fully using MCP servers right here in the chat, yes? Yes. So, it now has enough information. So, it wrote, it wrote us code. The code looks quite good, in principle. So, let's copy the code, go to 1C, paste it. Let's see what it wrote. Your file. So, let's copy. Copy the path. The path to the file, yes? So, let's see what it did. It calls combjects. Well, okay. Although I would prefer not combjects. So, in general, everything is fine. Request. Well, it looks good. So, let's run it. So, error setting value. What is this? This is, this is the shipment date, yes? Ah, shipment date, payment date. Ah, so-so-so. Well, I think here, let's comment it out for now. Like this. And let's try again. In general, it doesn't look like a code error, but rather a data error. Hmm, no access, but that's not an error. Most likely, it's open somewhere. So, I closed it. Let's run it again with these combjects, always. What a story. Yes. Replace. Okay. The file is saved. So, let's try to open it. So, the file opened. Yes, we commented out the shipment date. Payment date is there. Well, we need to figure out what's with that column. There was some error. But in general, in general, everything worked, and with the headers it's fine, and there's no problem like there was in the browser, where it just appended to the end. In general, it works quite well, I would say. No worse than, no worse than GLM. Well, we need to test it further in IDEs. Well, so, I showed you a way how you can quite well, uh, just in the browser, yes, use your MCPs, your configurations, and write code. If you don't have the opportunity to buy a subscription, you can write code right in the browser. You can also connect this to, uh, ChatGPT in the same way, and to Gemini, and so on. And, by the way, as an option, let's try, uh, deepseek, deepseek, that is, deepseek is available from Russia. We can click start now. Well, we can choose MCP Super Assistant configure insert Autosubmit execute. So, copy attachment. Let's run it. Yes, I forgot to check if we have it. Yes, everything, we have our servers. That is, look, I just took, I, for example, ran out of some free access, yes, somewhere. I just went to Deepseek, for example. Let's add the same prompt here now. So, where is it? One second. So, here's my prompt. Here we remove the module BCL from it again and run it. Let's see what Deepseek does. Here it also went to our MCP to start. I don't understand why only autoexecute. Well, autoexecute is understandable, but why do you have to press the button? That's certainly difficult these days. You want everything to press itself. But, of course, the coolest thing is that you can add your own context like this. Your own context, but add it to the models with free web access, with your own local context, without publishing MCP servers externally, just through a proxy. This is very convenient when you want to test some model, but you want to test it first and then buy it. I am in no way dismissing solving all coding tasks in the browser. Well, since we're talking about Deepseek, yes, let's see what the result will be. Deepseek also asks a lot of things. Well, it's really gone into deep study. 1C, apparently, >> it knows 1C poorly. Well, I didn't expect anything like that from it, actually. Oh, everything, yes, it started writing code. So, well, it's again, uh, combjects. So, okay. It's trying something. So-so-so so-so-so. Aha, it decided to search for documents by the number attribute. Funny, yes? Not by number, but by attribute. And here it wrote a request. Look, you object. Everything is fine. So, and what is this? Another request. Why phone? Any phone. Ah, well, if, well, okay. Well, let's try. Find by attribute. If it's not there, you need to use a different approach. Document search. Ah, well, it immediately said that, damn, it assumed a method like this, which, well, it exists, find by attribute, but there's a number, well, how does it know that number is an attribute that is always there, and it's slightly different from other attributes. Okay. Let's create a new one. So, let's create a new algorithm. This will be our deepseek. I don't remember how to spell deepseek, unfortunately. So, let's copy, uh, paste the path to the file. Even like this. File name. Specify the file name. So, okay. We specify the file name. Let's see what we get in the end. So, and let's execute it. Well, naturally, it probably won't execute anything, I think. Yes, and I should have closed it, of course. Incorrect attribute specified, because here you need to find by number. Okay. One error. Find. Find by number, uh, document number. Well, and date would be good too, actually. Document date, yes, date. Okay. We'll count one error. Yes, you need to close Excel. Let's close Excel. Let's close Excel. So. Well, and let's run it again. So, an exceptional situation occurred. Line 57. Responsible. Well, yes. Line. Line. Yes, damn it. So, line responsible. We can't write an object there. Let's do it like this. What else? Well, I consider this the second error, yes? We'll consider this the second error. There might be something else here, but let's see. So, let's execute deepseek. Two errors so far. Well, in general, I'm looking at the code, it's, well, more or less. Well, yes, shipment date. Let's write string here too. By the way, I thought that Kmi probably had the same problem with types, yes, you could have written the prompt and it would have corrected it. In fact, that's all. That is, in principle, you don't need to know all of this to debug this code. You can just send it errors, and it will fix everything. I'm just counting the number of errors for now, yes, but essentially it's the same second error. It's all one error with types, with conversion to string. So, there are two errors so far. I think it would have fixed them in one go. So, everything, it has written. Well done. Uh, let's look, let's look at this file that it wrote. The file actually exists now. Plus it opens. And so. Yes, it wrote it. Well done, counterparty, responsible, phone, shipment date, payment date. Yes, in principle, here, please, it solved the same task with MCP. Thank you all. I hope it was interesting. Goodbye everyone. M.