Transcription: OpenAgents Episode 084 - Exploring Mistral Large

Source: https://twitter.com/OpenAgentsInc/status/1762189301192835112 Wiki source: https://raw.githubusercontent.com/wiki/OpenAgentsInc/openagents/Video-Series.md Media title: OpenAgents - Episode 084: Exploring Mistral Large We add support for the new Mistr... Upload date: 20240226 Transcription model: mlx-whisper/mlx-community/whisper-tiny Generated at: 2026-06-01T18:27:01Z

Machine-generated transcript. Review speaker labels and wording before using this as quote-grade source material.

[00:00] Christopher David: Hot off the presses.

[00:02] Christopher David: Mistral just three hours ago launched,

[00:04] Christopher David: Mr. Large.

[00:07] Christopher David: Vasily Superior to Mr. Medium.

[00:10] Christopher David: 32K tokens of context.

[00:13] Christopher David: API, blah, blah, blah.

[00:16] Christopher David: Jason Format and function calling.

[00:20] Christopher David: Let's give it a try.

[00:22] Christopher David: I swapped out GPT4 with Mr.

[00:29] Christopher David: An hour amazing chat interface here.

[00:34] Christopher David: I have the previous version with GPT4 there.

[00:40] Christopher David: And I want to try these side by side.

[00:43] Christopher David: Apparently with the benchmarks,

[00:48] Christopher David: there's benchmarks against mixed role.

[00:50] Christopher David: We've used mixed role in the past.

[00:52] Christopher David: And here we go.

[00:54] Christopher David: So apparently,

[00:58] Christopher David: I would interpret this as Mr. Large being very close to GPT4.

[01:05] Christopher David: But I don't know. Let's try it.

[01:12] Christopher David: So we've got Mr. On the left.

[01:16] Christopher David: And we've got GPT4 on the right.

[01:28] Christopher David: So I'm going to try this.

[01:32] Christopher David: This isn't going to be like a general purpose.

[01:36] Christopher David: Apple gold to everybody.

[01:37] Christopher David: I'm just going to be seeing what of these work best for how I'm using it.

[01:43] Christopher David: And as a little bit of background,

[01:46] Christopher David: what we're doing is building an open agents platform.

[01:54] Christopher David: And the most recent sub project has been building an agent

[01:57] Christopher David: that can replace my day-to-day usage of chat GPT.

[02:00] Christopher David: On a paste code,

[02:01] Christopher David: a documentation into this interface along with a request.

[02:06] Christopher David: And get a response.

[02:07] Christopher David: We just added image input.

[02:09] Christopher David: So that's multimodal now.

[02:10] Christopher David: That uses GPT4V.

[02:12] Christopher David: But it would be cool if we could use GPT4V for some things.

[02:17] Christopher David: The best open model for other things.

[02:20] Christopher David: You know, be able to compose different models.

[02:22] Christopher David: But have it be one experience.

[02:24] Christopher David: That's kind of where we're coming from.

[02:26] Christopher David: So let's try first just asking both of these

[02:31] Christopher David: to summarize the read me of our project.

[02:41] Christopher David: Fingers crossed live demo this word.

[02:44] Christopher David: Type in the exact same thing.

[02:50] Christopher David: So one thing to note is that

[02:54] Christopher David: we're going to use the project.

[02:56] Christopher David: We're going to use the project.

[02:58] Christopher David: We're going to use the project.

[03:00] Christopher David: We're going to use the project.

[03:02] Christopher David: We're going to use the project.

[03:04] Christopher David: We're going to use the project.

[03:06] Christopher David: We're going to use the project.

[03:08] Christopher David: We're going to use the project.

[03:10] Christopher David: We're going to use the project.

[03:12] Christopher David: We're going to use the project.

[03:14] Christopher David: We're going to use the project.

[03:16] Christopher David: We're going to use the project.

[03:18] Christopher David: We're going to use the project.

[03:20] Christopher David: We're going to use the project.

[03:22] Christopher David: We're going to use the project.

[03:24] Christopher David: We're going to use the project.

[03:26] Christopher David: We're going to use the project.

[03:28] Christopher David: We're going to use the project.

[03:30] Christopher David: We're going to use the project.

[03:32] Christopher David: We're going to use the project.

[03:34] Christopher David: We're going to use the project.

[03:36] Christopher David: We're going to use the project.

[03:38] Christopher David: We're going to use the project.

[03:40] Christopher David: We're going to use the project.

[03:42] Christopher David: We're going to use the project.

[03:44] Christopher David: We're going to use the project.

[03:46] Christopher David: We're going to use the project.

[03:48] Christopher David: We're going to use the project.

[03:50] Christopher David: We're going to use the project.

[03:52] Christopher David: We're going to use the project.

[03:54] Christopher David: We're going to use the project.

[03:56] Christopher David: We're going to use the project.

[03:58] Christopher David: We're going to use the project.

[04:00] Christopher David: We're going to use the project.

[04:02] Christopher David: We're going to use the project.

[04:04] Christopher David: We're going to use the project.

[04:06] Christopher David: We're going to use the project.

[04:08] Christopher David: We're going to use the project.

[04:10] Christopher David: We're going to use the project.

[04:12] Christopher David: We're going to use the project.

[04:14] Christopher David: We're going to use the project.

[04:16] Christopher David: We're going to use the project.

[04:18] Christopher David: I want an analysis.

[04:20] Christopher David: Why you said that.

[04:22] Christopher David: The speed of Mr. Lard here is amazing.

[04:28] Christopher David: OK.

[04:34] Christopher David: Hose nations minus one.

[04:38] Christopher David: GPD4 wins that round.

[04:42] Christopher David: Let's try something else.

[04:44] Christopher David: Explain what I'm trying to do here.

[05:12] Christopher David: It's feeling the response.

[05:16] Christopher David: How do I look at it?

[05:20] Christopher David: Based on the provided link, you're trying to create an agent using the open agents.

[05:24] Christopher David: OK.

[05:30] Christopher David: Let's get a little more specific.

[05:36] Christopher David: What I'm actually in a process of doing is writing

[05:40] Christopher David: API docs for the API.

[05:50] Christopher David: Doing that in a way that I can kind of reason about doing our own front end application using this API.

[06:00] Christopher David: One of the things that I'm doing is going to be really finalizing the API as like this coherent

[06:08] Christopher David: holistic unit that I can reason about now because we've done 80 of these videos.

[06:14] Christopher David: I can see how the plugins need to work together with everything.

[06:18] Christopher David: So I'm going to be having GPT or some model help me do documents like this that will be visible on beautiful

[06:32] Christopher David: docs pages like this.

[06:36] Christopher David: It's so clear that I can have it generate the controllers, feature unit tests, models, validation,

[06:44] Christopher David: all the pain in the ass stuff just converting these docs, markdown docs into code.

[06:52] Christopher David: Having open agents itself do that using this interface kind of bootstrapping itself.

[06:58] Christopher David: I'll show off some of that in future videos.

[07:02] Christopher David: So what I really want to see it do is take this and generate some code.

[07:16] Christopher David: So what I'm going to do is I'm going to do this.

[07:26] Christopher David: The way I have this working right now is it'll fetch and retrieve the content and pass that in as context.

[07:32] Christopher David: But I don't think it saves that augmented context to the message chain.

[07:38] Christopher David: I'll probably add that soon.

[07:40] Christopher David: But for now, I'm just going to start a new message chain.

[07:44] Christopher David: I'll just add a generate example, LeraVille controller route tests etc to fully implement this API doc.

[08:02] Christopher David: Where I include obvious placeholder data like the 200, 400 responses with junk data,

[08:12] Christopher David: logical replacements for those.

[08:16] Christopher David: Okay.

[08:18] Christopher David: Uh oh.

[08:20] Christopher David: I have a habit of dumping all of the code in because that's what I've been doing, but we're not doing it anymore.

[08:26] Christopher David: Just dropping the link.

[08:30] Christopher David: Copy.

[08:32] Christopher David: Okay.

[08:34] Christopher David: I should really do the GPT1 first since it's slower.

[08:38] Christopher David: Yeah, I still got to improve some stuff with this chatter face like it doesn't apply syntax highlighting until the end.

[08:50] Christopher David: But doesn't that look amazing?

[08:56] Christopher David: All right, still waiting on open AI.

[09:04] Christopher David: That speed is quite good.

[09:08] Christopher David: All right, here's kind of the measurement of truth.

[09:14] Christopher David: Sure, I'd be happy to help you implement this API documentation using LeraVille.

[09:18] Christopher David: Here's a basic example of how you might trust your code.

[09:22] Christopher David: Okay. File pass. Nice.

[09:24] Christopher David: No file pass on GPT.

[09:26] Christopher David: Based on the provided API documentation and the requested generate LeraVille resources, here's an example.

[09:32] Christopher David: Okay.

[09:34] Christopher David: Do not say it's set to it to me. I say it's set to it to you.

[09:36] Christopher David: Our agent controller.

[09:42] Christopher David: This is like standard super basic LeraVille.

[09:48] Christopher David: This is interesting.

[09:52] Christopher David: Agent service.

[10:02] Christopher David: Ooh.

[10:20] Christopher David: This is good. Let's keep looking.

[10:22] Christopher David: Test feature.

[10:32] Christopher David: Not a fan of this add other fields here thing.

[10:38] Christopher David: Right, I'm looking at feature and unit.

[10:44] Christopher David: Where is their feature test?

[10:48] Christopher David: Ooh.

[10:50] Christopher David: Okay. GPT for said unit test colon and then gave me a feature test.

[11:00] Christopher David: That's a big minus one there.

[11:04] Christopher David: This test is fine.

[11:08] Christopher David: It is an actual feature test and it's got the file path.

[11:12] Christopher David: I'm going to use the file path.

[11:16] Christopher David: I'm going to use the file path.

[11:18] Christopher David: I'm going to use the file path.

[11:20] Christopher David: I'm going to use the file path.

[11:22] Christopher David: I'm going to use the file path.

[11:24] Christopher David: I'm going to use the file path.

[11:26] Christopher David: I'm going to use the file path.

[11:28] Christopher David: Okay.

[11:40] Christopher David: This is a bit of a subtle point.

[11:44] Christopher David: But I, in my vibes, what I like of these better.

[11:49] Christopher David: I like the mistural response better.

[11:52] Christopher David: Here's why.

[11:54] Christopher David: It includes this concept of an agent service.

[11:57] Christopher David: Now, this is a bit beyond the very basic layer of L.

[12:01] Christopher David: Pred controller.

[12:03] Christopher David: This is the basic pattern.

[12:05] Christopher David: You validate it.

[12:07] Christopher David: You do whatever model operation is.

[12:09] Christopher David: You return the JSON response.

[12:11] Christopher David: Basic good.

[12:13] Christopher David: Yes.

[12:15] Christopher David: For what we're building here.

[12:19] Christopher David: It's going to be used by.

[12:21] Christopher David: These methods that are doing operations to create agents.

[12:26] Christopher David: They're going to be accessible via an API.

[12:30] Christopher David: But also from live wire.

[12:34] Christopher David: And we're going to need a lot of that code to be accessible either from a live wire component.

[12:40] Christopher David: Or an API endpoint without duplicating things.

[12:50] Christopher David: The code to interact with agents is going to need to be in some way.

[12:54] Christopher David: I'm going to need to be in some separate abstraction.

[13:04] Christopher David: My abstractions like that to this point have been within the app services folder.

[13:10] Christopher David: So the fact that it kind of abstracted that.

[13:16] Christopher David: In the exact way that I'm going to need that I knew that I needed.

[13:21] Christopher David: This is like one or two steps ahead of where am I thinking was at.

[13:25] Christopher David: But it's like yes, I know that I need that.

[13:29] Christopher David: It's like it gets me.

[13:31] Christopher David: And I give me the service.

[13:33] Christopher David: Part of me I was like, oh, then it just do that.

[13:37] Christopher David: What didn't have to do the work of figuring out how this should be implemented.

[13:41] Christopher David: No, it's right there.

[13:43] Christopher David: Okay.

[13:47] Christopher David: The headline here is that for the first head to head request of how I use.

[13:57] Christopher David: Catched you BTE or these types of language models.

[14:03] Christopher David: Giving the exact same request to GPT4 and Mr. Large.

[14:07] Christopher David: I strongly prefer the response from Mr. Large.

[14:11] Christopher David: And it was like two times as fast.

[14:15] Christopher David: More to do in terms of experimenting with function calling and stuff like that.

[14:23] Christopher David: But I will be heavily integrating Mr. Large into my workflow moving forward.

[14:29] Christopher David: Stay tuned.

[14:31] Christopher David: See ya.