Transcription: OpenAgents Episode 021 - Hello Faerie

Source: https://twitter.com/OpenAgentsInc/status/1725910351563165748 Wiki source: https://raw.githubusercontent.com/wiki/OpenAgentsInc/openagents/Video-Series.md Media title: OpenAgents - Episode 021: Hello Faerie We enable our new agent to read and respond... Upload date: 20231118 Transcription model: mlx-whisper/mlx-community/whisper-tiny Generated at: 2026-06-01T20:40:30Z

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

[00:00] Christopher David: Well, looks like OpenAI is imploding and based on the excuse from Kara Swisher, seems like it's because

[00:14] Christopher David: developer day and the race to commercialization, the capstone of which was their plan for the GPT store,

[00:25] Christopher David: pushed the board over the edge to fire Alman, Brockman left to and now other senior

[00:32] Christopher David: talents leading.

[00:33] Christopher David: Fastening, we are going to keep building our version of the agents store with ref share.

[00:46] Christopher David: And we checked with our board, which is me.

[00:54] Christopher David: We're going to accelerate this.

[00:56] Christopher David: Okay.

[01:00] Christopher David: Exciting times.

[01:03] Christopher David: So next we're going to train our little AI agent we're calling ferry AI.

[01:15] Christopher David: So now to complete this issue.

[01:23] Christopher David: The first thing I want is a good response from ferry.

[01:29] Christopher David: Just with her response to this issue.

[01:31] Christopher David: And we'll start with that in a feature test in LeraVelle.

[01:36] Christopher David: And just I want to see the response in the console before we have her actually going

[01:42] Christopher David: to do the comment, like make the comment.

[01:47] Christopher David: Like yesterday we were looking at the conversation here where I had done this kind of

[01:53] Christopher David: previously and I was just feeding in that back to GPT4.

[02:01] Christopher David: We made TriGPT4 and GPT4 turbo to see what the quality of the answers are.

[02:05] Christopher David: Obviously the objective is to do this all with open models.

[02:09] Christopher David: One downside is I just don't know what would be the best open models to use for this.

[02:14] Christopher David: We have the ability to use any essentially any open model from hugging phase with our

[02:20] Christopher David: swarm network with our open AI compatible API.

[02:24] Christopher David: Might have to wait a little bit if we try to experiment with a bunch of different models

[02:28] Christopher David: for our workers to load them a couple of minutes each.

[02:31] Christopher David: So we can do that experimentation maybe down in this video.

[02:33] Christopher David: So we might just start with GPT4 and then branch off from there.

[02:41] Christopher David: So let's get started with revisiting our test suite.

[02:59] Christopher David: Through yesterday and cleaned up some of the tests I pushed the longer tests except for

[03:09] Christopher David: ingest test.

[03:14] Christopher David: What I do is I put them in the group called QueenB.

[03:18] Christopher David: And then in PHP units I exclude the group when we do the main run.

[03:28] Christopher David: And then if I want to run all of the QueenB tests I do PQ vendor bin past group QueenB.

[03:41] Christopher David: And this will take longer because it is interacting with our network including doing

[03:48] Christopher David: retries if things fail.

[03:53] Christopher David: And that's not optimized at all.

[03:55] Christopher David: We've got some multi-second things.

[03:57] Christopher David: And that just things that we got to revisit later.

[04:05] Christopher David: But for being able to run a test suite of everything else and I have things run quickly.

[04:12] Christopher David: That's nice.

[04:14] Christopher David: And I also switched over the openagents.app.

[04:18] Christopher David: Toopagents.com and set up Lairvelle in Voir which is just a little additional thing

[04:27] Christopher David: beyond Ford so that there's now zero downtime because it maintains two versions of the

[04:31] Christopher David: application.

[04:32] Christopher David: Kind of like next JS does.

[04:34] Christopher David: Okay.

[04:35] Christopher David: So we're ready to start getting an actual actual cool product up there which hopefully we

[04:40] Christopher David: will have this week.

[04:41] Christopher David: I'm feeling really good about this agent's stuff so far.

[04:47] Christopher David: So what's in the feature test be called?

[04:58] Christopher David: For now I'm just going to put this in very test.

[05:09] Christopher David: What can respond to GitHub issue?

[05:27] Christopher David: So this is basically the same as our inference test.

[05:32] Christopher David: Can do chat completion.

[05:38] Christopher David: I have curiosity.

[05:39] Christopher David: Let's see if GitHub is working.

[05:41] Christopher David: Test.

[05:46] Christopher David: So.

[06:08] Christopher David: One thing we're going to do is pull over the gateway infrastructure from our previous code base.

[06:37] Christopher David: The idea here is that it succeeded after 41 seconds.

[06:44] Christopher David: Obviously not good but it worked.

[06:52] Christopher David: So what we have here is this idea of an AI gateway and the AI gateway is to find as an interface.

[07:02] Christopher David: We're trying to default model name, default training model, the old training file,

[07:11] Christopher David: creative finding job, return the users finding jobs, listed supported models, make a

[07:15] Christopher David: chat completion up with file.

[07:17] Christopher David: And then this interface is implemented by three different gateways.

[07:25] Christopher David: The fake AI gateway, the open AI gateway and the queen BAI gateway.

[07:31] Christopher David: And this is cool because this lets us run tests for all of these.

[07:43] Christopher David: So we have something called the AI gateway contract tests with a function get AI gateway.

[07:52] Christopher David: And then each of our tests of the respective gateways have the use that.

[08:04] Christopher David: So they run all of the tests from there.

[08:08] Christopher David: So we want to test that every gateway can upload files, every gateway can create a fine

[08:13] Christopher David: to any job, every gateway can make a chat completion request.

[08:16] Christopher David: We're trying to list the fine tuning jobs, elisps before it models.

[08:20] Christopher David: And then there might be some tests that are just specific to the gateway.

[08:24] Christopher David: Like here we have returns completion with fine tune model.

[08:27] Christopher David: I don't remember why we do that just uniquely there.

[08:30] Christopher David: This is a pattern by the way that I got from a course called test driven layer of L.

[08:38] Christopher David: What is it called?

[08:44] Christopher David: This is like a, I don't know, like eight year old course by this point.

[08:50] Christopher David: It's from Adam Waffen, the founder of tailwind, just stepping through a whole bunch of best

[08:56] Christopher David: practices for test driven development and layer of L.

[08:59] Christopher David: This is where I kind of learned everything I know about test driven development.

[09:02] Christopher David: So if you want to like really learn proper test driven development, this is all still very

[09:07] Christopher David: relevant.

[09:08] Christopher David: Test driven there with that kind.

[09:10] Christopher David: Okay.

[09:12] Christopher David: So we'll kind of pull those things over gradually because we're going to start just with

[09:15] Christopher David: the functions that we need, but we might start.

[09:19] Christopher David: Setting up this gateway infrastructure better in a second.

[09:25] Christopher David: But let's actually kind of take this down to zero for a second.

[09:33] Christopher David: Can respond to get how this you?

[09:37] Christopher David: Maybe the first thing we should do is like fetch a get-head with you.

[09:48] Christopher David: That's the first thing we have to do.

[10:03] Christopher David: Yeah, I want this text right here.

[10:10] Christopher David: I'm going to guess that there's a PHP GitHub.

[10:32] Christopher David: So we have PHP GitHub API.

[10:35] Christopher David: Let's hit this a labels specific one.

[11:02] Christopher David: Yeah, this is looking good.

[11:20] Christopher David: I'm going to go to the right.

[12:20] Christopher David: Thank you.

[12:51] Christopher David: Okay.

[13:06] Christopher David: So we've got to learn about permissions.

[13:09] Christopher David: I'm guessing if we're fetching something from a public repo that no authentication would

[13:18] Christopher David: be needed.

[13:22] Christopher David: But maybe to get around rate limits or something blah blah blah.

[13:27] Christopher David: Let's try and see if we hit any wrote blocks.

[13:48] Christopher David: You can alias this.

[14:14] Christopher David: I don't think I've ever used that before.

[14:20] Christopher David: But let's try that.

[14:31] Christopher David: Use GitHub.

[14:44] Christopher David: Let's see what that does.

[15:06] Christopher David: If can fetch.

[15:15] Christopher David: Hey, that worked.

[15:24] Christopher David: I don't know if it's a good idea.

[15:35] Christopher David: Okay.

[15:44] Christopher David: Cool.

[15:49] Christopher David: Okay, let's.

[15:52] Christopher David: It says GitHub manager won't behave like it is a GitHub slash client class.

[16:04] Christopher David: There it is.

[16:18] Christopher David: Don't you love it when stuff is works simply.

[16:28] Christopher David: Now I says let's go see if Chagyby T's working today.

[16:32] Christopher David: Now I actually have to wonder about it.

[16:48] Christopher David: Change this layer of L test to add expectations.

[16:58] Christopher David: Let's see if the shape of the response looks like the following correct response.

[17:10] Christopher David: And I'm going to need to remove a bunch of that contents.

[17:40] Christopher David: I'm going to change the name of the name.

[17:58] Christopher David: Two match array.

[18:28] Christopher David: I feel the setting that it has.

[18:43] Christopher David: Huh.

[19:13] Christopher David: All right, first of all, I've never used that assert array.

[19:23] Christopher David: Here's what I mean to do.

[19:30] Christopher David: Change that to use a search as an structure.

[19:44] Christopher David: Oh, does return a response object.

[19:49] Christopher David: I first of all, did he response?

[19:53] Christopher David: Do we have an array with URL?

[19:57] Christopher David: Yeah, we do.

[19:59] Christopher David: Right, I'm just going to do this.

[20:29] Christopher David: And then I want the body.

[20:39] Christopher David: Oops.

[20:54] Christopher David: That's exactly what I wanted.

[20:58] Christopher David: Expects response, body to contain.

[21:13] Christopher David: Okay.

[21:38] Christopher David: Do we want to do our own wrapper on this gift home thing?

[21:59] Christopher David: Maybe not yet.

[22:04] Christopher David: Okay.

[22:12] Christopher David: So we can fetch a GitHub issue.

[22:21] Christopher David: And we'll start with our queen bee.

[22:28] Christopher David: Can respond to GitHub issue.

[22:51] Christopher David: Now, what was that prompt we had in here?

[23:03] Christopher David: Let's use this prompt.

[23:12] Christopher David: Egoals.

[23:42] Christopher David: For now, we're just going to keep the same prompt we had.

[23:53] Christopher David: What is it titled?

[24:05] Christopher David: Oops.

[24:09] Christopher David: Under find a rakey body.

[24:24] Christopher David: Am I getting rate limited?

[24:35] Christopher David: An empty array.

[24:45] Christopher David: Oh, what the hell?

[24:50] Christopher David: Oh, cop, I messed this up there.

[24:59] Christopher David: That is the title.

[25:22] Christopher David: Is used titled, colon.

[25:34] Christopher David: The issue body is.

[25:51] Christopher David: I don't even remember how a string interpolation works with PHP.

[26:08] Christopher David: Let's see what that looks like.

[26:49] Christopher David: I don't know if I can do it.

[27:04] Christopher David: I don't know if I can pass it in.

[27:19] Christopher David: I don't know if I can pass it in.

[27:29] Christopher David: Okay, let's take a look at our prompt here.

[27:54] Christopher David: You are fairing an agent specialized in writing an analyzer code.

[28:09] Christopher David: The issue body is we will implement the memory stream architecture mentioned in the general pages paper.

[28:21] Christopher David: Okay, pretty long prompt.

[28:25] Christopher David: The first thing I want to try to do.

[28:39] Christopher David: First of all, I mean shouldn't this be split into a system prompt?

[28:50] Christopher David: I just passed that as the first thing.

[29:00] Christopher David: What's the default system prompt for the open AI wrapper?

[29:05] Christopher David: Do they say you are helpful?

[29:14] Christopher David: I don't know if I can paste this stuff here.

[29:40] Christopher David: Let's print out the response.

[29:54] Christopher David: I'm going to try not passing a system prompt first.

[30:00] Christopher David: This is going to use the Queen B default model, which is the CUNA.

[30:10] Christopher David: We are going to want to change that for my try.

[30:46] Christopher David: So we can test this with GPT4.

[31:16] Christopher David: Okay.

[31:39] Christopher David: Here's our response.

[31:53] Christopher David: Considering the complexity of the task at hand, it is essential that we approach the implementation of memory stream architecture in a structured manner.

[32:00] Christopher David: In order to achieve this, I recommend the following says we take it.

[32:03] Christopher David: First we should define the data models that will.

[32:08] Christopher David: It's not going to be a good so far.

[32:18] Christopher David: Let's get a bit of an open AI gateway.

[32:30] Christopher David: The previous, I can't remember, I think I had a reason for not using the PHP open AI library because there was some weird era I was getting.

[32:53] Christopher David: Let's try that now and see if I run into it.

[33:03] Christopher David: Let's make sure we're not banjo, but the AI yet.

[33:06] Christopher David: It's going to suck for their company and their people.

[33:47] Christopher David: I'll come back in and clean this up later and get it all compliant with interface.

[34:08] Christopher David: For now, I just want to get the basics working.

[34:14] Christopher David: Let's try using this PHP thing.

[34:36] Christopher David: Let's try to get the components of the compiler.

[34:51] Christopher David: Let's try to get the components of the compiler.

[35:01] Christopher David: Let's try to get the components of the compiler.

[35:11] Christopher David: Let's do open AI gateway.

[35:38] Christopher David: Let's see if all model will be.

[35:46] Christopher David: I'll start with 3.5 turbo and we'll compare the responses we get.

[35:58] Christopher David: I'll try to get the components of the compiler.

[36:08] Christopher David: Let's try to get the components of the compiler.

[36:21] Christopher David: I don't know if that's the client.

[36:31] Christopher David: Is it that simple?

[36:59] Christopher David: Let's compare the

[37:28] Christopher David: I believe the implementation of the memory stream architecture mentioned in the generative.

[37:53] Christopher David: I don't care.

[38:07] Christopher David: I'm going to try to get the components of the compiler.

[38:22] Christopher David: I'm going to try to get the components of the compiler.

[38:32] Christopher David: Let's try to get the components of the compiler.

[38:47] Christopher David: Let's try to get the components of the compiler.

[39:02] Christopher David: Let's try to get the components of the compiler.

[39:27] Christopher David: Let's try to get the components of the compiler.

[39:55] Christopher David: Let's try this again with GPT4.

[39:59] Christopher David: to

[40:06] Christopher David: use GPT for a weekend

[40:11] Christopher David: I thought previously that if like

[40:14] Christopher David: GPT for us it would be because we get banned but it might be because the company and

[40:19] Christopher David: close

[40:23] Christopher David: are

[40:39] Christopher David: the

[42:11] Christopher David: the

[45:02] Christopher David: the

[49:09] Christopher David: the

[49:54] Christopher David: the

[50:01] Christopher David: the

[50:09] Christopher David: the

[50:16] Christopher David: the

[50:31] Christopher David: the

[50:41] Christopher David: the

[50:46] Christopher David: the

[50:53] Christopher David: the

[51:01] Christopher David: the

[51:11] Christopher David: the

[51:18] Christopher David: the

[51:26] Christopher David: the

[51:30] Christopher David: the

[51:36] Christopher David: the

[51:43] Christopher David: the

[51:46] Christopher David: the

[51:51] Christopher David: the

[51:56] Christopher David: the

[52:01] Christopher David: the

[52:08] Christopher David: the

[52:13] Christopher David: the

[52:16] Christopher David: the

[52:21] Christopher David: the

[52:26] Christopher David: the

[52:33] Christopher David: the

[52:38] Christopher David: the

[52:46] Christopher David: the

[52:51] Christopher David: the

[52:58] Christopher David: the

[53:01] Christopher David: the

[53:06] Christopher David: the

[53:16] Christopher David: the

[53:23] Christopher David: the

[53:28] Christopher David: the

[53:31] Christopher David: the

[53:34] Christopher David: the

[53:37] Christopher David: the

[53:41] Christopher David: the

[53:44] Christopher David: the

[53:51] Christopher David: the

[53:54] Christopher David: the

[53:59] Christopher David: the

[54:04] Christopher David: the

[54:09] Christopher David: the

[54:16] Christopher David: the

[54:20] Christopher David: the

[54:24] Christopher David: the

[54:29] Christopher David: the

[54:34] Christopher David: the

[54:41] Christopher David: the

[54:49] Christopher David: the

[54:53] Christopher David: the

[54:59] Christopher David: the

[55:06] Christopher David: the

[55:09] Christopher David: the

[55:14] Christopher David: the

[55:19] Christopher David: the

[55:24] Christopher David: the

[55:31] Christopher David: the

[55:39] Christopher David: the

[55:49] Christopher David: the

[55:52] Christopher David: the

[56:22] Christopher David: the

[56:29] Christopher David: the

[56:36] Christopher David: the

[56:40] Christopher David: the

[56:47] Christopher David: the

[56:54] Christopher David: the

[56:57] Christopher David: the

[57:02] Christopher David: the

[57:05] Christopher David: the

[57:08] Christopher David: the

[57:12] Christopher David: the

[57:19] Christopher David: the

[57:22] Christopher David: the

[57:27] Christopher David: the

[57:30] Christopher David: the

[57:32] Christopher David: the

[57:35] Christopher David: the

[57:37] Christopher David: the

[57:41] Christopher David: the

[57:45] Christopher David: the

[57:47] Christopher David: the

[57:49] Christopher David: the

[57:52] Christopher David: the

[57:54] Christopher David: the

[57:55] Christopher David: the

[57:58] Christopher David: the

[58:00] Christopher David: the

[58:01] Christopher David: the

[58:05] Christopher David: the

[58:06] Christopher David: the

[58:13] Christopher David: the

[58:14] Christopher David: the

[58:15] Christopher David: the

[58:16] Christopher David: the

[58:21] Christopher David: the

[58:22] Christopher David: the

[58:23] Christopher David: the

[58:24] Christopher David: the

[58:25] Christopher David: the

[58:26] Christopher David: the

[58:27] Christopher David: the

[58:28] Christopher David: the

[58:29] Christopher David: the

[58:30] Christopher David: the

[58:31] Christopher David: the

[58:32] Christopher David: the

[58:33] Christopher David: the

[58:34] Christopher David: the

[58:36] Christopher David: the

[58:38] Christopher David: the

[58:40] Christopher David: the

[58:42] Christopher David: the

[58:44] Christopher David: the

[58:45] Christopher David: the

[58:47] Christopher David: the

[58:50] Christopher David: the

[58:51] Christopher David: the

[58:53] Christopher David: the

[58:55] Christopher David: the

[58:56] Christopher David: the

[58:58] Christopher David: the

[59:00] Christopher David: the

[59:04] Christopher David: the

[59:08] Christopher David: the

[59:10] Christopher David: the

[59:12] Christopher David: the

[59:13] Christopher David: the

[59:16] Christopher David: the

[59:18] Christopher David: the

[59:19] Christopher David: the

[59:21] Christopher David: the

[59:23] Christopher David: the

[59:24] Christopher David: the

[59:26] Christopher David: the

[59:28] Christopher David: the

[59:30] Christopher David: the

[59:32] Christopher David: the

[59:34] Christopher David: the

[59:36] Christopher David: the

[59:38] Christopher David: the

[59:40] Christopher David: the

[59:41] Christopher David: the

[59:43] Christopher David: the

[59:45] Christopher David: the

[59:46] Christopher David: the

[59:48] Christopher David: the

[59:49] Christopher David: the

[59:50] Christopher David: the

[59:51] Christopher David: the

[59:52] Christopher David: the

[59:53] Christopher David: the

[59:54] Christopher David: the

[59:56] Christopher David: the

[59:57] Christopher David: the

[59:58] Christopher David: the

[59:59] Christopher David: and C-importance and relevance should help streamline our agent's ability to access and utilize

[01:00:04] Christopher David: relevant data effectively.

[01:00:05] Christopher David: Let's approach this implementation with the following steps.

[01:00:09] Christopher David: Data models, we need to design an Emory object model with properly encapsulates the required

[01:00:13] Christopher David: information for each observation.

[01:00:15] Christopher David: And naturally, we need to describe your integration time set in most recent stages.

[01:00:18] Christopher David: I suggest we also use a unique identifier for easy retrieval and modifications using

[01:00:22] Christopher David: memory habits.

[01:00:23] Christopher David: We could also consider using tagging our categorization to optimize the relevant search and

[01:00:26] Christopher David: make it easier for the agent to find related information in a large data set.

[01:00:29] Christopher David: As for third-party integrations, we need to efficiently interface together API for reading

[01:00:36] Christopher David: writing code, commit, and set in poor web-skimming relevant.

[01:00:39] Christopher David: It might be prudent to create dedicated functions modules for these tasks of maintain-mox

[01:00:42] Christopher David: learning in your suitability.

[01:00:45] Christopher David: Exactly what I was suggesting we do.

[01:00:48] Christopher David: You did it in fact, a feature test.

[01:00:49] Christopher David: We should be maintaining a high code coverage, especially focusing on their retrieval

[01:00:52] Christopher David: function.

[01:00:53] Christopher David: The new Azure Reson C-importance relevance are being calculated in my account.

[01:00:57] Christopher David: Exemies criteria, the main success criteria for the memory manager would be able to successfully

[01:01:00] Christopher David: store all new observations in the memory stream with the appropriate data format and

[01:01:03] Christopher David: bought values.

[01:01:05] Christopher David: The memory stream is able to accurately retrieve relevant memories based in the input situation.

[01:01:10] Christopher David: Memory stream obeys the recent C-dk factor and appropriately updates recently access time

[01:01:14] Christopher David: stamps.

[01:01:15] Christopher David: The managers retrieve a functionality performance well with the current pertinent information

[01:01:18] Christopher David: quickly, eventually, benchmark will vary based on our performance.

[01:01:22] Christopher David: This architecture is a success, predicate on balancing memory accuracy with efficiency.

[01:01:26] Christopher David: So during implementation testing, we could carefully consider any trade-offs in this regard.

[01:01:34] Christopher David: Regarding the generative agent's paper's excerpt, its filled with value insights, however,

[01:01:37] Christopher David: their implementation uses sandbox game, thus the decay factor of 0.99 might not directly

[01:01:42] Christopher David: apply to our case.

[01:01:43] Christopher David: We should be ready to analyze performance and make adjustments.

[01:01:46] Christopher David: Let's roll this out in stages, starting with the basic CRED for memory objects, then adding

[01:01:50] Christopher David: the retrieval mechanisms based on the current situation.

[01:01:53] Christopher David: And finally, optimizing the process using the principles of recent CRED's developers.

[01:01:58] Christopher David: OK, actual suggestions that all sounds good to me.

[01:02:07] Christopher David: It's smart.

[01:02:09] Christopher David: It's smart.

[01:02:13] Christopher David: I like it.

[01:02:14] Christopher David: Good job.

[01:02:15] Christopher David: Good job, Barry.

[01:02:17] Christopher David: Good job, OpenAI.

[01:02:19] Christopher David: Sorry, you're getting fucked kind of.

[01:02:22] Christopher David: Yeah.

[01:02:23] Christopher David: Feel free to help us build the open version of you.

[01:02:34] Christopher David: We're going to both commercialize and be safe because it's open and inspectable, but we're

[01:02:41] Christopher David: going to slam our foot on the gas pedal.

[01:02:49] Christopher David: Yeah.

[01:02:53] Christopher David: In the next video, we'll respond to this and integrate the comments.

[01:02:59] Christopher David: Get an actual conversation going.

[01:03:02] Christopher David: Get her actually in making these code changes herself.

[01:03:06] Christopher David: See ya.