Building a ChatGPT Plugin for Lex Fridman Podcasts

00:00:00.000 | Today, we're going to take a look at a ChatGPT plugin

00:00:04.640 | that is going to let us search through podcast episodes

00:00:08.700 | and basically answer questions based on those episodes

00:00:13.360 | and a few other things.

00:00:15.440 | So let me show you what we have so far.

00:00:19.000 | So here we're in ChatGPT.

00:00:21.160 | You can see up here, we have a little Lex Friedman logo.

00:00:26.200 | That's because we are going to ask Lex some questions.

00:00:30.360 | So we're going to ask Lex about the future of AI.

00:00:35.360 | So ChatGPT is thinking.

00:00:37.360 | It decides that it needs to use the AskLex plugin.

00:00:41.200 | And then it's kind of just giving us a summary

00:00:43.120 | from a few different episodes here.

00:00:46.040 | Okay, so we can see that it's also included

00:00:47.720 | the source of these episodes, which is pretty nice.

00:00:50.580 | And it's given us a few options.

00:00:52.480 | Okay, and then we also get this nice little,

00:00:54.760 | these cards at the bottom

00:00:56.320 | showing us where these are coming from.

00:00:58.720 | So we get a pretty good summary of AI being discussed

00:01:03.480 | in Lex Friedman podcast episodes.

00:01:05.660 | And also this as well.

00:01:06.720 | This is, I think, one of the courses he did at MIT, right?

00:01:11.620 | So that's pretty cool.

00:01:13.800 | Now, what if I, maybe I want a little more information

00:01:17.320 | from a particular one of these episodes.

00:01:20.440 | Okay, so let's come down here.

00:01:22.280 | And I'm going to say, can you tell me more

00:01:26.680 | about the Eric Brian Jolson episode?

00:01:31.680 | Okay, see what comes up.

00:01:34.460 | So it's again, referring to the plugin,

00:01:37.200 | even though I didn't say AskLex this time,

00:01:39.120 | it knows that I still want to use this plugin.

00:01:41.640 | Okay, cool.

00:01:42.480 | Okay, so there's a lot more information

00:01:46.400 | in this kind of summary of that particular episode.

00:01:51.400 | Right, that is really interesting.

00:01:53.360 | And basically what it's doing there

00:01:55.680 | is searching through our entire database

00:01:58.920 | for that particular episode,

00:02:01.120 | rather than up here where we were searching

00:02:02.800 | for a particular topic,

00:02:03.920 | now it's searching for a particular episode,

00:02:06.040 | which is pretty cool.

00:02:07.240 | Now, let's just continue this conversation a little bit.

00:02:11.160 | So interesting, just let's talk about space exploration.

00:02:16.160 | That's all.

00:02:20.000 | Let's see what we get.

00:02:21.120 | Okay, cool.

00:02:23.640 | Getting a few summaries like we did the other time.

00:02:27.600 | So we get, so it's two here.

00:02:30.960 | One with Darwin Newman and the other one with Ariel Ekblor.

00:02:35.960 | Now, that looks pretty cool.

00:02:38.960 | So this one's interesting talking about

00:02:42.360 | building space megastructures, that sounds interesting.

00:02:45.520 | Let's ask about this.

00:02:48.560 | Can you tell me more about building space megastructures?

00:02:53.560 | Okay, and then it's kind of honing in

00:02:57.680 | on that specific episode.

00:02:59.720 | So it's seen, we haven't specified

00:03:01.760 | that we want to talk about that,

00:03:04.080 | but it's seen the previous interactions

00:03:07.200 | and decided that, yeah, we're probably talking

00:03:10.000 | about this particular episode,

00:03:12.000 | or at least this information can be found

00:03:15.080 | in that particular episode.

00:03:17.120 | So again, it's kind of filtered down there,

00:03:19.920 | or maybe just from search, I'm not sure,

00:03:22.600 | but we definitely get a very interesting

00:03:25.680 | and relevant answer here.

00:03:28.160 | Now, that is just kind of a glimpse

00:03:32.000 | of what this sort of plugin can do.

00:03:35.200 | And naturally we can apply this

00:03:36.680 | to a ton of different podcasts, videos,

00:03:39.960 | or just anything you want really.

00:03:43.400 | Let's kind of dive into how I actually built this

00:03:48.400 | and what is actually happening

00:03:50.560 | when ChatGPT is deciding to use the AskElects plugin.

00:03:55.560 | So let's come up to the top here

00:03:58.600 | and we have our first question, okay?

00:04:01.400 | AskElects about the future of AI.

00:04:03.120 | We can open this to see what is being sent to this plugin.

00:04:06.960 | And we see that there's these queries,

00:04:08.440 | query the future of AI.

00:04:10.840 | All right, that's cool.

00:04:11.880 | So we're querying the future of AI.

00:04:14.400 | This is a semantic search.

00:04:16.760 | So we're using a vector database here.

00:04:19.080 | If that doesn't make sense, I'm going to explain it.

00:04:21.640 | But essentially you can think of this as being our question.

00:04:24.920 | And these are the parts or the chunks

00:04:29.160 | of transcribed audio from Deluxe Freeman podcasts

00:04:33.960 | that are the most relevant

00:04:36.400 | or the most semantically meaningful

00:04:40.080 | to our particular question, okay?

00:04:42.880 | These are pretty big, but if we go through,

00:04:47.320 | we should be able to find somewhere

00:04:48.680 | where they're talking about AI, right?

00:04:51.640 | You see there's a ton of things being spoken about here,

00:04:55.520 | but ChatGPT is actually managing to kind of find,

00:04:59.120 | okay, here, like when you start again,

00:05:01.360 | truly superhuman artificial intelligence,

00:05:04.200 | kind of by definition,

00:05:06.360 | I'll be able to do a ton of things

00:05:07.880 | that I couldn't have thought of in a greater world

00:05:09.760 | I couldn't imagine, you know, all this sort of stuff, right?

00:05:12.040 | So that's the way it's getting that information

00:05:13.720 | from it then summarizes and gives us in that answer, okay?

00:05:18.720 | And we can see that there's three entries there, right?

00:05:22.480 | So we have that the Eric Brian Jolson episode,

00:05:27.080 | the Deep Learning State of the Art

00:05:28.720 | and the Andrew Ng episode, okay?

00:05:32.040 | Which is what we see down there.

00:05:34.280 | And then there's this one.

00:05:35.120 | So I'm saying, can you tell me more

00:05:36.760 | about the Eric Brian Jolson episode?

00:05:39.280 | Here, if we take a look,

00:05:41.080 | you can see that the query has changed slightly.

00:05:42.960 | So I'm saying I want to know more

00:05:44.360 | about this specific episode.

00:05:46.240 | And what we have done is actually given specific instructions

00:05:51.240 | to ChatGPT saying, you know,

00:05:55.120 | if someone wants to know more about a particular episode,

00:05:57.960 | you can search for that particular episode

00:06:00.640 | by adding in this filter

00:06:03.160 | and then filtering based on the title, okay?

00:06:06.800 | So the title, it actually knows the title

00:06:08.800 | from the previous interaction, okay?

00:06:11.720 | So up here, it got the title here

00:06:15.360 | in the metadata of that result, right?

00:06:18.520 | So then it can search

00:06:20.280 | for that particular podcast episode, right?

00:06:24.880 | So I think that that is pretty cool.

00:06:27.840 | And you can see that the results here,

00:06:29.880 | then we're only returning chunks of texts

00:06:32.920 | from that particular episode.

00:06:34.800 | And that's basically what is happening

00:06:36.200 | as we go through the rest of this.

00:06:38.000 | So here, it will just be a search, the query.

00:06:41.120 | And then down here, I'm not sure what it's doing actually.

00:06:45.440 | So, okay, so it's doing a search

00:06:47.680 | and it's also filtering for that particular episode

00:06:49.800 | because it knows the information is in there.

00:06:52.400 | All right, so you can see a little bit

00:06:54.800 | of how this is working,

00:06:57.000 | but let's maybe dive into it a little more detail

00:07:01.080 | on how I actually built this.

00:07:03.680 | Okay, so let's just try

00:07:05.680 | and visualize what we have actually built here.

00:07:09.360 | So you just saw the chat GPT,

00:07:12.680 | you can think of it as almost like front end

00:07:14.880 | of this whole thing.

00:07:16.600 | So we have chat GPT over here.

00:07:19.920 | Now, chat GPT, now these plugins

00:07:24.160 | that you can connect to chat GPT

00:07:27.720 | and the way that I think they work

00:07:29.480 | because the code for that isn't actually public

00:07:33.280 | is that you are basically passing your plugin,

00:07:38.280 | which is like a tool for your chat GPT instance to use.

00:07:43.960 | You're basically passing that into the prompt

00:07:47.080 | or the initial prompt of chat GPT,

00:07:49.200 | maybe even like every prompt that is,

00:07:51.920 | maybe it's reformatted on its way to chat GPT,

00:07:54.480 | I'm not sure, but basically somewhere in the prompts,

00:07:57.280 | there is something that says,

00:07:59.280 | hey, chat GPT, you can use these plugins.

00:08:04.160 | Okay, and then it will list each of these plugins

00:08:06.240 | with a little description about each one of these plugins,

00:08:09.920 | which we actually do write ourselves.

00:08:13.160 | Okay, so in our prompt,

00:08:16.200 | you know, may say whatever it normally says,

00:08:19.280 | and then it says, you can, by the way,

00:08:20.600 | you can use the Lex Friedman,

00:08:24.680 | I'm just gonna put Lex F database,

00:08:27.320 | and it will help you answer questions

00:08:31.920 | about the Lex Friedman podcast.

00:08:34.640 | And when you see the words, ask Lex in the user's prompt,

00:08:39.240 | you should 100% use this.

00:08:41.880 | All right, that's basically what we wrote.

00:08:44.120 | So chat GPT sees this, sees your question,

00:08:47.840 | and it will sometimes decide that it needs to use this.

00:08:50.560 | If it does decide to use this,

00:08:52.200 | it is going to create a prompt

00:08:55.120 | to send to the Lex Friedman database.

00:08:58.880 | Okay, and that prompt is going to be like,

00:09:01.680 | actually, you already saw it, it contains queries.

00:09:04.600 | And then within that, you have like a list of queries.

00:09:08.760 | Okay, so it can actually pass multiple queries in there,

00:09:11.800 | but we just have one query here.

00:09:14.200 | It's just how the thing is built.

00:09:17.240 | And I'll explain that later.

00:09:19.400 | And yeah, so that is actually being passed

00:09:22.280 | to this ask Lex endpoint or plugin.

00:09:26.120 | And then we have this response from the plugin.

00:09:30.120 | Now, how do we get that response?

00:09:32.080 | That's probably the more complex bit

00:09:34.320 | that we will need to figure out here.

00:09:36.320 | So once we pass that query, okay,

00:09:40.680 | so we pass the query from chat GPT into our plugin,

00:09:44.840 | we are now in a space

00:09:47.240 | that is not necessarily anything to do with chat GPT.

00:09:50.720 | Okay, so in this space is our own code or API,

00:09:55.360 | whatever that we have set up.

00:09:56.960 | The API that I've set up here

00:10:00.040 | is based on the chat GPT retrieval plugin API

00:10:05.040 | that was created by OpenAI.

00:10:09.160 | And you can actually see that here.

00:10:10.720 | So we go to github.com/openai/chat-gpt-retrieval-plugin.

00:10:16.560 | And what kind of tells you here

00:10:19.240 | that chat GPT retrieval plugin lets you easily search

00:10:22.280 | and find personal or work documents

00:10:24.400 | by asking questions in everyday language, right?

00:10:27.960 | Sounds pretty familiar to what we have done.

00:10:30.920 | It's just rather than personal work documents,

00:10:32.960 | we've used Lex Friedman transcribed text.

00:10:37.640 | Cool, so we have that retrieval plugin,

00:10:41.160 | we'll call it the chat GPT retrieval plugin.

00:10:46.480 | CRP, okay, right?

00:10:49.360 | We've set it up to use OpenAI's text embedding model,

00:10:54.000 | ARDA002, okay?

00:10:57.000 | So ARDA002, and you can basically think of this

00:11:01.120 | as the same as the GPT models that you've seen,

00:11:04.800 | but rather than generating text,

00:11:06.920 | this one actually generates numerical vectors

00:11:11.280 | based on the semantic meaning of the text

00:11:16.240 | that you're giving it, okay?

00:11:18.720 | So what I mean by that is,

00:11:20.880 | you're basically creating these vectors

00:11:24.040 | that go into this vector space here,

00:11:26.160 | and this vector and this vector here

00:11:30.560 | will have a similar meaning,

00:11:32.240 | whereas this vector all the way over here,

00:11:34.280 | because of this distance,

00:11:36.200 | this space between this vector and the others,

00:11:39.160 | this one will have some different meaning, okay?

00:11:44.000 | So it's like we are mapping human meaning

00:11:48.960 | into numerical representations

00:11:53.320 | that we can then use to perform searches.

00:11:58.120 | So how do we perform searches through this space?

00:12:00.600 | Okay, so imagine we have our user query over here,

00:12:04.840 | it's going to be, what was our question?

00:12:07.040 | It was, ask Lex about the future of AI, right?

00:12:11.280 | So the future of AI is basically our query.

00:12:14.760 | It's going to go into this R002 model,

00:12:17.720 | and that is going to create an embedding, right?

00:12:19.880 | And it's going to maybe place it over here, right?

00:12:23.880 | So then we know that these two items,

00:12:28.040 | these two documents that we already embedded,

00:12:32.440 | that they are very similar,

00:12:35.600 | and therefore probably highly likely

00:12:38.840 | to be relevant to our particular query, okay?

00:12:41.880 | And all of this,

00:12:43.560 | this whole thing that you have going on here,

00:12:46.040 | where you have all these vectors

00:12:47.480 | and you're searching through them,

00:12:49.040 | all of that is handled by another service,

00:12:51.920 | which is the Pinecone Vector Database, okay?

00:12:56.160 | So the Pinecone Vector Database

00:12:58.440 | is basically a way of searching

00:13:03.080 | through this numerical vector space

00:13:04.920 | in a very efficient way at scale, right?

00:13:07.680 | So you can have like hundreds of millions of these vectors

00:13:11.400 | or basically documents in there,

00:13:13.280 | and you'll be able to retrieve relevant items

00:13:17.040 | within around a 10th of a second

00:13:19.280 | or something super fast, right?

00:13:22.360 | So that is all handled

00:13:24.080 | by the Vector Database component there, okay?

00:13:27.760 | So we are then going to return those items out here, okay?

00:13:32.760 | So we bring those here,

00:13:36.840 | and we'll see these are our,

00:13:38.960 | we'll say that we're going to return

00:13:40.320 | the top three most relevant items, okay?

00:13:43.240 | So let's say that was another item over here, right?

00:13:46.880 | So we're going to return those,

00:13:48.720 | and that is basically,

00:13:51.440 | this is going to be the output of our CRP,

00:13:56.360 | so our chat GPT retrieval plugin, okay?

00:14:00.320 | So they go into here and they go back to chat GPT.

00:14:05.120 | Now, chat GPT has more information

00:14:08.200 | to answer the question, right?

00:14:09.800 | And it also, as well as the original text

00:14:12.720 | that we have over here, right?

00:14:14.440 | It also has where this text has come from.

00:14:17.160 | So it has the podcast episode titles

00:14:18.960 | and the URL to the podcast episode and everything else,

00:14:22.440 | which is exactly what you can see in here, right?

00:14:25.080 | So these results, we have the original query,

00:14:27.480 | and then we have the things that we returned, right?

00:14:30.400 | So we have that text and the metadata title

00:14:33.920 | and the video ID here,

00:14:37.600 | which I think is actually being used for the URL.

00:14:40.680 | But that's something we should actually fix.

00:14:42.320 | So in the source here, we should add,

00:14:44.120 | or in the URL, we should actually add the actual URL

00:14:49.120 | for the video.

00:14:51.280 | But that's something we can do another time.

00:14:52.960 | So what we have so far is,

00:14:55.280 | we've just basically augmented our original query

00:14:59.600 | with all of this other relevant information

00:15:02.200 | from the Lex Friedman podcast, right?

00:15:04.480 | So now, our query is going to be a ton of text

00:15:08.280 | based on all that stuff,

00:15:10.960 | and we've just returned it to JetGPT over here,

00:15:14.920 | and we're saying, okay,

00:15:17.320 | I want you to answer the user's question

00:15:21.560 | based on all of this information here.

00:15:23.800 | The user's question is, and then it's whatever you,

00:15:27.760 | whatever you put in here, right?

00:15:31.400 | So that gets fed into there.

00:15:33.360 | And then, JetGPT is going to answer your question,

00:15:37.400 | and we saw that, right?

00:15:39.320 | So we got this kind of response

00:15:41.560 | where it's telling you a ton of stuff, right?

00:15:44.920 | And it's sourcing the information.

00:15:46.760 | Look, we can actually click on here,

00:15:49.400 | and it's going to take us through to that episode, right?

00:15:53.080 | So this is, yeah, the Eric Brian Jolson episode.

00:15:58.800 | But there is still a little more to this whole thing

00:16:03.520 | than what I've just described.

00:16:05.400 | I mean, Wanda is actually the code to do all of this,

00:16:09.320 | which you can actually find if you go to GitHub.

00:16:12.840 | The plugin code itself is actually all here, right?

00:16:17.640 | So this is the forked JetGPT repo.

00:16:22.440 | And yeah, I mean, this is,

00:16:25.600 | it basically contains the same code

00:16:27.760 | as what you would find in the JetGPT retrieval plugin.

00:16:32.200 | It's just I've modified it a little bit

00:16:34.960 | for our particular use case.

00:16:37.280 | So for example, if you go to the well-known here,

00:16:40.480 | this is basically where you sell the instructions

00:16:43.680 | for JetGPT to understand your plugin.

00:16:47.880 | And you go to the AI plugin here, right?

00:16:50.760 | So I've changed the name for the model, okay?

00:16:53.640 | So the name that JetGPT we use to understand this model,

00:16:58.520 | the name that we understand it with, right?

00:17:01.360 | We also have, okay, you can use this tool

00:17:04.400 | to answer user questions using Lex Friedman podcast.

00:17:07.800 | If the user says, "Ask Lex," okay?

00:17:10.040 | So we saw that "Ask Lex" earlier on.

00:17:12.760 | Use this tool to get the answer.

00:17:14.840 | This tool can also be used to follow up questions

00:17:17.000 | from the user and filters can be used

00:17:19.560 | to grab more information from specific podcasts

00:17:23.440 | like by filtering for a specific podcast title, right?

00:17:26.960 | So that's the primary context,

00:17:29.200 | the primary instructions I've given JetGPT

00:17:31.560 | on how to use this plugin, which is they're pretty light.

00:17:34.560 | And then down here,

00:17:36.920 | we've just set up where our API is hosted, right?

00:17:41.000 | So we need to include that so that JetGPT knows

00:17:44.760 | where to find the opening API spec for our app.

00:17:48.720 | So these are basically going to be the instructions

00:17:51.440 | that JetGPT will use in order to understand

00:17:53.880 | how to actually programmatically interact

00:17:57.240 | with our API or our plugin, okay?

00:18:00.920 | And there were some changes made there as well.

00:18:03.480 | Nothing significant, okay?

00:18:06.880 | So just put in the server, you have to do this.

00:18:11.120 | And then there are some instructions

00:18:14.160 | around how JetGPT can use this again.

00:18:18.720 | So say this is an array of search query objects,

00:18:22.840 | each contain natural language query string, query,

00:18:25.560 | and an optional metadata filter, filter, okay?

00:18:29.680 | So we also say here filters can help refine search results

00:18:33.120 | based on criteria such as document title or time period.

00:18:36.440 | So we can say these are particularly useful

00:18:38.560 | when a user asks for more information

00:18:40.080 | about a particular record.

00:18:41.400 | For example, the user may ask for more information

00:18:44.040 | after being provided with information

00:18:46.400 | from a specific podcast with the title,

00:18:49.680 | Our AI Future with JetGPT, right?

00:18:52.000 | So I'm using an example here

00:18:53.640 | so that JetGPT knows exactly what to do.

00:18:57.680 | In that case, the filter field can then be used

00:19:00.280 | to return more information by using title,

00:19:03.400 | Our AI Future with JetGPT, right?

00:19:05.960 | So we've been very specific and explicit

00:19:10.520 | on how JetGPT can use this.

00:19:13.760 | Okay, so we build that.

00:19:15.960 | Let me show you some very quick code

00:19:18.880 | that just shows how I actually populated

00:19:21.240 | our database with all this stuff.

00:19:23.760 | So there'll be a link to this.

00:19:25.760 | You'll be able to, in fact,

00:19:27.400 | it will be somewhere around the top of the video right now.

00:19:30.280 | To this notebook, you'll just be able to run it through

00:19:32.400 | and do exactly what I'm doing here.

00:19:34.720 | Okay, so there are a few items that you need,

00:19:38.120 | and we need this in general

00:19:39.480 | for the rest of the plugin as well.

00:19:42.720 | So there's the open AI API key,

00:19:44.720 | which we get from platform.openai.com,

00:19:48.960 | the Pinecone API key,

00:19:50.720 | which we get from app.pinecone.io,

00:19:54.400 | and these other things as well.

00:19:56.080 | So we run through this.

00:19:58.040 | We're using a pre-built dataset in this example,

00:20:01.240 | but actually, if you wanted to do the download

00:20:04.080 | and transcribing yourself, you just open this,

00:20:07.000 | and you can actually see all of the code that I use there.

00:20:10.520 | So I actually built this PodGPT library to help with that.

00:20:15.520 | It's just easier, okay?

00:20:18.160 | So in reality, there's just a few lines of code, okay?

00:20:22.320 | Cool.

00:20:23.840 | But that takes some time, right?

00:20:25.880 | Downloading all of these, like, mp3 files

00:20:29.000 | from the YouTube videos

00:20:31.360 | and then transcribing them using OpenAI's Whisper,

00:20:34.800 | you know, it takes a while.

00:20:35.760 | So there's also this option, right?

00:20:39.280 | So I've already done all of this, obviously,

00:20:42.040 | for the Lex Rubin podcast.

00:20:43.440 | So you can actually just use my dataset here.

00:20:46.480 | This is hosted on HuggingFace datasets.

00:20:49.400 | There's 499 videos in there, okay?

00:20:53.200 | But they've already been transcribed.

00:20:55.800 | So then all you need to do is you initialize your --

00:21:00.320 | this is your Pinecone index

00:21:03.160 | and also the embedding model that you're going to be using

00:21:07.440 | to actually create those embeddings

00:21:10.640 | and store everything in Pinecone, all right?

00:21:12.920 | So that is going to be using OpenAI

00:21:16.920 | and specifically the OpenAI text embedding R002 model, okay?

00:21:21.520 | And then here, I'm just looping through all of the podcasts.

00:21:24.520 | We're indexing everything, and that is actually it, okay?

00:21:28.480 | That's all I did in order to index all of this text.

00:21:32.480 | Then from there, I just went over to DigitalOcean.

00:21:35.080 | So DigitalOcean is just like a service

00:21:38.360 | where we can host APIs,

00:21:40.560 | which is pretty ideal for what we have, right?

00:21:43.440 | So we have this James Callum AskLex plugin.

00:21:49.120 | I basically just went over here,

00:21:52.520 | went to Create, Use, which one, Apps, okay?

00:21:57.400 | Clicked on Apps, clicked on Deploy from GitHub,

00:22:01.240 | and just deployed from this repo.

00:22:03.520 | And then it deploys, and I end up with this, right?

00:22:06.400 | So I have the AskLex app here.

00:22:08.200 | You can click on there.

00:22:09.600 | This gives you the URL for your app.

00:22:13.840 | So you can copy that.

00:22:15.520 | You can open it, and we will probably see this, okay?

00:22:20.520 | So detail not found, right?

00:22:22.400 | That's fine.

00:22:23.520 | It's just like the URL of the API.

00:22:26.920 | Without the endpoints or without any file extensions,

00:22:30.040 | you're going to find this, okay?

00:22:32.520 | So then I took that.

00:22:35.960 | I went over to -- let's go to New Chat.

00:22:42.000 | I went to Plugin Store.

00:22:45.520 | I said, "Develop your own plugin."

00:22:47.680 | My manifest is ready.

00:22:49.400 | So this is the -- we saw this, the AIPlugin.json file.

00:22:54.080 | And then I just pasted in the URL, okay?

00:22:57.800 | So from there, we can see, okay, we've got ValidateManifest,

00:23:01.360 | ValidateOpenAPISpec.

00:23:04.080 | And then you can see, okay, we have, like,

00:23:06.160 | all the information that I include in there.

00:23:10.760 | Okay, so I've already installed it.

00:23:11.920 | I'm not going to install it again.

00:23:13.600 | But then from there, all you do is you go to your Plugin Store.

00:23:16.640 | You go to Unverified Plugins,

00:23:18.560 | and you have your actual plugin here.

00:23:22.440 | And then you just go ahead and use it, right?

00:23:25.480 | So it's incredibly easy to pull this together.

00:23:31.320 | Obviously, there are a few steps.

00:23:32.520 | It's not the most straightforward thing

00:23:35.800 | in the entire world,

00:23:37.280 | but it's definitely not the most complex.

00:23:38.800 | And given what you can do, it's pretty cool.

00:23:41.080 | So now I can just ask about random things.

00:23:43.560 | Like, you know, I can ask Lex.

00:23:45.960 | I want to know what he thinks about World War II, right?

00:23:53.880 | I know this is, like, a favorite talking point of him,

00:23:57.160 | so there's probably plenty of things to talk about there.

00:24:03.600 | Okay, so, yeah, we get all this information.

00:24:05.920 | We're getting, like, pretty similar thing to before.

00:24:10.640 | And then we can kind of, like, draw down.

00:24:13.000 | We can ask, okay, cool, who's, like, this guy?

00:24:15.640 | I have no idea.

00:24:16.960 | So we can ask about that.

00:24:19.000 | Like, tell me more about the thingy episode,

00:24:23.200 | this episode, and it will actually be able to do this.

00:24:27.120 | You know, it's just going to refer back.

00:24:28.840 | And we can also just use this.

00:24:31.080 | Like, we don't have to use it as Lex plug-in

00:24:33.480 | every single time, right?

00:24:35.680 | So maybe I'll let this generate,

00:24:39.240 | and then we can ask another question.

00:24:41.800 | Maybe, you know, something about this makes us think,

00:24:45.440 | oh, I kind of want to know a little more about that.

00:24:48.000 | So maybe we can say,

00:24:49.760 | so there's this Ian Kershaw biography on Adolf Hitler.

00:24:53.760 | We can search that, see what it is.

00:24:56.880 | So what do you know about that?

00:25:04.560 | Now, I'm not sure if it's going to use the SLX plug-in.

00:25:07.640 | Okay, so here it's not using the SLX plug-in

00:25:10.200 | because it doesn't need to.

00:25:11.120 | It already probably knows some stuff about this.

00:25:14.280 | Now, my internet isn't great, so it just cut out,

00:25:17.680 | but maybe I can try and regenerate that.

00:25:25.200 | Okay, so we get a ton of information

00:25:26.920 | about this particular book that was kind of inspired

00:25:30.880 | by us reading through the Lex Friedman information.

00:25:35.240 | But this, in particular, we didn't need that

00:25:38.920 | in order to talk about this book.

00:25:40.720 | So Chad G. Petit knows this and has decided,

00:25:43.920 | okay, I don't need to ask Lex this question.

00:25:47.840 | I can actually just give you all this information

00:25:50.960 | from my training memory, essentially.

00:25:54.360 | So, I mean, I think that is all fascinating

00:25:58.440 | and it's really cool how quickly

00:26:01.400 | we can just build something like this.

00:26:04.320 | It doesn't take that much.

00:26:06.000 | Now, I haven't dived super deep

00:26:08.200 | into the technical side of things here.

00:26:11.960 | I showed you some of the things

00:26:13.520 | or I showed you some of the code.

00:26:15.120 | I showed you the GitHub repo.

00:26:18.600 | I think with that, you can pretty much do all of this.

00:26:22.400 | But if you are struggling to kind of follow along

00:26:26.240 | or you just want a bit more of a technical deep dive,

00:26:28.600 | I do actually have another video on Chad G. Petit plugins

00:26:33.480 | and specifically the OpenAI retrieval plugin

00:26:38.480 | that we saw at the start of this video.

00:26:41.080 | So if you like and you want to go into technical details,

00:26:45.240 | I would definitely recommend following along with that video

00:26:48.520 | as I actually go through, like from start to finish,

00:26:51.680 | building a plugin, not specifically for podcasts,

00:26:56.680 | but for basically online code docs.

00:27:00.360 | But yeah, for now, that's it for this video.

00:27:03.320 | I hope this has been interesting and insightful

00:27:07.080 | as to the sort of things that you can build

00:27:08.920 | with these Chad G. Petit plugins.

00:27:11.280 | If you are right now looking at where to find these plugins,

00:27:16.280 | there is right now still a wait list to get access to those.

00:27:21.400 | So I will make sure that the wait list link

00:27:24.480 | is somewhere in either the description of the video

00:27:26.520 | or even at the top of the video right now.

00:27:29.600 | So you can go and sign up for that

00:27:31.800 | if you don't have access already.

00:27:33.880 | For now, I will leave it there.

00:27:35.640 | Thank you very much for watching the video.

00:27:38.560 | I hope this has been helpful

00:27:41.040 | and I will see you again in the next one.

00:27:43.720 | Bye.

00:27:44.560 | (gentle music)

00:27:47.980 | (gentle music)

00:27:50.560 | (gentle music)

00:27:53.140 | (gentle music)

00:27:55.720 | (gentle music)

00:27:58.300 | you

Building a ChatGPT Plugin for Lex Fridman Podcasts

Chapters