AutoShow and Local LLMs with Monarch Wadia

Episode Description

Anthony and Monarch discuss AutoShow's CLI features, task orchestration tools, local LLM options, and AI's societal impact, joined by Fuzzy Bear for a spirited debate.

Episode Summary

Anthony Campolo walks through recent developments on AutoShow, his open-source CLI tool for generating transcripts and show notes from video, audio, and RSS feeds. He demonstrates how the project now supports multiple input types, configurable Whisper models, and modular architecture built with Commander.js. Monarch Wadia introduces the concept of task orchestration using tools like Spotify's Luigi, and the two explore Node.js equivalents such as Bull, Agenda, and Bree.js for managing complex pipelines with checkpointing and parallelization. The conversation shifts to running LLMs locally via node-llama-cpp and browser-based options like WebLLM, weighing trade-offs between API costs, privacy, and hardware requirements. They touch on LLM output evaluation challenges, the DSPy framework's approach to using logged outputs as training examples, and the evolving copyright landscape around AI-generated content. Guest Fuzzy Bear joins to argue that LLMs represent diminishing returns at the model level and that the real work now falls on developers to deliver the technology efficiently, emphasizing energy costs and the importance of open source as a counterbalance to big tech. The three settle into a philosophical exchange about whether interacting with LLMs constitutes coding in natural language, ultimately finding common ground that the technology functions as a scaled inference tool regardless of how it is described.

Chapters

00:00:00 - Introduction and AutoShow Overview

Anthony and Monarch catch up after a week off, touching on side projects and Monarch's interest in DAG workflows. Anthony then introduces the main topic: AutoShow, his open-source CLI project for processing video, audio, and podcast content into transcripts and show notes. He shares some AI-generated logo concepts and the two discuss which designs best capture the tool's identity.

Anthony walks through AutoShow's expanding input capabilities, showing how it handles YouTube videos, playlists, and now RSS feeds using the Fast XML Parser library. He demonstrates the modular project structure, including separate modules for processing individual items and full feeds, and explains how the Commander.js-based CLI accepts options for different input types and Whisper model configurations.

00:07:41 - Task Orchestration and Pipeline Management

Monarch introduces task orchestration as a solution for managing AutoShow's increasingly complex processing pipeline. He describes Spotify's Luigi framework, which provides DAG-based workflow management with checkpointing, parallelization, and failure recovery. The two discuss how processing an entire RSS feed with 94 episodes creates a multi-step pipeline where any stage could fail, making retry logic and checkpoint resumption essential.

They explore Node.js equivalents, evaluating Bull, Agenda, and Bree.js. Bull requires Redis, Agenda appears unmaintained, and Bree.js emerges as the most promising option since it avoids external database dependencies and supports parallelization. Monarch frames this as a broader pattern in AI development, comparing modern information pipelines to factory assembly lines and suggesting the industry is experiencing an industrial revolution in software development.

00:14:20 - Local LLMs, APIs, and Cost Trade-offs

Anthony demonstrates node-llama-cpp, showing how it enables running models like Llama 3 locally through a pure Node.js interface without Python dependencies. He walks through a basic hello world example using an eight-gigabyte model, and the two discuss whether tools like SageMaker are even necessary when node-llama-cpp can manage model deployment directly. The conversation turns to the spectrum of deployment options: fully local, self-hosted on a server, or API-based with associated costs.

Monarch shares that generating content for their side project currently costs five to ten dollars per run across multiple API calls, which becomes unsustainable at scale. He explains his motivation for refactoring with Luigi — to gain visibility into per-task costs so they can make informed scaling decisions. Anthony discusses the various transcription and LLM API integrations he's building, noting that services like AssemblyAI and Deepgram offer faster transcription but weaker summarization compared to feeding transcripts directly to language models.

00:24:06 - Browser LLMs, WebLLM, and the DSPy Framework

The conversation shifts to running LLMs in the browser through WebLLM, which downloads models locally and runs them through the WebGL layer. Monarch explains the privacy and cost advantages of client-side inference, while Anthony initially questions how much can run in a browser before understanding that the model is simply stored on the user's machine. They compare this approach with node-llama-cpp's backend-only model, noting WebLLM's limitation to a smaller set of pre-compiled model formats.

Earlier in this stretch, Monarch mentions LM Client, a library built around the DSPy paper's approach of using logged input-output pairs as examples to improve LLM performance. Anthony connects this to his own workflow in AutoShow, where he feeds example markdown files alongside prompts to guide model outputs. The discussion highlights how few-shot examples and structured prompts remain among the most effective techniques for improving LLM results.

00:31:13 - AI-Generated Content, Copyright, and the Semantic Web

Anthony and Monarch explore the broader implications of AI-powered content pipelines, including a classicist using translation workflows to unlock thousands of pages of untranslated ancient Greek medical texts. This leads to a discussion about copyright in the age of LLMs, touching on the New York Times lawsuit against OpenAI and how major AI companies have built legal protections into their terms of service for user-generated outputs.

The conversation turns to Google's shift toward AI-generated search results and the potential death of PageRank as AI-generated content floods the web. Monarch and Anthony realize that LLMs have effectively delivered the Semantic Web that was promised two decades ago, making data machine-readable through natural language understanding rather than structured metadata. They reflect on how information access is becoming fundamentally more semantic, with implications for search, content discovery, and the distinction between human and AI-generated material.

00:42:02 - Digital Empowerment, Accessibility, and Societal Impact

Anthony and Monarch debate whether technology is truly empowering people, with Anthony noting that most people he knows feel disoriented rather than empowered by digital tools. He argues that the chat interface is a breakthrough because anyone who can hold a conversation can use it, praising OpenAI's focus on that interaction model. Monarch compares LLMs to the calculator and the printing press as tools that democratize capability.

The discussion touches on personal anecdotes about family members discovering ChatGPT, the gap between technical and non-technical users' experiences with technology, and a chat participant named Fuzzy Bear who has been expressing skepticism about AI hype. Anthony and Monarch set the stage for Fuzzy to join the conversation, while also sharing their goals for the coming week: Monarch plans to refactor his pipeline and generate more artifacts, while Anthony aims to integrate a Node-based Whisper wrapper and build out additional model API options.

00:55:06 - Fuzzy Bear Joins: Energy, Open Source, and Big Tech

Fuzzy Bear joins the stream and immediately acknowledges the value of open-source AI tooling while raising concerns about diminishing returns at the model level. He argues that the real innovation happened in the research papers and neural network architectures, and that current development is iteration rather than true innovation. He highlights the energy crisis facing LLMs, noting projections of a thousand terawatt hours annually by 2030, and draws parallels to blockchain's efficiency problems.

Fuzzy also warns about big tech acquiring small AI startups to suppress competition and emphasizes that open source must serve as a counterbalance to corporate monopolies over AI technology. He references the Star Trek computer as a prescient vision of where computing is heading, and Monarch agrees, noting that people with no coding experience are now building integrations and agentic workflows within months using ChatGPT. The discussion frames the current moment as one where the barrier between developers and non-developers is rapidly dissolving.

01:17:13 - Language, Intelligence, and Natural Language as Code

The three hosts engage in a philosophical debate about what it means to interact with an LLM. Fuzzy argues that prompting is fundamentally coding in English, making natural language a domain-specific language for AI systems. Anthony pushes back, insisting that conversational interaction with ChatGPT is qualitatively different from coding. Monarch mediates, pointing out that both perspectives describe the same system from different angles and that the disagreement is primarily linguistic rather than substantive.

Fuzzy introduces the idea that AI should be called "additive features" rather than "artificial intelligence" to avoid anthropomorphization, and advocates using "inference" to describe what these systems actually do. Monarch finds alignment by framing the human brain itself as an inference engine, drawing parallels between biological neural networks and artificial ones. The conversation closes with all three acknowledging common ground: regardless of terminology, the technology scales human inference capabilities in ways that are genuinely transformative, even if the underlying mechanism is far simpler than popular narratives suggest.

Transcript

00:00:02 - Anthony Campolo

All right. We are live. Hello everyone. Welcome back to AJC and the Web Devs with our recurring guest, Monarch Wadia. What's up, man? What've you been up to? We were off last week, so this is our first time in two weeks. Is it warming up near you yet? It's warming up near me.

00:00:26 - Monarch Wadia

It's getting a lot warmer. Yeah, it's nice. Therefore feeding my Coke habit. Cheers. I'm really enjoying life. It's summer, it's a nice day. I've been getting into DAG workflows and task orchestration. We're not going to talk about that today, I don't think, but yeah, stuff's good.

00:00:55 - Anthony Campolo

Yeah. And we have a super secret side project we're working on, which I don't think we're going to talk about, but super secret side project.

00:01:04 - Monarch Wadia

Maybe we should build a few more of those things that we're building, maybe ten of them, and then share it.

00:01:12 - Anthony Campolo

No, totally. I mean, a bunch of my friends were like, this is very cool. So people are into it, seems like. Nice.

But today we're going to talk about AutoShow, which I've been working on a ton. Been just building it out, making it more legit in terms of project structure and docs and building out the capabilities and making it idiomatic to what a CLI should do. We got Fuzzy in the chat. What's up, Fuzzy.

So yeah, anything else you want to chat about that you've been learning or working on recently before I dive into my stuff?

00:01:57 - Monarch Wadia

No, let's dive into it. Do you want to do a quick rundown of what AutoShow is?

00:02:05 - Anthony Campolo

Yeah, totally. So AutoShow is basically my workflow that I've translated to work with any kind of content, whether it's video or audio, and creating transcripts. And I'm using the Stage Manager now instead of the Docs. I've been getting used to a new workflow, but check these out. I generated some potential cool images for AutoShow. Do any of these strike you?

00:02:55 - Monarch Wadia

I like the first one. I think that was really cool. The one after that one, this one, I think this one's my personal favorite because it's just so cool. I don't know, it's got this nostalgic, retro-futuristic, but AI-generated vibe to it. And I like this one.

00:03:16 - Anthony Campolo

Yeah, I like the colors. It's actually in line with my rainbow parakeet also, which is nice.

00:03:25 - Monarch Wadia

That's true.

00:03:27 - Anthony Campolo

Yeah. I like this one because I always liked this mic image for podcasting in general. That just always strikes me as a quintessential podcast kind of image. But it's not just for podcasts. And then this one is nice, but it's also kind of generic looking.

00:03:46 - Monarch Wadia split

I don't like this one. So rule it out.

00:03:48 - Anthony Campolo

Yeah, exactly. It's the easiest, the simplest thing. Or this one, this has a little personality to it.

00:03:54 - Monarch Wadia split

Totally.

00:03:55 - Monarch Wadia

I like this one. Okay. I like the other one too. It has that video reel in the background too, you know, it's not just audio.

00:04:03 - Anthony Campolo

So it's powerful. It reminds me of Boogie Nights. Have you seen Boogie Nights?

00:04:11 - Monarch Wadia

No. But just by the name, I can kind of imagine.

00:04:15 - Anthony Campolo

It's a great movie. It's about the porn industry in the 70s. And Mark Wahlberg is the main character, and he comes up with this name for his porn name. And he imagines it in big lights, like AutoShow. So it reminds me of that. It's a great movie.

00:04:33 - Monarch Wadia

Gotcha.

00:04:34 - Anthony Campolo

So the cool thing about AutoShow is that it works with a lot of different input types. So when I first was doing this, I wanted to do it for my podcast, FSG, which is here, but I ended up going through a circuitous route where I first started doing it with YouTube videos and YouTube playlists, because the yt-dlp was a really nice tool to manage that part.

But now you can feed it anything you want, even for stuff like the video. If you just feed it an actual file, it can still basically figure it out. And so I've broken it up into actual modules. And if we look at this, we have the process video. This is the base one where it takes in the video from YouTube and creates the front matter and then runs transcription and then basically attaches a prompt to it that can get fed to an LLM. And there's also process playlists, which will run on multiple videos.

[00:06:03] And then now I have process RSS feed, which uses this Fast XML Parser library, which is pretty sweet. It just gives you a really simple way to turn XML into JSON data. And then you can work with that in a similar way to how I'm working with the outputs of it.

So that would be this one. Yeah, so this is process RSS item which pulls out all this stuff from the podcast feed and then runs all the same stuff on it. And then here it processes the full feed and iterates over the items.

So yeah, there's different structures for RSS feeds that would have also blog content or even YouTube channels themselves have RSS feeds. But I find for most that you would actually listen to through a podcast player, they basically standardize on this kind of format. They give you a very specific breakdown into items. Each item has an enclosure which has a URL, and then there's a title and all that kind of stuff.

[00:07:31] So this is cool because there's so much podcast content on RSS feeds and now you can run all those through.

00:07:41 - Monarch Wadia

Gotcha. And then so could you... oh go ahead.

00:07:46 - Anthony Campolo

Well, as you say, the last thing is I also added the ability to configure which model you're running for Whisper transcription, because if you want to do a quick demo, you could do the base one. It'll run ten times faster than the large but with not as good accuracy. And then these are sort of the more nitty gritty file management type stuff.

00:08:12 - Monarch Wadia

Nice, nice. So does the RSS code call the code that you already had, that just runs the other commands under the hood? Or how's that structured?

00:08:32 - Anthony Campolo

Yeah. So now Autogen, which used to be this one monolithic file that did everything...

00:08:40 - Monarch Wadia

And this is the Commander. So this is the argument parser. So this is where the CLI entry point starts. Yeah, exactly.

00:08:46 - Anthony Campolo

So this is where you're actually defining now each of the options. And so there's an option for video, playlist URLs, RSS, and then model. And then a short alias option as well.

00:09:01 - Monarch Wadia

So you can do like autogen dash R and then the URL for your RSS feed. And that's how you can actually parse the RSS feed, right?

00:09:10 - Anthony Campolo

Yeah. And then I'm at the point now where I also have a couple of things that aren't fully integrated yet, like different transcription APIs. So we'll be able to feed those to the CLI if you want to pay to get something faster without running Whisper yourself. And then same thing for LLMs, like I've got Claude integrations and ChatGPT integrations.

So yeah, there's a lot of things that are kind of half built out that need to be connected still. But it's all coming together. It's really cool, and it's giving me a lot of different things to work with in the AI space right now, because you can do everything through APIs and it has a cost, or you can do everything local, but then it's this crazy local dependency nightmare. And you have to download massive gigabytes of stuff. So it's all these trade-offs with all these different options.

[00:10:04] So eventually it gets to the point where I'm going to run a crap load of benchmarks that are going to be mixing and matching all these different options, like APIs versus local and all the trade-offs of cost and speed. And then just how the outputs compare to all that, and I'll have like 20 outputs to look at and be like, okay, which of these suck and which of these are pretty good? And which are all kind of in the middle. So that part is coming soon.

00:10:33 - Monarch Wadia

Nice dude. That's awesome. I think you're doing a lot of things in a very similar way to what I'm building out for us. We're sort of going in parallel tracks, so we should talk about that stuff later, because I don't think we're ready to share what we're building. But we should talk about it because you're also doing stepwise operations. And it sounds like you can make stuff even more and more complex. I can see a place here for agents, for example.

00:11:10 - Anthony Campolo

Yeah.

00:11:10 - Monarch Wadia

So have you considered using any workflow or task management, like a task runner or anything like that?

00:11:19 - Anthony Campolo

No, not yet, because it's kind of a clearly defined amount of things that need to be run right now. But there will be a point where obviously it'll make sense to have that kind of stuff. What would be an example of that that I would look at?

00:11:44 - Monarch Wadia

So I'm using this thing called Luigi. And my use case is different because I want the stuff that I'm building to not be a CLI tool, but be a server component. But there's lighter weight stuff too. So what Luigi is... yeah, it's by Spotify. That one.

00:12:08 - Anthony Campolo

Yeah. And I was telling you how I'd heard this through podcasts back in the day, but had no idea what it did. Let me look at some code. I want to see some examples.

00:12:22 - Monarch Wadia

Gotcha.

00:12:23 - Anthony Campolo

These are not helpful actually.

00:12:27 - Monarch Wadia

I can give you a quick rundown. So actually if you go lower there, that's sort of what this gives you. What it gives you is a graph of tasks. And those tasks can be dynamically defined or statically defined, but it gives you a workflow management engine so that each task can be controlled and logged and monitored separately and even parallelized.

So if you're taking text and you're generating 500 different images from a piece of text, then you want to parallelize that and you want to also see that all those 500 image generation calls go successfully.

00:13:09 - Anthony Campolo

So this is what I need then for when you want to run on a playlist or an RSS feed, like you just hit every single episode at once.

00:13:19 - Monarch Wadia

In theory it would help, because what this also does is it gives you the ability to have checkpoints. So let's say you got the RSS feed. The first thing you want to do in your case is you want to fetch the RSS XML, which is for those who don't know, it's a big XML text object, and it has a list of all the different episodes that are on the RSS feed. Then you want to go through all those RSS feed episodes, and then you want to download the YouTube video for all of those.

00:13:54 - Anthony Campolo

So here's an example. Go to this RSS right here. So we have individual items for each episode. There's title. And then here we got the URL. So you got 94 of these, right?

00:14:20 - Monarch Wadia

So you have 94 different URLs that you want to first parse. First you got to get the RSS. And that might fail, that call might fail. Then you want to parse those 94 episodes out and that parsing might fail.

Then, after you've parsed it, for each of those 94 episodes, you want to download a YouTube video for each of those. And that's definitely going to have failures. If you're going to try and download 100 YouTube videos, you're going to have one or two failures that need to be retried because of rate limits or because of some issue with your network, whatever it is.

Now after you download those, you want to generate transcripts from those. And then after you generate the transcript, you want to generate summaries. So you have this complex, beautiful pipeline already. And having a task runner gives you the ability to, you know, if you've already downloaded the YouTube video, you don't want to download it again just because you're running the command again.

[00:15:21] You just want to use the darn YouTube video. So a task runner could give you that. Luigi definitely gives you that. So you don't have to download that YouTube video over and over again. It just gives you a checkpoint and you just continue from checkpoint forward. And that's what I'm using it for. I wonder if there's something like that for Node.js.

00:15:45 - Anthony Campolo

That's what I'm doing right now. I just went off screen because I want to show my ChatGPT history, but I wanted to ask it, what is the Node.js equivalent for Luigi? The task manager or task runner?

00:16:06 - Monarch Wadia

Yeah.

00:16:16 - Anthony Campolo

Bull and Agenda. Interesting.

00:16:21 - Monarch Wadia

Okay. Or a full workflow automation tool.

00:16:27 - Anthony Campolo

I feel like I've heard of this. This is AI native. That's nice.

00:16:33 - Monarch Wadia

I don't know what that means.

00:16:34 - Anthony Campolo

I know, right? Everyone's saying that.

00:16:40 - Monarch Wadia

Yeah. So how these things work is these might be overkill for a CLI. And maybe you want to ask which one of these is appropriate for a CLI that's only going to be run locally. Because this is supposed to be server stuff.

00:17:04 - Anthony Campolo

I'm building a CLI with Commander.js so it will be a good fit. It's suggesting Bull or Agenda. For Bull...

00:17:13 - Monarch Wadia

You'll need a Redis server running, which is kind of a bummer.

00:17:17 - Anthony Campolo

Yeah. Let's see. Let's check out Agenda. Says lightweight job scheduling. Except this website is not secure. It's not a good sign. Let's see. Oh yeah, this hasn't been updated in two years. I just don't like using open source projects like that. Honestly.

00:17:41 - Monarch Wadia

I think in JavaScript, that's a fair concern. Yeah.

00:17:44 - Anthony Campolo

Yeah. That's like six Node versions behind. CI says Bull's currently in maintenance mode. Okay. So I feel like this is just the type of thing that Node.js hasn't really solved as a package. And a lot of people use services to do this kind of stuff. And then, you know, maybe overkill. If it does the thing I need it to do, then why not?

00:18:18 - Monarch Wadia

Yeah, I'm still looking into it. Bree.js is another one that came up. B-R-E-E. It has cron, so I know that it's robust. I just wonder if... oh yeah. Bree might be it because it looks like it has parallelization.

00:18:45 - Anthony Campolo

Yeah. That's cool.

00:18:49 - Monarch Wadia

I think... let's see. [00:18:54] - Anthony Campolo Bree does not force you to use an additional database layer of Redis or MongoDB. Great. That's exactly what I wanted. Yeah, no, I feel like I've heard this before a while ago, but I never got to a point where I actually had a project where I needed this kind of thing.

00:19:11 - Monarch Wadia

Yeah. It'll give you... you could always build checkpoint logic yourself. But what a framework like this gives you is a lot of things that are unit tested and robust out of the box. And you don't have to worry about the architecture of how to build something like a DAG from scratch.

Yeah, this is I think going to be a common problem in the world of AI. Having just a multi-hop agent or agent workflows is not going to be enough. We need pipelines to do stuff that's very well defined, almost like a factory pipeline. I think we're kind of living through the industrial revolution of software development. We're building factory pipelines for information.

00:20:06 - Anthony Campolo

Have you used SageMaker?

00:20:09 - Monarch Wadia

I have not. No. But isn't SageMaker more for training LLMs and stuff?

00:20:15 - Anthony Campolo

Yes. It also deploys them, but it's probably just one of the parts. And then there's model monitoring and debugging.

00:20:29 - Monarch Wadia

Gotcha.

00:20:30 - Anthony Campolo

These are all just... have you used Weights and Biases?

00:20:33 - Monarch Wadia

I mean, like...

00:20:36 - Anthony Campolo

I don't really know that form of it.

00:20:38 - Monarch Wadia

Is that a platform?

00:20:40 - Anthony Campolo

I think it's general tooling. They have MLOps solutions, LLMOps solutions.

00:20:56 - Monarch Wadia

Now I'm a Grug brain developer man. I don't train LLMs. I just work with them, you know? So for better or worse, I'm a systems integrator now. I'm not really the one building the brains.

00:21:14 - Anthony Campolo

Yeah, I feel like the things that need to be tested and verified is once you start outputting stuff with custom data, how do you make sense of whether you're getting good outputs or not, and which kind of models you should be using. Which is not traditional model testing and iterating and training. It's just you're mostly working with a foundation model, but you're tweaking different prompts with different sets of data you're querying on. And so that seems like the thing you want to test around and iterate on. But that's not an easy thing to have a one-size-fits-all solution like training models.

00:21:56 - Monarch Wadia

Yeah, the testing is something that very few people have figured out. I think a lot of people are just paying lip service to the need to test LLMs. A lot of companies are just paying lip service to it.

00:22:11 - Anthony Campolo

Well, everyone's...

00:22:12 - Monarch Wadia

Running...

00:22:12 - Anthony Campolo

These benchmarks that are considered the benchmarks. And they're useful signals for people who are working with these things. They don't really tell you how good these things are at anything.

00:22:27 - Monarch Wadia

Yeah. We're so early on in LLMs that I don't really think a lot of companies have cracked the output evaluation yet.

00:22:43 - Anthony Campolo

Mostly it's just that's where the human in the loop is. Check the output. That's the one job that humans are going to have left.

00:22:51 - Monarch Wadia

Right? Yeah. I was looking into this thing called the DSPy framework, and one of the hypotheses over there was if you can log the outputs of your pipeline, then you can use the outputs as examples. And if you can make input-output example pairs for your LLMs, that's the best way to actually instruct them. You can always tell your LLM, "Hey, I want you to write in the style of Shakespeare," but if you actually gave it examples of Shakespeare, then it'll do a lot better.

00:23:31 - Anthony Campolo

That's the thing I do with AutoShow. I give it a prompt and I tell it what I want, and then I give it an actual markdown file with the example of the structured data and output that I want it to give me.

00:23:42 - Monarch Wadia

Exactly. So I was talking to the guy who made LM Client. He's at Dasco on Twitter, a really smart guy, really nice guy. And he was telling me about the DSPy paper. So he built LM Client around the DSPy paper.

00:24:05 - Anthony Campolo

Oops.

00:24:06 - Monarch Wadia

So full disclosure, I haven't played with LM Client yet, but it looks promising. It's sort of like an LLM library. It's lighter weight than LangChain, for example.

00:24:21 - Anthony Campolo

It's been around for almost a year now.

00:24:23 - Monarch Wadia

Yeah.

00:24:25 - Anthony Campolo

That's good.

[00:24:26] - Monarch Wadia

00:24:27 - Anthony Campolo

Cool. Yeah, I'm pretty sure you're the one who first told me about this. I haven't tried it. I've kind of shied away from anything that runs in the browser because I'm not sure how... because I know how much the actual models that work really well, how huge and ridiculous they are. And I feel like you could just do an API call for a really basic model and get a similar output, but maybe I'm wrong because I haven't actually used them that much. So I need to try and see what it's like.

00:25:01 - Monarch Wadia

Yeah. Well, I mean, it might not be a good use case for mobile, but for desktop, I have a gigabit connection and I can download a model in less than a minute. It's not a huge lift. And those models can be downloaded right into the browser.

So once you have... the nice thing about having an LLM in the browser is that now everything is private by default. You're actually not doing any API calls. Your prompts are not going over the wire anywhere. So you have privacy and security that way.

And because everything is local, you as a developer, you're not paying anymore for the LLM. The user is running the LLM on their own machine. Also because they're running it on their own machine, if they have a powerful machine like an M3, for example, we both now know someone who has an M3. And having powerful hardware helps.

00:26:05 - Anthony Campolo

Well, in that case, that's not really running in the browser and that's just using the browser to call out to the underlying whatever someone's hardware is.

00:26:15 - Monarch Wadia

Well, I mean it goes through the browser though. It goes through the WebGL layer, not the OpenGL layer, but it...

00:26:22 - Anthony Campolo

So at the end of the day, it's saved a three gigabyte model on your machine.

00:26:28 - Monarch Wadia

Yes, yes.

00:26:29 - Anthony Campolo

Okay. So yeah, then it's all kind of the same thing at the end of the day then.

00:26:34 - Monarch Wadia

Yeah, exactly.

00:26:34 - Anthony Campolo

In my mind I'm thinking if you're running a model in the client, how much can you fit in the client? But what you're saying is it's just giving you another way to download the thing onto your machine, just through a different API interface.

00:26:47 - Monarch Wadia

Yes, exactly. And it also, you can npm package your code so the library itself is just an npm package. You just include it as a dependency.

00:27:02 - Anthony Campolo

And that's exactly why I was looking at node-llama-cpp, because this is a similar thing where it's just a package you install and then you can run all sorts of... you can pick a model and then just download whichever model you want and then run your prompts and whatnot.

00:27:24 - Monarch Wadia

Gotcha. Now node-llama-cpp is backend only though, right?

00:27:29 - Anthony Campolo

Yes. But my project right now assumes everything is Node.js anyway, so there is no client code at all in what I have right now. So I figure once I get everything running on the backend with Node.js, and then I'm not making weird C++ or brew library stuff, I need to first get rid of that. And then once it's running in all Node, then just flipping all those to browser compatible stuff.

00:27:55 - Monarch Wadia

Gotcha.

00:27:56 - Anthony Campolo

Which is easier and easier now with Node.js because it uses fetch and all this other stuff.

00:28:01 - Monarch Wadia

Gotcha. Yeah. One of the downsides of WebLLM is that it only takes a very specific file format for its weights. So the model file has to be pre-compiled by WebLLM. And they only have a small, relatively small list of LLMs that are available. So they just added Llama 2 and Mistral 7B. But I don't know if they have Llama 3 yet.

00:28:30 - Anthony Campolo

So yeah, if... let me go to the... so node-llama-cpp gives you all of these available right now. So you have Llama 2, Llama 3, Code Llama. You have Mistral, some crazy Dolphin thing that apparently is an uncensored Mistral mixture of experts thing. That's super awesome apparently. And then someone I've never heard of, Gemma and Orca.

So this is all available in node-llama-cpp and you can run a start command. This is technically the version 3 beta, which is not out quite yet, but it comes with this suite to create projects where you scaffold up projects that use whichever model. Let me actually show you that.

00:29:19 - Monarch Wadia

That's really neat.

00:29:21 - Anthony Campolo

So just run this guy.

00:29:25 - Monarch Wadia

You don't have to run Python at all. This is all Node.js.

00:29:28 - Anthony Campolo

Yeah, it's right through. So let me pick a small one because this is going to take a long time. Let me do this one here.

00:29:41 - Monarch Wadia

So do you even need a special LLMOps deployment for this? I mean, isn't this just going to manage all of it for you and deploy it on a server? So do we even need to worry about SageMaker at this point, or can we just use node-llama-cpp?

00:29:58 - Anthony Campolo

Yeah. And this is kind of where I've been going with all this stuff. I haven't needed to reach out for a more complicated thing to manage, just finding ways to do each part incrementally.

And the big thing is basically whether you're running things locally or you're going to want to... because a lot of this can then be ported to if you wanted to run this on a server. And a lot of these will have Docker integrations and stuff like that. So if you wanted to spin up your own thing that then you're still not making API charges, if you have the ability to spin up a Node server, which is hard or easy depending on who you are, that's a cool middle ground too that I'm eventually going to want to try out for myself, because I would like to have a $5 DigitalOcean droplet that I can just have constantly running transcriptions and stuff.

[00:30:48] For me, that would be pretty cool. And then once you build in the APIs, it's just going to run and give you an instant output, but you're going to have to pay. That's where I can then start to have paid options to give you access to these APIs and a really fast instant output with very little technical knowledge required. But then they'll be paying to cover the margins of what it costs for me to use the APIs.

00:31:13 - Monarch Wadia

Yeah. That's interesting. Right now to generate one of those things that we're working on, it costs us like five bucks to ten bucks to generate one.

00:31:25 - Anthony Campolo

Oh, really? Interesting.

00:31:28 - Monarch Wadia

Yeah. Because there's a lot of calls going out to a lot of different LLMs. And right now I haven't done any deployments of models. I'm just calling APIs of various sorts, and it costs 5 to 10 bucks, which isn't bad unless you scale it up. And once you scale it up, it's like, okay, well, identify...

00:31:50 - Anthony Campolo

If you do a video every day that gets expensive pretty quickly.

00:31:53 - Monarch Wadia

Oh yeah. We're using audio, video, and we're using traditional text-based text-to-text models. I can see scope for using multimodal models too. And I don't want us to pay for that over and over and over again, you know?

00:32:13 - Anthony Campolo

Sort of how you're managing costs would be nice.

00:32:17 - Monarch Wadia

Exactly. So right now I'm refactoring because I want to make sure that we know what we're paying for. And once we know what we're paying for, we can make decisions like, okay, if we want to scale this up to 500 artifacts generated a day, then how much is that going to cost us per day? And how can we reduce those costs?

That conversation can't be had until we know what's costing us, and we can't know what's costing us until we have discrete jobs in Luigi, where those discrete jobs are being run separately.

00:32:56 - Anthony Campolo

Gotcha. Okay. Yeah.

00:32:59 - Monarch Wadia

It's a good time.

00:32:59 - Anthony Campolo

So I stopped the downloading because it was making my whole thing lag. And I have one that's already downloaded here. So just so people can get a sense of what's happening here.

This is a quick getting started thing with node-llama-cpp. So I've got a really bare bones project, just an index.js file. And then the models directory with this model right here. So this is Llama 3, 8 billion parameters, instruct, one of the Q8_0s for the quantization level. And then there's just two dependencies. This is originally a TypeScript thing. I stripped out all the TypeScript and now it's just Chalk and node-llama-cpp.

And this is not the best example because it's an instruct model. So you have to prompt it and create a back and forth conversation between user and AI. So I'm actually going to just pull this out and just keep a prompt to a chat interface. But this is the bare bones output it gives you. So I'm just going to run this and hopefully it works.

[00:34:15] Okay.

00:34:16 - Monarch Wadia

And I haven't seen Chalk in so long. I love Chalk. When I first found it, it was like wow, this is cool.

00:34:24 - Anthony Campolo

So what this is doing is... oh there it goes.

00:34:29 - Monarch Wadia

And all this is local?

00:34:32 - Anthony Campolo

Yeah. So this model is like eight gigabytes or something. Let me see here. "Hi there, how are you?" "I'm doing well, thank you for asking." And then you ask it to identify verbs in a sentence.

So I was actually listening to an interesting interview with one of the main guys at OpenAI. He was talking about how this is the big unlock for ChatGPT, going from these instruction things to just being able to prompt the chat models directly without having to feed them this back and forth to make it simulate a user-chatbot kind of interface.

00:35:15 - Monarch Wadia

Yeah.

00:35:19 - Anthony Campolo

So yeah, that's your node-llama-cpp hello world. And then AutoShow also has right now a couple APIs available that you just string together. And then these two transcription things, which we talked about last time, AssemblyAI and Deepgram.

Their summaries and chapters are way worse than just feeding a transcript directly to the model and prompting it. So the real thing that's nice with these is you get way faster transcripts and also fairly good speaker diarization, which is only experimentally supported right now in whisper.cpp. And that's really important if you don't want to spend a ridiculous amount of time actually editing the transcripts. Having to put each individual speaker and separate those manually defeats the whole purpose. You might as well write the full transcript anyway.

00:36:18 - Monarch Wadia

So are you building transcripts or are you also doing summaries?

00:36:22 - Anthony Campolo

So it starts with the transcript. And then the transcript gets fed to the LLM with the prompts. That creates every part of the show notes, which is configurable: different lengths of summaries, you can get titles or not, you can get chapters and descriptions of the chapters, you can get key takeaways, you can get tangents.

You can do whatever you want, but the basis is the transcript with the timestamp. So you just feed that whole thing with a prompt, and then you have the transcript and the output that you can put together and create the total SEO optimized web page.

00:37:03 - Monarch Wadia

Gotcha. Wow. Okay. And then you could feed that and turn that into a video by itself. But this is really neat. The nice thing about all these things is that because it's multimedia and now you can actually do programming on the text itself, all of a sudden the difference between text and media just doesn't exist anymore in a way.

00:37:35 - Anthony Campolo

No, it's really nuts. Just being able to have AI watch all this content for us and summarize it and make sense of it and create new things off of it. It's a huge unlock.

It's just a really interesting interview actually, with this classicist who was talking about Galen, one of the most published Greek doctors of all time, and how he has 20,000 pages of writing with 10% of it translated into English. And so all these classics can now have this pipeline where 99% of it is translated into English. And then you just have a couple grad students all of a sudden with this whole canon of new translations of old ancient texts. Stuff like that. The possibilities are endless.

00:38:28 - Monarch Wadia

I wonder, the copyright gray area is concerning, just legally, not ethically. Because ethically, whether you translate by hand or use a machine to translate, I personally don't really put any ethical value to one way of doing it or the other. But legally, it's an interesting question, right? Because who owns the derived materials? Is it you? Is it the original author?

00:38:54 - Anthony Campolo

Well, it's a lot simpler when you're talking about Greek texts. That's public domain. No one can claim to own Plato on a monetary basis, whereas someone who wrote, like you're a sci-fi author, or if you're like Harry Potter, that's a whole different claim.

00:39:14 - Monarch Wadia

That's true. I guess if you're translating Galen, who seems to be an ancient Greek doctor, then it's going to be in the public domain. But then, if you generate a piece of content using an LLM, is that copyrightable? Have courts decided on that yet?

00:39:38 - Anthony Campolo

So this is an interesting thing. This is working its way through the courts right now because the New York Times is suing OpenAI, and that's going to decide a whole bunch of the parameters around copyright and training and things like that.

But what I've heard is that OpenAI and Anthropic and all these companies have actually written in their terms of service, and this is pretty cool, that anything you generate with them is not copyrighted to anybody. It's something new and you can use it yourself and you cannot be sued. Or if you get sued for copyright infringement, they will defend you in court.

This is what I've heard on podcasts at least. I thought it was really interesting when I heard that. And there was one LLM company that was not going to do that, and it was this big thing because all the other ones had already done it. So I did a whole conversation about this. It was really fascinating.

[00:40:31] This was like six months ago or so, so things could be totally different now. All this stuff is moving so fast. But at the time, they kind of built in some legal protections because they knew so many people would be generating so much content and using it for everything.

00:40:45 - Monarch Wadia

Yeah.

00:40:45 - Anthony Campolo

That would be a huge Pandora's box. All of a sudden, everyone could be sued for anything they created with ChatGPT.

00:40:52 - Monarch Wadia

Yeah. Did you hear about this? It's related and I'll tie it back. But Google recently now answers with AI-generated results, and it's kind of putting its web results into a separate tab altogether. So if you wanted just the traditional Google search results, which have been around for 20, 30 years, then you actually have to click on a separate tab. And that's how you can get the old-fashioned results.

But now everything is a mish-mash of videos and AI responses and web results and stuff you can buy and sponsored results. It's a mish-mash now after you do a search. I think PageRank is probably dead or dying at this point because there's so much content that's relatively high quality and on topic that AI is going to generate. Everybody's going to be generating AI content on all sorts of materials, and who knows what's AI-generated and what's not, and who knows what's real and what's fake.

[00:41:57] The PageRank algorithm is probably effectively done.

00:42:02 - Anthony Campolo

I wonder how much inertia you get from people just going to Google to look for things. Because if you're someone who searches for websites and things like that, you still go through Google, but a lot of people are just in their social apps. They'll search for things on YouTube or whatever.

00:42:24 - Monarch Wadia

Oh yeah, that too. It's a weird world. The way we access information is becoming more semantic.

00:42:37 - Anthony Campolo

The semantic web. Are you telling me we got to the Semantic Web 20 years later?

00:42:43 - Monarch Wadia

I suppose I am, right.

00:42:49 - Anthony Campolo

Semantic again?

00:42:53 - Monarch Wadia

Isn't that...

00:42:53 - Anthony Campolo

What semantic?

00:42:54 - Monarch Wadia

So semantic web, correct me if I'm wrong, but semantic web was where people decided to put metadata all over the HTML and use special HTML tags that were meaningful instead of using divs everywhere, so that the web content would actually be machine readable in a meaningful way. Am I right or am I wrong?

00:43:13 - Anthony Campolo

Yes. And it was always coupled with the idea of Web 2.0. They always said Web 2.0 is a semantic web, and they would say it's about user generated content instead of just website generated content. And those two things were always kind of coupled together, but really were different concepts actually.

And then the semantic web just kind of went away as a concept. And Web3 was in relation to the user generated content, and they're saying now the user generated content has permissions as well. So the whole semantic thing had nothing to do with Web 3.0. They kind of just threw that out.

But the semantic web now with LLMs is real. We actually have the semantic web because of LLMs. It makes all our data semantic in a way that you couldn't with knowledge graphs.

00:43:59 - Monarch Wadia

Yeah, it's sort of weird because there's all this stuff out there. I'm sure everybody has something on the internet with their real name on it that they're not proud of, and all that stuff is now mineable.

00:44:17 - Anthony Campolo

I posted a lot on message boards in high school with accounts that, since I was a teenager, I'd be like, yeah, obviously I said that as a teenager. If I said anything stupid, I'd be like, I'm a different person, I'm grown now. But I just wonder, there's thousands and thousands of messages.

00:44:39 - Monarch Wadia

The law is going to have to keep up with it, because can you imagine the number of frivolous lawsuits? I'm afraid to even put this idea out there. Can you imagine the number of frivolous, tort style lawsuits that will come up from people who will mine your history? Lawyers will just go mine your history, see if they can get you for something, and then just sue you over that.

00:45:06 - Anthony Campolo

It's like patent trolls. It's gonna be patent trolls on an individual basis. It'll be like ransomware.

00:45:13 - Monarch Wadia

Yeah, I can imagine. Especially if the lawsuits themselves can be generated using an LLM.

00:45:20 - Anthony Campolo

That'll empower lawyers who use more LLMs to handle more cases.

00:45:26 - Monarch Wadia

We're seeing the empowerment of each individual, really. I mean, that's what LLMs are, right? With technology, everybody starts getting superpowers. Now in 2024, everybody's telepathic. Everybody can talk to each other using nothing but a little box that's sitting on their desktop. I can see you across the world. I can start a business just sitting in my pajamas over here. Everybody is just way more empowered by default, whether people use that or not.

00:46:02 - Anthony Campolo

People don't feel that way, though. That's what's so crazy about this time. So few people that I know personally, friends or family, feel that way. They feel very unempowered by this constant encroaching digital stuff. It feels weird and alien and disorienting.

And I understand. When you don't actually know how things work, you can't cope when you don't understand what's lying behind them. When things are broken, you don't know why. You have no idea how to fix things. It's just like, things are broken, what do I do?

We still have that experience, but we can dig down a layer deeper, like, hey, it's broken for a reason because there's a system that's connected to things that are connected to other things. When I give it an input, it's doing this chain of things that comes back to me. That whole conception, most people don't have that with tech. So it's hard to feel empowered around it.

00:46:52 - Monarch Wadia

That's fair. Although it's weird because on one hand there's the whole disempowerment and that's super valid and not enough has been said about it, and more has to be.

00:47:05 - Anthony Campolo

This is...

00:47:06 - Monarch Wadia

Also why...

00:47:07 - Anthony Campolo

I still think there's something to be said for the chat interface. Some people are saying we need to go beyond the chat interface because they want AI to just be built into all their applications. But I feel like forcing you to sit down, talk to a thing, and get a response back, no matter who you are as a person, you have that ability. If you have the ability to talk to someone, you know how to do it.

So I really like that. And I feel like there's such a brilliance to what OpenAI did about really honing in on that interface and making that the value prop. Because some people still don't spend the time to learn how to talk to it, right? But anyone who is curious, it's all there.

00:47:47 - Monarch Wadia

I have people our age in the family who are just getting on to ChatGPT, even though I've been telling them for years, hey, you guys should get on AI. And now they're just getting into it, and their minds are being blown by how powerful this thing is and how useful it is.

00:48:02 - Anthony Campolo

My dad really likes ChatGPT. He uses it for all sorts of things. It's always funny, kind of who will get into these things. It's like the very old and the very young sometimes end up on the same side, like the horseshoe thing.

00:48:15 - Monarch Wadia

Oh yeah. It's a funny world. I think it's incredibly empowering. It's about as empowering as a calculator. You no longer have to do all this crazy math in your head. You no longer have to be an engineer with a slide rule to figure out the height of a building. You can just do it with a bunch of relatively easy calculations.

The calculator is sort of one of those things. The printing press was one of those things, right? Before, people didn't have the money to buy a book, which could cost you a year's salary. But after the printing press, you could go out and buy a book for a week's salary.

00:48:56 - Anthony Campolo

Yeah, I was gonna say a week.

00:48:59 - Monarch Wadia

And it's like...

00:49:01 - Anthony Campolo

It's like Fuzzy as our resident contrarian calming the GPT hype.

00:49:08 - Monarch Wadia

Yeah. I find it difficult to... I don't know.

00:49:16 - Anthony Campolo

I think people compare the internet to the printing press and to me that's a metaphor that makes sense. And LLMs, I think if LLMs could only translate between languages, I think that alone is a world changing invention. So I think there's still so much to be done in that area, and I 100% disagree that we're in diminishing returns.

I think people got very used to this brand new thing that showed up and then it's gonna probably behave in ways that are somewhat expected at this point. But the interesting thing is, what do you actually do with it? How do you integrate it into all this other stuff? How do you actually build out use cases with this thing?

So how can you prove it? What is your rock solid logic and empirical research, Fuzzy? I'm open to hear it. I know Monarch is as well. We're always open minded.

00:50:21 - Monarch Wadia

We should have Fuzzy on video, man. Seriously.

00:50:24 - Anthony Campolo

Oh yeah. Fuzzy always loves to throw down. Obviously not a huge manifesto, but if you want to come on and chat, it'll be fun.

00:50:34 - Monarch Wadia

Yeah, and I get to meet Fuzzy.

00:50:38 - Anthony Campolo

Fuzzy is a hoot and a half. He's the most Scottish person I know by a wide margin.

00:50:45 - Monarch Wadia

You know, I found out something about Scottish culture. The national flower of Scotland is a thistle, which is this spiky flower. I see it everywhere in Ontario. It's a spiky flower that's really tough. If you touch it, it'll poke you back. It's like a cactus basically. And that's the national flower because it symbolizes endurance, defiance, and strength. I thought that was just the most amazing flower to have for a national flower. That's amazing.

00:51:16 - Anthony Campolo

Okay, my Discord is finishing updating. We're about to get in it. All right, check your Discord messages, Fuzzy.

00:51:29 - Monarch Wadia

He's coming up.

00:51:33 - Anthony Campolo

The Fuzzy Bear, the one and only. For people who don't know, Fuzzy Bear was on episode 89.

00:51:49 - Monarch Wadia

Astro community with Fuzzy Bear.

00:51:51 - Anthony Campolo

Yeah, he worked for Astro. I'm pretty sure he was a full on Astro employee for a long time before he went to the Linux Foundation.

00:52:00 - Monarch Wadia

Gotcha.

00:52:03 - Anthony Campolo

But yeah, he's been someone who's definitely one of the web developers I know who's been more bearish on AI. And yeah, I get where they're coming from, but I just disagree.

00:52:17 - Monarch Wadia

Linux Foundation must be such a fun job, honestly.

00:52:24 - Anthony Campolo

Do you have things you want to work on between now and our next show?

00:52:30 - Monarch Wadia

Yeah, I'm gonna be doing that refactor. Hopefully I can get it done today, although I am sort of sleep deprived, so we'll see how far I get. Maybe with caffeine. But definitely by tomorrow I'll be done with the refactor. Then I'll generate more of those artifacts.

I think by next week, if I can generate like eight more, which should be possible with a cleaned up pipeline, then hey, maybe we can demo what we built next week.

00:53:07 - Anthony Campolo

That'll be exciting.

00:53:09 - Monarch Wadia

Yeah, I think a lot of people are sci-fi fans. I think they'll be excited about it.

00:53:17 - Anthony Campolo

Super cool. For me, I really need to look at this Node wrapper for Whisper. I need to try and see if that's gonna make sense. That might help mitigate some of the weird integrations with Whisper.

And then just kind of actually refactor the transcription stuff and the different model APIs and give different options for those, because pretty much all that seems to be kind of connected. So that's what I plan to have done by next week. That's my goal for myself.

00:53:57 - Monarch Wadia

Yeah, maybe we can make the two worlds meet, because the pipeline that you're building is very adjacent to the pipeline that I'm building in many ways. So I wonder if there's a way. I think you're the one who mentioned that idea. You might have mentioned something along those lines, or maybe not.

00:54:20 - Anthony Campolo

Well, I was saying how it would make sense to bring in Ragged, because with Ragged I would have an easy way to switch out different models. I want to run an output on ChatGPT and Claude and Cohere and get three outputs really fast, and then compare them and compare the different costs.

If Ragged is a universal interface, then I want to bring that in for the LLM layer, because right now what I'm doing is I'm installing client libraries. I have an OpenAI client library in its own file. I have a Claude Anthropic library because I don't want to, at this point, bring in a LangChain or a LlamaIndex. I've used those things, I know how to use them, but I want to be close to the metal for each of these different things.

All right, we got Fuzzy, the man himself.

00:55:06 - Fuzzy Bear

There's my dad. Where's my dad? There's my dad. Hey, how are we?

00:55:13 - Anthony Campolo

Monarch, Fuzzy. Fuzzy, Monarch.

00:55:16 - Monarch Wadia

Hey, nice to meet you.

00:55:18 - Fuzzy Bear

Nice to meet you guys. Good to meet you both.

00:55:21 - Anthony Campolo

What you been up to recently, man? How's life?

00:55:23 - Fuzzy Bear

Loving life since summer's coming in. It's my favorite time of the year. I legit love it. Days don't end, man. Between now and June, literally the sun doesn't set.

00:55:40 - Monarch Wadia

What time does it go?

00:55:43 - Anthony Campolo

I'm in Alaska in the summer where actually the sun doesn't set.

00:55:47 - Fuzzy Bear

I don't know. Cool.

00:55:48 - Monarch Wadia

We get sunset at about 9 p.m. and sunrise at about five-ish, I'd say.

00:56:04 - Fuzzy Bear

Whereabouts in the world are you?

00:56:06 - Monarch Wadia

Toronto.

00:56:07 - Fuzzy Bear

Toronto. Alright. Canadian, eh?

00:56:10 - Monarch Wadia

A little bit, a little bit.

00:56:13 - Fuzzy Bear

Oh.

00:56:15 - Monarch Wadia

It's cool, it's cool. I think these days Canada is not as cool anymore. There's a lot of politics and election year next year, so.

00:56:26 - Fuzzy Bear

Yeah.

00:56:27 - Monarch Wadia

It's just.

00:56:28 - Fuzzy Bear

Like I said to my Canadian pals, when it comes to Canadian politics, it's literally just like The Simpsons did it. When they drew the Australian Parliament, that reminds me of the Canadian parliament. The one I'm talking about where there's no clue and Bart gets the boot.

00:56:53 - Anthony Campolo

There's a scene in The Simpsons. Yeah.

00:56:56 - Monarch Wadia

Yeah. I haven't watched that one. No.

00:56:58 - Anthony Campolo

Because Bart runs up a collect call in Australia and he won't apologize. So they're going to have the king of Australia, whoever, kick him with a boot.

00:57:14 - Monarch Wadia

We have a king here too.

00:57:17 - Fuzzy Bear

He's not my king. I just want to say, it's really good. I've been loving the last couple of episodes these guys have been doing. Legit, it's fantastic. It's good to see library authors and developers in the new AI space.

I just want to say, before I go into my rant about diminishing returns, which is why I came out here, there is still a vacuum in the space. For players like Vercel and such to be the only ones creating SDKs and toolkits isn't exactly what we should be looking forward to as a community of open source and citizens' intentions.

I really respect what you're doing, Monarch, with Ragged. I took a look at it.

[00:58:19] It looked really interesting. I'd like to see where it goes. How you're going to bring in tool calls and the streaming UI elements into it as well, because this is where I see the ChatGPT interface going forward. The ability to bring in UI components and not only that, be able to interact with React, Svelte, Solid components and have that update and interact with the model itself. Just to give you a fully rich experience is definitely the way that it's going to go.

But this is where I hit the wall of diminishing returns. You could have these tools, but all we're doing is iterating on GPT. We feel like we are innovating, but the innovation was done with the neural nets and the papers and the research. That was the innovation.

[00:59:19] All we're doing is iterating upon the speed that is being given to us, the opportunity. So when we look at the models themselves, they've already cannibalized pretty much the petabytes worth of human information and data that has ever been generated.

You guys mentioned the leap between the printing press and the internet. Put it this way, between when the Gutenberg press came online and all the print media up until '95 when Berners-Lee turned on, when the internet became available to the masses, not when it was turned on, but when it was available to the masses, that was like summarized into a petabyte. Now we're generating that almost on an hourly basis.

01:00:14 - Anthony Campolo

For myself, just personally, the reason why it at least feels different for me in terms of what I've done is that I'm generating, I have a giant transcript, I feed it to a model and I want to create show notes for it. And over the last year, I have seen improvements. First, the models couldn't even take a transcript a year ago. Now they can. So that's something new.

And even just with different models, I now use Claude because it gives me a better output than ChatGPT does. Now is it 100x better? No, it's like 20% better. But it's better, it's different. And so I see improvements being made. Maybe they're not as world changing. But I think once you have cannibalized, like you say, all of the data, it's like, okay, how do we actually create feedback loops so that these things get better and better? Can they learn what you want better and better by learning your patterns or having more data to work on, having longer context length?

[01:01:19] So that is where I see the interesting work, and I see where gains can still be made at the foundation model. You're right that that stuff is the next 20 years. We're going to have foundation models, but it's hard to say where that's going to go.

01:01:35 - Fuzzy Bear

No, I mean that is it. What we are now looking at is pretty much the underlying tech at its max. What we now have to do as developers, as engineers, the onus is on us to deliver this technology to the masses in the ways that we can.

Like for example, with transcribing transcripts, this functionality was not known to us as developers, but that behavior is available. So what I'm trying to say is that I think the way that we're going to be creating our programs is going to be more implicit. So that way we have it available as a JSON response that an LLM can also pick up instead of just the response value, the return value that we want to get.

If you get what I mean, we're going to have to build our programs and our APIs a bit more complex to take in additional inputs from these LLMs, or be able to be called from these LLMs, because that is where we're seeing the real value.

I was speaking to my physio today.

[01:02:47] The guy served on the modern battle lines of history. And he basically was saying that all this compute is not for him. We've heard this, "computing is not for him." But then, Anthony, you said that your dad uses ChatGPT as well.

01:03:08 - Anthony Campolo

He's almost 70. Yeah.

01:03:11 - Fuzzy Bear

And what we're now seeing is that computers are going to become like clay. We used to be gatekeepers on barriers of entry when it came to this stuff.

01:03:22 - Anthony Campolo

Here's a 2023 Linux Foundation report on generative AI.

01:03:28 - Fuzzy Bear

Yes. The Linux Foundation, before I get to the LFW, their research team and department is fantastic. Hilary took it over and she is literally, I met her when I was at the All Hands. She's the director of the department. She is just an amazing person, a pure force. She just loves and breathes data all the time. I haven't seen this one. I think this was shown to us. Can't remember it.

01:04:07 - Monarch Wadia

You were going on an interesting tangent. I just want to put a pin on that and come back to you about computers becoming more like clay, because I'm getting that sense too. I have a lot more to say about it.

I met some people recently. They didn't know how to code two months ago. And now after two months, they're building integrations into Amazon's FBA service and into Excel. And they're automating stuff using agents, not just AI, but agents and agentic workflows. And they're using ChatGPT to build this out. Two months to go from not knowing how to code to actually building that is insane.

01:04:48 - Fuzzy Bear

Exactly. That is it. But the thing is, you gotta also understand that for a very long time, for as long as computers have been around, only a certain sliver of the population has been able to utilize and maximize the ability that a computer as a tool can give you. And so now when we come to this clay-like thing, not everybody is a potter. But what I'm trying to say is that.

01:05:18 - Anthony Campolo

And it's great though. That's what's empowering about it.

01:05:21 - Fuzzy Bear

Yeah. I mean, wow, these are some interesting stats. There's the whole self-empowerment thing, like you were saying earlier.

01:05:38 - Anthony Campolo

What is to you the direction that we should go with LLMs? Where do you think this should go in your ideal world?

01:05:48 - Fuzzy Bear

Looking at making them efficient. Because this is the one thing that stops us as a species from going up the Kardashev scale, and that is energy and propulsion.

It's the same thing with LLMs in the sense that they're extremely expensive computationally. Energy-wise, we're already seeing energy demands increase exponentially by magnitudes of tens to hundreds. Already, the energy demand they're anticipating equals Japan. By the end of, I think it was 2025, they say that it'll be the equal output of Japan. They estimate a thousand terawatt hours a year just to run LLMs by 2030. Wow. And that is just so we could do small little things on the screen, give them this feedback ability, "oh, it's talking back to us."

[01:06:59] I think we need to get smarter on that side. I think we really do owe it. If we're going to claim a crisis.

01:07:08 - Anthony Campolo

It's more useful than computing power for blockchain, though.

01:07:13 - Fuzzy Bear

Oh hell yeah. But blockchain hit that precise problem. It couldn't get efficient energy wise. It became more expensive to run and do the compute. LLMs are heading towards the same crisis of energy that blockchain hit.

01:07:36 - Anthony Campolo

So buy Nvidia.

01:07:39 - Fuzzy Bear

Pretty much. Buy Nvidia, buy shares in every power company under the sun. That is it. Energy, energy, energy.

To bring propulsion into the conversation, that is more on how I feel like the momentum is with big tech in this particular period. All the small, quick AI startups got picked up and acquired. That is simply so they can sit on it so you can never put your technology out there. They're not going to put it out. That's just so they don't see it out there. That's why they bought it.

We're now at this point where big tech is the driving force. Open source is in a crisis state in itself that needs to get rectified quick. Otherwise we're just going to let big tech dominate the entire scene. Open source has always been the counterbalance between corporate centric technologies and monopolies and the digital democratic domain that is decentralized information.

01:09:02 - Anthony Campolo

Totally. Yeah. That's why I know I have, and I think Monarch has as well, leaned into doing as much of this open source as we can and making this stuff work with open source tools. So I think that's cool and a huge priority.

01:09:16 - Fuzzy Bear

Totally. For me, I keep thinking like the Star Trek computer. If anybody ever envisaged a future, it was Gene Roddenberry. That man saw it happen. We're literally living in his wake.

01:09:44 - Monarch Wadia

That's wild. It's so true. I think about this often. You're maybe the second other person who's said that in those words. Because the holodeck.

01:09:55 - Fuzzy Bear

[unclear]

01:09:58 - Monarch Wadia

What's that?

01:09:59 - Fuzzy Bear

I'm the biggest Trekkie you'll meet.

01:10:02 - Monarch Wadia

What's your favorite series?

01:10:08 - Fuzzy Bear

I would have to say recently, just for nostalgic reasons, Picard Season Three. For me, that was just the most tear-jerking moment. I wanted to relive every moment of that season again and again and again. There's so many good bits to that. Like when Data and Geordi finally met, I could go into it.

01:10:33 - Monarch Wadia

No spoilers, no spoilers.

01:10:36 - Fuzzy Bear

Oh dude, Season Three of Picard, legit, it is hands down a fantastic show.

01:10:48 - Monarch Wadia

I got turned off by Season One a little bit, I gotta admit. But every single Trek...

01:10:55 - Fuzzy Bear

But the thing is, I love what they did with Picard. They tried to tie up a lot of loose ends. So Season One was the whole Data's daughter kind of thing. Season Two is about Q.

01:11:16 - Monarch Wadia

Oh, really?

01:11:17 - Fuzzy Bear

Yeah.

01:11:18 - Monarch Wadia

You have my interest. Okay.

01:11:19 - Fuzzy Bear

Yeah, so it's no...

01:11:24 - Monarch Wadia

Spoilers, no.

01:11:25 - Fuzzy Bear

Spoilers.

01:11:29 - Monarch Wadia split

I gotta watch this. I'm a huge...

01:11:32 - Monarch Wadia

DS9 fan personally.

01:11:34 - Monarch Wadia split

Oh.

01:11:36 - Fuzzy Bear

I love the intro to Deep Space Nine. When the comet comes by, you just follow it right through. The space station, the wormhole. Edge of the universe, I love it.

01:11:50 - Monarch Wadia

I hated that one too in Season One. I was like, what is this? There's no warp core engine. We're not flying through space. We're just stuck in one place. What is this? And then DS9 just blew up after that. I was like, whoa.

01:12:05 - Fuzzy Bear

Right when Worf came over, and then the Defiant, and then the whole Bajoran arc. There's an underlying story to DS9, which is the Founders, the Changelings. So Odo was a Changeling, but he was part of the Founders, which are like the primordial creatures who exist.

01:12:28 - Monarch Wadia

Spoiler alert, spoiler alert, spoiler alert. If you haven't watched DS9, go ahead.

01:12:33 - Fuzzy Bear

Oh yeah, it's all about the Founders. That whole thing was all about the Founders. And that was a shout out back to the original series where the Changelings first appeared. When did they appear? Oh dude, you're going to have to ask me to bring out my encyclopedia. I'll show you this.

01:13:00 - Anthony Campolo

It's on now.

01:13:05 - Fuzzy Bear

I don't know if it shows you this, but I have three of these.

01:13:09 - Monarch Wadia split

Whoa, man. Whoa! Oh, that is so cool.

01:13:13 - Fuzzy Bear

I have literally three of them.

01:13:17 - Monarch Wadia split

That is so cool.

01:13:19 - Fuzzy Bear

Yeah.

01:13:22 - Monarch Wadia

I think Fuzzy, you and I can sit down and talk about Trek forever.

01:13:27 - Monarch Wadia split

Forever.

01:13:29 - Monarch Wadia

I did not know that the Changelings were there in the original series. They even have the... is that an urn? Wait, what? This is interesting. I'm seeing pictures of the episode The Changeling.

01:13:49 - Anthony Campolo

Did you guys ever watch the Enterprise Star Trek series?

01:13:54 - Monarch Wadia split

Yeah.

01:13:55 - Anthony Campolo

Not good, not very good.

01:13:57 - Monarch Wadia split

I love TNG.

01:13:58 - Fuzzy Bear

I'll eat it because it came with a canon, right? And it was that period when we had Star Trek Advanced and they tried to work back. It's the prequel, the prequels. I liked how they tried to do it, but at that point the whole show kind of got repetitive in terms of the screenwriting and all that stuff.

01:14:26 - Anthony Campolo

Right before TV got really good. It was early 2000s. TV budgets went up 10x.

01:14:35 - Fuzzy Bear

When they're still using actual physical props. You could tell.

01:14:40 - Anthony Campolo

Yeah. It's like watching Friends.

01:14:42 - Monarch Wadia split

Yeah.

01:14:44 - Monarch Wadia

You just have to squint a little.

01:14:47 - Fuzzy Bear

Has anybody seen the stuff from Disney? The Disney labs? What they're putting out in terms of The Mandalorian and all that stuff?

01:14:57 - Anthony Campolo

Mandalorian. I've heard it's awesome, though. I don't have the Disney streaming service, though.

01:15:02 - Monarch Wadia split

Right.

01:15:03 - Fuzzy Bear

Got torrents though.

01:15:05 - Anthony Campolo

That's true. Yeah.

01:15:07 - Fuzzy Bear

Cool. Legit, watch it, right? And then look at how they made it. How they shot The Mandalorian. It is honestly going to change cinematography, the whole way it's done.

It's pretty much like this in a nutshell. Imagine a domed room, a semi-domed room. But it's all 8K screens around you. They use Unreal, the engine, and then they basically immerse the landscape. They put the landscape right there.

01:15:41 - Anthony Campolo

Every movie is 300 now.

01:15:43 - Fuzzy Bear

Pretty much.

01:15:45 - Monarch Wadia split

Is that how 300 was made?

01:15:47 - Anthony Campolo

Yeah, because that was the first. It was just everything was green screen. They just put a bunch of green screen and projected digital images for all the scenery. 300 was that.

01:15:56 - Fuzzy Bear

The six packs? Every single six pack was hand drawn on those screens.

01:16:01 - Monarch Wadia split

Really?

01:16:03 - Fuzzy Bear

I'm not making that up. No, those are real.

01:16:09 - Monarch Wadia split

That's the funniest thing I've heard.

01:16:10 - Monarch Wadia

But what's this 8K monitor thing? Are you just using a metaphor, or did they actually use 8K monitors for that?

01:16:16 - Fuzzy Bear

8K monitors right around you. It completely wraps you. It's like a single screen kind of thing. And it completely wraps you, so you get the apex at the top and it completely goes down on all sides.

And then they basically just drop Unreal Engine to create the scenery. And then using the lighting, because The Mandalorian was all a tin can, you get to see the external environment reflected directly off them. That's how they did that.

01:16:49 - Monarch Wadia split

Ah.

01:16:50 - Monarch Wadia

So they actually had 8K monitors and the reflections on the armor would actually be real? Wow.

01:17:01 - Monarch Wadia split

Yeah. Wow.

01:17:02 - Fuzzy Bear

Yeah, it's stuff like that. And they got movable tiles and everything. It's really cool.

01:17:10 - Monarch Wadia split

Ah.

01:17:13 - Anthony Campolo

For now, actors are still people. For now.

01:17:18 - Monarch Wadia split

You know.

01:17:21 - Fuzzy Bear

I want to get your thoughts on this. How best can we use this? For me, language is a big thing, right? Because we anthropomorphize AI and GPT and LLMs and all these things when in actual fact they're just technologies. They should be just considered like network layers, sockets and all this stuff, pure primitive levels of technology that's available to us.

When we start describing AI in terms of these superfluous phrases, we describe it in a human-like quality and quantity. We're describing that AI with intelligence, artificial intelligence, when there's none actually there. Because we all know it's not really intelligent. All it is is additive inference. It gives us the ability to scale ourselves and infer more quickly and do more, faster.

[01:18:39] If anything, it allows me to scale to ten times me, right?

01:18:46 - Anthony Campolo

That's 10x of you, then. You also then could get in a room and have a conversation with all of them.

01:18:52 - Fuzzy Bear

Yeah, but that's not ten times the intelligence though.

01:18:55 - Anthony Campolo

Maybe a committee of ten Fuzzies is exactly what the world needs.

01:18:59 - Monarch Wadia split

Maybe. Oh hell, the Republicans.

01:19:02 - Fuzzy Bear

Nothing will get done. Nothing.

01:19:06 - Anthony Campolo

The way I think about it is they always said when you got a bug in your code, rubber ducky. You work it through. I see AI like a rubber duck that talks back. It's not intelligent, but it's mirroring back to me in a way where I can kind of glean useful bits of information from it because it understands what I mean well enough to give me back a response that is useful to me. That doesn't mean I think it's intelligent.

01:19:32 - Fuzzy Bear

You're the intelligent agent in this respect.

01:19:35 - Anthony Campolo

I'm able to talk to it and get a response back. That makes sense to me.

01:19:39 - Fuzzy Bear

But what you're talking to, in a sense you're effectively coding with it, but you're coding in natural language.

01:19:46 - Anthony Campolo

I'm not coding, though. I'm writing English language sentences.

01:19:50 - Fuzzy Bear

You're writing English language sentences, right? But here's a repurposing of that statement. You're coding into a prompt using natural language, natural English.

01:20:04 - Anthony Campolo

Yeah, I know, but that's different. And that's why it needs to be a different thing, because it's a good thing. It's where we want computers to be because it allows us to actually talk to them.

01:20:17 - Monarch Wadia

I have a totally different take on this.

01:20:19 - Monarch Wadia split

Yeah.

01:20:21 - Monarch Wadia

It's sort of a meta take. Stepping back from the debate, watching the debate happening. On one hand, Fuzzy is saying that this is just a technology. It's not really intelligent, and we're coding with it. Anthony is saying, well, I'm talking to it, and it's doing things in natural language. And that's different from coding.

Now you're both talking about the same system from two different perspectives, and we're trying to narrow down on what exactly this thing is. But the truth is that both your perspectives are valid, and there is no actual difference in how Fuzzy perceives the LLM and how Anthony perceives the LLM. There's no difference. It's just a difference in the language we're using.

So if we sort of put the differences aside, can we agree that talking to this thing in natural language gives you a natural language response?

[01:21:25] Is that something that we can agree on?

01:21:28 - Monarch Wadia split

Yeah.

01:21:29 - Fuzzy Bear

Yes. But I wouldn't say talking.

01:21:32 - Monarch Wadia

What would you say? [01:21:35] - Anthony Campolo Communicating.

01:21:39 - Fuzzy Bear

Yes, but like...

01:21:44 - Monarch Wadia

Manipulating.

01:21:45 - Fuzzy Bear

When you're talking, I'm not instructing you. I'm not giving you explicit instructions or commands, right?

01:21:52 - Monarch Wadia split

You might be.

01:21:53 - Anthony Campolo

If we're working on something together.

01:21:55 - Fuzzy Bear

Oh, but I might be doing it subliminally, right? But the thing is, it's one of those where we are as a form of communication. But I would say it's more explicit instructions.

01:22:09 - Anthony Campolo

I can shoot the shit with ChatGPT too. I've had lots of interesting conversations with ChatGPT that have nothing to do with building anything, like hours and hours of interesting conversation about all sorts of topics. That's just me. I'm weird. I like talking to it.

01:22:20 - Monarch Wadia split

Found it instantly.

01:22:21 - Anthony Campolo

Fascinating.

01:22:23 - Monarch Wadia split

After all my many...

01:22:25 - Anthony Campolo

Conversations. You have no idea.

01:22:28 - Monarch Wadia split

Who needs friends? I've replaced all my...

01:22:31 - Fuzzy Bear

Friends with...

01:22:31 - Monarch Wadia split

ChatGPT.

01:22:33 - Fuzzy Bear

I'm on the phone to one very...

01:22:37 - Anthony Campolo

I get one new, very interesting friend who's different from all my other friends. He doesn't replace our friends, but I got a new, very interesting, different friend now.

01:22:45 - Monarch Wadia

I think all three of us are seeing the same thing and using different words, right? So whether you say communicate...

01:22:53 - Monarch Wadia split

Or code.

01:22:54 - Fuzzy Bear

That is probably at the heart of what I'm trying to get at, right? It's the language that we're using to describe it. Like you're saying, we're discussing the same thing around the same parameters, but the language and the heuristics of what we're trying to get across becomes fuzzy in the nuance.

And whereas there was a Google I/O conference in Glasgow at the weekend, right? And there was this lassie talking about prompt engineering from a coding perspective. And she was like, codify your prompts. And if you could codify your prompts, you would ensure a level of accuracy through it. So it goes from a one shot to a few shot to a cohesive shot.

01:23:53 - Anthony Campolo

And with AutoShow, I've had my base prompts that everything is based off of.

01:23:57 - Fuzzy Bear

Exactly. So you codified it, right?

01:24:00 - Monarch Wadia split

Yeah.

01:24:00 - Fuzzy Bear

And that's...

01:24:01 - Monarch Wadia split

What...

01:24:01 - Anthony Campolo

I did. Yeah.

01:24:02 - Fuzzy Bear

And that's what I'm trying to get at. When we're using these prompts and these models, regardless of whatever form that they take, we are codifying our intent, right? Same way as we would code in JavaScript or TypeScript or any other domain specific language.

01:24:28 - Monarch Wadia split

Right.

01:24:28 - Fuzzy Bear

It's just that English is the domain specific language for these GPTs. So as a benefit for the English language and its speakers, it's just going to accelerate the importance of that language itself. It could become a universal language we might see one day, right? But right now we're coding in English when we interact with ChatGPT.

01:24:56 - Monarch Wadia

I see what you're saying. So the implication...

01:24:58 - Anthony Campolo

A lot of other languages, though. Sorry.

01:25:02 - Monarch Wadia split

Go ahead. For sure.

01:25:03 - Monarch Wadia

No, you're right. So most of the Western languages I think are in there. I know Hindi is in there too. So there's a lot of different languages. Let's stick to English for a second.

So we're speaking in English. That means going back to your "computers are becoming like clay." The implication there is that every human who can speak English, or one of these languages, is now also a semi developer.

01:25:29 - Fuzzy Bear

Exactly. And that's what Uncle Jensen was trying to get across when he said that in the next 15 years, there's not going to be any programmers or need for programmers because everybody will be one. And I could totally see that happening.

But I think if we were to get smart about it and not run away with the fairies when we talk about GPT and models and all that stuff... Falsely dressing up a Hoover and taking her out on a date is effectively what's happening. I mean, I've just put a dress on a Hoover and just taken it out for a date, right?

01:26:16 - Monarch Wadia split

That is like...

01:26:17 - Monarch Wadia

A vacuum cleaner.

01:26:18 - Fuzzy Bear

Yeah, like a vacuum cleaner.

01:26:21 - Monarch Wadia split

Yeah. So you're putting...

01:26:23 - Monarch Wadia

A dress on a vacuum cleaner and taking it out on a date?

01:26:25 - Monarch Wadia split

Yeah, yeah, yeah.

01:26:27 - Fuzzy Bear

And so we've got to see it like that. We've got to discuss it like that is what I'm trying to say, right?

Because if we don't get lost and run with the fairies, coming back to the heart of it, the people who don't know how to use computers or just don't use computers at all are going to eventually be using this technology. If we can figure out the compute and the efficiency side, and we could break down the barrier of entry.

So it's not like away with the fairies where people perform anthropomorphic interpretations of what this technology is. Whereas if they had a clear understanding of what the technology is, you get what I'm trying to say.

01:27:23 - Monarch Wadia

I do. You're right, but maybe the anthropomorphization is the evangelization. Maybe the anthropomorphization is the interface, and maybe there's no need to overexplain past it.

So if I'm driving a car and I'm in automatic, I don't really need to understand how gears work. It would help if I did, but all I need to do as a person who doesn't know how to drive manual, I only know how to drive auto. I just put it into drive, and then I push the gas pedal, and then I turn my wheel. And as long as the car doesn't break, I'm happy.

It does what I need it to do. I don't need to go look under the hood. I don't need to understand how to drive manual. I don't even need to know what fabric my seats are made of. I can just sit on it and I can go, and that's all that you really need.

01:28:17 - Fuzzy Bear

But that's an active skill, isn't it? It's an adaptive feature. The automatic gearbox, right?

01:28:23 - Monarch Wadia split

Yes.

01:28:24 - Fuzzy Bear

And that is at the heart of what I'm trying to say. These are adaptive features. It should be AF, not AI. Additive features.

01:28:44 - Monarch Wadia split

Intelligence, huh.

01:28:47 - Monarch Wadia

Interesting. I think intelligence itself is an anthropomorphized word. So if you want to move away from anthropomorphization, I think we should move away from the concept of intelligence itself.

01:29:00 - Fuzzy Bear

That's why I use inference. That's why I use the term inference more so in this case. And these use cases over the past year, because if anything, it has allowed us to scale the ability that our minds, our brains, are nothing more than a giant inference engine themselves.

It's allowed us to scale that inference ability to a complexity that I've yet to find a suitable measurement that I could put my hat on.

01:29:31 - Monarch Wadia split

You know.

01:29:32 - Monarch Wadia

Now that you put it that way, I'm completely on board. If you consider the brain to be on equal footing with the LLM or a more advanced version of an LLM, I'm with you. I think there is barely any difference.

01:29:47 - Fuzzy Bear

Yeah, if you see it as an inference engine, then you could see the similarity between the neuro... Let me just think of the words I'm trying to say. If you were to look at the synapse of a brain and see the dendrites and the...

01:30:13 - Anthony Campolo

On-off. McCulloch-Pitts neuron, 1945.

01:30:17 - Monarch Wadia split

Right.

01:30:17 - Fuzzy Bear

And so when you look at that, you see complexity of just unimaginable scale. But when you look at how it works under the hood, you see a very basic dendrite. You know what I'm trying to say?

01:30:35 - Monarch Wadia split

Yeah.

01:30:35 - Fuzzy Bear

Oh, I need to go. I got a call.

01:30:38 - Monarch Wadia split

Right.

01:30:40 - Monarch Wadia

Amazing, Fuzzy.

01:30:40 - Monarch Wadia split

Nice meeting you.

01:30:42 - Anthony Campolo

Thanks for coming on, man.

01:30:44 - Monarch Wadia split

Peace, love. See ya.

01:30:48 - Fuzzy Bear

How do I get out of this?

01:30:50 - Monarch Wadia split

Oh, just...

01:30:51 - Anthony Campolo

Cut the window.

01:30:51 - Monarch Wadia split

Just leave. [unclear], man. Thank you.

01:30:55 - Anthony Campolo

Alright, with that I think we'll call it there.

01:30:58 - Monarch Wadia split

Yes.

01:30:58 - Monarch Wadia

We need to have one more. I'll see you, Anthony.

01:31:01 - Anthony Campolo

Yeah, that was fun.

01:31:03 - Monarch Wadia split

Right?