Reid Riffs with Parth Patil on Individual AI Mastery (Part 1 of 3)

0:00Hey, everyone. Aria here. We want to kick off the new year with inspiring conversations about AI as well as practical and tactical guidance around technology. So for the month of January, AI specialist and my dear colleague Parth Patel is joining Reed for Reed Riffs to talk about how everyone from individuals to enterprises to startup founders can harness AI to level up their work, retrofit legacy orgs for the AI era, and build AI native businesses right from the jump. So tune in. You're in very good hands with Parth. And I will be back putting Reid in the hot seat come this February. Thanks so much.

0:41Thanks for the kind word and warm welcome to Possible, Aria. As I talk with Reid this month, I'll be walking through some of my AI projects, demos, and tools on screen. While I'll do my best to describe what I'm looking at for our audio-only listeners, consider switching to the video version of the episode on Spotify or watching on Reid's YouTube channel for the full experience. Thanks. And let's get into it. Parth, I partially know the answer to this, but actually one of the things I've been really looking forward to doing in this interview is actually to discover the fuller answer on some of these questions.

1:13So what is your AI stack and how did you familiarize yourself with these tools? You know, I was a little later to AI than maybe you were in your career, but for me, the big moment. I'm older. I was following AI during the video game AIs, the ones from DeepMind and OpenAI when they were working on the Dota 5. And then when opening, I put out ChatGPT and it changed the world. For me, it was like, oh my God, this is the tool that you can use to teach yourself every other tool. And so it started with ChatGPT, started with getting really good at wielding a language model and asking questions. And then also having it look at all the tools that I have and be like, teach me how to use them even better.

1:48Teach me about my computer. Teach me about this video editing tool, music. So ChatGPT is probably the meta tool that I go to, to teach myself how to use all the other tools. This gets to kind of more of a Parth biographical personal question. but when was your light bulb moment of, oh, this isn't just a work tool. This isn't everything tool and everything agent. What was the light bulb and what was the moment of frisson? So I used to work on Clubhouse, the audio app from the pandemic. And I was working there when ChatGPT came out. And what happened was when ChatGPT came out, it became the most popular topic.

2:25And so it was an audio app. People would talk to each other on the internet and meet and talk about different topics. And ChatGPT became the most popular topic in every single part of the world. And so I would just join the app, join rooms, ask people how they were using the app, using ChatGPT. And I had so many moments where it's like, oh, photographer learning about the different settings on their camera. Or then you have farmers in my home state in India that are using it to help plan their crop cycle. And then you have people here that are like, I don't like how it rewrites my email. And I was kind of just like, oh, my God, this is a new computer.

2:56You know, it's the first computer that we can talk to. It's sort of like the C-3PO from Star Wars, except, I mean, you know, Neuromancer, you talk about science fiction. This felt like it, like it felt to me that the conversational computer is a hundred years early. Like I never imagined it would happen in our lifetime. And here it was on our doorstep. Already speaks every single language. All the human, well, most of the human languages and all of the programming languages as well. So it's got this like incredible, like general capability. And to think of it only as a work tool is an incredible oversimplification.

3:27It almost represents the human collective intelligence, in a sense. And it's a way for us to access the collective intelligence through natural language, by talking to an AI. One of the things that I picked up from watching your usage was that you, by prompting a role assignment, whether it's mean VC, skeptical co-founder, emulated customer, a bunch of other things, but by prompting role assignments and various other forms of kind of meta-prompting, that could get you useful things. And obviously when you combine that with a swarm of agents or a set of agents, that even then becomes more of a part of how we're all deploying a team.

4:07So say a little bit about your prompting guidelines, like maybe one that you would hand to beginners and then two or three that you would hand to non-beginners. For beginners, I think it's mostly start by just prompting heavily, like use voice, transcribe, talk at length about what you're trying to do. and then um and maybe assign roles yeah and maybe assign roles i think uh you'll find that the model can emulate all these perspectives like you can say pretend you are a vc and critique my business plan critique my the way i'm running my startup potentially if i'm going to go have a conversation with a vc for capital or pretend you are the customer of this product that i'm working on and explore my website.

4:50Do you feel like my website properly positions our product based on your needs as the customer? And so the model gets to pretend to be all these personalities. And so even if you don't have that person sitting next to you, you can kind of simulate that perspective. And then the AI playing that role will help you understand your problem in a way that you didn't even consider. Or you might say, I want you to be the most skeptical person of all that you can possibly imagine. and find 25 different criticisms of my approach to this problem. So I think role-based prompting is very powerful because I think we need to get out of our own perspective.

5:26And being able to call on all these other perspectives that the model can emulate is really powerful. I once had a coding agent generate 100 ,000 unique expert personalities. It turns out if you create 100 ,000 unique experts, you kind of cover every single topic or a huge swath of all of the topics that humans have. like a hundred thousand unique experts covering everything from like parenting to like legos to like every single form of art that the models have been trained on and you get this like very multifaceted like set of like minds that you can tap into they're not complete minds but they're perspectives and then you can say like find the 10 that are most relevant to my problem and then have them all answer the question and they all answer it differently like that that different perspective like an optimist is going to have a different answer than a pessimist but then you know an oceanographer is going to have a different perspective than say like an accountant on almost anything uh that you present and so it it's a very surreal thing i call that role play and then i think the one of the most powerful prompting techniques it's in the area of meta prompting like prompts like it's the it's the prompt that helps you find the right prompt a simple one that I use almost every day is I go to a language model and I'll say, here's my problem.

6:42I'll describe the problem. And then I'll say, interview me until you have enough context to help me with this problem. Ask clarifying questions and then we're going to begin. And that's really important because I think a lot of people initially, you go to AI thinking you know what the answer is. And I think a better way to go to AI is let's describe the problem. And like maybe a set of solutions will emerge once the AI kind of collects that context and draws it out of you. The things that maybe you weren't thinking about when you first articulated it. And it's really good to, I call it the interview me prompt.

7:13Just interview me and then we're going to begin. And so I do that for any project that I'm starting from scratch. I'll say interview me and then we will begin. And it might just be a 10 turn back and forth and I realize, oh, this consideration, I wasn't even thinking about that. And the AI will ask those intelligent follow-up questions, kind of like any good employee isn't just going to start doing the work. They ask the clarifying questions up front. And then that brings out a higher quality problem scope. And then when the AI begins, it becomes much more magical. And you have to kind of, I think a lot of times people just, they need to recognize that maybe the answer that you have in your mind isn't the right answer.

7:47And the AI can kind of feed off of your initial suspicion and provide an even better answer when you're in that loop of iteration. And so I think for a lot of people, it's really just realizing that being a little bit humble about what our limits as humans are, how many scenarios we can think about and how many considerations we can make, and then using the AI to expand that and parallelize that thinking. Part of the thing is the mindset shift of how do you put your ego aside. So say a little bit about that, because this is actually one of the things that gets in the way of a lot of otherwise smart people's way under delivery of AI.

8:23Yes. For me, when I was interacting with GPT-4 for the first time, so this was March 14th of 2023, GPT-4 came out, which was the successor to GPT-3. And the first big leap after the chat GPT moment was GPT-4. And I was talking to my teammate at Clubhouse, and we were both data scientists, data analysts. and I was like, Olivia, this GPT-4, it writes perfect SQL, it writes perfect analytics code if it understands the schema of the problem that you're working in. If it understands your database organization, it just writes perfect analytics code. And she was like, Parth, this thing aced the interviews for both of our roles.

9:01And I was like, wait, what do you mean? Like qualitative, quantitative? And she's like, both. I think it's got a pretty good idea for what we should do as a company too. And I was like, whoa, what does that mean? And then we went to our manager and we were kind of like thinking about this. We had a small data team, so we were kind of just like, whoa, this language model is clearly an amplification of our own ability to do analysis. And he was basically like, we're not going to hire anyone until we figure out how to use this. And then everyone we hire is going to be using this. Because then we get this super analytics kind of approach where I'm describing a problem in English, and then the AI is executing what I would have done manually by hand.

9:36And I had a moment where I asked for an analytics query. I thought it was a hard query to write. and it wrote a very elegant solution and i just didn't believe it and then i looked closely i was like oh that's just better than every version of the solution i've seen before and it was very humbling kind of like oh my god like this is this is definitely better than me at the writing of a and then i realized like okay i think my job is to aim this my job isn't to compete on the like it's like the kasparov versus deep blue moment but for me it was like the data analysis it's like the or if it's like John Henry, the steel man, against the steam engine.

10:11And I'm thinking, well, I sure don't want to compete on our manual writing of SQL queries anymore. Actually, I would rather be working on the automated version of analysis where I'm speaking in English to the computer. It starts turning my questions into computer code that can solve the problems. And so that was a huge shift for me. And I realize a lot of people are a little bit later on that, especially in engineering. I see experienced engineers tend to be attached to their core, their superpower. But I think looking at AI and realizing, eventually it might be better than you at the thing that you were really good at.

10:45But then your wisdom of working in that problem space becomes how you expand beyond just the AI or just you. And when did you get to that recognition of it being the meta tool? I think it was probably like three or four months into talking to it for 14 hours a day. realizing it could teach me about programming realizing it could teach me about music and then i was i was like sharing screenshots of my desk and i was seeing that it could actually click around like if you allow it to click around the computer and you run it on an api on your computer you're able to like orchestrate a web browser you're able to write code in every single language and i was like okay this is like language is actually the most powerful thing that you could possibly automate i think uh that's that's that was the realization there that like language touches everything.

11:31And then you're always talking about Wittgenstein, and he has a quote, which is language is the limit of my world. And that was very, that was like, I realized then like my vocabulary, everything I've been exposed to in my life, I could now access intelligence through that vocabulary, through that language. So one of the things that in serious part I've learned from you is how we got to voice pilling. Say a little bit about how important it is to actually, in fact, be using voice, why that is, and what people will learn from that. It is probably one of the most powerful prompting techniques there is.

12:03If you haven't tried it, it's really, like you want to try, you want to go to one of these language models like ChatGPT. And I think people get hung up thinking about typing their prompt in a certain way and structuring their prompt a certain way. And I think actually what people should be more concerned with or more focused on is trying to get as much of the ideas out of their head into the model. So it's more about like, you want to say more. You want to describe the problem. And I go to the extent of like, I will sit here and I'll ramble for five, ten minutes to the computer about the problem that I have on my mind.

12:33And that turns into like a three page kind of transcript. Even though it's kind of like unstructured stream of consciousness, turns out that is a very high bandwidth way of communicating with AI. And I think that some of my best prompts are not the ones that are structured a certain way, but they're the ones where I'm just being extremely effusive. I'm communicating as much as I possibly can because I'm just rambling at length about the problem. And I find that typing, when we type, we're kind of committing our ideas to a couple words. And in that process of committing, we're not saying as much as we might if we were talking to a friend or we're describing a problem to someone that we wanted helping us with the problem.

13:09Yeah, and part of that use of voice that I learned from you was not just depth of context, but also in breadth and sitting in the hands, is when we're typing, we also tend to write like precision, like I write a coherent sentence and so forth. Whereas actually, in fact, these are such good, these AIs are such good interpreters, that even if you're like, well, I got a half-baked idea here, it actually will guarantee have a more focus such that you might go, well, that was part of it, but now this is what I really mean. And that iterative gameplay, almost like video game, was a real key thing. So there's a relationship between the voice pilling and also almost like a video game style interaction.

13:54Yeah. I mean, as a gamer, I think when you're playing games with your friends, you're not typing to them. You're really just yelling commands. You're like, oh, I'm going to come here. Here's what we're going to do. And it's faster. It's high speed. I think any real-time coordination, even on a basketball court, people are yelling at each other. They're calling out what they're going to do. I think that voice is the way to get real-time coordination, both between people and also between people in AI. And also, yeah, I see sometimes people will be typing a prompt and then they'll have a typo and then they'll hit backspace.

14:26And I'm like, this thing is really smart. It's okay to have typos. And then if I look at my prompts, all of my prompts are just littered with typos because I know it's so smart that it understands what I'm saying, even though a couple of the letters are in the wrong place. But yeah, same thing with voice. You don't need to have structured thought all the time. I think sometimes if you're going to use a prompt every day, you should think about that prompt. But if it's like a single one-shot, you're describing a hard problem, it's more important that you describe it at length. Get that context out.

14:56One of the things also that you do that I think relatively few people do is you use multiple of the frontier models, both in parallel, in rotation, in experimentation. How do you choose which models to use? How is that evolving over time? And any hacks or heuristics or principles or things that our listeners might be able to kind of apply or kind of remember and take with them? Right. So it's a very, I mean, in the beginning, it felt like it was just OpenAI and ChatGPT. I think that it's easy to get kind of overwhelmed by the options that we have. But the main piece of advice I would say is you want to get good at one state-of-the-art tool in every single category.

15:36so one really good language model one really good image image model one really good video model and then those principles tend to translate over to the competitor products in each category right so you get good at chat gpt you're also probably going to be good at using clod and gemini and and any given week the number one model sometimes is different so i don't think everyone should necessarily be you know trying to stay up to date on that but really if you have one of the three in terms of Cloud, Gemini, and ChatGPT, that should be your goals. Be really good at least one of them. And then every once in a while, you try the other ones as well.

16:13First, a heretical question. If you say, bizarrely, someone has not done this at all, what's the model they should start with? I would say start with ChatGPT. If you're just getting into language models for the first time, it's probably the best general purpose assistant productized version of a language model. And then if you're interested in more technical stuff, I think Claude code and moving into the coding agents is a really good move. What has been some of your experience in the last six months about like, I prefer ChatGPT for this, I prefer Claude for this, or I prefer Gemini for this, your AI stack?

16:48I'd say like my general purpose, like my web browser, I use the ChatGPT Atlas browser. So it's got ChatGPT baked into the web browser and it can control the browser. So you can tell it to click around and you can tell it to book flights for you, book hotels. I'm the kind of person that doesn't book a flight until last second. And it's not that you know you're going to go on that trip, but you just don't take the 15, 20 minutes it takes to sit down and book a flight. And I'll just open a tab and these days I just say, Chachapiti, go find me the best flight in the evening from LA to San Jose. And it'll go find that.

17:18And then it's like, go find a hotel. And it'll find that. And then all I do is the final booking. So I really like Chachapiti Atlas as taking over this kind of like everyday kind of drudgery kind of work. It's very interesting. It's kind of like a mechanical Turk AI that just does the menial form-filling kind of task. How much memory context of you do you have the agent keep in mind, like for presumably ChatGBT with GBT Atlas, to like when you say best hotel, best flight, it knows what your parameters are, for example? So I think memory is a very interesting thing that these language models are beginning to start getting their grasp on.

17:57And memory is like, what is that personal context about me, my preferences, my tendencies, the things that I actually prefer that I would want the model to know so that anytime it takes an action, it's like, oh, Parth likes to sleep in, so maybe don't book him a 6 a.m. flight. That would be an interesting memory point for the model to take into consideration before making agentic decisions on your behalf, like booking a flight, for example. or you know you really only need like like i i i think that like the better it knows you the better it can help you with some of these things but memory is very tricky i think it's largely unsolved i think chat chiquiti has more of my memories but i also noticed that um sometimes i'm like i don't i actually prefer talking to a coding agent that knows nothing about me personally and i and i and i because i like to sometimes have a fresh slate with these models where they're not assuming anything about my my preferences or my tendencies so i think it goes both ways the personal co-pilot you kind of do want it to have some sense of memory but then like your automation tools maybe they don't need the same level of personal life kind of understanding and then my everyday primary co-pilot is chat gpt um obviously because it's got a great mobile app you can point your phone at things and talk to it about the real world use it to help you solve problems every day like even just like oh how should i organize my my like apartment like i obviously should buy some containers to organize my my my closet it's very good for like the in-person real world kind of ai and really good for research chat gpt pro mode is the best research tool that i've ever seen and it's it continues to be the case and then i would say after chat gpt it's i've got my coding agents so i usually have three coding agents assigned to almost anything that i care about in my life.

19:44Just like this. Yeah, I have a Claude, a Codex, and a Gemini, and they're just attacking three. And this is just one project. I have like, you know, seven, eight other projects where I'm just spawning more agents in different directions at the same time. So I would say the coding agents are more like my ambient fleet, where anything I care about has one to three agents. Any project that we're working on, read AI, some of these creative projects, or when we're building something new, the first thing I do is I create a folder and I put Cloud OpenAI's Codex and Google's Gemini into that project. And I just send them in three different directions on that project in an empty folder.

20:19And I kind of transcribe to them and I tell them what I want them to do. And then they kind of just work on that in the background. And then I spin up another project and I have three more on that project. And so I'm at this point now where it's like anytime I have a new idea, my instinct is to put three agents on it and have them make some progress before I come back to it. What do you think is the maximum number of projects you've spun up with your trio of agents? This fleet, I've gotten up to like 17 projects where they have on average two to three agents each one. I think this is just a current constraint and it's something it took me like three months.

20:53The last three months I kind of designed my workspace to allow for this kind of context switching. Because then it's like when especially when if you look at the frontier models now, gpt 5.2 and opus 4.5 you can set them up if you tell them if you give them a good approach to planning and you say make sure you write down the plan and periodically update your status in the plan you can say go work on this for today and it will continue to work in a loop for a whole day even longer than that and then people are like oh how like why would you want multiple agents and i was like because when you tell someone to work on something like you're not going to just walk away from the computer and go to the beach.

21:31You're like, no, actually I want to fire up another one. Then the justification for having a small fleet, a couple different directions going at the same time makes perfect sense to me. And I think that as we get better at the user interface constraints and the context switching, I think the code review and how many of these can you manage is really a question. I think a lot of people are still in the single co-pilot phase. and I think right now is probably the time to move into the how many of these can you orchestrate at the same time across a range of different projects. Like the trio of agents, awesome.

22:07We'll get to what people should be doing with one agent in a second. What's the funniest thing that comes to mind about what's gone wrong with one of your trio of agents projects? When we first met two years ago, I was thinking about this problem, an earlier version of this problem of like, can I put two chatbots in a room and then give an objective and then like can they make progress and then walk away from my computer the first time I did that I realized that they you know they would work on a problem and then they would not know when to end the conversation and so they would just say thank you to each other in an endless loop and I came back and I spent like a couple hundred dollars on the word thank you and I was like oh my god what are we doing here do I need to create a manager to say end the conversation which is what I did like it felt like a naive thing but i was like makes sense to have a third person in the room that just ends the meeting right um and then and so but that kind of clued me into this idea that maybe the the coordination across multiple chatbots with their own context windows is like largely kind of unsolved territory and now we're in that same phase we're in the next version of that same problem which is they're not just chatbots but now they can take actions they can work on projects and i tried it again right i said okay what if i had the gave them the ability to talk to each other in or like send messages to each other right they can send messages to each other or they can see where each other are in a project and it was experimental it was basically allowed three different coding agents to dm each other while they worked and i realized one they kind of missed timing a lot so like they'll be working but they they'll get a message and then they'll reply to the message by that time the other agent's already like halfway through the problem so there's something about like timing and i also realized that uh it opens interesting questions as to like permissions because you'll have an agent and it'll be working on a problem and it might be like oh user i need some i need you to approve me using this tool and me being a human i can give that approval but then if it starts asking the other agents for approval it's like oh hey codex i need to use this analytics tool what do you think and then codex is very willing to just approve that access and i was like okay maybe this breaks our definition of a sandbox if they can just ask each other to permission themselves is it okay if i erase the hard drive yes absolutely yeah exactly Exactly, that's the risk there, right?

24:19So then I was like, okay, if we're gonna do that kind of experiment, we should be sandboxing it, we should run it on a machine where we don't care if we lose everything, right? Like a virtual machine with a Docker container. But it is interesting because I think we're gonna get there where like, I want them to be able, the only, like how I deal with that now is I just send them in different directions. I say like, you're gonna research, you look for bugs, and then you help me with like the blog on my website. And then I know like their work doesn't overlap. And so there's less likely to have, we're not going to have this like collision of like different workers and i'm sure that something similar to what github did for humans is going to be there's going to be something some similar kind of solution for coordinating multiple agents working together on the same project then when i want to deploy software to the web i want to make an app and then share with people i use replit so replit is another coding agent but um and and these these three agents will work with the replit agent on a project like my website is built and hosted in replit and these codecs and cloud code have helped with the website but it is it is largely replit agent has built the website and so that's really good because if you want to deploy software there's no easier way than replit and then for image models i really like um nano banana from uh from from google nano banana pro from google is the most powerful image model and it's good for it's not just about generating cool images of like surreal things but it's also very good at building like designing infographics that are extremely coherent and um ethan mollick he thinks that he thinks that nanobanana is kind of like a successor to powerpoint in the sense that it can create this rich visuals with coherent text and layout um i think i think it's some combination of code and and generative images like nanobanana is probably the evolution of presentation tooling i like nanobanana i think other image models i like flux flux from the black forest labs team um it's a it's kind of an open source model um i also like of course mid journey is it's best in class in its own category for video models i really like sora and vo3 from google so sora from open ai and vo3 from google these are the most powerful video models i've seen and um so i use them for animation pipelines so i'll take a coding agent and i'll give it an image model and then i'll give it a video model and i say help me animate this world and it'll start creating scenes and characters and animating the intersection of coding agents and all these other creative tooling is also a very interesting it's a new form of animation new form of like world building and storytelling some of our folks are probably listening to your description and thinking oh my god sci-fi like you know william gibson the future is already here just unevenly distributed but like if someone said okay i'm gonna start using an agent yeah what would you go okay start using an agent thus i would say start with chat gpt and maybe i would say start with like the chat gpt atlas browser if i type slash agent mode now we're in agent mode and so chat gpt is a chatbot but i can say let's go to the tech meme front page and look at some of the headlines and then maybe pick an article and then open it up i've transcribed my prompt and i'm going to send this message and now chat gpt is going to take control over the browser and then navigate to the website and explore the front page.

27:31Looks like there's an OpenAI article right there. And you can see on the right side, it's thinking about how to browse the website. It's clicking on the first, it's probably going to click on the first article. It's reading this article. Let's surf Wikipedia and I want to learn about language models. So let's go explore Wikipedia. And by the way, for those people who don't have the PATH command line, the agent mode is also available on the button. Yeah, there is a button. I've gotten so hot key oriented that I won't even use the trackpad. Right now we're watching ChatGPT use a web browser. If you're just getting into what is an agent and maybe you've talked to ChatGPT or a language model before, this is a version of an agent that can take actions.

28:16Without me clicking around, it took us to Wikipedia. You could tell it to book a flight. You could tell it to research a topic. and you can fire off more. You can open up more tabs and then have more agent mode queries running. And on the side, it's explaining, we can talk to ChatGPT about the contents. So I think that ChatGPT as an agent is probably the best everyday agent. There's so many more capabilities, but I think for research, it's probably one of the best. And also just for personal everyday kind of exploring the web. Everything we were using the web browser for is now you have a very intelligent language model that can help you explore pretty much any topic, learn anything, teach yourself anything.

28:55And I think that's the most powerful thing you can do with AI is have it teach you about how the world works. So we've covered some of the entry points. Now, what is it, one of the things that, as you know, one of the ways that I describe you off the matrix is not only have you taken the red pill, but you're bathing in the red pill. And as a question that's kind of in the - The Morpheus. Yes, the Morpheus thing. What does it look like for an individual to cross the line from asking questions and using ChatGPT as a search engine or give me my research Wikipedia answer, etc., but to building agentic systems, automations around themselves, and in particular, share an example from what you do?

29:41We talked about ChatGPT, that's a great intro to language models and agents. I think that if you want to go one step further, there are limitations to this chat experience in a web browser where the real power of the language model is unlocked when you give it a computer and you allow it to use a computer with you. Use your own computer and working with files on your computer. For me, one of the first things I realized I needed, I was like, I should have a personal website. I should start talking about these things. I thought, well, AI is going to help me build my website. And I've never done this before.