Tibo, a successful entrepreneur who sold Tweet Hunter and Taplio for $8 million, shares his journey and insights into leveraging AI for product development. This interview covers his early adoption of GPT-3, his current AI tool stack, cost management strategies, prompting techniques, and thoughts on the future of AI in gaming. He emphasizes a scrappy, iterative approach, focusing on building tools for personal use and rapidly shipping products. He is currently building revid.ai, feather.so, superx.so, and outrank.so.

Insights

• Early Adoption Advantage (and Misconception): Tibo started using GPT-3 in 2021, focusing on analysis rather than generation. He highlights the misconception that it's "too late" to enter the AI space, emphasizing the vast untapped potential.
• Build for Yourself: Tibo builds tools for himself to solve his own problems and streamline workflows, which allows for deeper understanding and faster iteration.
• Leverage High-Level APIs: He prioritizes using readily available APIs like Replicate, AssemblyAI, and others to save time and avoid the complexities of managing servers and GPUs.
• Character Consistency in Video: He discusses techniques like masking and inpainting to maintain character consistency in generated videos, addressing a significant challenge in the field.
• AI Costs and Margin Management: Tibo openly discusses the high costs of AI video generation and the challenges of balancing expenses with user expectations and rapid growth.
• One-Day Sprints for Rapid Iteration: He shares his strategy of dedicating one day a week to building and shipping a completely new project, highlighting the importance of rapid iteration and experimentation.
• Mega Context Prompts: Tibo uses TypingMind to create and manage project-specific mega prompts, ensuring consistent context and eliminating the need to repeat instructions.
• Generated by AI, Polished by Humans: He champions using AI for generating the initial draft and then refining it with human input, optimizing both speed and quality.
• Structured Outputs with JSON: He emphasizes the reliability of using JSON for structured outputs from language models, simplifying parsing and integration.

Frameworks

Build for Yourself

Tibo's core philosophy revolves around building tools that he would personally use. This allows him to deeply understand the problem he's trying to solve and iterate quickly based on his own needs. It also bypasses the complexities of traditional product development processes.

One-Day Sprints

Tibo dedicates one day a week to developing and shipping a new project. This rapid iteration cycle allows him to experiment with new ideas and technologies, increasing the likelihood of stumbling upon a successful product. It also fosters a mindset of continuous learning and experimentation.

Generated by AI, Polished by Humans

Tibo advocates using AI to generate the first draft, which he then refines manually. This framework allows him to leverage the speed and capabilities of AI while maintaining quality control and adding a human touch. It optimizes for both efficiency and quality in content and code creation.

Transcript

00:00:00 Tibo: All the people that think this today, I think they are so wrong too. Like, there's so many things to be left. Model that just got out a few weeks ago, it's much better at creative thinking and it's 10 times cheaper. Replicate is an incredible example. So from day 0, I want to save time by use super high level APIs. It was way better compared to what a human can do to analyze, summarize, cluster, and categorize stuff. It honestly saved me 1 hour by creating the bones of the component.

00:00:47 Greg: Meet Tibo. His first product had $0 in sales. His second product hardly made any cash at all. But on his 15th product, Tibo sold Tweet Hunter and Taplio for $8,000,000. In 2022, he was awarded maker of the year by Product Hunt. In this interview, Tibo walks us through which AI tools he has bookmarked in his browser, which techniques he's exploring for character consistency in video generation, and his thoughts about when to use o one and ClaudeSonic together, or maybe not together, when generating blog posts. Let's get into AI show and tell with Tibo. Tibo, I'll tell you what. So I was looking at your Twitter during my research here, and in 2021, you had a very interesting tweet.

00:01:27 Greg: And the tweet said, do you want me to have GPT 3 analyze your tweets for you? And what blew my mind was this is 2021. So when did you start with AI? How would you walk me through that?

00:01:40 Tibo: I think that with that I think it was, like, one of my first who got some decent engagements this night. And I think it was the very, very beginning of my successful startup. The thing is, I think at the time, I already had a few project failures with GPT 3 because, like, I think, like everyone else, I was trying to generate stuff. Like, I was trying to use GPT 3 to generate tweets, generate copyright copy. And at that time with these tweets, I realized that GPT 3 was super, super good at analyzing stuff. Like, it was it was way better compared to what a human can do to analyze, summarize, cluster, and categorize stuff.

00:02:31 Tibo: And so I I did this tweet where I would just use the a p the Twitter API behind the scene, getting some information, dumping everything into a g p three prompts, and just just tell it's like, analyze stuff. Tell me what's what's interesting about this profile. What's what's the main topic that's generate engagement for him and it works. And I was I remember that I was manually, like, launching my function, my GPT 3 function, like copy pasting the handle of the guy putting it putting the handle on my code, launching the function, and copy pasting the results back to Twitter as a reply.

00:03:19 Greg: Yeah.

00:03:20 Tibo: I did, like, 20 or 30. And every time I had the like, I had a mistake in my code, like, I was seeing something bad, I was I was actually fixing the prompts on the go and just trying to improve it. It was so fun. Like, I was I was seeing the likes going up and my prompts getting better and better. It It was it was so fun.

00:03:42 Greg: That's wild. So your likes go up, the prompt and performance goes up. But then also probably your motivation and your excitement goes up at the same time.

00:03:48 Tibo: Yeah. Exactly.

00:03:49 Greg: Yeah. Yeah. Yeah. And that was back when it was still magic. Because, like, 2021, that was pre chat GPT. So this stuff was still, like, relatively new all over the place. I feel like it wasn't a big crowd that even knew what GPT 3 was at that point.

00:04:02 Tibo: You know what? Like, when we joined, I think I think we needed, like, 6 months to get a GPT 3 key, like an an open AI key because it was so restricted at the time. And when we finally got it, it was like everything was done. Like, all the cool use cases was built and it was too late. Like, we were we missed the train. Yeah. And it was so wrong. And all the people that think this today, I think they are so wrong too. Like, there's so many things to be left.

00:04:33 Greg: Well, so let let's keep on going down that path. And so I love that metaphor that that was your mindset back then in 2021, and you still kinda feel the same way today. So, like, what's your advice for people who say we're too late? Like, we've built everything we need to so far.

00:04:46 Tibo: I would like, I think my the thing that worked for me and that I want to repeat at every new thing that I that I build is I want to build something for myself. Like, I I want I want to be maybe not the main user, but I want to be the user of my own product because I don't have the discipline or I don't have this case to, I think, to manage a big team and to build the right processes so that the correct product manager can learn from the clients and categorize the feedback in a good way so that another team can build the thing. Build the thing. I don't have these skills. What I know is that when I when I use a tool, I know how to fix my problems.

00:05:26 Tibo: I know how to take my daily workflow and to make it simpler, thanks to some code that I will just dump somewhere in an app.

00:05:42 Greg: Yeah. Yeah. Yeah. Absolutely. Well, I'd love to hear more about the tool stack that you're building with. And so as you're building these tools for yourself, you had a tweet that said or it it kind of showed your p and l, so it showed with your expenses. And you were talking about Deepgram, OpenAI, Fall, LUMA, Sieve, 11 Labs, all of this. Where is your tool stack settled from an AI perspective right now?

00:06:05 Tibo: I think I have a are you a are you a ARC user? No, I

00:06:10 Greg: no. I do promise.

00:06:12 Tibo: ARC is my browser and in ARC I have like folders. Like like, it's a little bit like a bookmark, but it's it's better organized and have a folder that's Could you show us? Yeah. Yeah.

00:06:25 Greg: Sure. That'd be really cool to see.

00:06:28 Tibo: So this is like my all my cool AI APIs that I'm gonna use. Like, it's if if you combine the power of file, the power of, like, replicates and and save too where where is it? This one. You have, like, you have tons of AI models that are, are ready to use and super powerful. I think, like, most people that are trying to like, I see tons of people that are trying to get a server up and up and running, like, set up a few GPUs and scroll hacking phase, find coolant models to use, and spend so much time setting them up, fixing crash, fixing configuration issues. And I really don't see the point of that. Like, those providers, they're going to be more expensive, but they're going to save you so much time.

00:07:30 Tibo: Like Yeah. Replicate is an incredible example because they have so many models. There are so many that you can't, you cannot even like explore them. You have to search for something because you have just so many things. So from pretty much day 0, because I want to have a super lean team, I've always wanted to, like, use, like, ready to use things. Like, I I want I want to save time by use super high level APIs Mhmm. Like them together with code and and make a product out of that.

00:08:08 Greg: Yeah.

00:08:08 Tibo: And I think when you do when when you are able to chain a few APIs, like chain Chargept to then generates an AI image to to then put that in the video, it's, like, so satisfying. You know?

00:08:22 Greg: Yeah. Yeah. Yeah. Yeah. Well, I took so I tell you what. I would love to hear how do these tools chain together? So what do you use replicate for? What do you use c for? What do you use Luma for? And, like, how how does your stack look like?

00:08:35 Tibo: I would try to, like I think I think file is very good as so you have, for example, like I'm using Flux a lot in in file. I think it's, like, one of the best provider for for Flux. I'm using I'm just generating images there. Then I'm going to, like, let's say, minimax, this one, or Runway. They have a few image to video models. And so you are like that, you are able to, like, have a prompt with gpt3 or Sonet or anything else, like, that would generate the image. Then you use Flex to generate the image with the prompts. Then you take this image and you put it into a runway. And just like that, you have a 3 step workflow where you're able to generate a video a short video throughout, like, basically anything.

00:09:37 Tibo: Yeah. One thing that is critical with file, and I haven't explored that, but they they make it quite easy to do comfy UI workflows, which then lets you build super specific and complex AI workflows by just chaining models between them. Like

00:10:03 Greg: Sure.

00:10:04 Tibo: Something like one thing that I haven't nailed yet, and I'm working a lot on that, is I want to be able to, like, use this kind of stuff, like train a Flux LoRa, meaning that I would be, like, training training an actual model from a few images of, like, someone or something. Mhmm. And then use that model to then generates generate a video of the same

00:10:37 Greg: person,

00:10:38 Tibo: with a strong consistency in the video itself. So, like, a a video is, like, 1 or 2 minutes long.

00:10:45 Greg: Uh-huh.

00:10:46 Tibo: Right now, it's incredibly hard to have character consistency in a video because you have to generate, many, many images and and short videos.

00:10:56 Greg: Sure. And what what what's the secret to consistency for those things?

00:11:00 Tibo: I don't so there's there's one secret that I'm, like, I'm just scratching the surface of that. That's you I I I think that is, like, a a strong like, a huge value in using mask and in painting. I mean, like, basically, you will generate an image. You will with a character, and you will just try to create a mask from about this character. You will extract this character and regenerate the correct image with by using this custom model with a tech called in painting where you just Sure. You just tell the the the model to just generates the missing piece of of the image. Does that make sense to you?

00:11:45 Greg: For sure. It's like, let the model run wild and everything in the background and the body and the sky and everything, blah, blah, blah. Because that's not what you want the consistency for. But then mask and then maybe the face because then that's what people really care about. And that's what you really need to get right is the face piece of it.

00:11:59 Tibo: And and this step is super important if you want multiple characters. If you want the the issue with fine tuned models when you're trying to generate images is that if you are using if you combine multiple fine tune models where each one has been trained on a specific character Mhmm. The image generation model will just merge all the characters into 1. And so all the characters on the image will look the same and will be a merge version of all the fine tune models.

00:12:36 Greg: Sure.

00:12:36 Tibo: Which is interesting. This is Harvard.

00:12:38 Greg: Yeah. Yeah. Yeah. For sure. It's interesting. It's like you want a model for every subject that you want consistent. And so it's almost like you have the parent image, and then you have a tree of all the subjects in And so it's almost like you have the parent image and then you have a tree of all the subjects in there that you need to have remain consistent. Yep. And so that's that's some I I'm just I'm just I'm

00:12:49 Tibo: blown away by by the fact that this has not been properly handled by like, this is a super important problem in my opinion, and there's a big value in fixing that.

00:13:05 Greg: Well, I tell you what. I mean, look at where text was even just 6 months ago. You couldn't do it. And now it's like a non issue. So I think that these things will just come around.

00:13:12 Tibo: Yeah. Sure. So

00:13:14 Greg: You had another tweet that was super interesting, and you said that AI is eating your margins. So, basically, I I read that as there's it's super expensive, paying a lot of money for it. How do you manage costs, or how do you track costs across all these tools?

00:13:28 Tibo: I have a big issue right now. It's that's like with my projects, like Revit dotai. It's it's the first time that I'm I'm working with videos. And AI videos is the most expensive stuff that I have ever seen. If you, like, if you are if you're trying to use if you're trying to, like, the the top level models, like Runway, Luma, or Cling, they cost you between 20 to 50¢ per 5 second video.

00:14:04 Greg: Wow. Yeah.

00:14:05 Tibo: And so when you have a user that is trying to create, like, a a 1 minute video, it's it's easily gonna cost you, like, more than a dollar, $2, maybe 3 if you add some some, like, voice duration on top of that. And so there is a huge disconnect between the price of those API and the expectation from the user. Like, the the users, they don't understand why it would cost that much because it's just 1 minute video. And so they are constantly expecting us to lower the price or they're just screaming at us because of the price of the products. But at the same time, our AI costs are so high. Like, they are I haven't like, more than more than 60% of the projects.

00:14:54 Greg: That's wild. That's wild. You know, one of my favorite things I don't love everything about Elon, but one of my frameworks I took from him was his idiot index, which is the price of something the price of the raw materials to actually do it. And then the delta between the two says how much of, like, you know, buffer there is from that. Do you know how much it costs hardware providers to generate those videos versus what they're actually charging you? Like, what are their

00:15:17 Tibo: margins? I have no idea. Like, this is not public. It's it might be that they are not even profitable. Like like, touch gbt is not like, OpenAI is not profitable maybe because

00:15:28 Greg: Sure.

00:15:28 Tibo: I think I think they want to, like, they want to, like, get the markets. They want to get the market share. And so they are, like, lowering their price to their very minimum they can. It's it's likely that they are not even making profit out of that.

00:15:46 Greg: Sure. Well, it's like the rumor about Grok having unprofitable tokens as well. The the, not the xai one, but the, the hardware provider.

00:15:54 Tibo: Yeah. Yeah. I see. Like, it's it's it's possible. The Yeah. Their generation is so fast. Like, it's it's likely that's it's the case. Yeah.

00:16:03 Greg: Yeah. Yeah. Yeah. For sure.

00:16:04 Tibo: For this project that I talked about, like, Revi dotai, the the issue is, like, it's there there is something that you, learned at beta school. And, for, like, for 10 years, I thought that it was not relevant to me, but it seems like it is. Like, the thing is our AI costs are more than half of what we are getting every month. And we are growing so fast, like, 50, 60, 60% month over month.

00:16:35 Greg: Wow.

00:16:36 Tibo: And and our revenue we we get the payout of our revenue 1 month later compared to what when we are getting it. Yeah. All that summed up makes us very hard to sustain the growth because we are paying right now the expenses of this month, but we are getting right now the revenue of the month before. And so since the growth is so hard, we are just barely getting enough money from the sales to cover the expenses of this month because of this delay.

00:17:11 Greg: Yeah. That's wild. It's funny how there's no better motivator to learn business topics than having too many expenses and too much money to try to deal with it.

00:17:19 Tibo: That's that's a good problem to have, I guess, but it's still surprising.

00:17:23 Greg: Yeah. Yeah. Yeah. Absolutely. You know, one thing that you were describing as you were coming through, you were talking about fall with Comfy UI, then you can do complex workflows for that. Was the mask example that you gave an example of one of those complex workflows? Or what's complex to you that you would be, you know, coding up in there?

00:17:40 Tibo: Yes. It is. Like, this this process where Meta just released some an awesome mother called called the, like, SAM 2.

00:17:51 Greg: The object detection 1.

00:17:52 Tibo: Yeah. Exactly. It it's it's able to detect the very specific edge of an object on an image and even a video, and it's able to label it. And so using that plus some in painting tech and some other advanced labeling technology, like, all that together, yes, using Convy UI is a nice thing to use. I'm not, unfortunately, because I'm not comfortable with that yet. Mhmm. I'm too I guess I'm too scrappy to to use it, so I need to, like, see my code and be able to, like, fully control it. But I'm pretty sure I would be able to do great things with it.

00:18:37 Greg: Yeah. Well, I tell you what. You you strike me as somebody who, when he wants to learn something, he has a pretty good way, a determined like, a very determined way of going into it. So say you're gonna approach Learning Comp UI. What's your personal learning plan to go do so?

00:18:52 Tibo: I would try to be like the bare minimum thing. Like, the just just I I would try to, like, get the first ID that I have and start working on it right away and try to ship something to the public as soon as possible. And I would I would treat about that even if it's, like, even if it's shitty and even if it's not, not that useful, I would just tweet about it and expect a few of my friends to test it and give me feedback. So 2 years ago, I was living in Bali for like 6 months. Nice. And at the time, I was working on Tweet Hunter, one of my most successful business.

00:19:33 Greg: Mhmm.

00:19:34 Tibo: And what was super cool is that I thanks thanks to a small group of hackers that I met there, I was able to, like, all the week work on Twitter Hunter, my main projects. But one day per week on Thursday, I would meet the group of hacker. And during one day, I would try to build something totally different. And it it can be related to Twitter. Like, many times I build, like, some me tool that was different code base, different setup, different server compared to the main projects, but could serve the growth of the main products by just being a firm thing aiming at going viral. And and I found this I found this to be quite healthy, like, to have, like, 4 days per week dedicated to my main thing and one day per week where I would try to read and ship something in just one day.

00:20:33 Tibo: Yeah. And I I did, like, many times. And I think my most viral products my my most viral tool Uh-huh. I really did that way. Doing one That's

00:20:42 Greg: so cool.

00:20:42 Tibo: One day sprints. That's it.

00:20:45 Greg: That's so cool. It just what stands out to me about that is when Sam says that the entrepreneur's way is to basically have a lot of shots on goal because you just need to be right for one of them. And so having one day a week where you can I I think the important part about that is is 2 things? 1, you have the execution to actually do it and actually commit to it, which is really nice and the consistency. But 2 is you package up the entire thing that you can ship in one day. So if you can't do it in one day, then it's it's really tough to do.

00:21:12 Tibo: I think I think you're so right. That's when we started, like, I think in, in 2021, like a few weeks after the initial tweet that you mentioned, we, we, we were in this process of building 1 project per week for and we did that for like 4 months. During 4 months, we tried to, like, ship 1 product per week because 9 of 9 product out of 10 fails. And it was, like, the 10th one who eventually got successful. Yeah. And that's I think that's how you build something truly valuable, like, by creating 10 things without knowing the one that will be valuable. If you're up for that, I want to talk about this all one thing because Sure.

00:21:59 Tibo: Because I I kind of changed my mind.

00:22:03 Greg: Oh, yeah. That's great. So let's talk about o one. What what are your thoughts on it?

00:22:06 Tibo: So the thing is I'm I'm working for, like, a few weeks on this new SEO project, where, basically, we have an interesting approach where we're trying to find the super relevant keywords to to rank for for your business. It's not launched yet, but I'm onboarding people right now in the beta phase. And it's it's working very well. And at the very beginning of the project, what's unlocked the project was the reason of o one. Because on this project, we were we're willing to spend a lot of money on prompts and LLMs and to to generate high quality blog posts. And o one was quite good at that. Like, it it's we thought that it would be the the the the piece of tech that would unlock the high quality contents.

00:22:57 Tibo: And the thing is it's very good at math. It's very good at logic. It's surprisingly good at complex problem solving. But the thing is the new release of, like, the new Sonet model that just got out a few weeks ago

00:23:18 Greg: Mhmm.

00:23:19 Tibo: It's much better at creative thinking.

00:23:22 Greg: Mhmm.

00:23:22 Tibo: And it's, I think, it's 10 times cheaper. And so we reverted to this one. And and and, yeah, we just by by prompting, by getting more in-depth about what we want, by providing some example. Like, I think I think this is super important. People don't do this enough. It's by by providing good context and example about what we wanted. I think we are able to work with SONETs and to get better result than o one. That's specific to our workflow, by the way. But I think creative thinking is a very, very interesting way where all one is not very good.

00:24:06 Greg: For sure. So I would imagine when you're doing, AI generated blog posts, there's many steps in the pipeline from candidate ideas, pick the idea, pick the title, write the actual thing, come up with the outline, write the come up with the outline, write the actual paper, etcetera, etcetera. Like, there's a lot that goes on to it. Will you use Sonnet for every single one of those steps? Will you one shot it? How how will you break it down? Like, what's what's your thought process with that?

00:24:30 Tibo: Not not entirely. Like, there is one thing where when you're trying to generate very long stuff, like, very actuates and in-depth contents, SONET can sometimes try to limit the length of the generation. The one model that I found that is not trying to do that is Gemini. Gemini 1.5 Pro is very good at generating long content.

00:24:58 Greg: Cool. Beautiful. So will you have Gemini actually do the writing for you or will it write the outline for you? Or how's that work?

00:25:06 Tibo: I cannot go into that.

00:25:07 Greg: Okay. That's fine. Alright. No. I I understand. Yeah. I totally understand. Okay. So that's the o one side. Beautiful.

00:25:15 Tibo: The maybe the one thing that's I think it's interesting and to be cool to show up is I would I would highly encourage I I will share my screen.

00:25:28 Greg: Yes. Fabulous. I love this.

00:25:31 Tibo: The I I think I think, like, you you said that you are working on many projects, and I am too. And I think that's the norm. Like, I think I think right now, most people are working on many, many projects because they have unfinished stuff. They have the new stuff. And the one thing that I do, which I think is quite useful, is I'm using typing mind. It's from it's basically like another chat gpt UI, but it's very convenient for me to to make it to work on multiple project because I have those folders. And those folders, they are that like, my all my project that I'm running right now. And if we take this one, for example, I would be able to, like, set a specific set of instruction for these specific projects.

00:26:26 Tibo: And so then every time that I will run a new prompt, on on on this project, it will have this, like this context about this project. I will not have to put everything back here.

00:26:43 Greg: Mhmm.

00:26:44 Tibo: I think this is a very simple step. Like, it's it's not crazy workflow that you will spend weeks to build, but just spend spend the time to work on this mega context prompts for your projects, do it for each of your projects. And then you have something super specific, super useful where you can ask anything anytime.

00:27:09 Greg: Yeah. That's very cool. You didn't build that project, did you?

00:27:13 Tibo: No. I didn't. I would love to if I did, but I'm not. It's a guy called Tony Dean.

00:27:20 Greg: Okay.

00:27:21 Tibo: He's he's pretty big on Twitter.

00:27:23 Greg: That's very cool. That'd be cool if after each chat session that you have with it, it's like, hey, here's the new stuff that I extracted that's not in your system prompt already. Do you want me to add it to it? And then it just comes to the top.

00:27:33 Tibo: Amazing. Yeah. That's be amazing. I I think TagePT is trying to do that, but

00:27:38 Greg: Yeah.

00:27:38 Tibo: It's it's so it's way too b to c to do it in a in a good way for us.

00:27:45 Greg: Sure. It's almost it'd b to c. Yeah. You must need you want the power user use case is Yeah. Exactly. Yeah. Yeah. Yeah. Yeah. That's the side that you're on for it. That's cool. Nice. Well, do you think that the regex demo is interesting? Like, would that be cool to walk through? Is that kinda small?

00:28:01 Tibo: No. I don't think so. But Okay.

00:28:03 Greg: That's fine.

00:28:03 Tibo: Basically basically, basically, I'm using cursor. Do do you know about cursor?

00:28:09 Greg: Oh, yeah. Big time. I'm a I'm a huge cursor user.

00:28:12 Tibo: And basically, every time I have to every time I have to use a cron Mhmm. I I triple check my my cron with with cursor. Like, I just I have I have these tools where I just press command t by selecting the the text, the the regex or the the Chrome and and just tell me if it's good or not. Something that I do a lot too after testing, I select my function or my code and I I do the command t again and I will ask it if is there a mistake here? Like, is it is there an edge case that I didn't think about? And that's it. And and just doing that allow me to go to be to be way lighter on the on the testing, like, on the manual testing and push the production very fast.

00:29:07 Greg: Interesting. And I was just speaking with Sully Omar who says that when he uses cursor and he writes with it, he actually has it write the test first, And then it's a lot easier to actually write the function to execute against that test because you have a validated test that comes with it.

00:29:21 Tibo: Yeah. That's amazing. This this is this is super interesting work for the thing is I've never I've never written a test in in my life, and I don't really I don't really plan in starting. But when you when you used to do that, I think that's very powerful.

00:29:38 Greg: Yeah. Yeah. Yeah. That's cool. That's beautiful. I wanna jump back over to the prompt side real quick. And so you said you're using a lot of Sonnet. And I know that Sonnet is works better with XML. So what sort of special formatting do you put into your prompts with regards to headers or XML or anything like that?

00:29:55 Tibo: I'm doing JSON every time.

00:29:57 Greg: JSON?

00:29:57 Tibo: Basically, I think I think that the sentence that I type the most is output in JSON. Do not output anything else. Uh-huh. Uh-huh. That is

00:30:09 Greg: So it's basically structured outputs only that come out from it.

00:30:12 Tibo: Yeah. And it's super reliable. Like, I I don't really understand why people say that it's not, but it is. Like, it's very reliable in in, like, being following this format that I'm going to I'm I'm showing it to you.

00:30:27 Greg: Yeah. Do you have a validator on the other end? Or do you just, like, accept the output no matter what and see what hopefully, it comes out alright?

00:30:35 Tibo: I have a custom made function where I validate the just not boots. And I retry if it failed to do so. But it was super useful 2 years ago when I was using a previous version of GPT 3 and GPT 4. But right now, it's not that much useful. Like, honestly, it's it's very reliable.

00:30:59 Greg: Cool. That's beautiful. Back in 2022, you said you're never gonna write you're never gonna write another commit message again. My question, are you writing commit messages still or no?

00:31:09 Tibo: Yeah. Because I'm working back with people. At the time, when I wrote that, I think I was I was using that so I was I think I was I was being a a douche, but let me tell you why. When I would when I when I wrote that, I was working with a team of people that I hired. And so I I made them agree to me using AI in my commit message, meaning that every single one of my commit message, it was it was basically me pressing common enter. And when I was pressing common enter, it would auto commit and auto push with an AI generated message.

00:31:58 Greg: Uh-huh.

00:31:59 Tibo: And right now, I'm working with people who are not not not employees. They are partner. Mhmm. And they really didn't like it. And so I had to change and and try to, like, try to be nice with them and Yeah. Write my own commit message.

00:32:20 Greg: Yeah. Yeah. Yeah. For sure. Well, that leans into one of the philosophies I saw that I really resonate that you had mentioned, which was generated by AI, polished by humans. And really, I think that that simple sentence alludes to a whole lot more about where AI fits or should fit into folks workflow, where it's you have it do the first 95% and then humans are on the other end doing the validation to make sure it's good. How does that saying fit into your workflow?

00:32:47 Tibo: That's that's how I call it right now. That's that's how I I program every day. And it's I think it's it's the beauty of a cursor is when I want to create a new feature, I would just tell it what to do, what I want to do, how I want it to work, and it will create the first draft of the components. Mhmm. And if if you set it up correctly, if you have these master prompts, if you if the project is somewhat well structured, it will create decent things. And and, like, I just I just did, like, a few hours ago. And it honestly saved me 1 hour by creating, like, the, like, the bones of the component. And I really didn't have to change, like, few lines.

00:33:39 Tibo: Sure. And I then ship it in production. It worked very well.

00:33:44 Greg: And then just ship it there. Are you using v zero or anything like that for front end components?

00:33:49 Tibo: I'm not, but I think I should. Yeah. Yeah. I I should. It's it's super powerful. And from what I'm seeing on Twitter, it's getting better and better every day. So yeah. I I should read that.

00:34:00 Greg: Well, what's amazing is having seen v zero from the beginning, they weren't investing it that much. I think it blew up. It started going nuts. And now they're throwing like a real team and real money

00:34:12 Tibo: behind it, which is really cool to see. Yeah. Yeah. Then that that's alright. Like, it's it's super valuable. Yeah. I I should I should give it a more, like, a a fair try, I guess.

00:34:19 Greg: Yeah. Yeah. Yeah. Yeah. For sure. We just need to get cursor to import v zero and have it just

00:34:26 Tibo: right before. There's actually, like yeah. I I have this. So I'm not using it, but I have in my project. Mhmm. So here I'm in cursor. And if I want to, like, create a new comp using 0, I would reference my master prompts with 0.

00:34:50 Greg: Oh, cool. Is that prompt public? Have you shared that?

00:34:56 Tibo: Yes. I think I I I got it from Twitter, so it should be

00:34:59 Greg: Oh, cool. Nice. Yeah.

00:35:03 Tibo: I'm gonna select for countries. Poof. And and by doing that, I think it should be able to, like yeah. It it creates a link. And this link, it's it's a v zero link with the prompt embedded in the query. Uh-huh. Oh, maybe it's not working anymore.

00:35:27 Greg: Wow. Oh, either way, I was gonna say, I did not know that, and I like that a

00:35:31 Tibo: lot. Yeah. But maybe the like, it seems like it's not working anymore. So maybe they they removed the support for the the query. Like, they re Interesting.

00:35:42 Greg: Well, call to builders out there. This is needed. We want this. Somebody please go do it. Just a few more questions here as we start to wrap up. What that type right or the the the the chat gbt alternative that you showed me, that was really interesting. What other mini tools like that are you grabbing for?

00:36:01 Tibo: That's a good question. Not not that much actually. I'm I'm using so I'm using TypingMine, the tool that I showed you earlier. I'm using Glip to, like, to handle customer supports, 200 tickets. It it's not that much deep into AI, but it has just the minimum stuff that, is useful to me. And I want to start experimenting with sidegpt. Sidegpt, yes, it's it's the it's a completely new SaaS that is supposed to totally, like, remove the need for a support team. So that's by by, like, by training it with the the correct documents, help center, it should be able to fully fully answer to your users with these unique capabilities that it it can take action.

00:37:10 Tibo: Like, if if a user want to cancel, it will not only show the user how to cancel, but it will actually be able to cancel the subscription for the user. And the way it it it's doing that is pretty simple. Like, it's it's quite easy to set up. And so I really want to dive into that and see if it can just help me save time.

00:37:31 Greg: That's cool. And this is sitegpt.ai. Right? Yes. Cool. Well

00:37:37 Tibo: it's it's quite it's quite old, but the maker completely revamps the software. And it's like it's the new version is completely new and mind blowing.

00:37:48 Greg: That's very cool. Tell me about what else is on your to learn

00:37:56 Tibo: list. Let me check. I have one.

00:37:58 Greg: Oh, I love this. This is so cool.

00:37:59 Tibo: It's very long. It's like it's a mix of things to explore.

00:38:06 Greg: Can you show us too?

00:38:08 Tibo: No. Because it's it's it's, like

00:38:12 Greg: Personal.

00:38:12 Tibo: So I I've I've seen that, like, the the AI provider that we mentioned earlier, it has some new cool models, and I want to explore them. Cool. I'll try to it's it's a very cool provider, and it has some things that I'm already using. But things like eye contact correction, for the spray. It's it's quite powerful. And this is just a way to, like when, like, you take this video where the you see the eyes sing around. And it's able to correct the video and to make the eye very static, always looking at the camera. Yeah. This is the kind of man that's it it might not change the game for me.

00:39:05 Greg: Mhmm.

00:39:05 Tibo: But it's it's still quite useful to improve the quality of some generated video for Revit.

00:39:11 Greg: Yeah. They

00:39:13 Tibo: have a lot that I haven't explored yet, and I really want to dig down.

00:39:17 Greg: Yeah. That's beautiful. You had another tweet that caught my eye. In July of 24, you said, I still haven't found a suitable use case where AI agents could help me for real. Have you found use cases where agents help you for real?

00:39:30 Tibo: No. And I so may maybe maybe it kinda work now with so I haven't I haven't tried with a new sonnet, but there was a few times where I built an entire agent infrastructure, meaning a set of actions, like, this this kind of infinite loop where an agent would try to achieve a goal by splitting this goal into multiple, like, smaller tasks. And the thing is because because those models that we have right now, they are, like, 95% accurate. If you if you like, if you take the those 5 percent errors that you have and you multiply them by the number of steps that the agent will take to achieve the task

00:40:24 Greg: Uh-huh.

00:40:24 Tibo: It's it it makes them unsustainable. Like, it it doesn't work because, like, most of the time, one error in the chain will mess up with the entire thing. And I understand that it's super interesting, but I just I just found that right now, I was not able to plug the correct context and set of actions to generate value. But for sure, it's happening. And probably in 2025, we'll have very interesting stuff related to agent. Like, Sam Atman Sam Atman himself said that's gonna be, like, the the the time for agent is coming. So Yeah. Yeah. It's something to watch, definitely.

00:41:09 Greg: Yeah. That's beautiful. Well, I tell you what, on that one, one question I'd like to ask at the end is you're on Twitter a lot. You see a lot of AI talk. You see a you see a lot of smart people talking about AI. What are the topics that the smart people in AI are talking about right now?

00:41:25 Tibo: I I've seen a lot of things about gaming. It's it was surprising because the I I kind of think that there are tons of people working on text for, like, for, like, 4 years right now. Pretty much the same for images. You have, like, very incredible models with 2 images. Videos is kinda works. Like, we have we have things. Like, we have very good providers. It's expensive and probably is gonna get disrupted. But we already know some of the main players, and they are out there. And gaming, it it it was, like, it was pretty silence for like for since the AI revision started, we haven't seen any good, good games and good game engines being disrupted.

00:42:17 Tibo: And maybe, maybe it's gonna come in 2025. Like, I've seen those demos about, like, Counter Strike and Minecraft being run by a neural net and not a game engine. And I found it super impressive. Like, it's how does it burn my mind? Because thinking that from just a prompt, you could create an entire game without all the game engine that would support it. It's it's such a huge change in and the thing is people playing video games, they already have GPUs. So

00:42:58 Greg: Uh-huh.

00:42:58 Tibo: It could be very easy to, like, to to them. If if it works, you can just get the games on quite powerful PCs from gamers and they could just run it without many things that makes game complicated. Totally. So again, we could see an entire industry being disrupted. Tons of people becoming, like, quite useless. And huge companies like a Red Engine or Unity 3 d be being completely useless.

00:43:29 Greg: Yeah. You know, I love in life when you have kind of, like, a frame breaker moment where you think one way, but then you hear something that's, like, completely different and completely Yeah. Shift your frame. And when I thought of games, I thought of, sure, you're gonna walk into a virtual and the the bartender is gonna have, like, a chatbot behind it and you're gonna have, like, an engaging conversation. Just some like, a chatbot basically in a game. But then when I saw those Minecraft demos, I thought, holy cow. I haven't even thought about this before hand.

00:43:55 Tibo: Exactly. Exactly.

00:43:57 Greg: Just the neural net generating the game. So it's like whole it's like this is different. Wait a minute. There's something interesting going on here.

00:44:04 Tibo: Yeah. We are so we're so like, we were thinking about how text based add ins would disrupt video games and it's really not happening that way.

00:44:14 Greg: Yeah. Yeah. Yeah. Yeah. It's always fun in life when that happens. Well, I tell you what, I wanna end on that one. Tibo, thank you very much for the conversation today.

00:44:21 Tibo: Thank you, Greg. It was awesome.