Pieter Levels
speaker
696 appearances
1 recordings
1 series
first heard Aug 2024
last heard Aug 2024
Pieter Levels’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsNo recordings in the last 12 months.Older appearances are listed below; set an alert to hear about the next one.
Appearances
So now I mix full body photos in the training with face photos, face crops. And it's all automatic. And I know that other people, they use, again, AI models to detect what are the best photos in this training set and then train on those. But it's all about training data and that's with everything in AI. Like how good your training data is,
is in many ways more important than how many steps you train for, like how many months or whatever with the GPUs, like the gold.
Like the photos should be diverse. So for example, if I only upload photos with a brown shirt or green shirt, the model will think that I'm training the green shirt. So the things that are the same every photo are the concepts that are trained. What you want is your face to be the concept that's trained. And everything else to be diverse, like different.
Yeah, outside, inside. But there's no like, this is the problem, there's no like manual for this. And nobody knew, we were all just, especially two years ago, we were all hacking, trying to test anything, anything you can think of. And it's frustrating. It's one of the most frustrating and also fun and challenging things to do because with AI, because... It's a black box.
And like Carpati, I think, says this. Like, we don't really know how this thing works, but it does something, but nobody really knows why, right? Like, we cannot look into the model of an LLM. Like, what is actually in there? We just know it's like a 3D matrix of numbers, right? So...
It's very frustrating because some things you think they're obvious that they will improve things will make them worse. And there's so many parameters you can tweak. So you're testing everything to improve things.
In a very vain way. Like me, you know? Like, I want to look good in your podcast, for example. Yeah, for sure.
Yeah, and look at you. You're one of the top podcasters in the world and you get millions of views and it works. And the people that spend so much money on optimizing for the best sound, for the best studio, they get like 300 views, you know? So what is this about? This is about that... either you do it really well, or also that a lot of these things don't matter.
Like what matters is probably the content of the podcast. Like you get the interesting guests. Focus on stuff that matters. Yeah, and I think this is very common. They call it gear acquisition syndrome, like gas. Like people in any industry do this. They just buy all the stuff.
There was a meme recently, like what's the name for the guy that buys all the stuff before he even started doing the hobby, right? Um, marketing, you know, marketing does that to people. They want to buy this stuff.
But like, man, you can make a, you can make a Hollywood movie on an iPhone, you know, if the content is good enough, it's, it, and it will probably be original because you would be using an iPhone for it, you know?
And you should probably do it yourself. Like you can use Photia, but it's even more fun if you do it yourself. So you train the models. You can learn about like control nets. Control net is where, for example, your photos in your podcast are usually like from the angle, right? So you can create a control net face pose that's always like this.
So every model, every photo you generate uses this control net pose, for example. I think it would be very fun for you to try out that stuff.
Do you play with lighting, with pose, with the... Man, actually, like this week or recently, there's a new model came out that can adjust the light of any photo, but also AI image with stable diffusion. I think it's called ReLight. And it's amazing. Like you can upload... Kind of like a light map. So for example, red, purple, blue, and use that light map to change the light on the photo you input.
It's amazing. So there's for sure a lot of stuff you can do.
I think you need to join Twitter. X, you know, X is amazing now. And the whole AI industry is on X. And they're all like anime avatars. So... It's funny because my friends ask me this, like, who should I follow to stay up to date? And I say, go to X and follow all the AI anime models that this person is following or follows. And I sent them some URL and they all start laughing, like, what is this?
But they're real, like, people hacking around in AI. They get hired by big companies and they're on X. And most of them are anonymous. This is very funny. They use anime avatars. I don't. But those people hack around and they publish what they're discovering. They talk about papers, for example. So yeah, definitely X. It's great.
Yeah.
Yeah, 100%. I like Twitter for that most, like building stuff, like seeing other, because it inspires you to build. And it's just fun to see other people share what they're discovering. And then you're like, okay, I'm going to make something too. It's just super fun. And so if you want to start going X, and then I would go to replicate and start trying to play with models.
And when you have something that kind of, you manually enter stuff, you set the parameters, something that works, you can make an app out of it or a website.
Showing 221–240 of 696 · page 12 of 35
← Previous
Next →