Meet Ethan - AI Developer!
AI researcher and developer working on Leonardo.Ai--we love his non-linear journey to Machine Learning.
“I originally had taken a course in high school on some really basic Java, which was fun. Then, the following course was a huge jump in difficulty,” Ethan says.
“The final project of the first course was to make a simple hangman game, while the first project of the second course was to run a simulation of a wildlife population over time, accounting for animals on a grid that may reproduce or eat another at each timestep.”
“So, I dropped out and I didn’t think about coding for a long time. Then, I became fascinated with image synthesis, at the start of 2022 and loved getting my hands dirty, messing around with the different kinds of settings there were. I did all sorts of comparisons to see how certain adjustments affected the image output, but it wasn’t enough for me. I had to know why this was all happening, and I wanted more control over it,” he says.
“After avoiding having to learn code for a while, I finally wrote my first line of Python in maybe mid-July, when I wanted to make some changes to Disco Diffusion. The changes I made didn’t require a ton of knowledge, just some light math and logic adjustments, and a sort of vague intuition of why what I wanted to do might work.”
“After several months post-release of MathRockDiffusion, I realized the changes I made actually had some theory and math to back them up, which was pretty neat. The journey is a bit frustrating but addicting,” Ethan adds.
Ethan also says, “I experience ADHD, I’m not great with organization, so some of the strict rules of code structure can be annoying, but also I get really fixated on working through problems, so I kept pushing myself to try more.”
A Code of A Different Color
Ethan says, “I’ve learned machine learning code is really different from other fields. I went in with an expectation that someone who has done general computer science may have an easy time picking up machine learning… but haven’t always seen that to be the case.”
In fact, the opposite has been true from Ethan’s experience.
“I’m not really good at web development, databases, and a lot of the other kind of stuff traditional coding revolves around. People laugh at this, but I like machine learning, because when I get an error, it makes sense to me,” Ethan adds.
“Like, I’ll see that some math doesn’t check out. I find that there’s a few good ways to visualize errors. With some of the other limited coding I’ve done on other tasks, the errors are so cryptic and abstract. They don’t feel grounded in anything for me to be intuitive about,” he says. “It’s just a personal opinion, of course.”
Leonardo.Ai & Respect for the Roots
Ethan is a developer for Leonardo.Ai, now in early access.
“I’m really excited about some of the stuff I’ve developed that we’re hoping to make live pretty soon,” he says.
He hasn’t ignored his roots though.
“I still have a soft spot for MathRockDiffusion. Just as it brought about a really cool community of fanatics for clip-guided diffusion. For those who aren’t familiar, it’s an older technique than Stable Diffusion. It takes much longer and tends to be a bit messy, but it has a special kind of charm to images that come out of it,” Ethan says.
“Something that I like to call “confidently incorrect,” where its messages might result in a face with misplaced features. It doesn’t look poorly photoshopped or glitched though, like all the correct shadowing and blending is done, as if it had meant to do that from the start.”
AI Upkeep & Keeping Up with AI
Ethan uses Twitter and Hacker News to stay up-to-date, mostly to see notable developments coming his way.
“It’s both riveting and exhausting. I’ll also read the paper on arXiv, a lot of times it’ll have some prior machine learning knowledge I haven’t heard, so then I have to track down some more papers,” he says.
Ethan shares, “there’s a lot of great resources I’ve been using to help me get the basics down so I can move onto more applied stuff. StatQuest with Josh Starmer and 3blue1brown have been incredible for learning the necessary math behind ML. There’s a few others that make good videos on more advanced stuff, like Yannic Kilcher.”
“I’m pretty well-versed now though, so I feel confident I can skim a paper to see the high-level basic idea of what they said.”
“It’s pretty impossible to keep up with all of the developments at the speed they’re happening,” he says. “Some advice might be to find a niche that suits you, try to be well-rounded enough to dabble in other areas or have decent intuition. I often just have to keep my head down and stay in my own lane instead of panicking about falling behind.”
“It’s also important to filter out what is useful and what isn’t. There are a lot of papers coming out with very little regulation—which I support because I think the more knowledge out there, the better. But also, not all of it is particularly helpful for your case, or something that can be built on,” Ethan adds.
“It’s important to recognize what has a future, and what will be obsolete by next week. How does this compare to alternative methods? How much does it cost? How can we reasonably create a simple UI for a user to make use of it?"
Thoughts on Society & Impact
“At this point, it’s a meme,” he says. “It’s on everyone’s mind, it’s pretty polarized. It doesn’t help that a lot of claims are being made on misunderstandings or misinformation.”
“There are a lot of people who feel reasonably threatened by AI as it may effect their livelihoods, but often I see it devolves into hatred for anything being labeled ‘AI’ as a whole. I don’t blame them, but AI is not just art machines. It will offer invaluable developments to the medical field, and I’ve gotten to witness that firsthand at the lab at my university,” he says.
“On the other hand, I see a lot of folks who have little regard for those who are concerned and have some sort of justification for every ethical issue,” he says.
“I can’t pretend I know what the right course of action is here. I am under the belief that the advancement of this technology is inevitable, and it will only become more pervasive in every part of daily life. Even if we put up regulations or one of those tech giants decides “we’re done with AI”—I think it will, at most, delay the revolution, but not stop it,” he says.
Ethan adds, “I think the best we can do is empathize with those who feel like their lives may be turned upside down, figure out compromises and ways to collaborate with AI, etc. That’s a big part of our mission at Leonardo.Ai.”
“In all honesty, [the impact on society] is something I’m unsure about,” he says.
Generative Art - Challenges and Discussion
For Ethan, the biggest challenge in his generative art experience is getting what he wants.
“I think when approaching generative art, you have to either let the model take the reins and ride the chaos or take as much intervention as possible to have it map your vision. There’s absolutely an art in both. I think there’s an art in creating with AI in unexpected ways, riffing off its weirdness. And then, there’s art in using AI to ideate, or amp up your workflow, absolutely,” he says.
"For commercial art, if there’s a way to quickly reach a desired logo or graphic, it’s definitely in the company’s best interest to use AI and take that action. For a lot of people though, art is a hobby or an enjoyable career. We like when AI takes over dirty work, like organizing files, or other things that are just flat-out nuisance,” Ethan says.
“It’s kind of surreal to think that one of the most human things about us, art, can almost be done totally autonomously by machine if you hook up language model to generate prompts. Aside from a few unique models and a simple few of those who take their AI creations to the next level,” Ethan notes. “I do often agree with the sentiment that it appears soulless.”
He adds, “Particularly from the outputs that are hyper finetuned on aesthetics. Everyone’s output looks the same to me, with an inhuman amount of detail, and showcasing almost like an average of the majority of the messier population’s ideals. There’s no room for error or imperfection. I’ve come to appreciation much messier, more abstract art styles, it’s especially refreshing with how flooded my feed is with Midjourney.”
Support Ethan’s journey on Instagram and check out Leonardo.Ai here.