Developer designs AI that creates Infinite Bored Ape NFTs

by James · October 10, 2022

Although the entire concept of the NFT market seems to elude us, the exponential growth of this market – which peaked at $11.6 billion in 2021 – has brought widespread attention to the subject. This in turn has opened the floodgates for new innovation and technology to drive the creation of these virtual works of art.

One form of innovation includes AI-powered and machine learning technologies that use data attributes from existing NFTs listed on publicly available websites to randomly generate thousands of new, unique NFTs per second that can pass off as originals. Experts predict that we will see the emergence of these AI-powered NFTs this year. This development may be a positive step towards inclusion, but may also have some negative effects on the appreciation of the current NFT population.

An example of one of these creative innovators is Yannic Kilcher, a machine learning researcher and engineer. He developed a method to generate infinite amounts of Bored Ape NFTs by training his own AI using readily available public web data. Kilcher has a YouTube channel with 133,000 subscribers, where he makes “videos about machine learning research articles, programming and AI community issues, and the broader impact of AI in society.” Through his Bored Ape project, Kilcher wanted to show that anyone could create their own unique Bored Ape NFT, for free.

By way of background, Bored Apes is a popular collection of 10,000 unique NFTs created by the Bored Ape Yacht Club. Each Bored Ape is “programmatically generated” from a selection of 170 traits, some rarer than others, and range in price from 60 Ethereum to 100 trillion Ethereum. The latter amount, at the moment, would put just under a $197-quadrillion dent in your pocket. How and why the prices of NFTs are set has little to do with AI and data and is a topic for another discussion. However, the recent incremental increase in the price of Bored Ape NFTs has been fueled by celebrities, such as Jimmy Fallon, Justin Bieber, Paris Hilton and others, who have recently purchased Bored Apes from the collection.

Kilcher described the process of creating the AI model in a joint video he produced on his YouTube channel with Bright Data. Among other methods, he used public web data collection and open source code and material found on GitHub to drive and design AI for public use. To feed the Bored Ape AI model, Kilcher collected online data from opensea.io, which contains the official collection of Bored Ape NFTs for sale on the site. He used online data collection tools to automate the collection of images, attributes, prices, when the NFTs were last downloaded or sold, and other information from all 10,000 of the listed Bored Ape NFTs available on the site.

Kilcher used this data and some open source code from Nvidia to train his model to create Bored Ape tokens. He did so by using a generative adversarial network (GAN) machine learning framework, which is effectively a complex set of checks and balances to create unique images. According to Kilcher, data is critical to building GAN frameworks. They automatically detect and learn from the patterns in the input data collected from the original NFTs – enough to eventually have the ability to generate new NFTs that can plausibly appear as part of the original collection.

GAN frameworks basically train two models simultaneously, one is the generator and the other is the discriminator. The role of the generator is to attempt to fool the discriminator by introducing new data into the images outside of the originally collected data set. Meanwhile, the discriminator’s job is to determine the authenticity of the image by estimating the probability that the image came from “training data” rather than original data from the generative model. Through this process, the generator and discriminator inevitably improve and end up balancing each other out over time. Eventually, a properly trained model would have the ability to generate new images that could pass as legitimate, which Kilcher realized could be used to create an infinite amount of NFTs.

Using the Nvidia StyleGAN2–ADA code based on the GAN framework, Kilcher trained his model step by step using the web data from original NFTs to feed the AI. Every few hundred steps, the AI produced a new series of images that allowed him to track his progress. The images were blurry at first, but eventually the AI began to learn and create apes as it correlated the data to achieve its designed purpose. And as the model became more advanced, it generated several different monkeys.

While the exclusivity and artificial scarcity of NFTs is usually what drives up their cost, Kilcher decided to publish his AI-powered application for public use – offering endless free Bored Apes to everyone. While this is bad news for Bored Ape collectors, it is good news for the development and promotion of new innovation.

Kilcher’s Bored Ape generator can be found at Hugging Face.

About the author

Or Lenchner, CEO, Bright Data. Ever since his appointment as CEO of Bright Data (formerly Luminati Networks), Or Lenchner has continued to expand the company’s market base as an online data collection platform dedicated to delivering complete web transparency. Over the past three years, under Lenchner’s leadership, the company has advanced its product offerings to include first-of-its-kind automated solutions that enable customers to collect and receive data within minutes. Among Bright Data’s thousands of customers are Fortune 500 companies, major e-commerce firms and websites, prominent financial firms, leading security operators, travel sites, academic and government organizations.