Episode interview with some of the Botto team (choobie, Mario Klingemann, Simon Hudson) find their Twitter or more info on their website
Time Stamps
[0:55 – 3:16] Mario is the artist behind and the project and he has been doing generative, algorithmic and AI art for about 20 years. He says that at some point when dealing with these mediums you start to wonder, can I pass this last step of choosing and learning to a machine?
With blockchain taking off in 2018, Mario felt that this could be the time for this project, but he couldn’t do it alone. So he built a team to help him with the other side of things.
[3:16 – 6:37] Kizu asks about decentralized art and other movements where artists were so engaged with the community. Mario laughingly states that he would never use the term decentralized art and that for him the most question for him is, if a machine can reach true autonomy. He feels that, if possible, decentralized systems and DAOs are the tools that would allow the machine to do .
Right now they are still working to get to a level where the machine could be so autonomous that it could make its own decisions, even fire/hire people. For now, Botto is still in a toddler phase, the AI needs help making decisions, and that’s why the community of stakeholders is so important.
[6:37 – 10:04] Simon started as a member of the community then decided to get more involved. Simon elaborates on the idea of decentralized governance and DAOs - the emerging subtleties of structures and principles that are needed to uphold them.
Simon was working in AI for the past five years and became fascinated with Mario’s work back in 2016. He was drawn to the idea of open, community oriented sourcing as a way to train AI as a reflection of taste. And within that, crypto as a great tool for getting the value to loop back to that same community.
[10:04 – 13:30] Sabretooth notes that Botto is using GPT3 neural network which is widely used in other NFT projects such as Pak’s LostPoets, he asks about the feedback loop and how the experiment evolves to be more autonomous over time.
Mario comments that Botto uses GPT3 for the description generation and that for the image Botto uses VQGAN + CLIP. Botto is a prompt miner, each image is generated from prompts. Botto finds good prompts which then produce interesting forms (aesthetic or mental).
Botto is constantly producing new fragments, about 3-4k every week, it then selects about 350 out of that week which get added to the voting pool. Botto pre-curates what the community sees and that process is influence from previous voting. This trains the model to mimic the voting behavior.
[13:30 – 17:34] The other way that the feedback comes in is that when particular images or topics receive more popular results, Botto will examine those areas for new ideas rather than generating new prompts. Botto is also built to...