A.I Innovation: Raising RadBots
||For the Press and Folx in the Tech Sector
Hello there! This article is part of a series about a radical experiment we have undertaken in modern collectivism and responsible crypto–the RadBots!
Today, we would explore the process that Dara’s team and the RadBot artists undertook to create the Bots, and what some of the challenges involved were. In case you aren’t familiar with our RadBots, go here to read the first article in this series.
Dara’s RadBots were created through a precarious process of trial-and-error, involving several moving parts; both human and mechanical. Dev Aggarwal, our very own Technical Founder, sat down with me to shed some light on the finer points. He explains that Dara’s team chose the GPT-3 AI model, as it is one of the largest of its kind that is commercially available. GPT-3 is a language model that uses deep learning to generate text in natural language. It is trained on a very large corpus of text from the internet, and utilizes a 175 billion parameter neural network to generate text that is nearly indistinguishable from writing by real people. It is developed by OpenAI.
One of the challenges faced by Dara’s team and the RadBots artists was deciding how much textual context to develop for GPT-3 before asking it to simulate a human interaction. This was necessary in order to ensure a human-like persona for each individual Bot. Another issue on this front was that the text-based prompts sent to GPT-3 had a character limit, so the RadBot artists that wrote these had to be extremely clever about how they framed them.
Dara’s team wanted to be dynamic with regards to how folx could interface with the Bots. This means allowing for both, text and video messages being sent inwards. In order to accomplish this, we set up a flow that would allow for videos being sent to be turned into text using Google’s speech-to-text API. Written messages would be treated as text as well. Once the text was generated, it would be forwarded to GPT-3, which would then generate a text response, and send it back to us. Here, the text would be converted to speech using Google’s text to speech API, and the result would be lip-synced to the Bot’s face being animated by a generative model (GAN).
Handling this entire process with the RadBot artists asynchronously was also a fairly challenging task. This was not made easier by the fact that multiple folx were participating in the project at the same time. This made performance optimization a close focus for Dara’s team.
However, the results of the project paid off: our Bots are an NFT collection of cutting-edge A.I that is highly conversational, and can also stand on its own merits as a group of opinionated and starkly original characters.
Dara’s RadBots were created with the generous support of Goethe-Institut / Max Mueller Bhavan Mumbai and BeFantastic, who are among our prestigious Partner network. Tezos India is also helping us by keeping the RadBots on the greener side of crypto.
Our Bots are being dropped here on Rarible, and you can find out more about them here. Don’t hesitate to reach out to us at support@dara.network if you have any questions. You can also directly message Dara’s team here, on the Dara app.
