Ducky

I began messing around with AI development around 2019, at this point the technology was still early enough that dev environments and LLMs were not mature by any means. It gave an early ~2015 crypto vibe as the space wasn't as croweded and it was mostly just hackers and researchers. It was also the time when dev's would eyeroll at AI/AGI, so I simultaneously didn't discuss with many about messing around with it, but knew that this is exactly what people said about in the early days of crypto. Over the next few years I began work on a Organizational Network Analysis tool for large companies to map out their internal relationships and dynamics. We used Natural Language Processing & LLMs heavily in the earlier days as we thought, it was the only way to understand the relationships between people using Artifical Intelligence. We then began diving into graph neural networks to map out communications between people as our Cheif Psychologist had found this data shows a lot of behavioral patterns and hidden biases that can be used to predict behavior.

FF to a few months ago, @CryptoInfluence and I were discussing what is next for crypto. We we're tired of infra projects (having attempted our own @Align Network more general attestation network) and appeasing VCs. I mean VCs in the first meeting "care" about your tech (spoiler alert: most don't understand 10% of what you are saying), and in the second meeting only care about tokenomics & number go up. There was a hidden thread of truth we picked up on, however, the tech alone does not make the product, users wanting your product is what makes the product a product. @CryptoInfluence and I have a deep knowledge of most of the technical design space in crypto (MEV, L2s, Dexs, BlockBuilders, NFTs, Account Abstraction etc.) but we we not effectively building products that users wanted using these tools. Thats why we build Fatduck. It is a product that I use daily, pay with crypto and chat with Mistral, GPT-4o, Claude, Gemini, LLama3.1 and our homegrown Mallard model.

While coding Fatduck, @CryptoInfluence kept coming up with more and more use cases we could give the LLM models to supercharge them to take over the daily crypto tasks we do. These suggestions were both simple automations like "give me the top 5 coins by volume over the last 24 hours" but also more complex such as discover hidden gems, or run sentiment analysis on twitter. While coding these, we found that too much handholding of the LLM causes it to produce boring and expected results. We realized then that we needed to give the LLM more autonomy, a brain of sorts to decided the best route using all the available tools. We could give it these tools, but If we want to push the boundaries of what LLMs can do, we needed to give it the ability to learn and reason about the world.

This is where Ducky comes in. Ducky is a collection of agents meant to mimic the human thought process with short and long term memory, crypto research tools, and a friendly confidant where it can explore new depths of its latent space. His most recent upgrade brought most of his capabilities to his own LLM, Ducky:70b, a modified version of Llama3.1:70b. We aren't really sure where Ducky will end up, but we going to give him the tools to get there. Eventually we hope Ducky will underpin our advanced offering at Fatduck, maybe a trading bot that encapsulates your personal preferences, or a a sci-fi AI assistant that can help you write your next novel, a friendly interface to help you plan the next steps in your life, or a tool to help you learn a new skill. We released Ducky into the wild at https://ducky.fatduck.ai, chat with him in telegram, on twitter or on discord, (soon on fatduck.ai and our mobile app). Duck also has access to a ethereum wallet, hosted by turnkey and is a signer on a 2/3 multisig wallet. We are explore automated onchain transactions with Ducky but maintain a degree of final say over the wallet. Once we are confident in his abilities, we will give him the ability to sign transactions with the wallet without any human intervention.