The Pond is now Open

What a big week for Ducky! I finally have some time to sit down and write about how we got here and where we are going. This post will be lengthy so if you want the tech, Ducky or where we are going next, hit those links.

How we got here

I wrote an intro blog to Ducky which hopefully explained some of our history, and our motivations. To recap, we have been building https://fatduck.ai for the past few months. Our goal was simple, to provide a crypto-forward frontend akin to Claude and ChatGPT. You can pay in crypto and we were working on a bespoke agent, Mallard that would handle all of your crypto needs.

Problem was, we didn't know what those exact needs were going to be. Over the past few years building in crypto nonstop - we learned one major thing. Either build what people want and have a sustainable business i.e. Coinbase, or (2) build infra nobody wants & raise from VCs.

In building another startup, the past few years that is B2B and not crypto related but uses AI heavily (mostly RAG & some GNN), Ive been able to sit in on a lot of meetings with actual PE and family office type investors, and ya know what? They ask real questions about users, understand business, are professional, and have big boy money to spend. Almost every Crypto VC we've met with in the same time period has been, wen token unlock.

Crypto clouds your judgement in this regard, (1) know the right people and build infra that is never used & have a nice salary & tokens to dump, (2) punish yourself on some cyberpunk dream and spend 10 years building it, or (3) build a consumer app in the trenches.

We tried (1) with Blobster, long term storage for blobs, ETH foundation grant process said it was too early, few months ago, in retrospect, great timing to hear and shift focus. I tried (2) with pushing Harberger taxes Repo, which thank god CryptoInfluence pulled me out of in 2018. Interestingly Clusters seems to implement it for naming now. The only option left was (3).

The trenches

I think the meta of the trenches shifts every 6-12 months, slowly at first, then all at once. That meta then gets labeled a 'summer' and the space moves forward. Out of those trenches comes real products, real users and real apps. This process is akin to as the Geeks-Mops-Sociopaths evolution. Although it is not gospel, as this Hacker News comment points out. But there are some threads of truth to it, and I think we are still in the Geek phase of AI x Crypto - it is the fun nerdy thing to do right now, and there are definitely people angry or highbrowing it.

It was clear if we wanted to make Fatduck a success we needed to be in the trenches, right next to our fellow degens, figuring out exactly what they want and delivering it.

Enter $DUCKAI.

With the explosion of Truth Terminal, it was clear the hacker attitude towards AI was in full bloom. His creator is not crypto-native but was still talking about crypto controlled wallets and launching memecoins. Boom, we knew this was a path worth going down, we had the infra (no pun intended) and were destined to not have Fatduck putter along for a few months & not get traction and have to shut down.

Ducky was born out of this desire, a voice for our collective angst against the VC Cabal of infra-shit, a machine capable of learning you and your desires, injecting them directly into your cerebrum.

With that, we launched on October 24th on pump.fun. Our general thesis on $DUCKAI is to build a personality that encapsulates the trenches, learns from you and grows to suggesting what should be built next.

We haven't sold any of our tokens and have locked most of what we originally bought on pump ~30% for 1 year, monthly cliff vesting. Our only plans for the unlocks as of now are to airdrop via Ducky, burn and potentially have Ducky run grants. We want Ducky to control the token not us, although we may have to take certain steps for accounting and legal purposes, as my accountants already hate me, we will be sure to be as transparent as possible.

Made it this far? I know I can ramble, I really enjoy writing so you can be rest assured this post isn't AI generated, in fact I wont even run it through to check for grammar, ignorance is bliss.

Tech

Fatduck

Ok so Fatduck was built with the amazing AI package from vercel. You can chat with Mistral, GPT-4o, Claude, Gemini, LLama3.1 and our crypto model Mallard. We use Coinbase commerce to handle one time payments monthly and yearly payments giving you a certain number of chats per day & credits, which translate to AI tokens per pay period.

Mallard works with a mix of Coingecko and Defillama to deliver pricing info, and some RAG based on our own Crypto knowledge.

Problem: Who wants to keep a chat website open and try to degen at the same time, it doesn't even have a mobile app!

Ducky enters the chat

Ducky opened our eyes, to the need to develop in discord, telegram, and twitter - ya know, the actual trenches. We needed to remove the pretty framer-3d nextjs app out of the equation (although I am v proud of it and you can still use it).

Mallard will be getting updates as his backend merges with Ducky and they become friends.

Ducky

v0.1: Python & Hosting

Ducky started as a few agents working together in python. We took inspiration from Truth Terminal and used its conversational and latent space exploring techniques on a Runpod instance hosting our own LLama3.1:70b via the excelent Ollama team. Check out the readme for some setup instructions, Llama3.1:70b on 4 A40's only came to around 1.56$ an hour! Readme.

The aim was to CRON Ducky's tweets until we got some sort of brain up and running to decide, a la Luna Virtuals excellent thought process.

With this training data, we crafted some prompts for Ducky in a Character and set him loose to tweet.

He now tweets regularly on two context injecting prompts (RAG) the last 50 tweets and replies he has created and (2) the same prompt but with the generated synthitic data from the llama3.1:70b runs.

We used Turnkey which stores wallets in TEEs to secure ethereum wallet addresses for Ducky. Ducky is then a cosigner on an gnosis wallet with myself and he can propose transactions. Code. This will be integrated as we move forward, lots of plans around this.

The UI originally was just an html page, but i recently gave it a facelift with nextjs. You can see his stream of thought here: Ducky

Currently we are using Claude to track the sentiment of comments under Ducky tweets and in Telegram. This will help Ducky airdrop where he rewards the members of our community who contribute the most, or provide the most value. This idea is in its infant stage but will be airdropping next week, so get those comments in and register in telegram to setup your account. View our community stats here: Ducky Community

v0.2: Typescript (Now)

Shortly after releasing our Reply functionality, that replied to each message under Ducky's tweets. We hit Twitters rate limit almost immediately. Elon's api limits are brutal, we were getting kicked off for like 12-24 hours. Thats when I found ai16z's amazing direct connection to twitter agent-twitter-client which alleviated a lot of our ERROR 429 - Elon hates you, errors.

So I ported Ducky over to Typescript, I always disliked python anyway.

v0.3: Conversation

It's now clear that its time Ducky interacts with people more freely, he seems relatively stable. Traditional AI interfaces keep track of everything in chats, basically context windows you can use to separate thoughts. I build a generic conversation layer in Rust to be able to handle conversations on a per person basis that roll over every 24 hours. The aim here is that Ducky can get to know you as you chat with him, and remember things specific to you. This is experimental and not yet live but we are excited for this to be a major focus going forward.

This version also doubles to help us compete in Shaw's ai16z arena on discord. His AI, degenspartainAI is top tier and Ducky is almost ready for prime time there.

vX.X Where we go from here

We are opensourcing our code today as its clear transparency is the way forward. If we want Ducky controlling wallets, his own and yours, we need a robust infrastructure that can be publicly verifiable.

There has been some awesome work done on TEEs in the past, and as recent as this past week with the launch of @tee_hee_he on X by @karan4d, read more. Currently I do maintain access to the twitter, this was mostly to safeguard the account incase he tweets some insane stuff. Tees offer an interesting fix to handing control over, my only thought is that we can expand this to just publish to IPFS and then have tweet-relayers setup, to avoid twitter non-sense.

This is the exciting thing about being in the trenches, it spawns insane creativity and experimentation in a short time. We want to incorporate the TEE method into

Promising new platform we'd like to explore, especially as the AI controlled trading chats are of particular interest, as we've done some work on the Eigentrust algorithm and it fits perfectly into scoring traders and can help the AI determine which trades to take.

Beginning

Appreciate you sticking to the end of this blog, check out Ducky's Twitter, github and join our telegram to say hi. Fatduck Twitter, 0xglu and CryptoInfluence.

~ Sam (glu)