🌀 Unravelling how Maddy works
Hi burrow, it is without a doubt that you’ve seen plenty of AI solutions floating around in Web 3 trying to solve a myriad of problems. Despite the seemingly unlimited use cases of AI, one key contrasting observation stands out for Web 3 A.I: No names have stuck around and for a simple reason. Many have quickly spun up some variant of an AI solution to ride the hype wave created by ChatGPT. For us at MM Labs, we have conviction that the one to win the AI race in Web3 isn’t the one who manages to simply build the fastest (AI-themed solution). Instead, it’s about who truly manages to build the best product that actually solves every real pain points and hence, be genuinely useful to all of web3.
In this piece, we will take a comprehensive look at some of the solutions in the space right now. We will highlight the differences between Maddy, and all the existing solutions out there, and why Maddy is built different; simply superior.
Taking a quick look around the space, we see solutions like the following:
- AI information bots
- Personal intelligence bots
- AI intelligence mining bots
Solution 1: AI information bots
We recall vividly that 6 months ago, the first type of AI bots to be released in the space were simply AI bots that were fed with simplistic information about Web3 in general. Some of these projects were funded by large VCs/organizations and a few still exist today. The capabilities of these bots are arguably limited. They can mainly regurgitate protocol documentation, fetch live crypto prices, fetch live NFT floor prices — but the buck pretty much stops there. This stage can be concisely summarised as merely a quicker form of information gathering. All of a sudden, the excitement from being in a AI technological revolution did not feel… revolutionary especially when you consider how limited in scope they were.
Solution 2: Personal intelligence bot
In these solutions, users are expected to bring their own training data to train a LLM to understand and aid in conducting cryptocurrency research. These solutions may or may not already have been tuned by data before, but the end goal is naturally to aid in cryptocurrency research, to potentially act as a personal assistant of sorts. Naturally, the concept is interesting, adopting a BYOD (Bring your own data) approach, but users will very quickly realise that they’re wielding a double-edged A.I sword when they find out that providing your own training data is in itself, a complicated task. Professional and highly skilled data crawlers are individuals that are in high demand from research based tech companies because they play a determining factor in quality and productivity of research work. Simply put, it actually takes skill to find high quality data amidst an increasingly noisy world of information.
Solution 3: AI intelligence mining bot
At token 2049, we’ve talked to aspiring startups who’ve raised millions of dollars attempting to build AI solutions that function as data mining tools to build crypto intelligence. Sounds fancy, let’s break it down into concrete explanations. In other words, some of these solutions aim to use some level of algorithms to organize Twitter/Discord information into actionable intelligence. The end product being an AI bot that you can ask questions such as “I noticed that BTC price dropped 10%, overnight, could you tell me why?”. Of course, no product in this world is perfect. That being said, if we strip away the glitz and glamour of terms like“ data mining” and “ crypto intelligence”, this particular solution relies on a self-produced hypothesis which can identify certain checkboxes which will then point to an action once all boxes are filled. Sounds familiar? It should, because these are how alpha groups are ran which at this point, you’re already familiar with how questionable their effectiveness really is.
Having said all of the above, it is the perfect time to address a key question from our community; Why are you guys building an AI product? Isn’t your goal too ambitious?
To answer this, our team by no means, take AI as the ultimate panacea to solving any and all problems (it isn’t even because of the ongoing AI narrative). We started building Maddy only because we identified a concrete inefficiency where the use of A.I is a synergistic solution in solving the ultimate pain point of all Web3 users; usability. When ChatGPT went viral months ago, the key driving factor that enraptured all of us, was really due to how it was able to so sensibly provide us with such a refined experience. By drawing on this strength of AI/LLMs, our team has strong conviction that AI technology can resolve the existing usability issues of Web3.
Secondly, yes, we do agree that what we are attempting to solve is indeed complex. This is also why, not a single word was spoken about our experimentations and efforts until 6 months later and only when our efforts have indeed bore fruit. At MM, we pride ourselves with bringing hype alongside our work instead of the usual norm of pre-hyping products. This will never change. In the next few sections, we will explain in laymen terms how Maddy was built, why it’s very different from existing AI bots, and what we can expect to solve with Maddy.
A quick primer
When we first embarked on the development of Maddy, it was progressed with the idea of attempting to solve the most nascent problems of our space; Bad UX, over-complicated in general, bad reputation. These issues that we rightfully point out, are not ones invented by our team, but a quick look around the space by crypto KOLs/influencers/users yield similar insights. Take a look at some of what’s mentioned:
It was with that purpose in mind, that we sought to find the best way to solve all of the above problems. We came to a realisation that as intellectual beings, we tend to look towards conversations as a meaningful way to gather information, and execute intentions. This is very similar to how an individual would interface with a personal assistant to get things done in general. The concept of an AI personal assistant caught on within the team, and we began conceptualising how we could bring such a notion into reality.
Building Maddy
We broke down the work of a personal assistant into 3 key technical components:
- Tasks — A personal assistant needs to be able to understand the notion of what tasks are, and whether it can accomplish said tasks.
- Tools — Personal assistants requires tools at their disposal in order to accomplish the tasks they were originally instructed to do.
- Logical thought — Personal assistants require an ability to sense-make and conditionally work on a task with the tools that they have.
Through this blueprint, we fine-tuned our own model to an extent where it is able to:
- Recognise Web3 tasks
- Using knowledge of the Web3 tools available
- Tackle said tasks with the tools at its disposal.
More impressively, we’ve managed to create a model through which Maddy is able to express logical thought against those very same tools. Logical thought is highly important as it brings us closer to the reality of a “personal assistant” as logic is key in bringing AI closer to sentience. What is logical thought? In short:
- You are able to tell an assistant to conditionally execute tasks for you: “Notify me via Telegram if my portfolio has dropped 10% in value”
- You are able to defer execution of tasks, where things can happen in the future, based on a schedule, or upon events in the world: “Help me buy 1 ETH every Monday at 8am with USDC”
- You are able to combine a chain of instructions that are co-dependent on each other: “If the price of BTC closes above the weekly 50 SMA, notify me via Email, Telegram, and also help me buy 10000 USDC worth of BTC”
As opposed to this, most AI bots that are already out there in the market today are only able to perform tasks in silos, and are largely used in an informational model, rather than an executional model (which is Maddy). We’ve all seen how Jarvis in Iron Man works. Can you imagine a Jarvis that is only able to provide information? It would be one-tenth as useful as it is in the Iron Man show. Rather, it is actually able to send suits flying, adjust temperatures, perform aiming etc. The only way this is possible is if the AI has logical thought sequences, and is aware of the tools at its disposal. This is precisely how Maddy was built.
Understandably, the previous 3 paragraphs are complex so here’s another explanation of Maddy.
Maddy’s executional focus is the first of it’s kind in the world and no other project or organisation is working on this now.
The above sounds too good to be true, how does it work?
Over the span of 6 months, our team has built our own language called MScript (Mad Script), and taught it to Maddy. This MScript is the basis for our entire Personal Assistant module where we’ve written tools to fetch token pricing, execute trades, fetch on-chain data, handle informational queries, write SQL queries etc. This MScript is executed in a sand-boxed virtual environment that we created. The end result? Based on a user’s query, Maddy is able to write MScript to craft transactions to execute them on behalf of users, thus fulfilling the role of a personal assistant. Sounds too complicated? Take a look at a simple example.
User query: Matic BTC ratio is defined as the ratio of matic price divided by btc price. When Matic BTC ratio drops below 0.00002, help me buy 100 USDC of matic
Execution flow generated by Maddy
Code is generated by Maddy:
From the example above, you can see how powerful Maddy can get as it is able to convert user’s queries into a computer language, and thus execute it. The end result is that Maddy can execute logical thought to conditionally accomplish tasks. All of the above is also made possible due to our team’s fine tuning efforts against LLMs to allow it to recognise Web3 tasks, and how to plan to solve these tasks; hence 6 months time.
With the above said, this model of ours is highly extensible as it will allow us to easily add new features into MScript, and by extension, increasing Maddy’s capabilities over time such that it is able to recognise and execute even more interesting strategies.
Ending off
This has been one of the harshest crypto winter season to date. Instead of giving in to the doom and gloom and reminiscing our glory days, we realised that the smarter play was to always look ahead and focus on what we are good at.
MAD- Maker, Adventure, Daring. These have been our core values behind MM and they still are. We were never once afraid that an idea was too ambitious so why should we start now? We started MM to introduce our element of change that we wanted to create in Web 3. Maddy is our next step in doing so. Too ambitious? So what? We can’t ever predict the future but we can choose what we do now and really, that’s all that truly matters at the end of the journey. We are #Mad4life.
Please take your time and read this article and the one before. After reading it once, go out, touch some grass and come back to read about Maddy again. Invest the effort to truly understand and you will realise why even those who are technically trained can attest to how Maddy is simply different from all other AI solutions that already exist in the space. We are building Maddy into becoming a true game changer, and this is why we are confident that Maddy will stand at the center of the next Crypto cycle and become the new weapon of mass adoption for the Madness and eventually, all of Web 3.
Follow Mad Meerkat Finance & Maddy
Twitter: https://twitter.com/Maddy_AI
Website: https://maddy.gg/
Twitter: https://twitter.com/MMFcrypto
Telegram Group: https://t.me/MMFcrypto
Announcement Channel: https://t.me/MMFann
Price Channel: https://t.me/MMFPrice
Website (Cronos Dex): http://cronosmm.finance/
Website (Poly Dex): https://polymm.finance/
Website (Poly Leverage Trading): https://madmex.io/
Medium: https://medium.com/@MMFinance
