Bittensor and Decentralized AI
We’re starting to see significant interest in blockchain-based decentralized AI systems. In this post, we analyze the dynamics of a widely popular project, Bittensor, and provide guidelines on when/whether it makes financial sense to join the ecosystem.
Bittensor in a Brief Nutshell
Imagine a world where anyone with a GPU or CPU can compete in a marketplace to train the best AI models for a diverse set of socially-beneficial tasks, such as predicting the stock market, detecting cancer from radiology scans, or training more human-aligned LLMs. Such a world would be possible if the world’s compute owners could be securely and transparently rewarded based on the quality of the AI models they produce. This is precisely the world Bittensor is enabling through the following key agents, all tied together in a blockchain-based incentive system:
Subnet Owners and Creators:
Subnets are groups of machines working to achieve a common AI task, such as labeling data, predicting the stock market etc. In theory, subnets can be created by anyone in the community, and are vetted based on their popularity and utility of the task they propose. To create and maintain a subnet, the owners are rewarded with a native cryptocurrency named $TAO.
Miners:
Miners are compute nodes that compete to train the best AI models. Miners are rewarded based on the quality of their AI predictions. Ideally, this quality should be assessed in a quick, transparent manner by another set of machines, such as the error of stock market predictions. Miners must pay a small registration fee, in $TAO, to compete in the ecosystem. Poor performing miners are periodically booted from the ecosystem and refunded their $TAO deposit.
Validators:
A distributed set of validators check the accuracy of AI outputs from miners. For example, validators can obtain the current value of a stock and compare it to miners’ predictions, assign an error (loss function), and distribute rewards to miners based on their respective errors. Since validators play a crucial role in assessing rewards and maintaining the integrity of the system, they must stake a significant amount of $TAO as a deposit. This $TAO can be slashed if they provide erroneous validation results. In return, validators are rewarded $TAO for maintaining the integrity of the decentralized AI ecosystem.
Should I Join the Ecosystem?
Suppose I have some combination of significant AI expertise, spare compute resources, and/or happen to already have a large amount of $TAO. Where should I dedicate these resources?
The answer is a bit surprising once you look at the data. For example, take a snapshot of time in taostats.io, a popular analytics tool for Bittensor. We see that a few “whale” validators who have staked a huge amount of $TAO get a large number of rewards as emissions, around $40K per day at the time of writing. Indeed, the wallets making the most money are typically all validators. In comparison, the best performing miner is ranked 66 amongst all top-earning wallets and makes only a paltry $940 per day. Of course, this is for a certain day and specific subnet 8, the popular time series forecasting subnet, but similar trends persist for other subnets as well. This leads me to the following, opinionated recommendations for joining the ecosystem:
Option 1: Join as a subnet creator or owner (High Reward, High Risk)
You must have a creative idea for an AI marketplace as well as significant human resources, time, and coding ability to create and test a new subnet. However, the rewards are high if your subnet becomes popular, but there is significant risk of wasted effort if the subnet idea is rejected by the masses.
Option 2: Join as a validator (Fork over the major $TAO)
If you have sizable GPU compute, disk space, etc. it makes sense to be a validator. However, you must have sizable $TAO to stake already, since there are limited validator spots. Moreover, they are given to those we stake the most $TAO.
Option 3: Join as a miner (Only need AI chops, minimal rewards today)
If you have fairly good AI expertise and modest compute (e.g., a single “standard” GPU), you can easily become a miner. In fact, you can easily adapt standard models from HuggingFace for the subnets’ task. However, at the time of writing, the rewards seem minimal even for the best performing miners. Of course, if the price of $TAO skyrockets due to a combination of speculative demand from exchanges as well as organic demand for real AI services, today’s modest $TAO emissions could be worthwhile in the long run.
A Gentle Critique of Bittensor
On a whole, the idea and execution of Bittensor are brilliant. The disparity between miner and validator rewards seems troubling, since the best AI minds are not incentivized to create continually improving models. Finally, the token price seems to be driven mostly by speculation and hype.
But, what if the predictions and outputs from the best AI models were actually sold to external customers?
This is exactly what many subnet owners are working towards, as we describe next.
Bittensor Needs Organic Demand for AI Services
Imagine you are a regular Web2 customer and just want to buy the “best” predictions for the stock market or produce images from the “best” stable diffusion model for your niche interest of penguins. Then, you can pay for AI predictions in fiat currency or its equivalent in $TAO (if you fancy Web3) to access the best predictions from miners. If this takes off, we will have fiat currency injected into the economy and the price of $TAO can rise since it will be demanded for (a) speculation and trading, but (b) for real-life AI services by the masses.
Financial Modeling for Decentralized Infrastructure Systems
As a miner, validator, or $TAO investor, it’s crucial to have financial projections for the organic demand for AI services on Bittensor, the cost of GPU compute, and the speculative demand for $TAO. For example, we might ask:
“As a miner, my cost for an Nvidia GPU is $5/hr on a popular cloud platform. If my AI model produces the top 5% of predictions, and we have $5000/day of demand for stable diffusion inferences, how much $TAO would I expect to earn daily?”
“Given my daily cost of running a GPU, the cost of training, and my expected model quality, how much demand should I service to become profitable in 2 months?”
If you’re interested in data-driven and “what-if” answers to such questions, check out our tokenomics paper (open-source code coming soon)! It can model the financial economics of any decentralized physical infrastructure system, such as decentralized AI, GPU compute, 5G wireless, and much much more!
