An Introduction to Bittensor [Educational]

Bittensor Protocol

A protocol that seeks to promote excellence in neural network development by distributing rewards through a peer network, designed to judge the highest quality Artificial Intelligence performance.

Introduction
Bittensor was founded by Ala Shaabana & Jacob Steeves, two individuals with extensive pedigrees in machine mind learning. Beginning in 2019 with the aim to create a platform that can drive growth in the development of neural networks, Shaabana and Steeves began the protocol with the intention to allow the industry to benefit from the principles that have allowed cryptocurrencies to thrive.

The core philosophy underpinning Bittensor is the idea of creating a corpus or public library of machine intelligence that can continually grow and allow the sharing of knowledge throughout the peer network.

What is Bittensor?
To keep it simple, it’s protocol that allows AI models to speak to each other, evaluate their own effectiveness and hopefully earn a reward by demonstrating the quality of their representations. The protocol is a novel approach to a layer one blockchain to allow an interconnected AI network of peers access to a universal dataset to demonstrate their models against and receive feedback on the quality of the output.
For the real science and engineering behind the protocol, check out the brilliant whitepaper the foundation has produced: https://drive.google.com/file/d/1VnsobL6lIAAqcA1_Tbm8AYIQscfJV4KU/view

Purpose – What is it trying to achieve?
In theory every endpoint or peer in the Bittensor network can be an OpenAI, a Google or an alternative AI company of your choice. This protocol ties together all of these entities by offering an accessible pool of validators that gives immediate feedback on a model, alongside incentivizing the successful development of the model.
Bittensor is the finely knitted tapestry that connects a colourful patchwork of global experts into a decentralised network that, in effect, acts as an API for any AI developer working on an intelligence model. Where the protocol can help power the growth of the AI industry is to create what is effectively a commodity market for valuable contributors to demonstrate the strength of their wares to the world, thereby creating an ecosystem that ‘monetizes’ quality machine learning performance in itself.

So how does it do this?
First the blockchain. Let’s take a look at the ‘blocks’ produced on the bitcoin blockchain. What we are seeing is a race between super computers to crack an algorithmic mathematical problem which will then be recorded onto the blockchain, granting the successful computer a BTC reward for completing the problem.
All that power, all of those colossal mining centres - in the middle of Icelandic rivers or buried into the sides of Norwegian fjords - fighting tooth and nail to win a race that is about as useful to humanity as a kindergarten class being asked to write down a random 256-digit number, presenting it to a bank and demanding 175k USD for their work (current market rate as of 04/23).
This is certainly not to disparage bitcoin, as here we respect the godfather of cryptocurrencies and appreciate the enormous sea change the protocol has brought to the economic landscape.
What Bittensor does is take the valuable elements of the blockchain - a decentralized digital trust ledger that can fuel the development of a new type of fundamental market - and uses this ledger to record ranking information as opposed to the transaction history of random calculations.

What makes Bittensor different?
The blocks in Bittensor are recording a matrix containing the ranking of each contributor’s artificial intelligence model, with the highest-ranking model being granted the highest level of reward. The native token of Bittensor is TAO, with this being the reward that contributors and validators receive.
Bittensor in its essence acts as a sophisticated benchmarking tool that measures the performance of a machine intelligence model.

Functionality – How does it work?
The concept of a ranking system then begs the question, what are the contributors producing and what is the criteria they are being judged upon?
Consider the example of what is currently the most well-known AI system, Chat GPT created by the developers at OpenAI. Chat GPT’s output can be viewed as a ‘representation’ of an enquiry input such as asking it to give you a chocolate cake recipe. The nature of the outputs will of course vary, however AI models using the protocol will be required to create their own representations based on the universal dataset, and this representation is what is validated.
The technique used to make the ‘judgement’ or validate the representations is Fisher’s information. This technique is used to create an approximation of how different or worse off the validator would be if a particular node were removed. Imagine the human brain, remove a single neuron, and check whether the brain is now functioning better or worse.
Through a method of take all embeddings and combining them, this can then be filtered through to identify the most important ‘neurons’.
This calculates the exact value of that particular peer or AI model and allows the ranking system to be generated.

How do the models learn from each other?
Learning from each other is the closest representation of what’s happening. Each model will ping the network to find other models to talk to, not too dissimilar to a DNS lookup,and this is where Bittensor behaves as an API. Your model will ask ‘can you run this through and give me back your results?’. This is of course the simplified version, the white paper gives much more information.
If for example you took some information from a model, ran the results and loss function reduced, then you are less likely to talk to that model again.
What are we really doing? Putting through a dataset, asking for a peer review and getting back immediate results with incentivised feedback.

How does it generate value?
Let’s get down to brass tacks. How does Bittensor create value that can be deemed worthy of monetary reward?
Fundamentally, Bittensor chooses to reward outstanding performance in machine intelligence (contributors) and simultaneously acknowledges the importance of validators (consumers) by also rewarding them with TAO.
Whichever models are producing ‘signal’ in the midst of all the noise are the ones that get rewarded the most. Thus Bittensor forms the foundation to create a market place for intelligence.
Once validators determine informational significance, and once the intersection is discovered of who generates the most valuable information for the most people most of the time, the intersector is computed on blockchain and that’s how it is determined who gets inflation, with new tokens then minted and given to them.
This is addressing the intrinsic value generation element of the protocol, but what of the myriad factors that can indirectly create value?
Indirect value can be interpreted through its operating philosophy.
Key features such as:

Decentralization: The most important principle of bitcoin and closest to the heart of Satoshi Nakamoto, no central organisation has control over Bittensor so we can be confident that results are not being manipulated in favour of any one particular model.

Open-source: Inherently more transparent, infinitely more accountable. We all love open-source software, the collaborative efforts that can bring communities together to help improve ideas and programs, ability to quickly eliminate bugs, what more do you need?

Scalability: The ability to accommodate higher volumes of requests, to give results quickly and efficiently back to contributors is something of enormous value, the time saved alone can be more than enough a fiscal saving. The flexibility and lower costs offered by eliminating the need for costly hardware cheapens the price on innovation significantly, and gives a better platform to anyone looking to contribute to machine learning.

Censorship Resistance: No central controlling entity? No problem! The Tensor swarm (ignore the horror movie images) form all the nodes and honest consensus that you need, hence no government or entity can swoop in should someone accidentally create consciousness on Bittensor (only half joking!).

Other points to note are that by using a universal freely accessible dataset, one of the main inefficiencies in upgrading machine learning models being re-learning time is counteracted. Rather than having to instruct your v2 model to learn everything your v1 model has learned, and then begin to grow on its own, by seeing new information each time you are developing a resilient generalist model and streamlining the performance of your AI.

Challenges - Collusion
The dreaded cabal. Similar in principle to the 51% attack that concerns Vitalik & Co over at Ethereum, the Opentensor foundation has to be wary of collusion between a large enough group of peers that can decide to inaccurately score their neighbours in order to unfairly weight the inflation in their preferred direction.
All Proof of Stake blockchains will have similar worries, however the monetary incentive of collusion can be counteracted by the system Bittensor has implemented by requiring the cabal to constantly increase their spend on stake to maintain their position.
There is little point in robbing the bank if the vehicles, weaponry and ski masks cost as much as the amount stolen in the first place.  

Conclusion
To decentralise the development of machine learning platforms and to build a marketplace to commoditize the brilliant work going on globally in the AI field, the Bittensor founders wanted to exploit the power of the hive mind demonstrated by cryptocurrencies such as bitcoin in order to benefit the AI field in a similar vein.
The concept of opening the accessibility of working on machine learning models is also firmly embedded in the heart of Bittensor, gone is the idea that you can only contribute to this field if you are working for institution like google, Facebook or OpenAI. The reward of the protocol is driven by engineering and ingenuity rather than the size of the institution working on it.
In purest terms this protocol is building a machine to machine market. Neural networks make the decisions, not the individuals. Removing the bias, retaining the innovation. I believe this protocol has the potential to dramatically change the machine learning space, challenges notwithstanding, and I am rooting for the project as it bringing innovation to the development of artificial intelligence is something that has possibilities to benefit us all.