Bittensor: A Decentralised Ecosystem for AI Innovation

Bittensor is an open-source decentralised protocol and blockchain-based ecosystem for AI that aims to democratise the development of AI by enabling anyone to contribute computing power, data, and expertise towards the development of new AI models, applications, and services, and earn rewards in the form of TAO tokens. The core of Bittensor's design is its innovative approach to incentivising the development of valuable AI models and applications through a system of interconnected subnets. This represents a fairly novel approach to AI development, challenging the current centralised model of AI innovation. To assess if Bittensor can really deliver on its vision, we need to understand its core value propositions, potential avenues for investment, and the open questions that could impact its long-term success.

The Core Components

At the heart of the Bittensor ecosystem are three key components:

• Subnets: These are individual, incentive-based competition mechanisms, each focusing on a specific AI-related problem or domain. They serve as the core of the Bittensor ecosystem and are designed to drive innovation and specialisation.

• Subtensor Blockchain: This is the underlying blockchain that supports the subnets, providing a decentralised, permissionless, and collusion-resistant foundation for the entire ecosystem. It ensures the secure and transparent operation of the network and the distribution of rewards.

• Bittensor API: This API connects the subnets and the blockchain, enabling seamless communication and interaction between the different components of the ecosystem.

Subnets: Decentralised AI Competitions

Bittensor's subnets are a critical aspect of its value proposition, as they provide a way to approach AI development with focus and specialisation. Each subnet operates as a Kaggle-like competition of unfixed duration, focused on a specific AI-related problem or domain, continuously evaluating and rewarding participants based on the solutions they submit. This is a crucial aspect of its value proposition, as it facilitates a decentralised and competitive environment for AI innovation and improvement.

Essentially, each subnet is designed around a particular problem, challenge or area within AI, acting as a focused competition. These can range from general tasks like text prompting and machine translation to more specialised areas such as pretraining large language models (SN9), verifying AI model outputs using zero-knowledge proofs (SN2), compressing language tokens to optimise AI inference speed (SN47), generating 3D assets for gaming (SN17), decentralised AI code generation (SN45), evaluating home prices in the housing market (SN48), and identifying optimal trading strategies using AI (SN53). At the time of writing, there are 64 subnets active on the network.

Within each subnet, "miners" develop and train AI models to solve the specific problem defined by that subnet. These miners are essentially competing against each other to produce the best performing models. "Validators" in each subnet then evaluate the miners' work, assessing the performance of the models and assigning scores based on the quality of the model's outputs. These scores are used to determine the distribution of rewards, which are paid out in the form of TAO tokens. The system incentivises miners to continuously improve their models, fostering a competitive environment that accelerates innovation. This structure is designed to ensure that resources are efficiently allocated to the most promising AI solutions. In essence, each subnet is an AI model competition subject to an evaluation system created by the subnet owner and administered by validators.

Subnets are also highly flexible, allowing owners to define their own incentive mechanisms and create specialised markets tailored to specific needs. As subnets evolve, so too can the parameters and constraints of each "competition," allowing the system to adapt to new challenges and technologies.

Yuma Consensus: The Heart of Validation

The scores assigned by validators within each subnet are aggregated using Bittensor's unique Yuma Consensus mechanism. This mechanism is used to reach an agreement on which models are most valuable and how rewards should be distributed. The interesting innovation here is in its design to handle the complexities of evaluating subjective or probabilistic truths, like the value and quality of an AI model's output, and to ensure agreement between the validators of individual subnets, incentivising miners to act in accordance with the consensus mechanisms defined by the subnet developers. It transforms the various incentive mechanisms of each subnet into an "incentive landscape," directing the network towards a consensus on the value of work done by miners, ensuring that the reward distribution is fair and transparent, based on the quality of work done by the miners.

Incentivised Improvement, Open Model Sharing, and "Rebasing"

The competitive nature of the subnets and the reward system create a continuous incentive for miners to improve their models. Miners are constantly pushed to refine their techniques to achieve better results and earn a greater share of the rewards. The system is designed to reward valuable nodes in the network at an increasing rate. Once a better model is added to a subnet, it generally becomes publicly available for other miners to use as a starting point for further training. This "rebasing" mechanism creates the incentive for miners to continually build on the best performing models. Successful solutions found in one subnet can be adapted and integrated into others, as the ecosystem encourages open-source collaboration

The Role of TAO Tokens

The TAO token is the native cryptocurrency of the Bittensor network. Apart from governance (voting on proposals and changes to the network) and staking (to participate in the consensus mechanism and earn rewards), TAO’s main function is to incentivise participation (rewarding miners for submitting winning solutions to the subnet competitions). A proposed but not yet realised function is to pay for AI services and applications built on the network. The use of TAO tokens in this way would create a closed loop of incentives, where miners are rewarded for producing valuable AI models, validators are rewarded for accurately evaluating the models, and the overall ecosystem benefits from the increased availability of machine intelligence.

Key Challenges and Open Questions

While Bittensor presents a compelling vision for decentralised AI development, several open questions remain. These questions primarily revolve around the real-world demand for its services and outputs, its overall competitiveness, the long-term value proposition of the TAO token, and the sustainability of the TAO token economy. Answers (or the lack thereof) to these questions will also have implications for subnets looking to commercialise their outputs, as well as startups and projects spinning out of the Bittensor ecosystem.

1. Real-World Usage and Adoption

The success of Bittensor depends on attracting a large number of developers and AI projects to build on the platform (i.e. participating in the subnet competitions), and those AI projects/solutions developed within subnets attracting a large number of users themselves. Are the AI models and applications built within subnets truly valuable? If the proof of that value comes down to adoption and usage, then is there solid evidence of actual demand for these models and applications? Given the network's nascent stage, most network resources are currently concentrated at the infrastructure level and subnet activities (to incentivise participation in competitions and development of solutions), and it’s not yet clear if solutions developed on Bittensor will be widely adopted by businesses or individuals. Over time, Bittensor needs to increase the number and quality of application end users to help enhance token value accrual and its relevance to everyday consumers.

2. Competitiveness of Bittensor’s outputs and approach

Bittensor must compete with established centralised and open-source AI solutions and development approaches. The success of Bittensor depends on its ability to offer better performance or cost-effectiveness compared to traditional options or approaches to AI research and development. How do the AI models and applications produced by Bittensor subnets compare in quality and performance to competitors developed outside of the subnets, and specifically, the state of the art? Are there any proof points indicating preference for or superiority of subnet outputs over external competitors?

Another dimension of competition that Bittensor must deal with is competition for talent. How effective is Bittensor at attracting and retaining top talent from the AI industry? Are the best AI researchers and developers even interested in “earning token rewards for open source AI model development”? For now, it doesn’t seem like many open source AI researchers and developers need more incentivisation than what is already driving the rapid advancement in the field, especially in open source AI. So where do we find proof points that Bittensor’s approach to AI R&D is not only better in terms of output, but also better in the long term based on the talent it's able to attract and retain?

3. Incentive Alignment for Commercialisation and Value Accrual to TAO

In general, a utility token is valuable if the services offered by or built on the network are valuable and the token is used to pay for those services, generating demand for the token. When it comes to real world usage and adoption of a miner's solution (the winning solution emerging from a subnet), while the TAO token is designed to be used as a payment method for accessing AI models and services built on subnets, this function is not yet widely adopted, and there appears to be no explicit requirement that TAO be used to pay for use of the winning solution.¹

This means there is a real risk that miners may release and commercialise their AI models and applications independently of Bittensor, TAO, and the subnet, generating revenue on their own, without any of that revenue and value flowing back to the subnet, TAO token, and Bittensor ecosystem. This could undermine the long-term value of TAO² and pose a challenge to the idea of a cohesive ecosystem where the value of the token is directly tied to the value (and therefore usage) of the AI solutions created on the network. In our opinion, this is a critical aspect for Bittensor to solve in order to be viable in the long term.

4. Sustainability of the TAO token economy

More generally, there is a question around non-speculative demand for TAO. What are the sinks for TAO? What would users spend TAO on, in order to drive demand for TAO? What can miners who earn TAO from subnets do with their TAO apart from cashing out by selling to retail investors? And in that event, what can retail investors do with their TAO? Ultimately, what is the external value inflow that backs the value of TAO? If it’s the effort, resources and expertise expended to create the winning AI models/solutions on subnets, then those solutions themselves need to have revenue/value inflow that links to the TAO token in order to close the value loop for TAO.

With TAO primarily being used to incentivise participation in subnet competitions and to reward winning solutions (does “Build AI Models 2 Earn” sound like “Play 2 Earn”?), and without meaningful sinks that drive non-speculative demand for TAO, there is a risk that TAO is unsustainably inflationary, with a growing disconnect between its current (speculative) price and its actual (utility-based) price, even if TAO is eventually used to pay for services. In fact, we imagine it will be difficult to gate access to winning solutions with TAO given that a) most solutions are open source, and b) for any that aren’t, as long as the IP is owned by the originating miner, it seems impossible to restrict their ability to offer their solution to users independently of TAO (for access) and the Bittensor network entirely, unless Bittensor brings to the table a large and valuable user base (i.e. potential users of the AI models and applications, not other AI developers/miners) for which it can justify restricting access or taking a cut for access.

While the long-term plan involves monetising the network by charging end users of applications that use its subnets, this appears to be a future goal rather than current reality. The value and adoption of TAO tokens seem to be driven more by speculative interest and potential future use cases rather than current practical applications in accessing AI services. Perhaps a generous perspective is that the network is still at a nascent stage, focusing on building its infrastructure and incentivising participation rather than facilitating widespread commercial use of its AI services. But it feels like this question needs an answer at some point, if the value of TAO is going to continue growing in a sustainable way.

5. Additional issues

Owing to Bittensor’s unique approach to incentivising AI development, it also faces a number of issues specific to its incentive mechanisms and participants attempting to game those incentive mechanisms due from the financial rewards at stake, including:

• Model Hoarding: Despite the intention to promote model sharing, there is a risk that miners might hoard high-performing models, limiting the network's overall progress. The network attempts to mitigate this risk by making models publicly available, but the incentive for hoarding may still exist.

• Collusion: As with any blockchain, Bittensor is vulnerable to collusion, where groups of peers may try to manipulate scores to gain disproportionate rewards. The network attempts to counter this by continually requiring colluding parties to increase their stakes to maintain their position.

• Weight Copying: The issue of weight copying, where validators copy the weights of successful miners, has been identified as a risk. Bittensor has implemented protocol updates to mitigate this issue, but its effectiveness remains to be seen.

Bonus Question

What does investing in the Bittensor ecosystem mean? Assuming the above open questions can be addressed, what are the opportunities to invest, and what exactly would one be investing in? Here are some possible angles:

• TAO token: The native token, TAO, used to incentivise AI development and participation in the ecosystem.

• Network infrastructure (miners and validators): Investors can participate in the network as miners by contributing solutions or as validators by evaluating the work of miners. These activities are rewarded with TAO tokens.

• Other supporting infrastructure and tooling: There is a need for tools and infrastructure to support the Bittensor ecosystem, such as wallets, analytics platforms, and delegation services. These needs create opportunities for investment in supporting services that add value to the network.

• Subnet ownership and management (subnet owners): Developers can create their own subnets with unique incentive mechanisms, creating specialised competitions for development of AI models and applications. This presents an opportunity to earn from TAO emissions minus distributions to subnet participants (subnet owners receive 18% of the TAO emissions generated through their respective subnets), as well as to build and commercialise specific AI solutions that emerge from their subnet. Here, it’s important to keep in mind that owning and managing a subnet (which involves the distribution of incentives to generate the best results) is a very different business from being a miner and owning the model/solution that wins the competition.

• AI Model and Application Development (miners): Bittensor provides a platform for building and deploying decentralised AI applications. This could lead to the creation of new and innovative AI-powered products and services. Miners primarily control the commercialisation process for their models and may commercialise their products independently of Bittensor.

The first four opportunities depend on value accruing to the TAO token, value growth beyond speculative value, and sustainability of the TAO token economy (for which we have open questions), while the fifth opportunity relies on commercialisation of AI services beyond the Bittorrent ecosystem (which may negatively impact TAO value accrual).

Conclusion

In essence, Bittensor is an economy of AI developers (miners), models, and applications sharing common infrastructure (competitions, validator network) and incentives to evaluate (“is the output better than competitors?”) and improve (“how can the output be improved?”) their outputs and performance. It leverages the blockchain to incentivise innovation and collaboration, driving the creation of potentially valuable AI models and applications, by providing a decentralised framework where diverse AI-related problems can be addressed through ongoing competitions, driving continuous improvement and innovation.

Rather than simply a decentralised AI protocol, Bittensor is a bold experiment in reimagining the future of machine intelligence and development of AI, offering a unique vision for the future of AI through a decentralised and transparent ecosystem for AI development. But to realise this vision, several critical aspects require careful attention:

• The need to establish clear evidence of real-world demand for AI applications produced by the Bittensor ecosystem. This means demonstrating that the subnets are producing solutions that are useful and that there is a genuine user base willing to pay for access. The value proposition might exist, but widespread adoption and clear metrics of demand are not yet evident.

• The need for a stronger link between value created by contributions to model development and the value derived from accessing those models. The current design primarily rewards participation in the process of developing models, but doesn’t address direct use of the models themselves.

• Ensuring that the value generated by miners ultimately benefits the ecosystem, instead of being completely captured outside of it through the independent commercialisation of solutions, in order to ensure long-term sustainability of the network.

• Identifying meaningful and non-speculative drivers of demand for TAO, in order to ensure the sustainability of the TAO token economy.

Addressing these open questions is vital for Bittensor to fully realise its potential as a decentralised ecosystem for AI innovation and to establish the long-term viability of the TAO token.

If you liked this article and are reading it on the web or received it from a friend, please consider subscribing to my regular newsletter (so you’ll get articles like this delivered fresh to your inbox) by clicking the subscribe button below.

1. In fact, as an open source protocol, AI models developed within the platform, including those emerging as winners in competitive subnets, are generally released as open source, allowing anyone to access, use, modify, and redistribute the models, promoting widespread adoption and collaboration.

2. It’s as yet unclear how "winning" AI models from competition-focused subnets are released for public use - the specific mechanisms for releasing winning models, e.g., licensing terms, quality control measures - or how their usage feeds value back into the subnet (e.g. how revenue generated from the use of winning models is shared or distributed within the originating subnet).