On Decentralized Intelligence Systems, Pt. I

Aligning Incentives for Compute

Modern technology is not merely a collection of isolated advances, but rather a symphony of intersecting innovations, each enriching the other in a confluence of multidisciplinary cross-polination. Among these, machine intelligence stands as a pillar, complemented by the advancement of database mechanisms and the growing power of computational resources (re: Nvidia’s earnings report this week). Together, the fruits of one discipline serve as the catalyst for growth in another.

Consider Bittensor, a protocol representing a novel approach to creating value and incentivizing the distribution of computational intelligence. It operates as a peer-to-peer intelligence market fueled by a digital ledger where the quantity and quality of contributions are appraised by peers within an ecosystem that rewards creation and collaboration.

It is a substantial undertaking to build, but a worthwhile one nonetheless. It represents the convergence of advances in many disparate technologies: blockchains and use-cases for decentralized databases/ open-source large language model creation, fine-tuning, and use/ availability of cloud compute. Since you can use a blockchain you don’t need to pay for an expensive data lake and schema. Since you can use GPUs in the cloud you don’t need to pay for extensive hardware and rent space for such hardware in order to mine or validate. Perhaps most importantly, you can use open-source large language models that are not riddled with ideology like Gemini or ChatGPT.

The Bittensor Model

At the heart of Bittensor lies a parameterized function that encapsulates the definition of intelligence, as proposed by Hinton et al. This function, denoted as f(x), is trained over a dataset D to minimize a loss L. In Bittensor’s architecture, numerous such functions (dubbed ‘peers’) collectively contribute to a stake-weighted machine learning objective [1]. The strength of the system is in its aggregation of computational entities: each agent in the system is incentivized to compete and collaborate in order to refine the overarching intelligence system.

The Incentives

Bittensor employs a market-driven approach where peers assign value to each other, ranking based on their outputs. These peer-assigned rankings are aggregated on a blockchain ledger, enabling the translation of higher rankings into monetary rewards. Thus, a feedback loop is created that inherently values the informational production of each peer, irrespective of the tasks or datasets employed for training [2].

One of the more novel contributions this architecture makes is the incentive mechanism designed to counteract collusion amongst peers, a common susceptibility within decentralized networks. The system updates rewards based on an incentive function I(W, S), which mitigates the likelihood of collusion by stipulating that rewards are limited to peers that have not achieved consensus within the network, assuming no single group possesses a majority stake. To facilitate this, a trust matrix T is derived from inter-peer ranking weights, thereby instilling a layer of trust and value validation that is decentralized and resistant to manipulation [3].

Thus, we identify the three most significant innovations as: 1) a decentralized intelligence marketplace, 2) a mechanism for rewarding niche contributions, and 3) a mechanism for collusion resistance.

The Usability

These are undoubtedly innovative mechanisms. What if you wanted to actually partake in the proposed architecture? What is the user experience like for someone trying to mine TAO and help secure the Bittensor network?

We think the OpenTensor Foundation (the team behind Bittensor) has much work left ahead of them. Unit Zero Mining Ops dove into pooling some GPU strength to run the Bittensor protocol and identified several key areas for improvement.

To begin, scalability. Network statistics are shown below. There are around 50 active validators and just over a thousand miners working to provide computational resources to underpin the protocol. This is while the market capitalization of the protocol’s token (shown prominently on their stats page) nears $4B USD. Remarkable.

However spinning up your own Bittensor mining unit is far more intense than mining something like Bitcoin. TAO mining in general is not user-friendly. Which impacts scalability; the complexity of mining TAO involves several hurdles such as necessary infrastructure, configuring software, managing stakes, and a deep understanding of the tokenomics involved. The OpenTensor Foundation has stated publicly in their discord that they are working on user interface improvements, which in our opinion is vital to truly test the viability of this complex structure. There is also a considerable lack of technical documentation for miners (although there is extensive software documentation). Offering managed services to less tech-savvy users might help to improve new user growth with respect to mining.

In short, the biggest takeaway is the lack of some sort of abstraction layer. A layer that handles the complex operations behind the scenes, allowing users to interact with the system through simplified tooling.

One of the better ways we can think of improving this system is adding more modularization. Separation of functions into distinct, more manageable modules, each with its own interface, would help users to interact with the system at varying levels of complexity depending on their expertise and areas of interest. You could implement built-in feedback mechanisms whereby users can readily report issues, suggest improvements, or request features rather than maintaining a chaotic discord channel.

The Offerings

There are 32 subnets available, offering a range of GPT-like services such as chat, data scraping, translation, and more. The protocol’s model outputs are distributed via Corcel (a separate entity), which maintains the Cortex Language Model, a library of fine-tuned Llama, Qwen, GPT, and DALLE models. This is a sound approach since the field of large language modeling is advancing so rapidly. Bittensor (or Corcel for that matter) can take advantage of fine-tuned publicly available models and deliver AI tools with ability near that of the highest performing models.

One of the major architectural decisions of the protocol involves the separation of model training (off-chain) from model inference (on-chain). There are obvious advantages to this, particularly when it comes to enabling the scale of resources required to train complex models. Performing this off-chain allows use of specialized hardware without being limited by computational capacities of blockchain nodes. Another critical feature of this is regarding privacy - something we’ll touch more on later.

Bittensor approaches this tradeoff by allowing models to train in a more adaptive off-chain environment, while the inference tasks (typically more lightweight and fewer in number), are processed on-chain. This hybrid approach combines the flexibility of resource utilization with the security of on-chain event logging, offering a pretty balanced model for a decentralized learning market.

While this general approach is likely the most advisable, there are a few areas of improvement to mention. An ongoing challenge, as noted by the OpenTensor Foundation, is validation of the off-chain training processes. There needs to be a reliable method for certifying that the models are trained as claimed. A reward mechanic for good behavior and performant models is one such path.

Another obvious challenge is optimizing gas costs, since delivering inference on-chain will always carry an associated cost to transact. Bittensor leverages substrate, a modular para-chain protocol framework built by former Ethereum co founder Gavin Wood. Identifying the correct L2 chain is essential to achieving system viability, considering the requirements for high throughput and minimal transaction fees. Finally, in consideration of all of these challenges, there remains the problem of balancing load, i.e. fine-tuning the balance between what processes run off-chain versus on-chain to maximize computation efficiency and system output.

How to Incentivize Data Sharing

The most important component to get right is the validation of off-chain training: if you could guarantee the security of an entity’s dataset within a complex system of like datasets, you could enable significant insights with much higher impact potential. Right now, businesses tend to operate in data silos within their industry; there is no Pareto optimal path for any individual corpus to share their company’s data, or even their data collection methods, with other companies operating within their industry.

But what if a protocol was able to align the incentives of companies competing within a single industry, homomorphically encrypting their data, seamlessly processing it together, and providing a single endpoint with outsized benefits for every individual company? That would be an incredibly compelling intelligence market offering.

It is a use-case Bittensor has probably considered for their own protocol, but something that would fail given the current implementation of their architecture. The lack of incentive for data provision and federation is the missing element. We anticipate competitors trying to take advantage of this path forward.

Work Cited

[1][2][3]Rao, Yuma. “Bittensor: A Peer-to-Peer Intelligence Market.” https://drive.google.com/file/d/1VnsobL6lIAAqcA1_Tbm8AYIQscfJV4KU/view.

Note: This is a free research article from Unit Zero Labs, and does not in any way constitute investment advice. We may or may not own assets we discuss in our research. We post premium market research monthly in addition to our research posts that you can subscribe to.