Around 24 hours ago, Bittensor ($1.3B) released a new proposal (proposal, twitter spaces) shaking up the entire network’s incentive flow and potentially innovation speed for new AI experiences. The network is moving from one core token TAO to specific AI subnet tokens for reasons we will discuss. 

As a recap, Bittensor is an AI coordination network between a few parties (model creators or miners, validators and users) to offer a transparent AI network. Bittensor does not create models itself. This is in contrast to something like ChatGPT where the models, training weights and inference all happen in a black box so our inability to know what’s going on is an obvious problem. The lack of open source transparency and continual model competition is the real opportunity for something like TAO. Bittensor is currently probably the most well known crypto x AI project that is liquid and it is the largest.

In layman’s terms, the thesis for Bittensor is similar to the thesis for Ethereum. Ethereum offers the world permissionless app deployment to explore every creative angle, while Bittensor offers permissionless AI model deployment for serving any use case and linking in a demand side of users. Betting on an open system generally offers more shots on goal for a killer use case. ChatGPT, while centralized, just launched its GPT store today to allow for innovation around ChatGPT which offers a marketplace for community fine-tuned GPTs.

Let’s go through a quick recap of the different parties involved with the Bittensor Network 

• Model creators train base models from scratch or fine-tune existing models (add your data to change the 0-1s in a model) to serve some purpose. One may have a killer model for conversation and upload that model to Bittensor’s chat Subnet to compete with other models.

  • As a recap, a large language model is just a zillion numbers each valued between 0 and 1. This is a model’s weights. Not every model is an LLM though.

• Validators: The nuts and bolts of making sure that miners (model creators who upload their models) are serving requests the right way. They run all of the blockchain aspects of the network.

• Users: This may be a person interacting (inference = interaction) with a subnet through a user interface (think chatGPT) or a smart contact itself querying a subnet (it’s getting wild out there fam).

Historically, Bittensor’s issuance (ser, I need TAO to pay my expenses) has been decided by a group of whales. This group is composed of TAO’s top 64 validators who decide which subnet the TAO issuance flows to. Visual learners can look at the subnets below; fellow nerds can go here for the data. 

TAO’s new token econ is an attempt to address Whales making all the decisions on value flows. Whales can just vote flows to a subnet forever, without any innovation or competition between subnets or within subnets (new models for better experiences) and milk the rewards until the network becomes outdated. 

In theory, it’s in the 64 whales’ interest to promote new subnets that offer a better user experience to drive usage but in practice it just becomes too much of a challenge for everyone to be an expert in everything even if they are acting rationally. We’ll go through it below, but the token changes move the speed of innovation to the edges. 

Under BIT001, each Subnet has its own token that can be converted into the main Bittensor token TAO. Also, each subnet has its own issuance (1 token per block, half to the TAO/Subnet token pool and half to miners/validators) schedule and Uniswap style LP pool for conversions between TAO and subnet token. 

The first order effect is that TAO issuance changes from being decided by that 64 whales to being programmatic in nature based on each TAO/Subnet Token pool. If a subnet token is being bought up (a semi-direct monitor of subnet usage), then more of the TAO issuance goes to that subnet’s pool programmatically vs manually. Now, when people go to convert a SubNet token to TAO, it would be worth more (and vice versa). This dynamic value flow should drive more competition between subnets in the form of subnets iterating on models to attract more usage in an effort to earn more TAO. 

As a second-order effect, the parties within each Subdao can compete more effectively to earn the subnet token itself. For example, if language translation is the subdao with the most demand and usage, it earns more TAO, which means more people submit better models to improve the subnet and earn rewards. How the subdao decides which model should be used, or the second order competitive dynamics beyond subnet competition, is still a little fuzzy to me and I’ll have to circle back here. 

Everything we discussed above covers how an increasing subnet token price drives more TAO issuance to that network. This all begs the question, how does increased demand for a subnet translate to a higher subnet token price? This is still something I’m a little shaky on as well (it’s hard to find really good resources on TAO right now). My understanding is that if you want to be a validator or a miner you need some stake, so those seeing demand they wish to serve must buy the subnet token in order to participate in that subnet. I don’t want to drag this on too far since I may be wrong, but having to buy a stake also limits the models that can be uploaded since there is a capital cost required to deploy a model. 

If you’re looking for more detailed information on the economics, I argued with Claude to get to the bottom of it (Bittensor draft -> copy paste to google docs -> export as PDF -> argue with Claude). Check the appendix below.

Conclusion

Overall, I really like the idea of moving value flow decisions that impact the network’s innovation to each subnet vs having a council decide. Any CEO knows the scale and speed unlocks of delegating responsibilities. Imagine if, as a 64-person council member, you had to be an expert in the ultra-hard tech of each specific subnet’s AI capabilities which vary substantially (image generation is obviously very different from text translation) that would be exhausting. 

I have a lot of questions, specifically around how demand for a subnet translates into demand for a subnet’s token since I think that is the crux of the system here. I’m also interested in how each model within a subnet is chosen or ranked in a competitive fashion. 

I would also flag that the introduction of subnet tokens will likely drive smart degens to ape into some sort of a TAO alt season. Imagine apeing a new use AI use case and earning tokens within that subnet. The cool part is unlike some crypto apps, the innovation cycle here creates AI use cases that the entire world can use at scale. My hope is this change allows TAO to move even faster with regard to innovation. 

Check out our podcast with Bittensor’s founders here.

Disclosures: Thank you Michael Rinko for feedback. I don’t currently own TAO. See our podcast with TAO’s founder here.

Appendix: Claude Conversations (I argue with LLMs a lot nowadays)

How does each subnet receive TAO rewards: 

What is the link between subnet demand and subnet token price: