Bittensor creator Const on Affine, dTAO, "mining reasoning," and more | E2326

Const, the creator of Bittensor, built the network by abstracting Bitcoin's proof-of-work concept beyond hash production and applying it to artificial intelligence, which he identifies as the most important computational problem of the 21st century. He argues that permissionless markets are purely meritocratic, blind to race, gender, nationality, and political affiliation, and that this property causes the efficiency of producing digital commodities to grow exponentially. Bitcoin's hash rate has grown exponentially and never declined even as price fluctuates due to speculation, regulation, and country-level bans, which Const cites as evidence of this dynamic.

The key technical challenge Bittensor had to solve was inventing a consensus mechanism capable of measuring high-dimensional outputs like a 1024-dimensional vector, far more complex than Bitcoin's binary hash verification. Bittensor currently has 128 subnets, each typically hosting around 256 participant computers, with discussion of expanding to 256 subnets. Each subnet has its own alpha token with a 21 million coin cap that functions as a staking token convertible back to TAO. TAO itself has a 21 million coin cap with one coin produced every 12 seconds before halving, and newly minted TAO is distributed to subnets as additional liquidity based on performance. Anyone can register a subnet permissionlessly by burning tokens through a Dutch auction mechanism that doubles in cost on each registration then decreases until someone pays, with the current registration cost approximately 600 TAO, equivalent to roughly 121,000 dollars.

Const describes dTAO as a meta-mechanism launched just over a year ago in which Bittensor applied its own permissionless mechanism design to itself, creating a system that selects among permissionless mechanisms. Under dTAO, all subnets received their own token paired to TAO and compete to increase ecosystem value as measured by price. Const says the increased competition has raised the quality of entrepreneur teams entering the ecosystem, though he acknowledges a significant number of subnet creators are scammers given the open nature of the system. He describes Bittensor's permissionless subnet mechanism as scaling capital more efficiently than human organizations because it routes money through a codebase rather than requiring hiring or physical resource acquisition, citing one subnet that tripled its compute in two months simply by tripling miner payments.

Const personally runs subnet 120, called Affine, focused on measuring intelligence as a commodity under the tagline mining reasoning. Affine miners produce reasoning models indirectly, where the quality signal is whether the reasoning generated causes other models to answer questions correctly. He claims Affine adapters could allow users to run a smaller model instead of a large model and achieve approximately 30 times faster inference. Bittensor produced a model last year that was better than the best available 35 billion parameter model at the time but was surpassed by a competing model before launch. Const states the network's ultimate goal is to compete with elite centralized AI labs at the level of intelligence rather than just compute, inference, or storage, and acknowledges that measuring intelligence as a commodity is significantly more abstract and difficult than measuring compute or inference.

Decentralized training is described as a holy grail of the AI field. The core technical obstacle is that merging model weights across the network requires transmitting approximately one terabyte of data per step, meaning hundreds of thousands of training steps require hundreds of thousands of terabytes of communication, exceeding what average home internet connections can support. A further unsolved problem is verifying that a miner trained on the specific required data subset and did not contribute corrupted gradients, since even one bad actor can destroy an entire training run. Const says training a trillion-parameter model is beyond what Bittensor can currently do until a decentralized training algorithm solving the bandwidth problem is developed.

Const describes the network as fundamentally adversarial by design, arguing that bad actors are always present and the system must force them to behave well regardless. He says adversaries can reach nation-state sophistication, citing North Korea as an example of actors who placed developers as remote workers at companies seeking six-figure salaries. In response to rug pulls, Bittensor built a conviction-locking mechanism into the chain allowing subnet teams to voluntarily lock their tokens, making large sell events a public signal to investors, and many teams have voluntarily locked their tokens for years after this mechanism was introduced.

The original Canadian foundation is no longer involved with development. Const now operates solely within the RAL Foundation focused on programming and chain development, while the OpenTensor Foundation handles marketing and outreach. The RAL Foundation intends to build governance systems directly into the chain within the current year. Const believes Bittensor would continue operating even if he stepped away, citing the active open-source community and the 128 subnet teams that understand the technology.