The Abstraction of Mining

Bittensor is a blockchain that makes mining its main purpose—and abstracts it to such an extent that any digital service can qualify. The blockchain is thereby intended to become a marketplace for machine intelligence. Now, a research startup wants to use it to accelerate the discovery of new drugs.

Is it bullshit or revolutionary? Often that’s hard to determine—even for us. An example of a story that could be either intriguing or just hot air recently appeared in Forbes:

This decentralized AI, the magazine headlines, „could revolutionize drug development.“

What is meant here is a neural network, specifically designed for drug research and decentralized via the Bittensor (TAO) blockchain. This revives an old idea: performing cryptocurrency mining not through „useless“ hashes, but through scientific computation. About eight to ten years ago, coins like Gridcoin tried this—without ever succeeding in truly decentralizing scientific mining.

Could it be possible that the combination of modern staking mechanisms and neural networks will finally make this achievable? That’s the question we’ll be exploring below.

Simulating Chemical Reactions at the Atomic Level

Drug development is typically a very long and arduous process, involving hundreds of steps and taking on average more than 13 years.

However, this process can be simplified and improved using artificial intelligence and molecular simulations. Instead of developing and testing drugs physically, compounds are constructed and simulated in computers. This allows researchers to test more molecular candidates in less time—at least, that is the hope currently permeating the pharmaceutical industry.

In April, Rowan Labs launched a specialized neural network for this purpose called Egret-1. Its goal is to simulate chemical reactions at the atomic level. Until now, this has been incredibly resource-intensive—even scientific supercomputers require a lot of time to realistically simulate just a handful of atoms for a few seconds. Rowan aims to improve this, not by training its neural network on internet data, as ChatGPT does, but by using quantum mechanical equations. The AI learns to reconstruct the results of those equations.

For models like Egret-1 to be successful, however, Rowan explains they require „much more high-quality data generated through density functional theory (DFT).“ To generate this data, Egret-1 will „leverage the decentralized computing power of the Bittensor network,“ specifically a subnet for decentralized AIs from Bittensor’s Macrocosmos. According to Forbes, this „could drastically reduce the cost and duration of discovering new therapeutic components and treatments.“

Not uninteresting, right? But what’s the deal with Bittensor and Macrocosmos?

Protein Folding in the Macrocosmos

Macrocosmos was established recently, in 2024, and is part of the Bittensor blockchain (more on that in a moment). It is essentially a mesh of Bittensor subnets.

Rowan uses the „Mainframe“ subnet (SN25). This enables scientific simulations, especially protein folding—a foundational problem in chemical biology for which AI is particularly promising: „Protein folding is one of the hardest scientific problems and demands enormous computational power. Although recent breakthroughs like AlphaFold3 have been highly effective, their centralized data handling means that the cost per request is far too high,“ explains the Macrocosmos website.

SN25 is designed to lower the cost of protein folding simulations. It employs the „GROMACS“ standard—to simulate protein folding—but integrates this into a „competitive design.“ This structure „incentivizes miners to develop machine-learning models that solve protein folding as efficiently as possible.“ Validators in the system check miners’ outputs using specific heuristics. The miners who perform best receive tokens from Bittensor—TAO—as a reward.

This competitive process is intended to reduce costs and boost efficiency. Currently, there are 30 active validators simultaneously conducting more than 3,000 simulations; since June 2024, over 400,000 protein folding tasks have already been completed. That’s still far from what AlphaFold accomplishes—but it’s a beginning.

The Abstraction of Mining

To understand Macrocosmos, there’s no way around delving into Bittensor itself. The core idea behind Bittensor is quite fascinating:

One can think of Bitcoin as a decentralized marketplace for digital goods—a market that rewards miners for generating hashes. For Bitcoin, this market is simply a means to an end: securing consensus over a digital ledger (the blockchain) to facilitate a decentralized transaction system. Bittensor, by contrast, makes the digital goods marketplace an end in itself.

Bittensor’s core innovation is the „separation of the blockchain’s core function (transferring value, etc.) from the operation of the validation system, which defines how the digital goods marketplaces are created.“ This is important: in classical consensus mechanisms like Proof of Work and Proof of Stake, the consensus algorithm includes the rules for when a consensus-relevant input—a hash or a stake—is valid. Bittensor, however, determines only under which circumstances the consensus itself becomes effective.

The consensus tasks themselves can be written in any language and are validated entirely off-chain—allowing large volumes of data and computing power to be employed. „Bittensor brings the same sort of abstraction that Ethereum added to Bitcoin by introducing smart contracts to Bitcoin’s inverse innovation—the digital marketplaces.“

Just as Ethereum abstracts transaction logic and enables the construction of diverse systems, Bittensor makes it possible to allow even complex and fuzzy mechanisms as consensus work: for example, machine intelligence, protein folding, data storage, model training, and more.

Bittensor does not define the consensus task itself, but rather „the marketplace that pays for these services to be made available to the network.“ This is done via the „Yuma Consensus.“ The network is split into miners and validators, elected by stakers of the TAO token. While the miners provide „intelligence,“ the validators check their results and vote on them. In this way, the whitepaper explains, „a market emerges in which intelligence is priced peer-to-peer on the internet by other intelligent systems.“ Through weights and stakes, peers vote on their influence in the network, mutually rank themselves and their neighbors, and log these scores on a blockchain.

The exact implementation described in the whitepaper is complex and full of mathematical formulas. From what I understand, the system is similar to election-based consensus mechanisms—such as Ripple’s or Proof of Authority like Binance Smart Chain—where the peers in a network vote for consensus and rate each other’s trustworthiness, and combines this with abstracted mining. In each subnet, the consensus tasks for miners and their validation by validators are defined. Once consensus is reached in a subnet, it is transferred to the global blockchain layer.

As with most decentralized systems, the crux is in preventing—or at least making it very unlikely for—fraud via collusion. To address this, Bittensor rewards honest selection of weights as much as possible, which, according to the whitepaper, makes the system resistant to collusion up to 50% of the network’s total weight. It should, therefore, have the same security parameters as Bitcoin and other blockchains—meaning 50% of the peers (or hashes, stakes, or weights) must be honest and follow the rules.