Why Proof-of-Training (PoT)



Crypto mining is a rapidly changing industry. In 2022, Ethereum transitioned from the energy-intensive Proof of Work (PoW) consensus mechanism to an alternative called Proof of Stake (PoS), in response to growing environmental and energy concerns. Consequently, this change led to a substantial reduction in power demand, ranging from 99.84% to 99.9996%. Ethereum's reduction in energy consumption could be comparable to the electrical power needs of a nation like Ireland or even Austria, whose advancement has a significantly positive impact on environmental sustainability. However, it has also resulted in a substantial amount of unused hashrate, equivalent to 1,126,674 GH/s, which now lacks a specific application. This brings the potential for miners to shift their computational resources from crypto mining to other areas like internet of things (IOT) and data services. This transition can remain fully within the blockchain space, by using these resources to run processes hosted on decentralized blockchain-based networks.

Meanwhile, with the integration of artificial intelligence (AI) into various sectors of the economy, the demand for computational resources to fuel this machine intelligence is experiencing rapid growth. Training a model like ChatGPT incurs expenses exceeding $5 million, and operating the initial ChatGPT demo costs OpenAI approximately $100,000 per day prior to the surge in its current usage. Due to the extensive number of neural parameters and significant GPU hours required, the high computational demands of model optimization present substantial challenges for academic researchers and small-scale enterprises, limiting the widespread use of artificial intelligence technologies.

It is therefore unsurprising that an increasing number of crypto miners are exploring ways to utilize their existing computation infrastructures to contribute to the advancement of AI, redirecting its previously mining-focused computational resources for machine learning and other high-performance computing (HPC) applications, as demonstrated by Hive Blockchain. The company's long-term HPC strategy involves shifting from Ethereum mining to HPC applications, including artificial intelligence, rendering, and video transcoding, with an anticipated revenue generation of approximately $30 million per month.

We believe that the emerging trend of combining and integrating these resources has the potential to significantly enhance the development process of AI tools in both technical and financial aspects. This would provide AI tool developers with a more affordable plan to monetize their innovations, including simplified training and marketplace access. Instead of exclusively commercializing their creations through major technology corporations, developers have the opportunity to contribute to the decentralization of technology by shifting their assets from centralized entities to a global commons.

Despite the considerable potential, the decentralization of software and hardware underlying AI remains in its early stages, due to the absence of well-developed consensus frameworks. In general, we identify the following major challenges currently hindering the progress and realization of a decentralized AI utility network:

  • Challenge 1: Reliable validation mechanism.
  • Challenge 2: Absense of efficient consensus protocols for delivering services.
  • Challenge 3: Ownership protection from model-stealing attacks.

We have a more detailed description of the problem and our proposed solutions in our recent archived paper. The mini kernel demonstrating the performance of the PoT protocol is also presented in our github repo, where we utilize the practical Byzantine fault tolerance (PBFT) consensus mechanism to synchronize global states. Our results indicate that the protocol exhibits considerable potential in terms of task throughput, system robustness, and network security.