Pearl pairs AI computing with mining as Van Ness calls it bigger than Bitcoin

Evan Van Ness said Pearl could use matrix multiplication, a core AI computation, for cryptocurrency mining, while cautioning that the additional computing burden varies by model.

Pearl could extend the uses of proof of work by combining cryptocurrency mining with AI computing, according to Van Ness. The approach would allow calculations used to train AI models or generate responses to serve a mining function as well.

In a November 11 interview with The Rollup, Van Ness, who previously held a leadership role at the Ethereum Foundation and ran a venture fund, described Pearl as “bigger than Bitcoin.” He said he discovered the project about 18 months earlier and had been absorbed in it for over a year.

His argument centers on useful proof of work: mining through calculations that also serve another purpose. Van Ness said Bitcoin’s mining computations benefit Bitcoin but have no other use, and that researchers had spent decades trying to make proof of work perform socially useful calculations.

Van Ness said Pearl mines through matrix multiplication, a fundamental computation in AI and large language models. He said it is used both in training models and in inference, the process of generating responses to queries. In his account, the appeal is that the computation supporting mining can also perform work needed by AI services.

Asked whether the approach made computing more efficient, Van Ness said it did not. Adding noise to the calculations can introduce some inefficiency, he explained. The aim is to run mining alongside AI inference while keeping the additional burden as small as possible.

He described Pearl’s software as a plugin for VLLM, software used to serve large language model inference. Operators could add Pearl alongside their existing inference workloads and mine while processing those requests, he said.

The additional burden nevertheless depends on the model, Van Ness said. He claimed that some models had shown no additional overhead across the full processing workflow, while acknowledging that skepticism about such claims was reasonable. Reducing that overhead as far as possible remains an engineering challenge, he said.