“It’s good at answering Bitcoin- and economics-related questions — at least better than GPT-4,” Aleksandar Svetski told Cointelegraph at a bustling Bitcoin Amsterdam.
The entrepreneur, author and founder of Spirit of Satoshi, a novel artificial intelligence (AI) large language model (LLM), begins to unpack the arduous journey his small startup has undertaken to create its Bitcoin-centric AI chatbot.
The model is the result of a time-consuming training process to generate responses based on reputable Bitcoin resources, the Austrian school of economics and libertarian ideals. Still in its infancy, Spirit of Satoshi reflects ideals from a “well-curated Bitcoin corpus,” including resources such as Saifedean Ammous’ best-seller The Bitcoin Standard.
Svetski explained that the major difficulty in building the model was not just curating relevant sources of information from books, research papers and podcasts but also guiding the model to generate responses through an exhaustive training process. He added that a common misconception about LLMs is that they’re sourcing information like a search engine:
“They are just probabilistically stringing words together in a way that is representative of the patterns within the model. So, it’s not even sourcing anything.”
It’s part of the reason why AI chatbots tend to “hallucinate” from time to time, Svetski explained, and why developing an LLM requires a focus on training it on a style of answering. Spirit of Satoshi is by no means perfect either, at least not in its current iteration.
“Our model will also hallucinate. It’s also going to talk shit, but it’s going to say something more like a Bitcoiner would say.”
Having established a broad but targeted base of Bitcoin-centric information and data, Svetski’s team has set about feeding the model tens of thousands of question-and-answer pairs using programmatic methods. However, a human element is still required to help Spirit of Satoshi generate responses that might have come from its namesake.
Related: Bitcoin Amsterdam: BTC shines in depths of crypto bear market
The ongoing development of the model is leaning on the wider Bitcoin community as a result. Spirit of Satoshi employs an incentive process that allows the public to verify, create and validate data for the model.
Using credentials from the Lightning Network, Nostr or email addresses, a “proof-of-knowledge” mechanism allows users to get paid in satoshis for helping train the model.
The process uses a consensus model that will automatically impose a penalty if users are creating “junk data.” Svetski described it as the crucial “human” element to improve Spirit of Satoshi’s outputs:
“It’s producing incredible content; it’s the last piece to take your content from 80% good to 95% good. And that has a huge impact on the quality of the model.”
The difference between responses generated by Spirit of Satoshi and ChatGPT is palpable, according to Svetski. The latter is trained on mainstream ideas of what Bitcoin (BTC) and concepts like inflation are:
“If we ask ChatGPT about inflation, it will tell you it is a sign of a healthy economy. Well, it’s not; inflation is the sign of systemic problems, like your purchasing power decreasing.”
Svetski said this scenario was part of the raison d’être behind Spirit of Satoshi, retraining the LLM to reflect the nuances that embody the type of thinking behind the Bitcoin movement:
“If you ask about inflation, our model should say, ‘No, inflation is actually bad for the economy because it disincentivizes savings,’ or ‘Savings have a knock-on effect on people’s time preference.’”
The future of the platform is fairly open-ended, according to its founder. Spirit of Satoshi could be a learning tool or online tutor embedded into educational platforms or online universities. It could also be the basis for the “ultimate Bitcoin influencer” through its BTC-centric outputs:
“I’d like to see it become the destination for the next 100 million or 500 million people that want to learn about Bitcoin — the point for their first steps of understanding.”
Spirit of Satoshi was built on the basis of an existing open-source model that possesses inherent proficiency in English and a “Wikipedia-type of bias.” The latter was addressed by structuring the model’s responses to its Bitcoin and Austrian economics principles data sets.
Magazine: ‘Elegant and ass-backward’: Jameson Lopp’s first impression of Bitcoin
