“It answers questions related to Bitcoin and the economy well, at least better than GPT-4,” Aleksandar Svetski told Coin telegraph during a lively Bitcoin event in Amsterdam.
The entrepreneur, author and founder of Spirit of Satoshi, a new extended language model (LLM) of AI, begins to reveal the arduous journey his small startup took to create its Bitcoin-centric AI chatbot.
The model is the result of a tedious training process aimed at generating answers based on reputable Bitcoin resources, the Austrian school of economics and libertarian ideals. Still in its infancy, Spirit of Satoshi reflects the ideals of a “well-curated Bitcoin corpus” including resources like Saifedean Ammous’s best-selling book. The Bitcoin standard.
Svetski explains that the major difficulty in building the model was not only gathering relevant information sources ranging from books, research articles to podcasts, but also guiding the model to generate answers through a process of comprehensive training. He adds that a common misconception about LLMs is that they search for information like a search engine:
“They simply string words together probabilistically in a way that is representative of the patterns within the model. So it’s not even about finding anything.
This is part of the reason why AI chatbots tend to “hallucinate” from time to time, Svetski says, and why developing an LLM requires focusing on training it on one response style. Spirit of Satoshi is by no means perfect either, at least not in its current version:
“Our model will also hallucinate. It’s also going to say shit, but it’s going to say something more like a Bitcoiner would say.
After establishing a broad but focused base of Bitcoin-centric information and data, Svetski’s team began populating the model with tens of thousands of question and answer pairs using programmatic methods. However, a human element is still needed to help Spirit of Satoshi generate answers that could have come from its namesake.
The continued development of the model therefore relies on the broader Bitcoin community. Spirit of Satoshi uses an incentivized process that allows the public to verify, create, and validate model data.
Using Lightning Network, Nostr credentials, or email addresses, a “proof of knowledge” mechanism allows users to get paid. satoshis for helping to train the model.
The process uses a consensus model that will automatically impose a penalty if users create “junk data.” Svetski describes him as the crucial “human” element to improve the results of Spirit of Satoshi:
“It produces amazing content, it’s the final piece that takes your content from 80% good to 95% good. And this has a huge impact on the quality of the model.
The difference between the responses generated by Spirit of Satoshi and ChatGPT is palpable, according to Svetski. The latter is trained on the dominant ideas of what Bitcoin is and concepts like inflation:
“If we ask ChatGPT about inflation, he will tell you that it is a sign of a healthy economy. Well no, inflation is a sign of systemic problems, like your diminishing purchasing power.”
Svetski says this scenario was part of Spirit of Satoshi’s raison d’être, reforming the LLM to reflect the nuances that embody the type of thinking behind the Bitcoin movement:
“If you ask about inflation, our model should answer ‘no, inflation is actually bad for the economy because it discourages saving’ or ‘saving has a knock-on effect on preferences people’s temporalities.”
The future of the platform is quite open according to its founder. Spirit of Satoshi could be a learning tool or online tutor integrated with educational platforms or online universities. This could also be the basis for the “ultimate Bitcoin influencer” thanks to his BTC-centric results:
“I would like to see it become the destination for the 100 or 500 million people who want to learn about Bitcoin, the starting point for their first steps in understanding.”
Spirit of Satoshi was built based on an existing open source model that has inherent English fluency and “Wikipedia-like bias”. The latter problem was solved by structuring the model’s responses to its dataset on Bitcoin and Austrian economic principles.