[Top Builder 2024]: Code Satoshi

[brenorb]

🏗️ Top Builder 2024 Application Form to track progress through Round 1 - 3 ~ Currently in Round 1

📝 Instructions

1. Only complete this form if you have been chosen for Top Builder, by PlebLab.
2. Provide detailed information to assist the judges in evaluating your project.

🧑‍🤝‍🧑 The Team

• Team Name: Code Satoshi
• Number of Founders: 2 - Benoit & Breno
• Team Background: Share the professional backgrounds of each founder.
  • Breno: Electrical engineer who switched to focus on AI while in college. Worked in traditional finance until launching own Bitcoin startup in 2017. Since then I have been 100% focused on working in the Bitcoin space. In 2023 I began working with the spiritofsatoshi.ai team, to help build a Bitcoin-centric Large Language Model.
  • Benoit: Stanford Graduate and former IT executive, I’ve been working on data and AI-related projects for a number of years. I joined the Spirit of Satoshi team in 2023 to help implement their data engineering framework for training and RAG data. In that capacity I’ve been tinkering with Breno on an AI chat assistant for Bitcoin developers.
• Bitcoin & Lightning Involvement: Describe your involvement and experience.
  • Breno co-created the Portuguese BIP39 seedword list, translated several books such as The Bitcoin Standard, and is now hosting BitDevs Brasília and will be launching a Bitcoin book in Q1 2024.
  • Benoit: long-time Bitcoiner, decided to leave the corporate world in part to contribute to open-source Bitcoin projects. I find the nexus between AI and Bitcoin most interesting and it gives me an opportunity to leverage my background in data and AI.
• Origin Story: What inspired your idea?
  • While working with the Spirit of Satoshi team on a general chat language model, our team did extensive research and product discovery with many Bitcoin companies. These calls made it evident that Bitcoin developers could not get the assistance they needed while writing code from mainstream code assistants like Github Copilot. Due to the niche nature of Bitcoin related coding languages, these models do not provide any meaningful support for these languages. Breno took on this challenge during a hackathon in Brazil, late 2023 and won first prize. Since then we’ve begun to explore the potential for a more serious “Code Satoshi” Bitcoin assistant.

💡 The Idea

• Project Overview: What does your project do? AI Code assistants such as GitHub Copilot have been rapidly adopted and have increased productivity for developers and accelerated the learning process for new and unfamiliar languages. While other AI code assistants already provide robust support for the most widely used languages such as javascript and python, they unfortunately are not trained on more niche Bitcoin specific languages. As such, there are no AI code assistants to help Bitcoin developers, creating both a barrier to entry for new participants and decreased productivity for current developers.

  Code Satoshi is a bitcoin-focused AI code assistant built to help engineers from all walks of life build Bitcoin and bitcoin-related applications better, faster and easier.

  Code Satoshi can:

  • Write code,
  • Correct code
  • Explain code

  Code is a broad umbrella term encompassing Bitcoin related languages, libraries and protocols. As described earlier, Code Satoshi already supports Bitcoin Miniscript. The goal will be to build out a larger library of supported languages, as a prototype initially.

• Website: (Optional)

  • http://codesatoshi.com/

• GitHub:

• Project Stage: Describe the current progress.

  • The Bitcoin Miniscript policy code assistant is currently in a closed alpha release. Feedback is being gathered from alpha testers to assess the quality of answers and further improve and plan out next steps, including further languages to support.

• Supplementary Materials: Provide any proof of work, workshops, or demo videos.

📚 Additional Information

• Additional Links/Info:
  • Satsconf Hackathon Pitch
  • Legends of Lightning Pitch
  • Flowchart Demo

📌 Submission Type

• [X] Initial Project Creation

🚀 Current Project Progress

• Progress Made:
  • Breno created a Bitcoin Miniscript assistant as part of the Satsconf Hackathon and won 1st place at hackathon. Breno was then invited to participate further in the Bolt.fun’s Legends of Lightning hackathon and received the BiFi award under the project name Minimaxis. The Model is currently implemented into the UI of the spiritofsatoshi.ai website for users to prompt the model and receive responses.
• Challenges Faced:
  • Training any kind of Large Language model requires that a large number of examples be provided to the model. Bitcoin related code is not easy to find, so much of it might have to be handcrafted just like Miniscript policies before. Given that we are not Bitcoin developers, we need to quickly learn how to code examples and evaluate the model answers, just as understanding what are the developers' pain points.

• Future Roadmap of what's next:

  • One of the larger issues currently facing Bitcoin users is the current mempool congestion and high on-chain fees. Blockstream’s Liquid Elements has been utilized by companies as a second-layer scaling solution to reduce fees and bypass mempool congestion.

  • We have already met with the Blockstream team to discuss adding support for Elements on Code Satoshi to help support existing developers and especially to accelerate the onboarding of new developers to Elements. We have spoken with several companies and are continuing discussions to identify key pain points and provide support for these challenges by Code Satoshi.

⏳ Round 1 Milestone

• Keep in mind that completion will need to be by February 10th, 2024.
  • Get solid feedback on the Miniscript alpha version and start building an early Liquid Assistant which can easily find and interpret the documentation.
• First-round milestone the team has set for themselves:
  • [ ] Get feedback from Miniscript alpha version
  • [ ] Make a fist version of a Liquid assistant.

📝 Submission Complete

• All required information has been provided above.

📅 Top Builder Program Structure

Registration Opens: Monday, December 18, 2023

• Application: Interested teams apply for Top Builder.
• Selection: Teams will be chosen based on their application. See FAQ for details.

Round 1: January 13 - February 10, 2024

• Workshops: PlebLab weekly workshops start, available live and online.
• Chat Access: Selected teams join the PlebLab internal chat.
• Mentorship: Begins for teams.
• Project Milestones: Teams are set and judged on their first-round milestones.

Round 2: February 10 - 24, 2024

• Team Reduction: PlebLab cuts half of the Top Builder teams. (Only half continued in Round 2.)
• Project Milestones: Teams are set and judged on their second-round milestones.
• Workshops: PlebLab weekly workshops continue.
• Mentorship: Continues.

Round 3: Starting February 24, 2024

• Team Reduction: PlebLab narrows down to 6 Top Builder teams. (All will be pitching in Austin.)
• Final Milestones: Teams set final project milestones for the last round.
• Preparation: Teams prepare their pitch in front of judges. Final teams are contacted to come to Austin for Startup Day.

Final Round at Startup Day: March 14, 2024

• Preparation: Teams work on their final pitches at PlebLab.
• Presentations: Final teams present before judges on Startup Day.
• Deliberation: Judges decide during a happy hour.
• Announcement: The winner of Top Builder is announced at Austin BitDevs.
• Celebration: After-party for Top Builder.

[brenorb]

Current status for Round 1

Market Survey

We contacted almost 30 bitcoin companies on how they use Liquid and Miniscript and their main issues with it, but also showed our product to get their reaction.

Some quotes from them about their issues:

I knew that we needed blinded keys, but I found it very very difficult and gave up quickly because it was way over my head.

I don’t have time to read all this documentation, and some of the time it’s confusing, so I end up just chasing someone from Blockstream.

If you could ask for code-examples it would’ve been useful.

Feedback when showed Miniscript Assistant and Liquid Technincal Assistant:

I definitely could see myself using this everyday.

There is someone from China who can't speak much English asking us questions all day, this could be perfect for him. Spanish would be equally important as English, Chinese would be nice to have.

Insights:

• We should check how well the model support other languages and also adapt the website.
• We could leverage our Lightning Enabled Crowdsourced-LLM tool to rate answers and solicit qualitative input in order to identify where the documentation gaps exist.
• We could improve UX generally, from emails to app UI.

Development

Miniscript Assistant

When PlebLab started we were just finished migrating the model to the Spirit of Satoshi website. It was not properly formatted in Markdown and also the flowchart was not always showing.

Now it shows the flowchart and we also automatically compile Miniscript policy into Miniscript code, Bitcoin Script and also show spending cost analysis, all of these in a deterministic way so there is no chance of hallucination and properly formatted.

Improvements and Problems solved

• [X] Flowchart showing.
• [X] Compiles Miniscript policy into Miniscript code and Bitcoin Script.
• [X] Show Spending cost analysis.
• [X] Markdown formatting.
• [X] Found the solution to reducing model hallucination in some edge case complex contracts, from 100% to only 20%.

Problems to solve and future improvements

• [ ] There is a bug in the flowchart generation that impedes it to render some small subset of more complex contracts.
• [ ] Implement in production the hallucination mitigation solution mentioned above.
• [ ] Improve UX for making it clear what model the user is working on and the invite email.
• [ ] Try a new approach that might reduce much further the hallucination with agents.

Liquid Technical Assistant

We noticed an early trend of people trying to use Liquid to escape from high fees onchain, so we decided to start a Liquid Technical Assistant to ride the wave. We already have a closed alpha version which we are testing.

What was done

• [X] Scrape all possible data for Liquid and Elements (Liquid documentation, community knowledge, Blockstream help, Elements tutorials, Elements Academy, etc.)
• [X] Index everything
• [X] Enrich RAG metadata (display source information)
• [X] try new embeddings
• [X] Optimize RAG storage and API

Future improvements

• [ ] Test more with people who need it in real life
• [ ] Scrape and index Liquid Dev telegram public chat (which seems to be the main source of help)
• [ ] Test Liquid Technical Assistant for the Chinese user mentioned above
• [ ] Test more with other languages (Spanish seems to be working fine)

Business

We are using a value4value approach with credits bought with LN, however since there's not many bitcoin developers right now, this can only go so far. We are focusing in capturing a great part of the market for the long term where lots of people will get a huge productivity boost from our assistants.

We intend to get closer with companies that work with Miniscript and Liquid it in the day to day life. AI research can get quite expensive, so in order to use and experiment deeper with open source models, and even to be able to open source them to public in the future, we are applying to grants to raise more funds. We are already applying to become a Liquid Federation Member, which will get us closer to Liquid developers to help us improve the assistant further.

[brenorb]

Minor update: I made this demo video to make it more clear: link