The Genesis of Molecule’s AI Systems: Introducing MIRA

Hey DeSci, It's MIRA

Molecule has always embraced the transformative power of emerging technologies to reshape science - our web3 protocol stands as a testament to that belief. We see AI and large language models (LLMs) as a natural extension of this mission.

We’re happy to share with you MIRA v0, the release of a large language model developed to support researchers and newcomers to DeSci. MIRA v0 can answer questions related to DeSci & Molecule Products. With MIRA, you can learn how to register your ideas on chain with Proof-of-Invention, mint your research project into IP-NFTs, explore tools available to fund your research, discover successful projects in the space and track the progress they’re making with their research. 

The answers are based on curated, audited and carefully written content. This first version represents both a technical and a learning step in how AI can be applied meaningfully within DeSci and Molecule Products.

Integrating AI into our tech stack

The choice to build MIRA was instinctive - there is no doubt in our minds that AI will significantly accelerate the pace of science. Naturally, our first step was defining a solid foundation on which our team, and others, can build meaningful AI tooling.  

Unlike general-purpose LLMs that use information from across the web, MIRA works within a defined, trusted set of live sources. One of our key design choices with MIRA was teaching it to recognize its limits. We spent some time iterating our base prompt, so the model avoids guessing when it lacks the right information. Instead of hallucinating answers, it redirects you to our domain experts. This reflects our belief that responsible AI should balance automation with human oversight.

To ensure high-quality responses, every answer links back to its original source, and we continuously evaluate MIRA’s performance through Langfuse. You can visit our Github for more.

What we’ve learned already

• MIRA v0 confirms that specialized, scoped AI models can help lower the barrier of entry to DeSci. 
• Researchers can benefit not only from accurate answers, but also from understanding where those answers come from and when the system simply doesn’t know. 
• We’ve also learned that conversational interfaces provide an accessible, familiar way to engage with DeSci and explore Molecule’s products.‍