This Bittensor Subnet Could Cut Drug Discovery Costs in HALF | E2267
This Week in Startups Podcast Recap
Published:
Guests: Michaela Baso, Pedro Pena, Tom Bliers, Max Sebti, Critically Lon
What Happened
Bittensor is a decentralized network employing crypto incentives to reward contributors who provide valuable AI models, compute, or results to specific subnets. The network consists of three main actors: subnet owners who design challenges, miners who solve these problems, and validators who select winners. Subnet 68 is focused on drug discovery, with a goal to cut down the time and cost of drug development, currently averaging 10 years and $2.6 billion, by leveraging decentralized contributions.
Subnet 68, launched as a proof of concept for decentralized drug discovery, started with a dataset of one billion molecules and expanded to 65 billion using combinatorial reactions. Miners submit molecules of interest and compete using chemical search algorithms, focusing on synthesizable molecules. The process includes validating these molecules in wet labs, often through contract research organizations, with the help of a partner in Shanghai, Yalatane.
The subnet is designed to be flexible and target-agnostic, allowing for co-development and screening as a service. This approach aims to improve the predictive power of drug development, significantly reducing wasted time and resources on ineffective drug candidates. The FDA's move towards accepting international trial results is expected to expedite the drug approval process.
Bittensor democratizes participation in drug discovery by enabling non-experts to contribute through a search problem format. The system encourages continuous learning and adaptation, with miners quickly identifying potential issues in scoring functions used for predictions. This iterative process ensures that submissions align with long-term value generation, leading to potentially innovative technology.
The Bitcast Network is another subnet where miners compete to generate social media views for brands, automating the process from creating briefs to measuring attention. Creators are rewarded based on engagement quality, as measured by watch time rather than views, which emphasizes the importance of meaningful interactions. This approach taps into the booming creator economy, projected to be worth $250 billion worldwide.
The SCORE project aims to give AI vision capabilities, allowing users to build vision AI apps by distilling large vision language models into smaller, specific skills that are more accessible and cost-effective. The project incentivizes miners to break down large models into specific tasks, such as person or car detection. This approach reduces model sizes from gigabytes to megabytes, making it feasible to run on local devices like a Mac Mini.
The company behind the SCORE project is building a community of enthusiasts and has a go-to-market strategy focused on vision vibe coding. A major partnership with a large corporation is hinted at, which could help with distribution to large enterprises. The Bittensor community is supportive of the project, and there is interest in investing in Tao as a learning experiment.
Key Insights
- Bittensor's Subnet 68 aims to reduce drug discovery costs by leveraging a decentralized network of contributors. The average drug development process currently takes 10 years and costs $2.6 billion, a timeline and budget the subnet hopes to cut significantly.
- The subnet's dataset expanded from 1 billion to 65 billion molecules using combinatorial reactions, allowing miners to submit and test synthesizable molecules. This process is supported by a partnership with Shanghai-based Yalatane, which aids in validating molecule candidates in wet labs.
- Bitcast Network automates social media engagement by rewarding creators based on watch time rather than views, emphasizing quality interactions. This model capitalizes on the $250 billion creator economy, aiming to democratize participation by alleviating administrative burdens for smaller creators.
- The SCORE project focuses on making AI vision capabilities more accessible by reducing the size of vision language models from gigabytes to megabytes. This allows for running these models on local devices, making them feasible for a wider range of applications and users.