Parasma
Parasma is a biological compute company that trains cultured human brain cells to perform computational tasks, offering an energy-efficient alternative to GPU-based AI processing. The company converts text context into electrical stimulation patterns delivered to neurons, then decodes their responses into tokens, achieving 90% accuracy in next-token prediction tasks. This approach aims to make advanced AI more energy-efficient and accessible by leveraging the inherent computational capabilities of living neural systems.
- Biotechnology
- Artificial Intelligence
- Clean Technology
Funding
Raised to date
$500KRaised to date based on public sources. This may differ from the amount the company actually raised and is based only on what is publicly available on the internet.
Founders
Founder details are not available yet.
Product
Problem
Modern AI is extraordinarily capable but extraordinarily expensive to run, with growth increasingly constrained by energy consumption and hardware limitations. Traditional GPU-based processing requires massive computational resources that are costly, energy-intensive, and environmentally demanding.
Solution
Parasma writes the algorithms and infrastructure that turn living neurons into programmable compute. The platform converts context words into electrical stimulation patterns delivered to cultured human brain cells, then measures and decodes their responses directly into tokens. Using reinforcement learning, the system provides feedback within changing environments to train neurons for adaptive control tasks. The company demonstrated this end-to-end by training cultured neurons to play Doom, and the approach offers a biological alternative for AI inference and decision-making workloads. This technology aims to make advanced AI far more energy-efficient and place it within reach of everyone.
Target Audience
Primary customers are AI researchers and compute-intensive industries seeking energy-efficient alternatives to GPU-based processing for continual learning and extreme energy and sample efficiency applications.
Features
- Converts text context into electrical stimulation patterns delivered to neurons, with responses decoded directly into tokens
- Uses reinforcement learning to train neurons within changing environments for adaptive control tasks
- Achieved 90% accuracy in next-token prediction tasks with full sentence context, with randomized controls scoring significantly worse
- Demonstrated 78.1% accuracy in non-linear token prediction experiments, exceeding the 75% ceiling of a silicon linear decoder
- Custom architectures designed for biological computing support language-based token prediction
- Built toward repeatable benchmarks across cultures, devices, and time to scale biology with advances in AI
- Treats questions of consciousness and suffering seriously, with safeguards guiding training of living neural systems