The brain, the body, every living system processes information electrically. We build models that predict these dynamics across organisms, starting from molecular data: transcriptomic and electrophysiological measurements in, network-level electrical behavior out.
The only confirmed instance of general intelligence runs on biological substrates. Those substrates process information electrically.
The constraints are now architectural.
Today's generative models, from transformers to world models, learn by predicting the next token or frame across enormous datasets. Brains arrive with structured priors and build an understanding of the world from far less data.
They also organize information differently from machine learning and conventional computer architecture: memory and computation sit together, activity is sparse and event-driven, and learning is continual. We are designing AI from those principles.
Generative models today
Brains
Learning signal
The next token or frame, across enormous datasets
Structured priors; understanding of the world built from far less data
Memory and compute
Separated, with data moved between them on every step
Co-located in the same tissue
Activity
Dense and clocked; every parameter on every pass
Sparse and event-driven
Learning
A training run, then frozen weights
Continual, for the life of the organism
Infrastructure
Racks of accelerators, bolted on after the model is designed
A body co-designed with the brain it carries
The approach
Predict the electrome. Decode the principles. Build from them.
AlphaFold treated protein folding as a prediction problem and transformed structural biology. We treat the brain's electrical network dynamics the same way.
This is not neuroscience for the sake of neuroscience. The brain is the data source, not the end customer. We study how biological systems organize information at the electrical layer, then port those principles to silicon, to build computing architectures that understand, not only predict.
Biological data
Transcriptomic and electrophysiological data across organisms, from molecules to whole brains.
Electrome prediction
Models that predict bioelectric network dynamics from that data.
Organizational principles
How living systems arrange memory, computation, energy and learning.
Computing architectures
Models and hardware designed from those principles, on non-carbon substrates.
Research
One chain, three stages.
Each stage is the input to the next. The electrome models produce the principles; the principles produce the model architectures; the architectures set the hardware.
01
Foundational models of bioelectric network dynamics
NEXUS. Research, code and trained models are open.02
Bio-intelligent models
A departure from transformer architectures and next-token prediction: computing architectures designed from how living systems organize information and energy, running on non-carbon substrates.
Not human-like AI. Not humanoid robotics. AI built as its own system, informed by biology at the level of principle, not imitation.
03
Neuromorphic hardware and infrastructure
Hardware and infrastructure co-designed for these models, from the operating system through the chip.
The brain's efficiency is not purely algorithmic; a significant component comes from the cohesion between brain and body. Co-designing model and hardware recovers analogous gains.
The data exists.The tools exist.The pipeline does not.
Data
- Whole-genome transcriptomics
- Connectomes at synaptic resolution
- Single-cell electrophysiology atlases
Tools
- Graph neural networks
- Differentiable simulation
- Large-scale compute
Missing
- The end-to-end composition pipeline from biological electrical data to computing architecture
That is what we build.
The input to that pipeline extends beyond human biology. Some animals process certain information more efficiently than humans. Earlier hominids were better at things that modern environments no longer select for. Cross-species study of bioelectric organization is a design resource, not a curiosity. There may be more efficient ways to organize computation than any single organism demonstrates[1].
The lab
Built in Austin.
s9m Labs is an AI and hardware computing research lab based in Austin, Texas. The lab draws on the talent density of UT Austin and greater Austin, and on a city with the power, the land and the hardware industry that a lab building its own infrastructure needs.
The research is open. NEXUS research and code are public. The models built from it are the lab's work product.
s9m is Sribhashyam, written the way a16z writes Andreessen Horowitz: the first letter, the nine letters between, the last.
Founded
2026, Austin, Texas
Scope
Electrome prediction, bio-intelligent models, neuromorphic hardware
Research
Public. Code and findings released as they are produced.
The mark
A neuron drawn as a circuit: dendrites to via pads, a soma, a split axon.
Start the conversation.
avsribhas@gmail.comIf you believe intelligence has a substrate, and that substrate can be understood and built from, we are the lab doing that work.
Researchers, engineers, partner labs and investors: write to us.