Frontier computing across biological, photonic and electronic substrates
Fig. 1 — Conceptual architecture: electronic orchestration, photonic processing, a living neural substrate and a biological operating system. A research target, not a built device.
Computation is not limited to one physical substrate.
Modern AI has largely scaled by increasing the size and speed of electronic computation. Thalamic Labs explores a different question: what happens when computation itself becomes heterogeneous?
Biological systems evolved adaptive information processing under severe energy constraints. Photonic systems process and move information with light at extraordinary bandwidth. Silicon provides mature control, memory and programmability.
We investigate what becomes possible when these substrates are designed as one computational system.
02 — Technology / Layer 01
Living neural systems as a computational substrate.
Neural organoids are three-dimensional cultures of living neurons. We treat them as adaptive, nonlinear dynamical systems, and build the closed loop around them: encode a task, stimulate, record, decode, feed back, and measure what is retained.
Research has demonstrated that living neural networks can take part in computational loops: cultured neurons adapted to a simulated game in closed loop, and a brain-organoid reservoir performed speech recognition and nonlinear prediction.
Fig. 2.1 — Neuron, myelinated axon and synapse
Demonstratedin the field
Contributes
Plasticity · nonlinear dynamics · adaptive state
Constraints
Batch variability · viability · life support
Sources
Kagan et al., Neuron 2022 · Cai et al., Nature Electronics 2023
02 — Technology / Layer 02
Light as a computational medium.
Photonic computing uses photons, not electrons, to carry and transform information through interference, wavelength, phase and multiplexing. Integrated photonic neural networks are an active, demonstrated research field.
We are investigating whether a photonic layer can work alongside living tissue: optical preprocessing and temporal encoding, plus optical stimulation and readout of neural activity.
Optical stimulation is an interface. It is not evidence that tissue computes with light.
02 — Technology / Layer 03
Silicon remains the orchestrator.
Silicon provides what biology and photonics cannot: precise control, memory, software, safety, communication and an auditable record of every experiment.
The first machine is hybrid: silicon orchestrates while the living substrate learns.
Demonstratedmature technology
Handles
Orchestration · memory · control · safety · logging · communication
02 — Technology / Layer 04
An operating system for a substrate that changes.
A conventional operating system manages fixed hardware. A biological operating system must manage living modules whose state drifts, adapts and ages, and must abstain rather than answer when a module falls out of calibration.
Module registry
Health & viability
Calibration & drift
Task scheduling
Uncertainty-aware output
Module replacement
Experiment ledger
Consent & access
Dual memory. Tacit adaptation lives in the tissue; explicit provenance lives in silicon. The ledger can retrain a replacement module, but it cannot copy the living state.
In developmentsoftware demonstrator, simulated substrates
02 — Technology / The system
Different substrates. Different strengths. One computational system.
BiologyAdaptation
PhotonicsHigh-bandwidth transformation
SiliconOrchestration
Quantum researchFuture information processing Long-term
Each component has been demonstrated independently. The unified closed loop is what Thalamic Labs sets out to build, and to benchmark against silicon-only baselines.
Hypothesisintegrated system
03 — About
We are interested in the point where a biological system stops being something a computer simulates and becomes something the computer computes with.
Why Thalamic Labs exists
Computing has become extraordinarily powerful, but computation is still largely constrained by the physical assumptions of conventional machines. Thalamic Labs develops computational architectures at the intersection of biology, photonics, neuroscience and computer engineering.
03.1 — Why biological computing?
Beyond silicon.
Silicon carried computing for seventy years. Living neural tissue computes on different terms: it is massively parallel, runs on a fraction of the power, and learns as it works.
SiliconBiological
Processing units
80 billiontransistors on one NVIDIA H100 GPU1
~86 billionneurons in one human brain3
Power while running
21.1 MWthe Frontier exascale supercomputer under load2
~20 Wthe whole human brain, roughly a dim light bulb
Energy to learn
~1,287 MWhto train GPT-3 once4
~3 MWha human brain running for 18 years (20 W × 18 yr)
How it learns
15 trilliontokens of training data, then the weights are frozen (Llama 3)5
~5 minutesfor living neurons in a dish to show learning during live play6
Azevedo et al., J. Comp. Neurol. 2009 (86.1 ± 8.1 billion).
Patterson et al., 2021, carbon emissions and large neural network training.
Meta, Llama 3 announcement and model card, 2024.
Kagan et al., Neuron 2022 (DishBrain, ~800,000 cells).
Published figures for each substrate. They describe silicon and biology, not a Thalamic Labs benchmark.
Silicon
Deterministic and programmable
Mature manufacturing and high reliability
An enormous software ecosystem
Excellent general-purpose compute
Expensive scaling for some AI workloads
Memory, compute and learning are usually separate
Biological
Inherently adaptive and plastic
Highly nonlinear and massively parallel
Stateful, and potentially able to learn from interaction
Difficult to standardise; variable and fragile
Requires life support and biological maintenance
Currently far less scalable and controllable than silicon
03.2 — Continuous learning
A computer that keeps learning from experience.
Today's neural networks and transformer models learn in a separate phase. They are trained on enormous datasets (trillions of tokens, weeks of compute), then their weights are frozen and deployed. What they know is fixed at that moment. Learning anything new means gathering more data and training again.
Thalamic Labs is building a different kind of system. Living neurons rewire their connections as they work; that plasticity is how brains learn. We are engineering a biological substrate that learns continuously from its own experience, where every input, response and feedback signal can reshape the network while it runs, with no separate training phase and no freeze.
Neural networks & transformers today
01Collect data
02Train
03Freeze weights
04Deploy
To learn something new: return to step 01 and retrain.
Thalamic Labs: biological continuous learning
01Sense
02Respond
03Receive feedback
04Adapt the network
↻ The loop never stops: learning happens while the system runs.
Hypothesiscontinuous learning in a living substrate
03.3 — Energy
Intelligence on twenty watts.
AI is driving data-centre electricity demand sharply upward. The human brain does its work on about twenty watts. Biological computing is our route toward that kind of efficiency, and we will measure the whole system to prove it, including:
Culture
Perfusion
Environmental control
Stimulation
Optical sources
Detectors
Photonic hardware
Electronics
Data processing
Cooling
Maintenance
03.4 — Evidence gates
Proving it, step by step.
Each frontier programme moves forward through clear experimental milestones, from first signal to a full system benchmark.
Photonic programme
Hypothesis▸ Thalamic Labs today: before P0
P0
Photonic baseline
A photonic neural or reservoir operation, without tissue.
P1
Optical interface
Controlled optical stimulation or readout of neural tissue.
P2
Optical closed loop
Encode a task optically, measure, decode the output.
P3
Hybrid computation
Photonic and biological layers each contribute measurably.
P4
Adaptive loop
Plasticity and photonic parameters jointly improve a task.
P5
System benchmark
The whole system against a silicon-only baseline.
Fig. P — Integrated photonic mesh
Light already performs neural-network operations on integrated chips. We are investigating how to couple that light to living tissue.
Biological quantum programme
Long-term▸ Thalamic Labs today: before Q0
Q0
Physical signature
A replicated, instrument-independent observation.
Q1
Information state
Prepare and distinguish states, ruling out classical explanations.
Q2
Coherent control
Reproducible manipulation with measured coherence.
Q3
Coupling
Controlled interaction between biological quantum elements.
Q4
Computation
A quantum operation with a clear classical comparison.
Q5
Scale
A reproducible multi-element architecture.
Fig. Q — Microtubule, 13 tubulin protofilaments
Microtubules are one candidate research substrate. A controllable biological quantum computer has not been demonstrated, and optical effects alone do not count as gates.
03.5 — Where it could matter
Applications.
Current research
Neurotechnology
New interfaces for studying and controlling biological computation.
Current research
Scientific computing
Biological substrates for nonlinear, temporal and high-dimensional computation.
Near-term target
Drug discovery & biological modelling
Computationally integrated living neural systems as research platforms.
Long-term possibility
Adaptive AI hardware
Computational systems that learn from interaction.
Long-term possibility
Edge intelligence
Potentially energy-efficient adaptive processing where power and latency matter.
Long-term possibility
Frontier computing infrastructure
A general platform for heterogeneous biological–photonic–electronic computation.
03.6 — References
The research we build on.
These results support individual components of the architecture. None of them establishes the integrated system.
