Research

The science
behind the brain.

OSCEN is grounded in decades of computational neuroscience. We build on established research, and we're honest about where we diverge.

Neuroscience foundations

Each mechanism in OSCEN maps to established neuroscience research.

STDP

Bi & Poo (1998), Markram et al. (1997)

The foundational discovery that spike timing determines synaptic plasticity direction and magnitude. Our implementation uses configurable time constants per synapse group.

Eligibility Traces

Izhikevich (2007), Gerstner et al. (2018)

Three-factor learning rules where STDP creates a 'tag' that neuromodulators later convert to permanent change. Bridges millisecond spike timing with second-scale reward signals.

BCM Theory

Bienenstock, Cooper & Munro (1982)

Sliding threshold for synaptic modification. Active neurons become harder to excite further. Creates competition and selectivity in neural responses.

Neuromodulation

Schultz (1998), Hasselmo (2006)

Dopaminergic reward prediction error, cholinergic attention gating, noradrenergic arousal modulation, serotonergic mood regulation. Each modeled as a global gain channel.

Synaptic Scaling

Turrigiano et al. (1998)

Multiplicative homeostatic plasticity that normalizes total synaptic input per neuron. Prevents runaway excitation while preserving relative weight distributions.

Critical Periods

Hensch (2005), Knudsen (2004)

Experience-dependent windows of heightened plasticity. Our adolescent phase models the reminiscence bump: a period of rapid pruning, myelination, and identity formation.

Energy efficiency

Spiking neural networks compute only when neurons fire. Sparse, event-driven activation translates to dramatic energy savings.

GPU Inference 50–200 J/s

LLM API Call ~1,080 J/query

OSCEN on neuromorphic silicon (projected) <5 J/s (estimate)

1 watt = 1 joule per second. Continuous systems are shown in joules per second; per-task energy in joules. An LLM query costs ~1,080 J (~0.3 Wh, Epoch AI 2025). The OSCEN figure is a projection from published spiking-network benchmarks on neuromorphic hardware, not a measurement of our system; today the brain runs as a simulation on a CPU server and we have not yet measured its power draw.

73×

less energy than CPU

Robotic peg-in-hole task on Intel Loihi: 52 microjoules vs 3,800 microjoules per inference on CPU. IEEE 2024.

109×

less energy than GPU

Keyword spotting on Intel Loihi: 109x lower energy per inference than GPU. Blouw et al., 2019.

SNN state of the art

Where spiking neural networks stand today, the capabilities and the gaps.

Capabilities

+ Continual learning without catastrophic forgetting
+ Event-driven computation (compute only when needed)
+ Temporal pattern recognition (spike timing encodes information)
+ Low-power edge deployment via sparse, event-driven compute
+ Real-time sensor processing at <1ms latency

Current Limitations

~ Best ImageNet accuracy: 82.39% vs ANN 91%
~ SNN LLMs max out at ~1.5B parameters
~ Training tools less mature than PyTorch/JAX ecosystem
~ Limited off-the-shelf pretrained models
~ Specialized low-power compute still maturing

Our position: OSCEN doesn't compete with LLMs on language benchmarks. We compete on continual learning, energy efficiency, and real-time adaptation. The exact capabilities robotics needs and LLMs can't provide.

Competitive landscape

How OSCEN compares to transformer-based robotics AI.

Company	Approach	Learning	Edge
Prometheus (Bezos)	VLA Transformer (Industrial)	Retrain	No
AMI Labs (LeCun)	JEPA World Models	Retrain	No
Figure AI	VLA Transformer	Static	No
Google DeepMind	Gemini VLA	Retrain	Partial
Physical Intelligence	3B param VLA	Static	Cloud
OSCEN	Spiking Neural Net	Continual	Yes

Honest assessment

Can OSCEN match GPT-4 at language?

No. And it's not designed to. OSCEN is built for embodied intelligence: real-time sensor processing, motor control, and continuous adaptation. The design lets it query LLMs as tools (patent Claim 3).

Is it training on one server really enough?

For proof of concept, yes. We're demonstrating that a million spiking neurons can learn from raw sensory data. Production deployment on next-generation low-power compute is designed to scale to billions of neurons at a fraction of the power.

What about accuracy benchmarks?

We don't yet have published benchmarks on standard tasks. The brain is currently in its adolescent developmental phase. Early results show cross-modal association formation and motor pattern learning. Formal benchmarks are in progress; a pre-registered retention pilot started in October 2026.

What about LeCun's AMI Labs and JEPA world models?

AMI validates our thesis: the future of robotics AI is beyond LLMs. But JEPA still relies on GPU-scale pretraining on millions of hours of video, then deploys a frozen model. It can't learn continuously on-device, can't run on milliwatts, and has no solution for catastrophic forgetting. OSCEN and JEPA are different bets: top-down (train big, deploy frozen) vs. bottom-up (learn continuously, adapt forever).

What about Bezos's Project Prometheus and the $12B physical-AI round?

Prometheus is a major validation of the physical-AI category: Wall-Street-scale capital (JPMorgan, Goldman, BlackRock) backing an 'artificial general engineer' for industrial design and manufacturing, with two of the original transformer authors as advisors. But it's the same architectural bet as AMI: video-language-action transformers trained at GPU scale. The contrast is the point. Prometheus and OSCEN aren't competing in the same market segment (industrial design vs. embodied continual learning); we're competing for which architecture defines physical AI. Theirs is the bet that data plus simulation beats biology. Ours is the bet that biology beats both.

Why not just use transformers for everything?

Transformers excel at static inference. But they can't learn from a single experience, can't run on milliwatt hardware, and can't adapt in real-time. For robotics, those aren't nice-to-haves. They're requirements.

Key references

STDP. Bi, G.-Q. & Poo, M.-M. (1998). Synaptic modifications in cultured hippocampal neurons. J. Neurosci.

Three-factor. Gerstner, W. et al. (2018). Eligibility traces and plasticity on behavioral time scales. Front. Neural Circuits.

BCM. Bienenstock, E., Cooper, L. & Munro, P. (1982). Theory for the development of neuron selectivity. J. Neurosci.

Reward. Schultz, W. (1998). Predictive reward signal of dopamine neurons. J. Neurophysiol.

Scaling. Turrigiano, G. et al. (1998). Activity-dependent scaling of quantal amplitude in neocortical neurons. Nature.

Critical. Hensch, T. (2005). Critical period plasticity in local cortical circuits. Nature Rev. Neurosci.

SNN energy. Published spiking neural network energy benchmarks, joules per inference (2024). IEEE IJCNN.

KWS energy. Blouw, P. et al. (2019). Benchmarking keyword spotting efficiency on neuromorphic hardware. NICE '19 / arXiv:1812.01739.

Be first in line for the OSCEN humanoid. It's free, and backing the brain moves you to the front.

Be first in line →

← Back to overview

The science behind the brain.