Our R&D program builds proprietary Neural Cognitive systems for the physical edge — compact multimodal models, modular AI edge boxes, robotic embodiment, and long-horizon hybrid thermodynamic compute.
The Laiss Neural Cognitive Architecture (NC1) is a proprietary ~4B-parameter multimodal model designed for real-time edge deployment. It fuses vision, audio, language, and robotic telemetry into one cognitive flow — so a machine can see a scene, understand an instruction, and act with low latency.
NC1 is not a cloud chatbot — it is built to drive physical systems. Our studies focus on a robotic arm controlled by stacked Laiss AI Edge Boxes: one module for kinematics, another for sensing/safety, another for high-level cognitive planning.
Demonstrators include precision manipulation and pedagogical use-cases (e.g. teaching painting technique), where the model observes motion, evaluates stroke quality, and provides corrective feedback in real time.
Vision, audio, force/torque & joint state
Unified latent context for the task
Low-latency motor plans on the edge box stack
A Multi-Agent Neural Cognitive Architecture via Hybrid Thermodynamic-Deterministic Computing.
Standard AI relies on static, feed-forward execution paths that lack the recursive nature of biological cognition. Laiss-NC-Thermal solves this via a "Society of Mind" approach, offloading stochastic Dreaming and regulatory Conscience functions to Extropics Thermodynamic Sampling Units (TSU).
By mapping agent interactions to an Ising model, we leverage the TSU's ability to naturally relax into low-energy states, enabling real-time, self-regulating artificial cognition at 10,000x greater energy efficiency than tradition MCMC on silicon.
Inputs are compressed into a high-dimensional spin-vector.
TSU hardware performs Gibbs sampling via physical noise.
Relaxed state is projected back to valid semantic outputs.
Our architecture distributes cognition across seven specialized agents, each operating as a nodal point in a thermodynamic graph.
Defines the global energy landscape (Hamiltonian). It ensures ethical consistency by penalizing high-energy configurations that violate hard-coded safety constraints, forcing the system to relax into "safe" states.
A symbolic verification layer that takes the relaxed states from the TSU and validates them against formal logic rules and mathematical axioms before final output commit.
Leverages the inherent thermal noise of p-bits to perform massively parallel Gibbs sampling. It explores the high-entropy state space to find creative associations and non-obvious solutions.
Utilizes a hybrid approach: Long-term Memory is stored in a high-density vector store (RAG), while Short-term Memory is mapped directly to the active spin configurations in the TSU.
Compresses multi-modal stimuli into a high-dimensional binary latent space (z). This vector serves as the initial "spin" configuration for the thermodynamic relaxation process.
Continuously ingests real-time sensory data, performing rapid feature extraction via deterministic "early-vision" layers before passing the results to the Interpretation agent.
Updates the inter-agent coupling strengths (Jij matrix) based on the divergence between the Conscience regulatory goals and the Dreaming agent's output.
Unlike monolithic Large Language Models (LLMs) that function as static probability predictors, Laiss-NC-Thermal utilizes seven distinct agents that engage in high-speed, parallel interrogation.
When a task is ingested, agents operate in sync through cross-channel resolution. The Interpretation agent compresses stimuli while the Dreaming and Logic agents cross-examine potential states across multiple output channels simultaneously.
1,024+
Interconnected agents1M/sec
Cross-agent feedbackInfinite
Multi-modal synchronization2^1024
Combinatorial complexity"By offloading high-entropy stochastic exploration to physical thermodynamic noise, we eliminate the computational overhead of software-simulated annealing."
Current generative systems face a prohibitive energy cost for running concurrent high-fidelity agents. Standard silicon architectures struggle with the recursive feedback required for genuine self-regulation.
Hybrid Thermodynamic-Deterministic Computing allows us to bypass this wall. Our TSU hardware naturally equilibrates into global optima, providing a physical solution to what was previously a massive mathematical bottleneck.
The journey from high-fidelity GPU simulations to native thermodynamic compute. We are currently in Phase I, validating our core stochastic exploration loops via massively parallel Gibbs sampling (MCMC).
Validation of HTDML algorithms using GPU-based Gibbs sampling (MCMC).
Integration of first-gen Extropic TSUs for specific agentic sub-tasks (Dreaming/Conscience).
Migration of all 7 agents to native thermodynamic hardware for 10,000x efficiency.
# NC-Thermal Core Initialization
# TSU Hardware: Extropic Z1 Liquid Nitrogen Core
[CORE_INIT] Calibrating p-bit temperature...
[CORE_INIT] Target Beta: 12.45 (Gibbs Saturation)
[CORE_INIT] Loading Jij coupling matrix...
def agent_loop(stimuli):
z = encoder.compress(stimuli)
# Perform physics-level relaxation
# Latency: < 250ns
z_relaxed = tsu.relax(
state=z,
thermal_noise=true,
iterations=10000
)
return decoder.project(z_relaxed)
| Metric | GPU Baseline (H100) | Laiss-NC-Thermal (TSU) | Improvement |
|---|---|---|---|
| Energy per Sample | ~Joules | ~Microjoules | 10,000x |
| Latency (Relaxation) | Milliseconds | Nanoseconds | 1,000,000x |
| Architecture | Sequential/Batched | Physics-Level Parallel | Infinite |
LAISS-NC-THERMAL IS A RESEARCH PROGRAM OF LAISS LABS. EXTROPIC TSU IS PROPERTY OF EXTROPIC AI.
Active development studies across Neural Cognitive software, cognitive libraries, and next-generation compute.
From-scratch transformer research (GQA, RoPE, QK-Norm, SwiGLU) and industrial NC1 studies for multimodal edge cognition and robotic control.
Cognitive Functions and States — specialized transformer nodes for attention, executive function, memory, language, perceptual-motor control, social cognition, and visuospatial skill.
Hybrid thermodynamic–deterministic multi-agent cognition. Full technical deep-dive above — jump to NC-Thermal.
Scientific and commercial study of synaptic-style connectivity and cognitive routing — bridging research narratives with productizable AI infrastructure.
Lightweight cognitive-affect research (e.g. fear regression pipelines) for safety-aware annotation and agent orchestration experiments.
Modular hardware/software units that host per-task NC models at the point of action — scalable stacks for arms, sensors, and high-level planning.
General-purpose cloud models struggle with latency, cost, and siloed modalities when placed on a factory floor or in a teaching lab. Our Neural Cognitive approach keeps intelligence local, multimodal, and embodied — so machines can complete skilled physical work with measurable operational ROI.
For the investment thesis and pitch deck, visit Investors. Partnerships: About Us or the home contact form.