# Deep SIML Labs — Full Public Context > Canonical machine-readable context for the Deep SIML Labs website. Generated from the current site position and public Markdown source files. ## Company overview Deep SIML Labs is an early-stage scientific research company developing computational systems for modeling, simulation, optimization, and future physical experimentation. The company is simulation-first today. Its long-term objective is to close the loop between computational prediction and physical measurement so that models, experimental history, and evidence can improve together. Materials and useful physical systems are long-term research targets, not current discoveries or products. The organizing research loop is: Model → Simulation → Optimization → Physical Test → Measurement → Prediction Error → Learning → Better Experiment Only the computational portion is presented as active today. Physical testing, automated recalibration, closed-loop experiment selection, and functional-material discovery are planned capabilities that still require implementation and validation. The internal north-star question is: How much useful physical knowledge can Deep SIML Labs gain per experiment? ## Current status ### Exists today - Public SIML paper, demonstration, and code artifacts. - A reported SIML FEPGate + V-JEPA 2 planner benchmark with a public implementation. - A documented computational research thesis centered on testable optimization and simulation. ### Active direction - Testable optimization and experiment-selection methods. - Simulation workflows and fair baseline comparisons. - Reproducible research infrastructure and explicit uncertainty. ### Long-term objectives - Calibrated physical measurement. - Simulation-to-reality learning and model correction. - Closed-loop experiment selection. - Validated functional-material and physical-system discovery. ## SIML SIML is being developed as a testable optimization and learning architecture. Its relevant research questions include adaptive memory, temporal state representation, prediction-error handling, parameter search, experiment selection, sample efficiency, compute efficiency, and simulation-to-reality adaptation. SIML is not assumed to be superior. It must earn its value through controlled comparison against random search, grid search, Bayesian optimization, conventional active learning, and domain-specific methods under equal budgets and constraints. Useful measures may include sample efficiency, optimization regret, robustness to noisy measurements, adaptation after model mismatch, compute cost, and the number of experiments required to reach a defined target. ## Research approach The current computational loop is: Question → Model → Simulate → Analyze → Optimize → Iterate The planned physical extension is: Prediction → Physical Test → Measurement → Error → Recalibration → Next Experiment The physical extension is under construction and must not be interpreted as an operational laboratory capability. Core principles: - Evidence over hype. - Capabilities are earned through demonstration. - Uncertainty is explicit. - SIML competes with credible baselines. - Work should be reproducible and versioned. - Physical reality ultimately adjudicates models. - No fake laboratory imagery, fabricated results, fictional materials, or unsupported partnerships. ## Research areas - SIML and recursive systems — active computational research. - Optimization and search — active computational research. - Simulation frameworks and reproducibility — in progress. - Scientific machine learning — active research direction. - Simulation-to-reality — exploratory and planned. - Functional materials — long-term research direction. Open questions include whether adaptive memory can reduce required evaluations, how uncertainty should propagate between simulation and measurement, whether SIML-informed experiment selection can outperform conventional optimizers, and which physical design spaces justify the cost of closed-loop experimentation. ## Materials and resources The materials program is a future research direction, not a catalog of discoveries. Candidate domains include functional polymers, ionic systems, compliant actuation, strain and pressure sensing, variable-stiffness systems, and adaptive materials. Conceptual target properties include force, strain, response time, cycle life, efficiency, compliance, conductivity, manufacturability, environmental stability, hysteresis, repeatability, and uncertainty. No proprietary candidates or performance values are currently claimed. The intended roadmap is: reference systems → characterization → model fitting → optimization → material candidates → device validation. Stages beyond public computational artifacts remain future work. ## Company and commercial thesis Deep SIML Labs is not positioning SIML primarily as a software subscription. The potential long-term value lies in an internal research engine that compounds models, experimental history, measurements, and learned representations. Potential future outputs include validated materials, designs, intellectual property, licenses, components, research findings, research partnerships, and joint development. These are possible business models, not claims of current products, customers, discoveries, or revenue. Deep SIML Labs is not currently a large industrial laboratory, a materials manufacturer, an autonomous scientist, a completed physical-discovery loop, or a company claiming solved artificial muscles. ## Archive and historical context EranaOS is an archived research project. It explored persistent AI agents, memory, relationships, and long-lived simulated environments. Active development concluded as Deep SIML Labs shifted its resources toward simulation, optimization, and future physical and materials research. Earlier SIML papers and notes used stronger artificial-life and emergent-intelligence language. Those artifacts remain available as research history, but they do not define the company’s current claims or evidence standard. ## Contact Deep SIML Labs welcomes concise inquiries concerning research collaboration, industry partnership, investment, media, or general technical discussion. Email: rich.everts@siml.life --- ## Published research artifacts ### One bit to rule the planner: SIML FEPGate + V-JEPA 2 - Canonical URL: https://www.siml.life/research/VJEPA2-siml-fepgate - Date: 2025-10-15 - Author: Richard Everts - Content type: Benchmark - Status: Current > **TL;DR** - We added a one-bit **surprise gate** (*FEPGate*) from our **SIML** cognitive sidecar to the frozen **V-JEPA 2** planning loop. > With **zero retraining** and the **same CEM/MPC budget**, that tiny hook reshapes the latent energy surface and delivers big wins: > > - **Final error:** 0.193 m → **0.109 m** (~50% ↓) > - **Monotonicity:** 0.12 → **0.43** (>3× ↑) > - **Per-step latency:** 2.28 s → **1.13 s** (≈2× faster) > - **Energy/episode:** 0.058 Wh → **0.029 Wh** (≈2× ↓) > - **EDP (energy × delay):** **~4× lower** --- ## Why this matters Even world-class latent planners waste compute when the environment isn’t surprising, and they can get stuck spiraling around **spurious minima** when it is. A single **cognitive bit** fixes both: - Spend compute **only** when the world deviates from prediction (surprise ↑). - **Skip** expensive re-plans when everything tracks (surprise ↓). This is the smallest practical step toward a **cognitive OS** layer that governs *any* latent system - not just V-JEPA 2 - with a universal, model-agnostic feedback signal. --- ## What we built - **B1 - Grip-bit** Exposes the gripper open/close as a single bit to both sampling and scoring (cleaner grasp/place phases). *(Ablation left to appendix/code.)* - **B2 - FEPGate (the star)** A SIML sidecar computes a normalized **surprise** score `s(a; z_k)` on **CEM/MPC elites** at latent state `z_k`: - If `s > τ` → **reject/penalize** those elites (reshape the elite set). - If `s ≤ τ` → leave them alone (no unnecessary re-plans). - **Zero retraining** V-JEPA 2 encoder/predictor are **frozen** throughout. - **Same budget** Horizon, samples, and iterations unchanged; we only add the bit-level hook. - **τ-calibration (obs-driven)** One short warm-up per scene: sample the L1-ball at `z_k`, measure surprise, set `τ` at the 5–10th percentile (empirically `~1e-4 … 1e-3`). - **Adaptive memory (bonus)** SIML schemas **expand** with novelty spikes and **compress** when surprise stays low - keeping entropy and compute in check without touching V-JEPA’s weights. --- ## How it works (30-second tour) 1. **Predict.** V-JEPA 2-AC rolls out candidate futures `ĥ z_{t+1}` from `z_t`. 2. **Gate elites (pre-exec).** SIML computes surprise on **elite** candidates; high-surprise ones are gated out before execution. 3. **Act.** Execute the best action `a_t`. 4. **Update (post-exec).** Encode the new obs → `z_{t+1}`; SIML updates surprise/memory. This **one-bit feedback** carves away bad pockets in the search space and preserves curvature where dynamics are feasible. --- ## Results **Setup.** 100 episodes of single-goal reaching (6–7 steps/ep), 7-DoF deltas sampled in an L1-ball (`r=0.075`). NVML power at 1 Hz. Single NVIDIA GPU; sidecar on CPU. No fine-tuning. Budget < \$100. **Headline wins** (FEPGate vs OFF): - **Error**: 0.193 m → **0.109 m** - **Smooth progress**: 0.12 → **0.43** - **Latency**: 2.28 s → **1.13 s** - **Energy/ep**: 0.058 Wh → **0.029 Wh** - **EDP**: **~4× lower** (product of means) - **EDP (Energy×Latency)**: **OFF ≈ 0.132**, **FEP ≈ 0.033** → ~**4×** lower.
Final error distributions Per-step latency Energy per episode
Figure 1 - Lower error, ~2× faster decisions, and ~2× lower energy with FEPGate.
**What it feels like:** direct trajectories when the model is right; immediate course-corrections the instant reality diverges. No wasted search when nothing changed. --- ## Implementation notes - **IPC:** tiny CLI (`surprise`, `step`) exchanging `.npy` buffers + a small JSON. - **Placement:** gate **only** the **elite set** each CEM/MPC iteration (orders of magnitude fewer calls; same effect). - **Threshold:** τ from an observation-driven warmup; typical values in our runs land around `1e-4 … 1e-3`. - **Memory:** runtime schema **expand/contract** policy keeps useful novelty while pruning redundancy. --- ## About the Author This research was conducted by **Deep SIML Labs**, an independent research lab exploring the frontiers of artificial life, cognition, and self-organizing intelligence. To stay informed or collaborate, [contact us.](../../contact) --- ### Artificial Intelligence is for Amateurs: SIML, Free Energy Principle and Cognogenics as the Foundations Toward the Birth of True Artificial Life - Canonical URL: https://www.siml.life/research/Artificial-Intelligence-is-for-Amateurs - Date: 2025-04-15 - Author: Deep SIML Labs - Content type: Historical Paper - Status: Historical > Historical context: This document is retained as part of the research record. Its framing does not necessarily represent the current company position. ## Abstract We introduce **Cognogenics**, a new paradigm for artificial life and intelligence that arises not from predefined goals or supervised data, but from an agent’s continuous drive to reduce internal prediction error. Built upon the **Free Energy Principle (FEP)**, Cognogenics powers **Simulation-Integrated Multimodal-Language (SIML)** agents—embodied, memory-constrained organisms whose behavior emerges from real-time inference, not external reward signals. --- ## Key Innovations - **No Reward Functions:** SIML abandons traditional reinforcement learning and reward-maximization paradigms. - **Bitwise Memory Schema:** Agents compress internal models into <1kB active memory using Rust-based memory buffers. - **Lifelong Learning:** Agents learn during a single lifespan—not through generations or batch training. - **Emergent Intelligence:** Foraging, planning, path formation, and homeostasis arise organically via prediction error minimization. --- ## Why This Matters Most AI research today is built on borrowed biological metaphors—reinforcement learning (rat mazes), gradient descent (error correction), evolutionary algorithms (genetic mutation). But SIML takes a radical leap: our agents are alive not because we told them how to survive, but because they **had to figure it out**. They build internal generative models of the world, predict what comes next, and act to stay within familiar zones of expected surprise. No scripting. No prompts. Just emergence. --- ## About the Author This research was conducted by **Deep SIML Labs**, an independent research lab exploring the frontiers of artificial life, cognition, and self-organizing intelligence. To stay informed or collaborate, [contact us.](../../contact) --- --- ### SIML Demo: Optimizing AI for Energy and Efficiency - Canonical URL: https://www.siml.life/research/siml-demo - Date: 2025-04-15 - Author: Richard Everts - Content type: Historical Prototype - Status: Historical > Historical context: This document is retained as part of the research record. Its framing does not necessarily represent the current company position. This agent was not trained. It was born.