Skip to content
The future of AI

From Thermodynamic Silicon to Decentralized Memory

Published by Zanno on 17 August 2026

Over the past few months, my R&D focus at Creepybits has crystallized around a singular architectural reality: the modern computing stack is colliding with hard physical and institutional limits.

Whether you look at artificial intelligence inference melting enterprise power grids, bloated context windows leaking confidential corporate data into centralized clouds, or legacy governance models failing under bureaucratic weight, the root cause is the same: centralized, brute-force architectures that ignore physical and structural incentives.

Instead of treating hardware optimization, cognitive memory, networking protocols, and political economy as isolated hobbies, I have spent this year formalizing them as the four modular tiers of a Unified Sovereign Stack:

┌────────────────────────────────────────────────────────┐
│  4. GOVERNANCE LAYER : Fiscal Direct Democracy         │
│     (Decoupling The Sword from The Purse / Pruning)    │
├────────────────────────────────────────────────────────┤
│  3. TRANSPORT LAYER  : iroh-socks-proxy                │
│     (Zero-NAT, Cryptographic P2P Wire in Rust)         │
├────────────────────────────────────────────────────────┤
│  2. COGNITIVE LAYER  : Associative Procedural Memory   │
│     (Air-Gapped HybridRAG + SQLite Knowledge Graphs)   │
├────────────────────────────────────────────────────────┤
│  1. COMPUTE LAYER    : SDSIE Runtime Engine            │
│     (Entropy-Routed Dynamic Precision & Hysteresis)    │
└────────────────────────────────────────────────────────┘

Here is where each project currently stands, how they fit together, and the roadmap ahead as we move from theoretical specifications into bare-metal hardware execution.

The Compute Layer: Software-Defined Stochastic Inference Engine (SDSIE)

Modern generative AI inference is fundamentally bottlenecked by the Memory Wall and thermodynamics. Running massive autoregressive models entirely in static, dense floating-point precision (FP16/FP32) is an immense thermodynamic waste: moving weights across memory buses consumes up to 80% of total inference energy, while every filler word and punctuation mark is processed with the exact same computational intensity as complex logical reasoning.

Earlier attempts to deploy Stochastic Computing (SC) in software collapsed under the “Emulation Tax”—where sequentially simulating random bitstreams and population counts on conventional CPUs/GPUs introduced a ~2,000x latency penalty.

To solve this, I designed SDSIE: Software-Defined Stochastic Inference Engine (Read the Paper on Zenodo).

Instead of simulating non-existent physical laws, SDSIE operates as a Hardware-Aware Adaptive Virtual Runtime (AVR)—an automated “bicycle transmission” for AI inference:

  • The Resolution Gear: Dynamically downshifts execution into sub-byte integer Tensor Core kernels (SageAttention / INT4 / FP4 microscaling) with per-thread outlier smoothing, reserving heavy FP16 pipelines only for high-uncertainty tokens.
  • The Temporal Gear: Compresses compute in spatial-temporal workloads (Diffusion Transformers / DiTs) via input-modulated L1 distance and Jacobian sensitivity caching (TeaCache / SenCache), skipping up to 50% of redundant forward passes.
  • The Computational Clutch: Utilizes real-time Shannon entropy monitoring (Ht) stabilized by a Schmitt-trigger hysteresis buffer (θhigh,θlow,h), permanently preventing GPU memory stalls and “gear hunting” during continuous batching.

    Current Milestone: The mathematical blueprint is published. Dedicated next-gen test hardware is currently arriving at the lab to begin executing bare-metal empirical benchmarks on CUDA/Blackwell architectures, logging exact Joules/tokenand throughput deltas against standard vLLM baselines.

The Cognitive Layer: Artificial Associative & Procedural Memory

Standard LLM agents are plagued by catastrophic forgetting, hallucinations, and context rot. The prevailing industry “solution” has been brute-forcing million-token context windows or deploying flat vector databases (RAG)—both of which introduce latency, memory bloat, and massive cloud privacy liabilities.

To give autonomous agents persistent intelligence without cloud dependency, I engineered a local Associative & Procedural Cognitive Memory Architecture (Read the Deep-Dive):

  • Procedural Memory via Hierarchical Task Networks (HTNs): Instead of storing flat text snippets, the system decomposes workflows into deterministic causal schemas. The AI remembers how a task is executed, what contingencies exist, and how outcomes are verified.
  • Associative Multi-Hop Knowledge Graphs: Combining an embedded vector store with an explicit SQLite relation graph, the runtime supports multi-hop traversal with temporal decay and reinforcement modeling.
  • Zero-Cloud Air-Gapping: The entire cognitive persistence pipeline executes natively within local environments (WSL2/Linux), making it completely impervious to SaaS surveillance and fully GDPR-compliant for confidential enterprise operations.

The Transport Layer: iroh-socks-proxy

Autonomous agents cannot coordinate if their communication layer is bottlenecked by centralized cloud servers, API rate limits, or restrictive ISP firewalls (CGNAT).

To secure sovereign communication at the wire level, I built iroh-socks-proxy (GitHub Repository | Read the Spec):

  • Bypassing Browser & Extension Bottlenecks: Decouples the network stack from Google’s Manifest V3 restrictions by running a lightweight browser interface hooked into a high-performance native background daemon written in Rust (Tokio).
  • Direct P2P QUIC Hole-Punching: Leverages the Iroh protocol to establish direct, authenticated, end-to-end encrypted peer-to-peer tunnels between independent nodes without requiring port forwarding, static public IPs, or centralized relay chokeholds.
  • Decentralized Agent-to-Agent Mesh: Provides the encrypted data plane required for distributed AI agents to synchronize memory states and execute cross-model verification across consumer and edge hardware.

The Macro Governance Layer: Fiscal Direct Democracy

A sustainable computing and technological infrastructure cannot survive on top of an economically captured institutional framework.

To address systemic rent-seeking and the “Principal-Agent Problem” in modern governance, I authored a comprehensive 57-page political economy paper: Fiscal Direct Democracy: A Proposal For Revitalizing Our Democracy (Read the Paper on Zenodo):

  • Separating the Sword from the Purse: Strips the legislative body of its discretionary budget mandate, reducing statutory politics to general, non-excludable law while returning capital allocation power directly to taxpayers via a portfolio-based system (adapting Sweden’s Premium Pension/PPM mechanics to the sovereign budget).
  • Algorithmic Sovereign Pruning: Introduces strict minimum-viability thresholds (Sthreshold) that automatically prune unviable or underfunded state ventures, reallocating capital to capped rainy-day reserves and permanently collapsing corporate lobbying.
  • Two-Step Qualification Engine: Re-anchors voting franchise in objective contribution (3×PBB) lifetime participation or military service) coupled with a dual-flanked epistocracy exam testing awareness of both State Failures and Market Failures with equal weight.

Strategic Collaborations, Enterprise Pilots & Financing

As these projects transition from open architectural specifications into empirical hardware validation and prototype deployments, Creepybits is actively opening discussions for strategic partnerships, pilot testbeds, and deep-tech backing:

  1. Enterprise On-Premise Pilots: We are looking to collaborate with organizations, legal practices, and institutions requiring air-gapped, local associative memory pipelines where client privacy, confidential data residency, and GDPR compliance are mandatory.
  2. Hardware & Datacenter Benchmarking: We welcome partnerships with infrastructure providers and GPU cluster operators interested in co-validating SDSIE thermodynamic energy-reduction telemetry on enterprise silicon (H100 / Blackwell B200).
  3. Strategic Deep-Tech Capital: We are opening exploratory dialogues with deep-tech angel investors, multi-family offices, and grant consortiums aligned with high-efficiency inference runtimes, sovereign infrastructure, and next-generation cognitive architecture.

If your organization is interested in testing these runtimes, collaborating on hardware benchmarking, or opening an investment dialogue:

📩 Direct Contact: business@zanno.se
🐙 GitHub: github.com/Creepybits

Sign up for our newsletter and follow our journey from theoretical specifications to bare-metal hardware benchmarks and autonomous agent systems.

Published inAIEnglishLocal Agentic AISoftwareTechWeb