/ Digital_Garden

A chronicle of pushing an 8 GB NVIDIA Quadro RTX 4000 with FreeToken and Qwen3.6-35B-A3B: patching Turing kernels, quant benchmarks, harness traps, and the final pragmatic choice.

How burning through 45 Antigravity accounts, discovering Ollama Cloud, cloning a rotator with an AI agent, and watching 80% overlap led to redesigning everything so the provider is just a plugin.

I built SLMarena to stop reviewing local-model responses one by one. The result was a practical benchmark that found a winner under my conditions: Gemma 4 E2B.

After seven iterations of fine-tuning Qwen3-4B for Home Assistant, safe rejection collapsed. The empirical lesson on why security and autonomy belong in a deterministic supervisor, not in LLM weights.

Training a local smart home specialist without teaching it fake JSON. Native tool calling redesign on Qwen3-4B, DDP on dual Tesla T4s, 4096 context budget, and the empirical roadmap toward 0.6B micro-models on 8GB GPUs.

A real-world GGUF quantization experiment (Q4 vs Q5 vs Q6) using Qwen3-8B for Home Assistant on an RTX 4000: why more bits don't win when crossing the threshold from VRAM to CPU.

Migrating Qwen3-8B QLoRA training from Colab's opaque quotas to Distributed Data Parallel (DDP) across two Tesla T4 GPUs on Kaggle.

A technical reflection on the economic divide in local AI. Navigating VRAM bottlenecks on budget GPUs like the Quadro RTX 4000, extreme compression (1-bit, TurboQuant, MoE), and open-source community power vs SaaS hyper-capitalism.

A technical and geopolitical analysis of the 'Open Weights and American AI Leadership' manifesto. Corporate hypocrisy from NVIDIA and Anthropic, cyberdefense with local models like GLM-5.2, and regulatory foolishness.

Speculative journalism: a fictional scenario about Xuanwu Intelligence and a hypothetical geopolitical response to an AI security incident, followed by a separate section on linked real-world events.

How I diagnosed and fixed the HTTP 400 'Unable to generate parser' error when connecting Bonsai 27B to Hermes Agent's MCP tool schemas via llama.cpp, using a simple proxy sanitizer.

How to squeeze an NVIDIA Quadro RTX 4000 8GB VRAM to run 27B models at over 27 tok/s using native 1-bit binary and ternary quantization in llama.cpp.

How to inject offline reinforcement learning into a Java-based GameServer's main tick loop without compromising performance.

Discover why using your National ID as your WiFi password exposes you to hacks, and learn how to secure your home or business network.

How to decouple XML generation, digital signing, and SOAP transport to expose Ecuadorian invoicing as native LLM tools.

How to structure a private AI environment using OpenClaw, Ollama, and LiteLLM to avoid depending on external APIs like OpenAI.

How to configure a fully local ecosystem to run open-weights language models and agent frameworks like Hermes-Agent without third-party API dependencies.

How to orchestrate tenant-specific MCP servers from WhatsApp to access accounting, payments, and business automation.

How to turn scattered routers into an operable fleet with centralized control, synchronized state, and remote deployments.

Architecture notes on model-based rotation, observability, and risk control in a proxy for Google Antigravity.

Insights on saving preferences, corrections, and project patterns so future agent sessions don't start from scratch.

How to turn an AI proxy into a spending control layer and prevent model usage from breaking the budget.

How to use WhatsApp as a conversational interface to operate a local agent without fighting a new UI.

Basic reverse engineering and CAN frame interception to extract cell-level metrics from lithium batteries without proprietary software.

Best practices for deploying LXC containers, managing storage pools with ZFS, and isolating services on local server infrastructure.

How to implement an asynchronous ZFS snapshot replication strategy to ensure homelab data survival.

How to tame Kali Linux and KDE Plasma to use it as your main operating system without breaking security packages.