AI & LLMs

3 min read

AI & LLMs: Intelligence Augmentation

The Instructor's Perspective

I am an avid supporter of AI and education. We need to understand the power of this new tool that humanity has created. However, remember that every tool can also be used as a weapon in the hands of the wrong people. In our labs, we use LLMs as “force multipliers” for learning and coding, but we never follow them blindly. Maintain your “signal discipline”—verify everything.

The Story: How we got here

Before you dive into the technical details, read The AI Homelab Story. It’s the cohesive narrative of our move from a “manageable mess” of manual builds to a high-performance, Local-First Agentic Framework.

Why use it?

  • Synthesis: Summarize large buckets of data and find hidden synergies.
  • Force Multiplier: Speed up coding and configuration tasks with Aider and Qwen 2.5 Coder.
  • Agentic Framework: Use OpenClaw to give AI “hands” via local tools.
  • Educational Peer: Use it to “rubber duck” ideas and ask the “stupid questions.”

Knowledge Gained: The Containerized SYCL Victory (Updated April 8, 2026)

For a long time, we were battling broken symlinks and failing services (ollama.service, llama-serve.service).

The Breakthrough: We’ve shifted entirely to a Containerized SYCL (ava-agentone) stack using the ghcr.io/ava-agentone/ollama-intel image.

  • Tactical Correction: The original intelanalytics image was archived in early 2026. Switching to the community-maintained ava-agentone build restored GPU acceleration and stability.
  • Single Source of Truth: All local AI tools now point to Port 11434 (Ollama Docker). The legacy llama-server (Port 8081) and host-level systemd services have been decommissioned.
  • Outcome: By aligning the volume mapping and explicitly passing DRM nodes (/dev/dri/card1 and renderD128), our Primary (P) local AI stack is finally “in the fight.”

AI Reliability (The PACE Plan)

Information Discipline

P (Primary): Local Ollama Docker Instance (gemma3:12b, qwen2.5-coder:14b, deepseek-r1:14b) with Intel Arc (SYCL) acceleration. A (Alternate): Gemini (Cloud) for complex synthesis and cross-vault strategy via OpenClaw. C (Contingency): Secondary local nodes (e.g., Sparse Page) or manual llama.cpp builds. E (Emergency): Human expertise and physical reference books.

SOP: AI Delegation & Usage

  1. Delegate Routine Tasks: Use the Primary (P) local AI for coding assistance, summarization, and formatting.
  2. Orchestration: OpenClaw acts as the Command and Control (C2) layer, linking your brains to your tools.
  3. Never Trust, Always Verify: Treat every AI output as a “draft” that needs human review.
  4. Protect Intel: Never feed sensitive passwords or private data into external AI.

LLM Knowledge Tree

TABLE join(sort(rows.file.link), " | ") as Documents
FLATTEN regexreplace(file.folder, ".*/", "") as lastPart
WHERE contains(file.folder, this.file.folder) AND file.name != this.file.name
GROUP BY lastPart as Folder
SORT lastPart

Recent Changes

TABLE 
  file.ctime AS "Created", 
  file.mtime AS "Updated"
WHERE (file.cday >= date(today) - dur(7 days) OR file.mday >= date(today) - dur(7 days))
  AND contains(file.folder, this.file.folder) AND file.name != this.file.name
SORT file.mtime DESC
LIMIT 10

Check for Understanding

  • Why is it important to have a Local LLM (Primary) instead of just relying on cloud services (Alternate)?
  • How can an AI be used as a “weapon” if not handled with proper “signal discipline”?

Related: AI System Prompt, Shell, GIT