25.0608 (0526,0512,0418) Doc URLs Stack URLs Lab notes (Gdrive) Git

Or (in my case) just landing a good AI job. :)

This video from Andrew NG ("ing") sums up AI opportunities. https://www.youtube.com/watch?v=KrRD7r7y7NY

But how to go from watching a video to actually doing the tech? You need to learn yourself. And you can with AI copilots.

This genius is doing incredible things without assistance (in his modest Berlin apartment). He sums up it up perfectly.

https://www.youtube.com/watch?v=x4F-jRK8Hoc

but how to connect the youtube hacking chaos with Andrew's big picture? where is the structured approach?

I found none. so I had to hack things out, figure them out. And I created this Wiki to organize it all. I spent the past 6 months building AI stacks (before that I spent 3 months building tech stacks (MERN), then 9 months building AI drones).

This wiki is now Your QS to taking the shortcut to navigating the AI jungle. see the next wiki page "Concepts" for details.

---

• A1 AI stack demos. Hands-on experience.
• A1b AI stack deployments. Create deployments based on demos.
• A2 AI stack docs. Generate end-user docs (not sure myself yet what kind of docs).

I started out building AI drones, then Tech stacks, and then the core focus settled on the topic of this section: AI stacks (LLM models, agents, tools, etc).

I am building this section as a logical organized framework that in the future I can

• understand and organize topics
• approach more in depth topics efficiently … exploration in a more methodical way to make sure I hae covered all the bases
• keep abreast of the latest changes… this is a living document

The challenge is that I have been learning from demos. Demos I found thanks in great part to youtube search algorithms, great demos, Anything I could get to run, with just me, Copilot and ChatGPT. But youtube authors do not usually organize things much conceptually.. so I’m guessing about how to org.

After a several months, its now time to inventory it all, make sense of it. rehash it. That’s a process I go thru often in projects I’ve worked on before, but in customer projects I have system experts to help and guide, feedback. things go quicker than this.

In any case, its an iterative process. in this iteration, the following is the structure that reflects my understanding. its prioritized by increasing (security) risks.

D0 (tech) Stack concepts is pretty simple (I stuck to MERN). But the AI stack gets a lot more complex.

Basic ideas of this section:

• LLM = brains
• Agents bring the deterministic workflow control
• Agents are many types and complex
  • usually only change API call to connect to model (model is just the "brains"?)
• UI is not so important, but put it in here because it will usually connect to agent
• RAG, tools, MCP, etc: Connection to the "real world" of digital tools
  • agent is probably tightly designed for the RAG, tools, ets
• A2A to other agents

An "AI stack" primarily means (these were separate wiki pages... just add them (just diagrams) to this page sometime)

• SECURITY??
• LLMs (concepts)
• Agents
• Agent-LLM connect
• UI (Gradio, Streamlit)
• Agent addons (RAG, MCP, etc)

MERN (left) and AI stack (right) (with Postman in the midddle).

• level 1
  • (left) React is the UI and workflow. Usually uses EN API, but can also implement the API itself and directly connect to DB.
  • (right) A0.1 UI (Gradio, Streamlit). Can directly connect to the model wrapper.
• level 2 Workflows
  • (left) EN usually has the API for access to the DB.
  • (right) A0.2 Agents. Agent usually runs the workflows.
• level 2b A0.2b Agent-LLM connect
• level 3 DB/Model
  • (left) DB.
  • (right) A0.3 LLMs. The base model (token generator) and the "wrapper" (all the logic that surrounds the base).

This is just my initial attempt at organizing the concepts.

See #498b_ai_deep_dive_KARPATHY_.docx for details (the docx is a very draft WIP).

Following diagram (some details may be wrong; still working on this) summarizes the

• P1 Base model: The main token generator (takes input tokenized text and outputs single tokens)
• P2 LLM model: Patch or “wrapper” around P1 that enables easier communication.
• P3 Agent: Binary program (python) that drives script/human interaction.
• P4 External: Various resources at disposal of agent.

---

• "1 AI stack diagram" below show core AI stack concepts.
• "2 MERN stack diagram" shows corresponding MERN stack concepts (comparing with MERN is my original idea, not sure it makes sense).
• "3 docx #499" shows the table of demo docx's in my #499 planning doc.

Note:

• MERN demos were complete; AI stack demos are not. all piecemeal. It will take a while to create a complete demo.
• AI stack demos do not include deployment. I need to figure this out.

#499_AI_API_deployments_devplan_.docx has a table that shows all the demo docx's and what parts they include.

---

I am trying to make the AI sections (A1-4) a conceptual mirror of the stack (S1-4) sections.

commments

• Maybe I am making this too complex, and i am an beginner with the AI stuff, but I instinctively try to find the parallels between MERN and AI.
• Seems to me that this concept is totally missing from the AI world.
• In general what makes the AI so confusing for me is this top level missing concepts.

Then again maybe i am running around in circles.. if so, this will become apparent.

Note

• For stacks,
  • S1 the biggest part of the demos was building the stack.
  • S2 the routing/state was setup while building stack.
  • S3 REST APIs were also built in demos while building stack
• but For AI,
  • A1 you dont build the model (too complex).
  • A2 chat/workflows are done separately
  • A3 as are RAG/tools/MCP (getting outside data)

NOTE FOR S1 below:

• i see the model as just a DB with some added AI functionality (like mongo).
• The AI agent manages the workflows (as does React).