← Home


The diagram below (v7) summarizes ZAI evolution:

  • 1 AI drones were the first project. The AI was CNNs (object recognition).
  • 2 AI models:
    • 2.1 Core NNs were next. Simple input/output “bare bones” NNs that you can train and run locally.
    • 2.2 CNNs (convoluted NNs for object recognition). The main AI used on drones.
    • 2.3 LLMs (with internal agent and transformer (TF)). The TF performs a similar process to convolution using “attention heads”, which basically “convolute” tokens related by meaning.
    • 2.5 Robotic AI (includes LLMs and JEPA). I worked on JEPA for a while, and was not impressed. The idea of being anywhere near a robot that is in any way controlled by JEPA or an LLM is scary (just like FSD in cars).
  • 3 Agents:
    • 3.1 Agent demos (non-AI) are classical control loops (written in Python, etc). They might use a model, but not for any control inputs.
    • 3.2 AI agent demos are basic demos that use AI for control input. This is tricky because (1) you can’t predict 100% LLM responses and (2) the control needs predictable responses.
    • 3.3 AI agent projects are meant to be working prototype projects (not simple demos).


AI evolution diagram



Previous versions of this diagram.

v6

AI evolution diagram

v5

AI evolution diagram

v3

AI evolution diagram

v2

AI evolution diagram

v1

AI evolution diagram


26.0527 (v1 26.0523)