2 AI (models)
Concepts
A “model” in this section is anything that has a neural network (NN) at its core. It may a lot of other stuff wrapped around it (for CNNs convolution, for LLMs Attention heads, etc) but the core pattern matching and statistical probability runs on the NN. The differences are mainly architecture, training, orchestration, and use case. All models are first (1) trained (internal NN params are SW programmed) and then (2) used to infer output from input.
Models ARE MECHANICAL. They run on clocked binary circuits. They have no intelligence. You could run even the most sophisticated LLMs on electro-mechanical relays (it would take years to generate a token, but it theoretically possible). AI intelligence simulation is based on (1) binary computing structures and (2) massive computing power/speed. The more you understand how AI models work, the better you can use them in AI projects.
Other sources about models (wiki pages):
“2 Models” in the Agentic AI evolution (diagram)
2.0 NN UFA concepts
The core of every model is the NN that implements a UFA (Univeral Function Approximator) that generates an output based on pattern matching of the input. The patterns were programmed (“trained”) by setting NN parameters (weights and biases) during training.
3 groups of models:
- 2.1 core (no convolution/tf)
- 2.2 convolution
- 2.3 tf
3 phases of model demos for each model group:
- P1 DIY (doit yourself) models (phase1) understand mechanics
- P2 OTS (offtheshelf) models (phase2)
- P2 OTS fine-tuned (phase3)
2.1 Core NNs
Predictive NNs are small customized NNs for specific recognition tasks (if you create a DIY model, the first one will be something like a simple predictive NN).
2.2 CNNs (convolution)
A CNN computes the most probable label (“dog”, “airplane”, etc) for a set of pixels. Studying CNNs is great preparation for studying more complex LLMs.
2.3 LLMs (transformers) (iAgent + TF)
An LLM (1) inputs a sequence of tokens, (2) computes the probabilities of all vocabulary tokens as the next token, (3) outputs the selected new token, and finally (4) adds the new token to the next input sequence. An LLM consists of an (1) internal agent and a (2) transformer (TF) NN. The TF is the core computational engine behind modern AI systems and the mechanism that enables modern LLM behavior.
Robotic AI = future topic
26.0526: This section does not really belong on this “2 Models” page, because this is not a different model type. I originally put it here because I believed LeCun’s hype about his version of JEPA that would be a fundamentally different type of AI platform that would fix the fundamental limitations of LLMs (it doesnt, and never will). I might move this section somewhere else in the future. Right now this does not interest me too much… robotic AI (that claims to be able to do what humans can do) is about as realistic as FSD or colonizing Mars.
A Waymo “self-drives” into a flooded street.
2.5 Robotic AI
I spent a couple of months doing demos for basic robotic AI. I was intrigued by LeCun’s claims about JEPA and his new venture that would basically go beyond the LLMs that were already becoming obsolete. After a few demos and chats with GPT, I felt like I had been scammed.
Drone AI is something that works, because drones operate in very forgiving environments (in the air, far from other objects; actually they dont even need AI to fly, they need it for object recognition, terrain guidance, etc). But self-driving cars (robots) are another story. Its been over a decade of (empty) promises that FSD was just a year away. Its simply too dangerous for unintelligent robots to operate on complex and congested roadways.
In any case, its inevitable that AI will pair up with humanoid robots in the home and workplace. Just like with agentic AI, there will be applications where the lack of any real intelligence can be tolerated.
26.0526