Why learn what AI is not - terrytaylorbonn/auxdrone GitHub Wiki

25.0627 Lab notes (Gdrive), Git

Contents of this page

1. AI tools (LLMs) simulate intelligence
2. AI "memorizes" (encodes) and computes probabilistic answers to your questions
3. The marketing hype

AI tools provide simulated intelligence that is incredibly valuable because it provides capabilities that humans do not have. I compare AI with an (all-powerful) librarian and a card catalog:

• (below left) Library front desk where you ask the librarian for assistance in extracting required info from books (unintelligent sources of answers to your questions)
• (below right) AI frontend (on browser) where you interact with an AI agent to extract info from the AI model (unintelligent source of answers to your questions)

The diagram below shows the model (the heart of AI). Note the following:

• P1: The base model is a binary token generator. It takes input and computes the most probable required output token (word). It then adds this word to the previous input, and this becomes the next input to generate the next token. This process is repeated until P2 stops it.
• P2: The base model "wrapper" (what I sometimes call the "nanny") is a bunch of glue logic that makes it easier to interact with the base model.

To summarize:

• AI LLM models do basically what I did to pass my Chinese drivers license test: "Memorize" the answers. The model is “trained” (weights are computed) on the answers, and can then answer the questions. That’s the core of an LLM.
• The LLM simply selects the most probable answer. Even if its garbage (the marketing term for a garbage answer is "hallucinations").
• Its not fully understood how AI works.

https://www.anthropic.com/research/a-mathematical-framework-for-transformer-circuits

Convincing people that they need the tech you create (and that they understand nothing about) is what any good tech businessman does.

Jensen Huang (Nvidia founder) recently said that kids shouldn't be learning to program. Having used Nvidia products (Jetson mini-PCs and laptop GPUs) and experienced the challenge of managing/installing Nvidia drivers), I consider good programming skills essential for using such AI tools.

So if AI (LLMs) really is an unintelligent source of info that could be compared to a librarian and card catalog, would you blindly trust AI with your life? On the previous page I stated that AI "when used correctly will raise your productivity". Its all about using it where it excels. Thats what this wiki is about.

The following is a Google AI explanation of AI "hallucinations". The word "hallucination" is a carefully chosen marketing term. AI runs on binary machines, and does not have any consciousness. So if you input something that does not compute, the "simulation" of intelligence breaks down. Fails. The simulation fails.

A more accurate description for this can be found in the old space travel movies of the 1950's. My favorite example is Robby the Robot's most beloved phrase "Does not compute, does not compute").