Consume Product

All it does is use words to predict the next words. That's it.

The rest is just a net of checks, triggers, and rules. For instance, when it generates an image, it is because threshold of word associations having to do with image generation is crossed, which automatically triggers a separate model to create an image. The LLM itself only predicted plausible word associations and plugged those into a prompt of its own for another tool to use, then it served the results. It doesnt "know how to generate an image". It can't "look at an image". Everything it does in excess of "guessing the next word" is handled this way.

It accurately hallucinates which words come next based on the strength of the relationships between instances of those words from its training data. A layer that reads its output will trigger other tools such as ocr image readers, search engine macros, etc as the need for it is detected via word association. The model accredited with intelligence, that generates words in your language, does ONLY that.

Thank you for attending my Tod Talk.