What the Model Cannot Know
Watch the assistant answer with nothing but its memory.
Tidewater's assistant answers students with total confidence. Find out when it is wrong.
A language model writes one word at a time. At each step it scores every word it could write next and takes a likely one.1
Those scores come from the text it read in training, which stopped at a fixed date, the cutoff. It never looks anything up. When it has not read the answer, it still writes the most likely words, fluently and with confidence. That is a hallucination.2
Watch Tidewater Bootcamp's assistant answer a student with nothing but its training.
From the references
- 1. Hugging Face LLM Course, Chapter 1, How do Transformers work?: language models predict the next token. Hugging Face, Apache-2.0, commit 8b893f0. Read and rewritten in our words.
- 2. Generative AI for Beginners, Lesson 1, Introduction to generative AI and LLMs: limitations and hallucinations. Microsoft, MIT, commit 3a9d9ef. Read and rewritten in our words.
Written for this lesson
Unnumbered explanations, the simulation and every number it shows were written for this lesson.
No FAQ entries. The model has only what it learned in training.
Question When does the spring cohort start?
Answer The spring cohort starts
The spring cohort starts
The next word: the model scores every option- on68%, mostly from training
- in32%, mostly from training
A student asks when the spring cohort starts. There is no FAQ in the prompt, so the model has only what it read in training. It opens by restating the question.
When you have watched the answer form, open the live model.