# AI answer review notes

Apply the LLM foundations course to one recent AI answer.
CodeBridge: https://codebridge-ai.com/en/courses/llm-foundations/

## 1. What kind of answer do you need?

- Question:
- Is the purpose factual information, current information, or creative work?
- Criteria for a useful answer:

## 2. What information was provided?

- References needed:
- References actually included in the input, and their dates:
- Missing or outdated information:
- Evidence that search actually ran:
- Actual result if a tool needed to run:

Adding a reference to the input is different from training the model's parameters.

## 3. What needs a separate check?

- Claims that need verification:
- Original or official sources to check:
- Does the answer match the source and its date?
- Is the source itself reliable?

A consistent or fluent answer does not guarantee factual accuracy.
Temperature affects the variety of token selections; it does not replace fact checking.

## 4. How would you check tokens?

- Model and tokenizer used:
- Expected differences if you change spacing or language:
- Measured token count (leave blank if you have not checked):
- Service documentation for length limits and cost:

Tokens do not always match words or characters. Token IDs do not rank meaning.
Check the service's current documentation for its prices and limits.

## One change to try first

- A reference to add or evidence to check in the next question:

Start with the input and supporting evidence, beyond how convincing an answer sounds.