He says LLMs generate new query code each time and do not keep what they learned, so rephrased questions get different answers. Listen
Mike said systems like Claude generate new code every time they run a query without pulling results back into memory to learn from them. He said a slightly different wording can produce a different answer, delivered with certainty. He presented this as a reason the output cannot be trusted without checks.
“It's not pulling anything back into memory and learning from it”