This is the next letter in Practical AI. The pin gave you the four questions. The OpenEvidence piece showed how to use a search without treating it like a consult. This one is the mechanism. If you get this wrong, every tool after it will feel smarter than it is.
I’m Travis Mendel. Radiation oncologist. I sign the note. I use these tools. Most of what I read about them comes from people who do not.
Plain version.
A large language model predicts the next piece of text from patterns in a huge amount of writing. That is the whole trick. It does not open PubMed. It does not fetch your guideline. It does not look at a database and return a row. It generates. Search finds documents. The model writes new sentences that sound like documents.
That distinction is the whole clinic problem.
Fluent is not the same as accurate. A model can invent a citation that looks real. It can misread a paper you actually pasted. It can fill a gap with confident language. Hallucination is not a glitch. It is fluent, confident, and wrong.
The fix is not a better model. The fix is a human who verifies the source.
A few terms you need, then we move.
Tokens are the chunks the model bills and counts. The context window is its working memory. If the note, the guideline, or the lab is not in that window, the model is guessing from training, not from your case. Paste what matters. Leave the identifiers out of consumer tools.
Training is school. The weights are frozen. There is a cutoff date. Inference is clinic. The model does not learn from your chat. Yesterday’s conversation is gone unless you put it back in the window.
Prompting is a consult request. Role. Context. Format. Ask for the source. Then verify. A vague “what do I do next” gets a vague answer. A hallway question with specialty, setting, and what you want back gets something you can check.
Nested words, in case someone throws them at you: AI is the broad field. Machine learning is how many of these systems learn from data. Deep learning is a thick stack of that. Generative AI makes new text, images, or code. An LLM is the text engine. An agent can take steps, not just answer. Useful does not mean autonomous.
Outside the clinic, the same rule holds for the agents I run on my own inbox and calendar. That lane stays off the chart. Clinic and life do not share data, tools, or memory. This letter is about the model under every tool you will meet at work.
Treat AI like a new resident. Fast. Fluent. Sometimes confidently wrong. Always supervised.
Input. Processing. Verification.
AI drafts. You decide.
The four questions before a tool touches a patient or a chart are here:
https://www.practicalaiforphysicians.com/p/automate-the-work-keep-the-judgment
More to come.