57 AI: Learning with an AI
The shift from using the assistant to produce code to using it to learn.
Back in week one, this course mentioned a second mode the assistant can work in, the Study and Learn agent, and asked you to file it away for later. Later is now. Everything so far has been about the assistant as a producer: of code, of explanations, of drafts you read and test. This week the assistant’s job changes to helping you learn something, biology as much as programming, and that shift needs its own version of the one habit you have been building all along.
A tutor, not an answer machine
Ask the ordinary assistant a factual question and, true to form, it predicts a plausible, fluent, confident answer and hands it straight to you. The Study and Learn agent is built to resist doing that. Ask it something and, more often than not, it answers with a question of its own, or a smaller example, or a hint about where to look, nudging you toward working the answer out rather than receiving it. This is not the assistant being unhelpfully coy. It is a deliberate design, because an answer you were handed and an answer you derived leave you in different places: one is a fact you are holding onto because a machine said it, the other is understanding you can rebuild yourself. The whole reason to reach for this mode is that the second kind is what you actually came here for.
The verification principle, applied to facts
You already have the habit for code: a claim about what a function does is plausible until a test or the running machine confirms it. Facts, including biological ones, work the same way, and the parallel is worth stating in full, because it is not a metaphor, it is the same principle wearing different clothes. A test is to a function’s correctness what evidence is to a fact’s truth. Code you did not write and did not test is plausible, not correct. A biological claim you read, from the assistant or anywhere else, and did not check against evidence or your own knowledge is plausible, not known. The assistant states biological facts with exactly the same fluent confidence whether they are true or false, for exactly the same reason it can hand you a wrong function with total confidence: it is predicting a plausible continuation, not consulting a source and reporting what it found.
Try this concretely. Ask the assistant, in its ordinary mode, why the third position of a codon is often called the “wobble” position. You will get a fluent, plausible-sounding paragraph. Do not file it away yet. Find it in a source outside the assistant, a textbook or a reputable reference, and check whether the explanation the assistant gave, base pairing flexibility at that position of the anticodon, matches what the source actually says, and whether any specific claim, which bases wobble with which, was stated more narrowly, more loosely, or differently than the assistant put it. That checking step is not optional extra credit. It is the entire difference between a fact you were told and a fact you know.
What to do when the tutor mode still just tells you
The Study and Learn agent nudges rather than answers, but it is still the same underlying model, still predicting plausible text, and nothing about tutor mode makes it incapable of confidently stating something false while it nudges you. A hint can be a wrong hint. The verification habit does not relax just because the assistant is behaving more like a teacher; if anything it matters more, because a wrong hint delivered patiently is more persuasive than a wrong answer delivered bluntly. Whatever you derive with its help, you still check the result against a real source or against running the code, exactly as before.
Exercise 57-1
Open the Study and Learn agent and pick one thing from this course you have not fully settled in your own mind, a concept from the Python side or a piece of biology. Note the badge: you are not asking it to explain the thing to you, which is what week one would have had you do. You are working on it together. Tell it what you already understand, tell it exactly where your understanding runs out, and let it nudge from there. Write down each hint it gave you and what you did with it. Then verify the understanding you arrived at against a source that is not the assistant itself: a textbook, a documented reference, or, where it applies, a small program you write and run.
Exercise 57-2
Now close the assistant. Tomorrow, with nothing open in front of you, write the same explanation out again from memory in your own words, and only then compare it with what you wrote yesterday. The second version is the one that counts, and the gap between the two is the honest measure of whether you learned the thing or merely watched it being explained well. Keep both versions. In Chapter 62 you will do exactly this under exam conditions, and it will go a great deal better if this is not the first time you have tried it.
For your logbook this week, describe one fact, biological or otherwise, that you checked against a real source this week rather than accepting from the assistant, and say whether the source confirmed it, corrected it, or left it more uncertain than either you or the assistant first assumed.