Track: research and learning
This is the track with the sharpest trade-off in it. AI is extraordinary at getting you from nothing to competent-sounding in twenty minutes, and that is exactly the state in which you cannot tell whether what you just learned is correct.
Used well it is the best learning tool most of us have had. Used badly it manufactures confident beginners. The difference is one habit, and it is in stage 2.
---
Stage 1 — the orientation stage
You know nothing about something and need to know roughly what it is.
This is where AI beats everything else, including search, and it is not close. A search engine gives you documents written for people who already know the vocabulary. A model gives you the vocabulary.
What to try first.
- "Explain X to somebody who knows Y but has never encountered X."
- "What are the five words I need to know to read anything about this?"
- "What is the argument this field is actually having?"
- "What would somebody experienced assume I already know?"
Anchoring it to what you do know — "I understand databases, explain vector search" — is the single biggest improvement available here, and almost nobody does it.
What goes wrong. Stopping here. Twenty minutes of good explanation produces a strong feeling of understanding that is mostly the feeling of having followed something, and following is not knowing.
When you have got it. You come out with a map and a list of terms, and you know it is a map rather than the territory.
---
Stage 2 — the verification stage
The habit that makes the whole track work.
Ask for the specific claims, then check three of them.
Not "is this right" — it will say yes. Ask it to list the factual claims it just made, pick the three the rest depends on, and check those against a source that is not a language model. Documentation, a standard, a paper, a person.
Twenty minutes of orientation plus ten minutes of checking three claims is a genuinely different outcome from thirty minutes of orientation. You will find something wrong perhaps one time in four, and the thing you find is usually load-bearing — a version number, a rule with an exception, a name that belongs to something else.
What to try.
- "Which parts of that are you least confident about?" (Imperfect — it is guessing about its own guessing — but the answers cluster around the genuinely shaky bits often enough to be worth asking.)
- "What would I search for to check that?"
- "What is the primary source here — a spec, a paper, official docs?"
What goes wrong. Asking the same model whether it was right. It will agree with itself, and if you push back it will agree with you instead — which tells you nothing either way. That agreeableness is a real property of these tools and it is covered in what to be careful about.
When you have got it. Checking is automatic and you no longer feel it as an extra step.
---
Stage 3 — the grounded stage
You give it sources instead of asking it to remember.
What to try first. Paste the actual material in — the documentation, the paper, the contract, the transcript — and ask questions about that rather than about the world. The failure mode changes completely: instead of inventing facts it now summarises real ones, and the worst it can do is emphasise the wrong part.
Tools that search the web and cite what they used are the same idea with the fetching automated. They are markedly more reliable than an unaided answer and the citations still need opening — a real link is not the same as a link that says what the answer claims it says.
What goes wrong. Trusting a citation because it exists. Check that a couple of them actually support the sentence they are attached to; this is the single most common way a grounded answer is still wrong.
When you have got it. Your default for anything that matters is to supply the source rather than hope it remembers one.
---
Stage 4 — the learning stage
Using it to actually learn, not just to find out.
The difference: finding out is getting an answer, learning is being able to produce it next time.
What to try first.
- "Ask me questions about this until you find something I have not understood." Genuinely uncomfortable and genuinely effective.
- "Here is my explanation of X. What is wrong with it?" — the strongest single technique on this page. Explaining it badly and being corrected beats reading a good explanation.
- "Give me a problem I can only solve if I have understood this."
What goes wrong. Reading rather than producing. Nothing on this page works if you only ever consume the output.
When you have got it. You use it as somebody to explain things to, more than as somebody to be explained to by.
---
The tools
A chat interface with web search, for anything where recency matters. All the major ones have this now and it is worth turning on for research specifically.
A deep-research mode if the provider offers one. These spend several minutes reading many sources and produce something with real citations. Good for a survey of an unfamiliar area; still needs the citations opening.
Whatever holds your notes. Writing down what you learned, in your own words, is the step that turns orientation into knowledge — and it gives you something to give the model next time (see writing and thinking, stage 3).
---
Where it genuinely does not help
Anything recent. Training has a cut-off. Without search it will answer confidently about the state of things as of some date it will not volunteer.
Anything contested. It will give you a balanced summary of a live argument in a way that flattens which side has the evidence.
Anything where the details are the point. Legal, medical, tax, regulatory, safety. It knows the shape of the answer and the shape is not the answer. Use it to understand what you are reading and to work out what to ask a professional — not instead of one.
Niche or new things. The less written about a topic, the more it fills gaps with plausible invention, and the fewer people there are who could tell you it is wrong.
---
Next: Coding if that is your field, or back to all the tracks.