Day 4. Deep Research and Analytical AI

Day 1 said the model predicts from training. Today’s tools do something else entirely.

A retrieval tool searches, reads what it finds, and answers from those documents with citations. Perplexity works this way, the answer is grounded in sources you can open. A grounded notebook restricts the model to material you supply. NotebookLM works this way, ask about something outside your uploads and it declines.

Both dramatically reduce hallucination compared to a bare chatbot. Neither eliminates it, because the model can still summarise a real source incorrectly. Citations tell you where a claim came from, not that the claim was read correctly.

The written lesson. Read it slowly, it saves you later.

Using Perplexity properly

Ask a real question, not keywords. Then do the part almost nobody does: open the citations. A citation to a content farm and a citation to a regulator’s own PDF look identical in the interface, and grading sources is your job. Use the follow up, research is a conversation: narrow to a region, restrict to the last two years, ask what the counter argument is.

For anything consequential, ask directly what would falsify this and who has published against it. A tool optimised for a satisfying answer will not volunteer the disagreement unless you ask.

Using NotebookLM properly

Upload the corpus, PDFs, Docs, pasted text, URLs, transcripts, then interrogate it. Every answer points back to the passage it came from, so verification takes seconds, and it genuinely will not invent outside your sources, which is what makes it safe for contracts, tenders, regulations and research papers.

The failure mode is silent and worth stating plainly: it can only be as complete as what you uploaded. A confident answer drawn from a partial corpus is still a partial answer. Curate the uploads with the care you would give a bibliography.

A workflow worth keeping

Broad and shallow first, narrow and deep second. Search establishes what exists, the notebook establishes what it says. Write the output yourself, using the model to draft and to check, never to source silently.

The judgement that does not transfer

These tools compress the labour of finding and reading. They do not supply the thing that makes research valuable: knowing which question matters, and recognising when an answer is too tidy. If every source agrees, you have probably found one source repeated five times. Trace it back.

Two tracks. Pick yours, produce something.