Discovery with Machines: A Field Guide to Research with AI
57 pages. 24 chapters. Four appendices. Written for the person using AI to do research alone, with a lot of leverage and almost no error correction.
PDF edition — typeset for reading and printing.
EPUB edition — for e-readers, phones and tablets.
Free, CC BY 4.0 — copy it, host it, print it, hand it to a student.
What is inside
- Why volume fails. 6,000 generated hypotheses, eight checked, eight already published.
- Search before you generate. The order of operations that saves the most time.
- Six shapes a testable claim can take — bound, count, identity, threshold, reversal, artefact.
- The counting rule. Identifiability is a question of counting before precision.
- Seven defects that survive peer review, with the checks that catch them.
- Techniques: uncertainty budgets, detection limits, test design, public datasets, reading papers suspiciously.
- Doing research without a laboratory — six access routes, what each really costs, and the three questions to ask before any booking.
- The money chapter most guides skip: services fund research, research keeps the services from being a commodity.
An excerpt
A completeness threshold was computed as bin_edge + 0.05. In binary that is not 4.4 but 4.400000000000001, so the selection m >= mc excluded every record sitting exactly at 4.4 — 2,060 of them, in the densest bin. With the bug: 6,351 records, exponent 0.980. Corrected: 8,411 records, exponent 1.252. A 28% difference in the headline number, from a comparison operator. Neither the data nor the estimator changed.
Support the research
Everything here is free and stays free. If the book saves you an afternoon, that is what the support links are for.