AI Systems · 9 min read
Why AI shouldn't just answer questions
Building research systems that investigate, connect evidence, and generate original insights instead of simply retrieving information.
Retrieval is not research. A system that fetches the five most similar chunks and paraphrases them is a very expensive search box.
Hypotheses over summaries
Research begins when a system can state something that might be false, then go looking for the evidence that would break it.
claim -> "Domain X is consolidating around vendor Y"
falsifier -> new entrants funded in the last 4 quarters
verdict -> hold / revise / discard
Structure that actually helps
- Planner decomposes a question into falsifiable sub-claims.
- Gatherers run in parallel, each restricted to one source class.
- Adjudicator scores evidence and, crucially, is allowed to return insufficient evidence.
The adjudicator is where most systems cheat. If a model is never permitted to say "I don't know", every answer it gives is worth less.
Cost of being wrong
Design the loop so a wrong conclusion is cheap and visible: cite everything, version every verdict, and re-run old claims against new evidence on a schedule. Research is a standing process, not a one-shot query.