“I re-ran that same query this week, and actually found an existing patent awarded in 2018 — so the data would have been there. They may not be getting any smarter in the core effort around something like patents, but it may be that they’ve got better hooks in how to do queries.” — Scott Delinger

Session context: 2026-07-23_Mastermind — Scott picked a project back up after a break, decided to re-run his early groundwork “just in case the tools are better,” and found the thing that changed his plans.

Core Idea

In January, Scott ran a patent search across Gemini and Claude for the invention he was designing. Nothing came back. In July, he ran the same query and found a patent filed in 2016 and awarded in 2018 — data that had been sitting in Google Patents the entire time, indexed and public, on both occasions.

Two lessons, and the second is the one that generalizes.

The first is the practical one: a negative result from an AI research query has a shelf life. It is not “this doesn’t exist.” It is “the retrieval path available on that date didn’t reach it.” Anyone who has ever cleared a decision on the basis of “I checked and found nothing” is carrying an assumption with an expiry date on it — and the higher the stakes, the more that matters. Scott’s null result was load-bearing for six months of design work.

The second is the mechanism, and Scott named it precisely: the model probably didn’t get smarter about patents. The hooks got better. Better tool integrations, better query construction, more aggressive traversal of the source. Capability improved without intelligence improving. That is a different axis of progress than the one everyone tracks, and it moves on a different schedule — which means the useful question after a model release isn’t only “does it reason better,” it’s “what can it reach now that it couldn’t reach before.”

Lou’s observation supports the same read from the other end: with several research tools installed, Claude increasingly routes research through third-party APIs rather than its own web access — sometimes when he’d rather it didn’t. The retrieval layer is where much of the recent movement has been, and it is largely invisible in the benchmarks people quote at each other.

The follow-through in Scott’s story is worth copying too. He didn’t stop at the finding. He asked whether the patent had been filed internationally, learned that multinational filings surface as links in the same source, sent the model to check, and came back with a usable position: US-only, held by a company that has done nothing with it, unfiled elsewhere. The setback became a map.

Practical Application

Keep a “null results” file. Every time an AI research pass comes back empty on something that matters — a patent, a competitor, a prior art search, a citation you couldn’t source — log the query, the date, and the tool. Re-run the whole file quarterly; it takes fifteen minutes and you are not re-deriving anything. Before any decision that leans on an old “nothing found,” re-run that specific query first. And when a search does hit, don’t stop at the hit: ask what the next query is, the way Scott asked about international filings.

Evolution Across Sessions

Builds on Insight - Tools Define AI Capability More Than Model Intelligence, which established that tool access shapes what an AI can actually do more than raw model quality does. The new development is the temporal consequence: because the tooling layer improves independently of the model layer, your past AI research results decay, and negative results decay fastest. This introduces a practice — the periodic null-result re-run — and a caution for anyone who has treated an AI search as due diligence. Future sessions should watch whether retrieval-layer gains keep outpacing reasoning gains, since that would change where practitioners should be paying attention on each release.