Atlas/Research
Research
Below are some of the research questions the atlas plans to work on next.
What is being worked on
Second-season survival
Which pilots return for a second season. The register records 11 of 299 cases as closed, acquired or wound down — either AI in wine almost never fails, or the failures are not being written down.
The cost of the gap
What the step from pilot to commercial operation costs. The register holds 54 cases on one side of that step and 129 on the other; the price of crossing it is what the sources leave out.
Does knowledge predict use
Whether a benchmark score on wine knowledge says anything about a model's usefulness in a cellar. On OenoBench the drop on questions that need contextual reasoning rather than recall is significant in 16 of 16 configurations, and the cellar's questions are almost all contextual.
If you hold that data
Each of those questions turns on records that never reach publication: spray logs, season data, tank measurements, what a pilot cost and what it saved. Estates, cooperatives, laboratories, vendors and research groups hold them; nobody else does. If that is you, write.
Proposals of your own for research or projects on AI in wine are just as welcome — through the form below, or by writing to nikitahudov@gmail.com.
Until there is something to publish, the measured work is in the Casebook and on the OenoBench leaderboard.