same wine, two labels, two scores

a few days ago on the show i asked a guest the question i'd been waiting all prep to ask. "so can you tell us about the two labels moment? like the one identical wine, two labels, wildly different scores."

she's a chemist turned founder with a master's in bioinformatics. years ago she was working at a custom crush wine facility, the kind of place where one winery makes wine for a lot of different brands.

one identical batch went into two bottles with two different labels. same liquid. and industry critics gave them wildly different scores.

when she dug in, she found the critics were less consistent than regular consumers. the experts were the noisy instrument.

that moment is the whole company. if you've ever wondered whether ai can predict taste preference, her answer is yes... but not by listening to what people say about taste.

words are a bad instrument for preference. a critic says "notes of blackberry and leather" and it sounds like data. it's poetry. her insight was that predictability lives in chemistry, not vocabulary.

think about the tasting wheel or the spider chart every food company uses. it captures about five variables. sweet, sour, bitter, that kind of thing. meanwhile more than eight hundred other compounds matter. and they're all connected. move one lever and the rest move with it.

so her model starts from first principles. it captures all the analytes in a product, every compound, and learns what people prefer from the chemistry itself. she says it's a method billion dollar companies can't easily copy, and the earlier attempts at sensory ai failed because they started from the words instead of the molecules.

here's the bigger idea that stuck with me. ai has eaten vision, sound and language. it can see, hear and talk. taste, smell and texture are the senses it mostly skipped.

and it skipped them for a reason. you can't scrape your way to chemistry. there's no internet full of flavor data to train on. you have to go measure it, product by product, in a lab. that's slow, and slow is exactly what makes it defensible.

she said the company "happened like boiling a frog." no grand launch. one wine puzzle, then another question, then a model, and one day she looked up and had a company.

her go to market is just as precise as her science. in us beverage she figures there are about twenty thousand potential customers. not millions. twenty thousand. so she doesn't do mass marketing, she does account based targeting... every account picked on purpose.

there's a lesson in that for every founder. when your market is small and specific, every touch matters. you don't need reach. you need to be the obvious name to the few people who decide.

it's like tuning a guitar by ear versus with a tuner. one musician says it sounds a little flat. the tuner tells you exactly how many cents. both are useful, but only one scales.

and the two labels story works on founders too. the same product with a different name, a different deck, a different first sentence, gets a different score from investors. her company exists because people trust the label more than the liquid.

long term i think the biggest ai companies of the next decade won't come from scraping more of the internet. they'll come from founders who go measure something the internet never captured.

i wrote about that kind of building, where your edge is data nobody else has, in the ai method.

so what does your industry measure with words... that someone could start measuring with chemistry?