My first real project was EDI at Heineken, in 1995 or 1996 — I’d have to check which. Electronic data interchange: moving orders and invoices between companies that had never agreed on what anything was called. What I ended up building, we’d now call a semantic database. We didn’t have the phrase for it at the time.
Open a JSON payload today and the meaning looks like it comes free with the format. It doesn’t.
Take the word we dealt with constantly. Order. WordNet lists twenty-four senses of it, fifteen as a noun and nine as a verb. A command. A sequence. A state of things. A purchase you’d like delivered on Thursday. Same five letters. In a message from a brewery to a distributor, exactly one of them is right — so which?
What tells you is context. Where the field sits, what came before it, who is on the other end of the line. Sequence and context turn a string into a meaning, and back then we encoded that relationship ourselves, one mapping at a time, in meetings. It took six to nine months. I couldn’t tell you now which standard we were working to, which tells you something about where the difficulty actually sat — the standard was the easy half.
Which is the part I keep coming back to now. It is the same problem. Attention across a context window is doing what we did with lookup tables and a great deal of arguing, except it does it statistically, at a scale we could not have imagined, and it works on language nobody standardized in advance.
So no, I don’t think AI showed up out of nowhere. It is the oldest problem in data integration, finally being solved by something other than a person with a spreadsheet and an opinion.