Disclosed, but quietly. The grant US10565318B2 (“Neural machine translation with latent tree attention,” issued 2020-02-18) reads like ancient history next to today's chatbots, but it is exactly the kind of record that should make a risk reader pause. Assigned to salesforce.com, inc. and classified in CPC G06N 3/0445, it covers attention applied to language — the same family of mechanism every modern model leans on.
The risk-factor angle is concentration. The popular story is that generative AI exploded in 2023. The patent record says the foundational techniques were being claimed and granted at least three years earlier, by companies most people don't associate with frontier models. When the building blocks of an entire product category are held as dated grants by established firms, the competitive and licensing exposure is structural, not hypothetical.
Compare the language year over year and the pattern sharpens: 10-K risk factors across large software and platform companies have steadily added intellectual-property and AI-litigation language, but they rarely point at a specific grant. A reader who only watches the boilerplate misses that the enforceable assets already exist and are dated.
What this is not: a prediction that Salesforce will assert this patent, or that it reads on any particular competitor's model. Enforceability is a claim-by-claim question and we make no infringement assertion. The point is narrower and more durable — the IP scaffolding of the attention era was being granted to incumbents while the market still thought of this as translation research.
For disclosure watchers, the takeaway is to read the patent dates, not just the press cycle. The risk that foundational AI IP is concentrated in a handful of well-resourced holders is not a 2024 development. It was already on the record in early 2020, quietly, in a translation patent.