Salesforce is dialing back LLMs, leaning on deterministic automation and guardrails after reliability gaps. The aim: predictable outcomes, clean data, and no missed steps.
I think it’s less of “marketing speak” and more of a general ignorance on what language is supposed to be, leaking into the marketing team. [With then me being the old man angrily screaming at the cloud, I guess.]
If you ask most people what language is, they’re talking about words. A few will talk about grammar (morphology and syntax), some about the raw material (sounds, characters, gestures). But the important part of a language is none of those. It’s meaning: we speak because we want to say something, everything else is just prep work.
And those models are really good emulating some layers of the language, like the morphology and syntax. But when it comes to meaning, they consistently shit bricks. It’s like they have the most visible part of the language, but what matters the most — just like a taxidermised critter.
So let’s call a duck a duck. If what they do the best is to identify and chain morphemes, they’re morpheme models. And since “morphemes” are called “tokens” in the jargon, “token models”.
I think it’s less of “marketing speak” and more of a general ignorance on what language is supposed to be, leaking into the marketing team. [With then me being the old man angrily screaming at the cloud, I guess.]
If you ask most people what language is, they’re talking about words. A few will talk about grammar (morphology and syntax), some about the raw material (sounds, characters, gestures). But the important part of a language is none of those. It’s meaning: we speak because we want to say something, everything else is just prep work.
And those models are really good emulating some layers of the language, like the morphology and syntax. But when it comes to meaning, they consistently shit bricks. It’s like they have the most visible part of the language, but what matters the most — just like a taxidermised critter.
So let’s call a duck a duck. If what they do the best is to identify and chain morphemes, they’re morpheme models. And since “morphemes” are called “tokens” in the jargon, “token models”.