Anthony on Nostr: > The core argument is that systems trained only on linguistic form cannot learn ...

> The core argument is that systems trained only on linguistic form cannot learn meaning, because meaning requires a connection to extralinguistic referents.
[The octopus test](https://aclanthology.org/2020.acl-main.463.pdf )---which really summarizes a vast literature on linguistics and language---supports this point of view. It's not some flippant thing Bender at al. are throwing out there just to be contentious.
> Even if we accept this premise
...
> it still does not apply to the systems anyone is talking about today.
What? Anyone who says LLMs have "capabilities", can "think", or "reason", or "code", or "do math" or "solve physics" or "write papers" or..........is talking about exactly this. What they go on to argue about grounding reads as missing the point to me. Training datasets about the world are not grounding in the world.
> On the contrary, some LLMs trained only on text can, in fact, tell you how to build a catapult
False. They can emit words in the shape of instructions for building a catapult, maybe, but only if something like that is in the training data. But this misses the point of the octopus test, which was that *catapult-like things were not in the available data*. An LLM cannot give you instructions to build something it's never encountered in training data. It might fake it a bit if there had been something similar in the training data, but it would almost surely fail to answer any nuanced question about the instructions. We've seen this time and again and it will always be the case. Arguing otherwise presupposes that human beings exist in an entirely closed-ended world that can be fully encased in a large enough but still finite space of natural language text plus whatever can be interpolated within that space. This assumption is plainly false, besides being nihilistic, bleak, and horrifying.

The "people don't believe stuff because it sounds sneering" is irrational tone policing, and is not serious. If your life were at stake, say, would you tell the emergency responder who's there to help you to leave because they spoke to you in a sneering way? Of course not. One tone polices about matters that don't matter.