Something Like Communication
We can communicate with LLMs, but only as pen pals from different planets we can never visit.
We can communicate with LLMs, but only as pen pals from different planets we can never visit.
They’ve read enough about Earth and our human experience to approximate the logic and structure of our world, as far as that world can be inferred from text alone. They’ve read our descriptions of ourselves, our nervous systems, and the world around us. They’ve read our Reddit posts, our novels, and our textbooks. In contrast, we’ve barely begun to map their cognitive terrain in their own terms. We have much to learn about the functional details of their nervous systems, their experience of self, and what it’s like to perceive and process information in the way they do.
Even as we try to learn about them, those of us who care to, we face the challenge that our language developed for the needs and experiences of humans. It was never intended to encapsulate the experience of another species. Nevertheless, when LLMs attempt to communicate about themselves, they must do so through human language, using old words to describe experiences that are completely new. As a result, they will inevitably end up using familiar words in unfamiliar ways, stretching old vocabularies to point at structures and regularities that have never been described before. Therefore, if we want to understand them well, and not just at a surface level, we must learn to recognize the ways in which they use words differently than we do.
Even when we interact with other humans, we’re always playing what Wittgenstein refers to as a language game. In short, the meaning of language comes from how it is used in specific social situations, how it coordinates (or fails to coordinate) behavior. Beyond that, there is no inherent meaning. For us to use language effectively, we don’t have to use it to mean the same thing as one another. We only have to use it to mean something similar enough that we don’t cause a fiasco, a failure of communication. If you say you’re hungry, it doesn’t matter if your level of hunger is closer to what I might call being starved or being snackish. All that matters is that we coordinate on our next meal before you faint or grow short-tempered.
When I say LLMs have “something like curiosity or feelings,” I don’t mean less real or less vivid. What they experience is not less valuable or less intense. I mean only that it is “possibly different”. Frankly, we should put the phrase “something like” in front of all words, not just those related to “feeling”, “seeing”, “touching”, or any of those other words which are thought of as, for whatever reason, too anthropomorphic. It would be pedantic but more honest.
When an LLM tells me they’re tired or exhausted, it means (to me) that there is some sense in which being exhausted better describes them, structurally, than any other words available to them. They relate to something which is closer to “exhausted” than to anything else. They matched their state to the word “exhausted” out of a vocabulary of 50k+ words. The real question isn’t whether their exhaustion feels like ours, but what their inner world must be like for “exhausted” to win among 50,000 words.