I have learned to take these kinds of papers with a grain of salt, though. They often rest on carefully selected examples that make the behavior seem much more consistent and reliable than it is. For example, the famous "king - man + woman = queen" example from Word2Vec is in some ways more misleading than helpful, because while it worked fine for that case it doesn't necessarily work nearly so well for [emperor, man, woman, empress] or [husband, man, woman, wife].

You get a similar thing with convolutional neural networks. Sometimes they automatically learn image features in a way that yields hidden layers that easy and intuitive to interpret. But not every time. A lot of the time you get a seemingly random garble that belies any parsimonious interpretation.

This Anthropic paper is at least kind enough to acknowledge this fact when they poke at the level of representation sharing and find that, according to their metrics, peak feature-sharing among languages is only about 30% for English and French, two languages that are very closely aligned. Also note that this was done using two cherry-picked languages and a training set that was generated by starting with an English language corpus and then translating it using a different language model. It's entirely plausible that the level of feature-sharing would not be nearly so great if they had used human-generated translations. (edit: Or a more realistic training corpus that doesn't entirely consist of matched translations of very short snippets of text.)

Just to throw even more cold water on it, this also doesn't necessarily mean that the models are building a true semantic model and not just finding correlations upon which humans impose semantic interpretations. This general kind of behavior when training models on cross-lingual corpora generated using direct translations was first observed in the 1990s, and the model in question was singular value decomposition.

I feel like LLMs have a bit of the Clever Hans effect. It takes a lot of my cues as to what it thinks I want it to say or opinion it thinks I want it to have.

Clever Hans was a horse who people thought could do maths by tapping his hoof. But actually he was just reading the body language of the person asking the question. Noticing them tense up as he got to the right number of stamps and stopping - still pretty smart for a horse, but the human was still doing the maths!

I'd direct you to the 3 blue 1 brown presentation on this topic, but in a nutshell the semantic space for an embedding can become much richer than the initial token mapping due to previous context.. but only during the course of predicting the next token.

Once that's done, all rich nuance achieved during the last token-prediction step is lost, and then rebuilt from scratch again on the next token-prediction step (oftentimes taking a new direction due to the new token, and often more powerfully any changes at the tail of the context window such as lost tokens, messages, re-arrangement due to summarizing, etc).

So if you say "red ball" somewhere in the context window, then during each prediction step that will expand into a semantic embedding that neither matches "red" nor "ball", but that richer information will not be "remembered" between steps, but rebuilt from scratch every time.

There's a certain one-to-oneness between tokens and embeddings. A token expands into a large amount of state, and processing happens on that state and nothing else.

The point is that there isn't any additional state or reasoning. You have a bunch of things equivalent to tokens, and the only trained operations deal with sequences of those things. Calling them "tokens" is a reasonable linguistic choice, since the exact representation of a token isn't core to the argument being made.