5 ms·
LLMs as clearly doing more than just modeling words. In order to predict word placement, they need to build some kind of model, their latent space, of the types
by jacobr1 3mo ago
LLMs as clearly doing more than just modeling words. In order to predict word placement, they need to build some kind of model, their latent space, of the types of things they are able to predict. Not really full world models yet - but they have decent "blog space" or "github project" models. And you can see this with multimodality, or non-text modality modals such as images and audio. They map from their latent spaces to the outputs. The fact that multi-modal systems can share the interior layers shows some kind of internal representation is created.
- adamzwasserman 3mo agoOr.. the LLMs are just "telescope" sufficiently powerful to detect the aligned internal representation that already exists. A more powerful telescope does not "create" new galaxies.
- jvanderbot 3mo agoLove the analogy. I a lot of people had their mind blown by vector space representation of words. The idea that Female + King averaged out around "Queen". Words might foundational to concepts or just really-well-designed ways to transmit them.
- sometimelurker 3mo agoYou have the word 'bot' in your name and judging by your comment history love using em-dashes. Are you a machine or not?
- jvanderbot 3mo agoI doubt I used emdashes, but I do us "-". I dont have emdash unicode bound to my keyboard yet. (edit: Hyphenation is not emdash, of course) you can view my profile details freely and decide for yourself, and a quick look might even explain the "bot" suffix.
- sometimelurker 3mo agoOh I figured there's a real human being named Josh Vander and youre his openclaw or something. why do you have the `bot` suffix?
- jvanderbot 3mo agoBecause I spent 15 years working in robotics, and I'm a nerd who over indexed identity on that :)
- sometimelurker 3mo agosorry abt that, its just there's a lot of bots on hn nowadays and I get irritated sometimes. mb
- adamzwasserman 3mo agoI can guarantee you that the comment from jvanderbot was not AI generated.
- mhjkl 3mo agoThis comment would go hard in a sci-fi novel 5 years ago
- kibwen 3mo agoExcept that the Chinese Room shows that the existence of a mapping from input to output, however emergently it might have been devised, is not alone sufficient to demonstrate understanding.
- PaulDavisThe1st 3mo agoIt does so only in the claims of its creator. Plenty of other people have pointed out fallacies in the claim. My favorite is the Dennett/Hofstadter observation that while the man in the room may not understand chinese, the room/system certainly does.
- Marha01 3mo agoThe Chinese Room as a whole understands.
- dinfinity 3mo agoThe man in the room is comparable to a human hand, and the magic rulebook to a human brain. This in the sense that we can easily retain human "understanding" by stripping away almost all parts of the human body or replacing them with fairly trivially made replacements, except for the brain. In the Chinese Room the equivalent is the magic rulebook: We have no idea how to construct/replace it, yet people somehow handwave that away whilst simultaneously confidently asserting it does not understand anything.
- kibwen 3mo agoThat's irrelevant to what we're talking about, though. The point here is not to assert whether or not the room understands, but to emphasize that our methods are insufficient to demonstrate this. You could posit that the room understands, or you could posit the reverse, and neither argument can be refuted.
- swid 3mo agoIf our own brains were insufficient to demonstrate the fallacy of the Chinese Room; the particles, molecules, cells, synapses, and so forth of our bodies cannot be believed to have understanding, but we recognize it in the sum of the part anyway.... ... we now have LLMs to even more pointedly show the deficiencies of the argument.
- sometimelurker 3mo agoAnd they do have internal latent vectors for their own states, so there's some kind of recursion/introspectiveness dynamic.
- dinfinity 3mo agoExactly. The translation back to words is the final step, so in a way very similar to what the post describes. Improvements in model performance have been made exactly by having intermediate steps stay in the form of internal representations rather than words.
- Fraterkes 3mo agoThe assumption of this sort of argument, if I understand it correctly, is that consciousness is just an ordinary byproduct that appears (or grows gradually) somewhere on the spectrum of complexity. If that’s true, mastering language is basically orthogonal to being conscious (as you’d maybe expect, GPT-2 was pretty good at language, but had a relatively tiny amount of “neurons”). The same is arguably true for world modeling: a math textbook has a very complex and coherent model of a world, and is, probably, unconscious. Then the question is: what reason do we have to assume that these models have some form of consciousness?
- lowbloodsugar 3mo agoLikewise, what reason do we have to assume that any given human is conscious?
- __MatrixMan__ 3mo agoOr that anything at all is not conscious?
- Fraterkes 3mo agoBecause we are, in our life, not always conscious.
- __MatrixMan__ 3mo agoWhen you say "we" you're referring to something besides our minds? Because sure, when we get knocked out, we're out. That's not an observation, it's a tautology. But there's nothing indicating that whatever remains doesn't have its own kind of consciousness.
- Fraterkes 3mo agoI don’t know why I’m consciousness, but I know that I am. Other people are outwardly very similar to me, so I assume the same is broadly true inwardly. A rock is outwardly pretty similar to me; in that it exists in the world, has obvious physical boundaries, and is affected by the passage of time. But it is not so outwardly similar that I assume consciousness. An LLM is also outwardly similar to me, in that it can express itself in language, and seems to ‘contain’ notions about the world. But it is again not necessarily so similar to me, outwardly, that inward similarity is obvious (to me)