6 ms·
> To sum up: because LLMs are fundamentally limited to i) using our vocabulary, ii) “understanding” concepts in the ways we do, and iii) “talking” in the ways w
by johndough 3y ago
> To sum up: because LLMs are fundamentally limited to i) using our vocabulary, ii) “understanding” concepts in the ways we do, and iii) “talking” in the ways we do, then, at best, a LLM can only mirror back to us the order we have carved, the truth we have honed.
That conclusion is not correct. Generative models can combine existing concepts in novel ways which have never been considered before. This is most easily seen with text-to-image models. For example the teddy bear swimming in an Olympic pool [1] or many other surreal examples [2] were not part of the training set. So as long as an answer can be expressed as a combination of primitive concepts, LLMs can generate them.
[1] https://imagen.research.google/main_gallery_images/teddy-bear-swimming-butterfly.jpg https://imagen.research.google/main_gallery_images/teddy-bea...
[2] https://imagen.research.google/ https://imagen.research.google/
- Capricorn2481 3y agoThat's not an LLM, though. We haven't seen evidence yet that an LLM can combine existing language in novel ways. In fact, we've seen over and over that LLMs are quite generic. Compared to text-to-image, which is seemingly impossible to use without getting something weird
- glenstein 3y agoIs there something about the nature of language and linguistic meaning that would make it difficult to combine language in novel ways?
- TuringTourist 3y agoI would posit the opposite, language exists to be combined in any number of ways easily and still communicate well. A language that cannot do this does not I think exist.
- glenstein 3y agoSame. I find it hard to makes sense of why anyone would claim that computers can innovate visually but not with language. It would make more sense if someone at least felt that they both could, or both couldn't.
- doug_durham 3y agoI have absolutely seen an LLM combine language in novel ways. With prompting they will create outputs that are entirely novel with concepts intermingled in new ways. It's nonsense when compared to observed reality, but it is novel nonetheless.
- bloaf 3y agoThere was a post here on hacker news not that long ago where GPT4 came up with "the confetti has left the cannon" when asked for an original phrase similar to "the cat is out of the bag". Other users confirmed Google could not identify any other use of that phrase. People saying GPT4 is unoriginal have an uphill battle, it is not the default assumption of anyone who has worked with it.
- lacrimacida 3y agoChatGpt is as original as us but cannot ever surpass us because it needs our feedback loop. It is a mirroring of sorts after all. If it could become extremely creative, push the limits and make ideas so advanced that we can’t understand then it failed at the task so to speak. It always remix idea into larger idea sallads and get our upvotes downvotes. It is a great tool but a tool nonetheless.
- melicerte 3y agoConfetti cannon is a thing. So there is an obvious link between confetti and cannon. Furthermore, These ("cannon" and "conffetis) are two very common concepts used in a single sentence. Excuse me for being genuine but I totally fail to understand how original the phrase is.
- bloaf 3y agoWhy don't you come up with an example of a new phrase with the same meaning that is actually original so we can compare your "real" originality to this "fake" originality?
- Capricorn2481 3y agoThe dazzling ribbons, stripped from the corpses of trees, shot out of the pressurized tube
- visarga 3y ago> We haven't seen evidence yet that an LLM can combine existing language in novel ways. How about BubbleSort written in Shakespeare style?
- cscurmudgeon 3y agoI would say coming up with BubbleSort or Shakespeare's writings from scratch is an example of novelty/creativity.
- vidarh 3y agoMost humans can do neither, so it's setting the bar far too high.
- cscurmudgeon 3y agoI think we should compare what best of LLMs can do with what the best of humans can do for a fair comparison. Else, we will be comparing what the best LLM can do with what the media/mean human can do. I can also point to tons of language models (both large and small) which don't do anything remotely useful.
- vidarh 3y agoWhat is fair about that when judging whether or not they can be creative? Most people whom other people would judge to be creative are nowhere near the best of us.
- osmarks 3y agoThis is mostly a property of the RLHF/instruction-tuned chatbot models most people use. Base models tend to be more varied.
- jncfhnb 3y agoThat distinction is so pedantic I don’t think it’s worth mentioning. LLM architecture can be extended to other contexts besides language.
- polio 3y agoHow much of the "novel ways which have never been considered before" is just the novelty effect of having your very own artist? A human being could certainly produce any of the works of Dall-E 2, given the same prompt. The change here is the cost, and not the capability. Of course, this is still significant, but it doesn't suggest to me that Dall-E 2 "thinks" differently or would be able to seriously alter the nature of our cognition, except to the extent that it allows us to realize the same ideas faster.
- epups 3y agoWhat would be an example of a work that a human being could not produce?
- simonh 3y agoA vast, intricate design with microscopic details at hundreds of levels of emergent forms, coloured in semi-transparency generating holographic effects, using spectra of light far beyond the human visual range, executed in three dimensions and occupying hundreds of cubic miles. Oh, and why not animate it? It would have to be virtual I suppose, or vastly too expensive to ever be created physically. In fact, why design it to be physically possible? Go nuts, make it 7 dimensional.
- civilitty 3y ago> How much of the "novel ways which have never been considered before" is just the novelty effect of having your very own artist? What's the difference and why does that matter? Sure, plenty of people may have had the idea of an anthropomorphized teddy bear swimming in a pool, but if none of them ever realized it in the real world, the AI couldn't have learned from it. > The change here is the cost, and not the capability. I strongly disagree. Scaling the artistic process with GPUs instead of meatbags is a huge change in capability, just like the mechanized tractor was a huge capability change over the ox. The change in cost is a side effect in the change of capability. You can now use a tool instead of outsourcing it to someone else and artists now have an automated tool that ostensibly replaces their manual labor.
- jrflowers 3y agoThis is a good point. No human could possibly have envisioned a teddy bear swimming in an Olympic pool, as we did not previously have the words or conceptual framework to describe that. That image could never have existed before the invention of transformer technology.
- WanderPanda 3y agoIt mirrors back the distribution, not the individual samples. It can interpolate but not extrapolate
- hyperliner 3y ago[dead]
- thesz 3y agoI remember reading what Lenat did with early versions of Eurisco: he ran it on some axioms from number theory and observed discovery of some number theory theorems, some of them pretty fundamental. As "time" progressed, the "theorems" Eurisco kept getting discovered became duller and duller, like expansion of square of sum into sum of products (I can't remember exact example from the book, but level of dulliness is about right). I believe this can be served as example of combination of existing concepts in novel ways, applied to math, which is very formal and, on the other hand, is the basis of many other areas of human knowledge. If this can serve as a such example, then you are wrong. Also, what you have shown as your examples are not very much interesting outside of entertainment.