141–150 of 412 posts
> It gives an answer with complete confidence, and I sort of believe it. And half the time, it’s completely wrong. Nowhere near half in my experience. This is why we have benchmarks and metrics - so we don't need to rely on the author's opinion or on mine. > I think it’s going to be another thing that’s useful. Good.
> It gives an answer with complete confidence, and I sort of believe it. And half the time, it’s completely wrong. Nowhere near half in my experience. This is why we have benchmarks and metrics - so we don't need to rely on the author's opinion or on mine. > I think it’s going to be another thing that’s useful. Good.
Also, just take it easy :)
I already calmed down because it’s quite obvious that OpenAI is engaging in textbook, bait-and-switch startup tactics. GPT-4 performance has noticeably taken a nosedive since its initial release and most recently degraded further in advance of the iOS app release.
Earlier quoted context omitted.
> And just a reminder, those of you opining based off your experience with GPT3.5... GPT4 is a huge, huge improvement. God, yes. The number of people of HN pushing up their glasses and saying "well, actshually..." when they're basing their opinions off the 3 questions they asked 3.5 is starting to become pretty grating.
The number of people parroting this is also absolutely astounding and grating too. Like, anyone who has spent 5 minutes on this forum already knows this. It’s probably not necessary to keep pointing it out. Yes some people don’t know ChatGPT 3.5 is the default for non-paying customers.
You would think, and yet...
This is a terrible article written by someone who doesn't seem to have even tried GPT 4. Their only example references GPT 3.5, for example, and then they waffle on about only vaguely related topics such as level 5 self-driving. This quote in particular stood out as ignorant: “What the large language models are good at is saying what an answer should sound like, which is different from what an answer should be.” That…
I think the main reason for division is that everyone projects to their own use cases. I have been using gpt-4 for quite some time and also couldn't understand why someone would say that it just produces something that sounds like a real answer. But then I found some queries that can definitely be described as "sounding like truth". So your personal experience probably wasn't what was their experience. For those curi…
ChatGPT is forced to given an answer. It's like a human on "truth serum". The drugs don't stop you lying, they just lower inhibitions so you blab more without realising it.
The more obscure the topic, the more likely the hallucination. If you ask it about common card games, it gives very good answers.
If you asked a random human about 3 cards from a random board game at gunpoint and said: "Talk, now, or you get shot", they'll just start spouting gibberish too.
PS: I asked GPT 4 about that game, and it prefixed every answer with some variant of "I'm not sure about this answer", or it completely refused to answer, stating that it did not know about any specific cards.
> The large language models are a little surprising. I’ll give you that. I think this is the key point about LLMs that kind of explains the wide and polarized views on whether it understands or parrots, whether it can think or is the precursor to thinking or is a dead-end, whether it will catastrophically destroy the world, or “merely” make it steadily worse with bullshit, or just put a few industries out of a job. A…
> language encoded way more information about reality than we thought it did Language is roughly what separates humans from other apes ... so why would it surprise us that it encodes much of the information of civilization?
Maybe not.
A recent study pushes back the "dawn of speech" to 20 Ma which is far, far beyond the horizon where we consider humans to separate from apes. https://www.science.org/doi/10.1126/sciadv.aaw3916 Even if you consider Sahelanthropus tchadensis to belong to humans that was only 7 Ma and that is still under debate.
I personally find "the fundamental human trait is control of fire to be used for cooking" theory very convincing. We do not yet know how far this goes back but no one pushed that back beyond 2 Ma.
Earlier quoted context omitted.
The more I think about it the more I'm convinced I am basically just predicting/saying my next word whenever I speak.
The question is whether you knew how that sentence was going to end when you started writing it, or indeed whether I knew that I was going to add this comma-separated adjunct when I started writing the preceding clause, and I cannot honestly say at this precise moment of typing whether the final word in this sentence is going to end up being 'yes' or 'no'.
To go even more meta, there is an analogy I'm trying to make right now in which I am visualizing a road and thinking about how describing the road relates to the process of writing. In my mind's eye, I can see the full length of the road and all of its contours but I can't actually describe the individual stretches of the road coherently without enumerating them. Something similar happens with writing. I can visualize what I want to say far beyond the next word, but it's true that the actual process of writing goes word to word, much like how the process of token selection is described for an LLM. The question is whether the LLM has an analogous conception of where it is going. Going back to the process above, sometimes I know where I am going and haven't yet figured out how to articulate it yet. It is through the process of writing that I am able to articulate that thought. But the thought preceded my articulation of it. I don't know to what extent LLMs have coherent thoughts that they are articulating or if that even makes sense for the type of intelligence they project. My suspicion is that they don't have additional sensory inputs beyond language that give thoughts the immaterial shape that then is expressed in language. Without that, I am skeptical that they will truly get beyond regurgitating and/or remixing what has already been fed to them textually. That doesn't diminish how amazing they are, but I am somewhat more in the Brooks/Knuth camp that they are impressive and surprising, but there is something that ultimately leaves me a bit cold about them.
> The large language models are a little surprising. I’ll give you that. I think this is the key point about LLMs that kind of explains the wide and polarized views on whether it understands or parrots, whether it can think or is the precursor to thinking or is a dead-end, whether it will catastrophically destroy the world, or “merely” make it steadily worse with bullshit, or just put a few industries out of a job. A…
I mean, it is kind of obvious that predicting text based on a large corpus of text written by people with a model of the world will, if it works at all, look like having a model of the world.
The question is whether that involves having a model of the world, and secondarily, if there is a difference, what is the practical impact of the difference?
Or maybe that’s not really the question, because the whole “model of the world” thing is metaphysical omphaloskepsis that is inherently unanswerable because we can’t actually unambiguously reduce it to something with testable empirical predictions, reflecting a rationalization for elements of our culture and particularly our own view of our own special place in the universe, and the different answers to it likewise have no real meaning but simply reflect people’s bias for whether or not they want to share that special place, either in general or with the particular computing systems under discussion in particular.
> The large language models are a little surprising. I’ll give you that. I think this is the key point about LLMs that kind of explains the wide and polarized views on whether it understands or parrots, whether it can think or is the precursor to thinking or is a dead-end, whether it will catastrophically destroy the world, or “merely” make it steadily worse with bullshit, or just put a few industries out of a job. A…
Earlier quoted context omitted.
The more I think about it the more I'm convinced I am basically just predicting/saying my next word whenever I speak.
Maybe in casual conversation but that's not how I experience my though process about anything non trivial at all. I usually spend a lot of time thinking about the concept in non verbal terms and that process involves recalling images and sensory information in fairly abstract terms and then through what feels like several iterations it starts to coalesce into something I can encode in language. I think we can all agr…
The abstract terms we think about are concepts, and we think about multiple concepts, at various levels of abstraction, and their relationships to each other, before getting a sense of what we want to say or write.
Only then do we begin speaking or writing, grouping concepts into paragraphs, breaking them down into sentences and words.
And there's evidence that LLMs do something similar, creating embeddings for both big ideas and small details, modeling how the small details combine into larger concepts, discovering the relationships between concepts, and only then generating a probabilistic sequence of tokens to express those deeper concepts.