That assumes that there is no effective way to filter good from bad. But there is - both automated heuristics and manual training does this.
LLMs absolutely produce mediocre output in some ways, but it’s not an inherent limitation caused by them “averaging” the internet. If that were the case there’d be a lot more typos, emojis and internet lingo by default. The fact that LLMs have these instantly recognisable stock ways of writing and stock phrases is a simple way of seeing that they don’t simply produce “average” output in that very naive sense.
The fact that they “average” their inputs is why there are comparatively few typos (different sources have different typos, so they average out), not too many emojis or internet lingo (again, different sources use different ones in different places, so they average away), and why they produce such tedious stock output (it’s an average of the inputs, so all the little quirks and idioms that make human communucation more vibrant have been blended away).
I’m sure there is some filtering on the inputs to try to remove the worst of it, but ultimately it’s still just taking the rest and building it’s probability tables from that, which leads to the homogenised outputs we see.
Mind you, having said there are fewer typos, the last time I bothered trying to get one to write some code, it managed to misspell a popular library name in multiple places, which gives some indication of how bad the inputs are, how bad the tokeniser is, or possibly both.
If “different typos” averaged out to “nearly no typos” the same logic would have different words average out to nearly no words. What actually happens is the model learns context, and can produce output which contains emojis in one context and not others. These contexts can be very far from the average context.
I’m afraid the upshot is you don’t understand how the models work. There is extensive filtering before training - they do not get “the entire internet” and average it. If you want to understand properly, there are a lot of resources that will let you, but I’m not going to try to do it here, so you’ll either have to believe me or be wrong, I’m afraid.
If “different typos” averaged out to “nearly no typos” the same logic would have different words average out to nearly no words.
No, because typos are irregular, so combining multiple sources does not reinforce them, whereas “words” (tokens would be a better term, because they’re not always full words) tend to be used in similar ways, reinforcing those patterns. As you say, context is relevant, an LLM isn’t just looking at the last token to decide the next, but at a much larger window. That does allow it to adjust to tone, as the probabilities of certain tokens, and so words, will depend on that tone, and the type of words used, and thus context, of a conversation. If emojis are used a lot in certain contexts, those patterns will tend to be reinforced in their training, and so produced more in their output.
As to filtering their input, at no point did I say they ingest “the entire internet”, so quoting it seems rather disingenuous. They scrape as much text as they can get, both online, and by OCRing books, as we’ve seen with the recent upset about the number they destroy. What the commercial models do with this afterwards is uncertain, as anything they say is likely to be misleading for commercial purposes. I think it’s a fair assumption that they want good quality data, however they define that, but filtering it all manually is obviously much too vast a project to do entirely manually, so it’s done heuristically, which has the obvious problem that it’ll let through low quality sources some of the time, lowering the quality of the overall data set. You only need to read the anodyne screed they produce to see how all of the little quirks and nuance that marks human communication tends to get left out, leaving LLM prose feeling rather vacuous and repetitive.
I was going off the statement, “If you are using a generator that combines everything (both bad and good) on the net”. I think “the entire internet” is an OK paraphrase of “everything on the net”, but maybe you meant something else than what I understood.
The original public release of ChatGPT used a small army of humans to generate and curate training data, and it’s still the rule today. So what the average of what the model actually sees is quite far from the average of everything on the net.
The ability of the model to learn different contexts and tones means that all it takes to not land “right on the middle” is a slightly different context, which can all be contained in the instructions given to the model. Sure, the models have their stupid phrasings and stock phrases, but the context was producing a first draft. A competent human can reword those very easily.
Do you even know what the word heuristic means? A heuristic is something that is just good enough. Not great, not perfect, just good enough to get the job done.
Heuristic algorithms were always going to result in LLMs that were only just good enough.
Heuristics are by definition imperfect, but they are not, generally, “just good enough”. In fact, heuristics may not be good enough for a given purpose.
Am I right that you’re not actually disagreeing with my comment?
That assumes that there is no effective way to filter good from bad. But there is - both automated heuristics and manual training does this.
LLMs absolutely produce mediocre output in some ways, but it’s not an inherent limitation caused by them “averaging” the internet. If that were the case there’d be a lot more typos, emojis and internet lingo by default. The fact that LLMs have these instantly recognisable stock ways of writing and stock phrases is a simple way of seeing that they don’t simply produce “average” output in that very naive sense.
The fact that they “average” their inputs is why there are comparatively few typos (different sources have different typos, so they average out), not too many emojis or internet lingo (again, different sources use different ones in different places, so they average away), and why they produce such tedious stock output (it’s an average of the inputs, so all the little quirks and idioms that make human communucation more vibrant have been blended away).
I’m sure there is some filtering on the inputs to try to remove the worst of it, but ultimately it’s still just taking the rest and building it’s probability tables from that, which leads to the homogenised outputs we see.
Mind you, having said there are fewer typos, the last time I bothered trying to get one to write some code, it managed to misspell a popular library name in multiple places, which gives some indication of how bad the inputs are, how bad the tokeniser is, or possibly both.
If “different typos” averaged out to “nearly no typos” the same logic would have different words average out to nearly no words. What actually happens is the model learns context, and can produce output which contains emojis in one context and not others. These contexts can be very far from the average context.
I’m afraid the upshot is you don’t understand how the models work. There is extensive filtering before training - they do not get “the entire internet” and average it. If you want to understand properly, there are a lot of resources that will let you, but I’m not going to try to do it here, so you’ll either have to believe me or be wrong, I’m afraid.
No, because typos are irregular, so combining multiple sources does not reinforce them, whereas “words” (tokens would be a better term, because they’re not always full words) tend to be used in similar ways, reinforcing those patterns. As you say, context is relevant, an LLM isn’t just looking at the last token to decide the next, but at a much larger window. That does allow it to adjust to tone, as the probabilities of certain tokens, and so words, will depend on that tone, and the type of words used, and thus context, of a conversation. If emojis are used a lot in certain contexts, those patterns will tend to be reinforced in their training, and so produced more in their output.
As to filtering their input, at no point did I say they ingest “the entire internet”, so quoting it seems rather disingenuous. They scrape as much text as they can get, both online, and by OCRing books, as we’ve seen with the recent upset about the number they destroy. What the commercial models do with this afterwards is uncertain, as anything they say is likely to be misleading for commercial purposes. I think it’s a fair assumption that they want good quality data, however they define that, but filtering it all manually is obviously much too vast a project to do entirely manually, so it’s done heuristically, which has the obvious problem that it’ll let through low quality sources some of the time, lowering the quality of the overall data set. You only need to read the anodyne screed they produce to see how all of the little quirks and nuance that marks human communication tends to get left out, leaving LLM prose feeling rather vacuous and repetitive.
I was going off the statement, “If you are using a generator that combines everything (both bad and good) on the net”. I think “the entire internet” is an OK paraphrase of “everything on the net”, but maybe you meant something else than what I understood.
The original public release of ChatGPT used a small army of humans to generate and curate training data, and it’s still the rule today. So what the average of what the model actually sees is quite far from the average of everything on the net.
The ability of the model to learn different contexts and tones means that all it takes to not land “right on the middle” is a slightly different context, which can all be contained in the instructions given to the model. Sure, the models have their stupid phrasings and stock phrases, but the context was producing a first draft. A competent human can reword those very easily.
Do you even know what the word heuristic means? A heuristic is something that is just good enough. Not great, not perfect, just good enough to get the job done.
Heuristic algorithms were always going to result in LLMs that were only just good enough.
Heuristics are by definition imperfect, but they are not, generally, “just good enough”. In fact, heuristics may not be good enough for a given purpose.
Am I right that you’re not actually disagreeing with my comment?
Heuristics that aren’t good enough aren’t used for that purpose… what kind of rebuttal is this even?
It’s not a rebuttal because I don’t even know what point you’re trying to make, as you may have been able to tell from my question: