(1) LLMs collapse and start outputting garbage after a number of tokens if you do not sample and just pick the "best token" each time. This is a consequence of how they are trained.

Citation needed

You know you can just try it and see on any inference system thst has this knob, right?

Related: if you don't have a limit on sampling (top-K or top-P), eventually you'll hit one of the really unlikely tokens by chance and then the model will switch to Japanese because the most likely completion after a random Japanese character in the middle of an English sentence is more Japanese writing, not a reversal back to English.

That could be how it works, but in practice it takes into account all previous tokens when producing the next-token distribution to sample from. So a switch back is more likely than your explanation supposes.

No, if you switched to Japanese the LLM wouldn't ignore it, it would "assume" there's a reason for that. The same if the previous iteration of the LLM switched to Japanese. Else you're expecting an LLM to ignore its own previous outputs and restart "thinking" from scratch with every token?

It's situational and I suspect there are situations where it would and others where it wouldn't. Would depend on the almost infinite variables of how training was done. You'd be right that it'd be likely to switch but while it's possible it's due to temperature, there are just so many things going on. But it would be one sensible explanation among many.

> LLMs collapse and start outputting garbage after a number of tokens if you do not sample and just pick the "best token" each time.

LLMs are likely to get stuck even with sampling if asked to generate tokens on their own long enough, though sampling does tend to stretch out the time before that happens (as do other techniques that don't involve sampling, like applying repetition penalties directly to token logits). But LLMs generally aren't left to infinitely extend their own output, and the length response typically needed in the use case is much shorter than the would result in collapse given the kinds of inputs expected in that use case, the existence of the theoretical eventuality may not really matter.

> which inherently compromises quality.

I don’t see how this follows? Tokens are chosen randomly. If you choose tokens with a different RNG in the same distribution, you’re still getting equally good or bad tokens.

Not all values of "equally good" are equally good.

Writing has rhythm, or at least it's supposed to, and synonym swapping compromises it.

Never mind metaphors and similes, which are even more tightly constrained.

LLM writing is still a long way from good. Sometimes you get lucky with the odd line, but there's a difference in quality between influencer slop, genre fiction, and literary fiction and/or best-in-class journalism.

LLMs are still somewhere between the first two, and nowhere close to approaching the third.

> Not all values of "equally good" are equally good. Writing has rhythm, or at least it's supposed to, and synonym swapping compromises it.

We already know that a non-zero temperature improves quality though with current models (particularly with creative writing). The assumption that always picking the 'best' token results in the 'best' output is not the current reality.

And if you are already intentionally putting in randomness, I can imagine that it would be possible to seed the randomness in a way that is detectable but results in the same quality.

This is obviously not true for queries where temp = 0, but at temp = 0 then it becomes easier to identify anyway. I assume this technique implies some level of temperature.

It's not just swapping synonyms. The way llms work is by predicting the likelihood of the next token. It's inherently probabilistic. Choices are made based on weighted random number generation, based on those probabilities. Changing how you generate the random numbers doesn't degrade the output.

wow even on HN people have no clue how any of that works? all LLM generation has some inherent randomness to it, if you replace part of that randomness to be deterministically random the result of the generation with the fingerprint and without it, is INDISTINGUISHABLE. This has absolutely nothing to do with "synonym swapping". Its also again people not understanding how anything works missing the real concern, which is that nobody can ever tell if it isn't also secretly a fingerprint with the user and session id.

1) we’re not discussing those systems. We’re discussing a chat AI product called Claude, which does not offer those knobs.

2) Claude’s PRNG having a P is immaterial

Claude has those knobs, they are just not exposed to the user. They could make Claude nearly completely deterministic if they wanted to (of course it would be a far inferior product then. But they could).

“Not exposed” = has no knobs. Of course all autoregressive LLMs can be operated this way but Claude, the product, employs LLMs but isn’t one.

Your original statement “LLMs use randomness in a fundamental way” is incorrect. LLMs have these knobs and randomness is not an inherent property of LLMs.

You're generating a pseudo random number one way instead of another way. How would that inherently compromise quality?

Same reason that you don't just replace your rand() implementation with "return 4; // chosen by fair dice roll". If you need randomness for whatever reason, biasing the generator is compromising quality.

( https://xkcd.com/221/ )

In case of LLMs, you can look at it from high and low level.

At low level - if you could do with less randomness, you can always lower temperature. You usually keep it (or for SOTA providers' chat UI, they keep it) at a level where it's about right level - high enough to allow for more creative leaps and interpretations, low enough that it doesn't go off into crazy land after the third paragraph.

At high level - creativity is driven by randomness. If you had an author (fiction or nonfiction) you like for their both broad and deep range of insightful thoughts, would you be happy if they suddenly developed an acute porn obsession and uncontrollably added lewd subtext to every other sentence? Still creative, still deep, but now with that one strong attractor that biases their every thought in a single direction? Would you trust/enjoy their output as much as you did before?

That, slightly exaggerating to make it more obvious, is what "loss of quality" means here.

You're missing the same point that the blog post is missing. What they're doing is much less like replacing rand() with 4 and much more like setting seed(4) before generating any numbers. There is no "loss of quality" unless you're already using a temperature of 0.

1. As watermarked text is added to the training data, watermark-related tokens will be associated more with AI outputs and thus lower quality outputs which will hasten model collapse. Especially because every provider has its own secret key and they are all training on eachother's outputs anyway.

I guess they can at scale filter the watermarked documents (by necessarily allowing eachother to at scale checked for watermarks, but banning the labs not part of the watermarking-cabal). Makes me wonder how useful the human quality filter is on AI output - if a human judges a given output as genuinely good and posts it somewhere for the scrapers to find and take into the training sets, will these types of outputs also be filtered out?

2. (raw, pre-watermarked) Output token probability situations where 1 output token has the majority of the probability mass associated with it, but it is not in the watermarked set, will force the model with much higher probability to walk a non-optimal latent space. E.g., if the next OBVIOUS token for a given sentence would be a point, but the model is in this way not allowed to output it, it might put a comma and start off on a whole different tangent just to make the initial non-optimal comma grammatically make sense.

There are no watermark related tokens, there is no watermarked set - read the paper, it's public and not that complicated. It doesn't change the distribution of completions, and won't change the situation where one output token has the majority of probability mass.

You're missing the fundamentals here!

I am going off the explanation in the declaude page (and related papers). But I see now anthropic mentions Aaronson's distortion-free watermarking.

Random watermarking functions colour the tokens based on (small) contexts and a secret key. Given watermarking functions are randomly chosen every single time they are used (so essentially not deterministically seeded by the context and secret key), then indeed the completion distributions are unchanged. However, they are deterministically chosen, a given string of text will always have the same corresponding watermarking functions. Tokens scored high by the function see an increased probability of being the chosen completion, those scored low see a reduced probability. I dont see it is different from merely talking about it as green/red and the points hold?

For a given short-context (hash function seeding) you do have detectable manipulation of the completion (how they can read the marking). But, because (assuming enough entropy in context) the hashing function is decoupled from the log-probs, the perturbations are independent from the underlying distribution, so you're still sampling from the same distribution quotient some noise.

The only way you'd notice this is they weren't independent, and the most plausible way that happens is if you're re-completing pre-fills (resampling the same hash function against the same log-probs).

Ok. I dont have a good feeling for the actual completion distributions. The noise sounds problematic. I can imagine it relates to the size of the context used for hashing. You want this as long as possible so that the entropy is higher, but you also want it as short as possible, because edits invalidate the hashing for all the tokens of which the edited tokens are part of the hashing context.

Anyway, there are lots of cases where text carries very little entropy. E.g. boilerplate code, exchanges of pleasantries, well-worn platitudes and jokes, etc. These are sequences of tokens that will be seen across many, many separate outputs. Watermarking here (on the token following the common sequence) would thus be easily detectable and noticed as a claude style. The longer the hashing context though, the lower the amount of pathological cases with low entropy. It would be interesting to understand the exact parametrization better!

These are services, so "how would <x> do this today?" is irrelevant.

The real question is: "What can manipulation of pseudo-random number generation do?"

We know that in the cryptographic world, attacking "randomness" is a key offensive capability. It will be here as well -- if Anthropic can watermark text as generated it's LLM, will it be able to watermark outputs as generated by "Spooky23/FooCorp"? Can I pay Anthropic to steer inquiries in a way that benefits my company or governemnt?

Pseudo-random to the end user appears random. Most treat it like a random chance. It is not.