Hacker Newsnew | past | comments | ask | show | jobs | submit | ohyes's commentslogin

You forgot the “co-authored by Fable 5.1” line on your post.

So the interesting thing is that this shrinking “linguistic diversity” is fundamental to how an LLM works.

The LLM is a big probabilistic statistical trick. It picks the next token based on certain words are simply “the best” because they are specific and well connected to other tokens. The is gives them a great overall cost function. (Basically a good score on “will it make sense in context” while also having specific meaning that makes it better than other options, unambiguous in common use and being a single token rather than several).

You can trim those tokens, but then you just get other tokens that are “the best” tokens (and you’re worse off because the output became less clear).

The cool thing is this seems to get worse the more powerful and accurate your model is, because it is picking technically / statistically perfect tokens, not tasteful ones.


It's not fundamental at all, that's what randomized sampling is for. Try tinkering with a base model and you'll be surprised how diverse it is. The semantic collapse happens in post-training that is using the current methods.

> Try tinkering with a base model and you'll be surprised how diverse it is.

Have you tried? I have. Not much different from RLHFed; full of tics and slop, similar but slightly different from intsruction posttrains.


Yes, although admittedly I haven't tried recent ones which have a bunch of synthetic data in them (mainly to aid the reasoning), and are usually only available after mid-training. One look at the logits/output distribution and it's clear the base is pretty different.

that's a very longwinded claim that inference providers are sampling with temperature T=0, but is that even true? a sufficient explanation would be merely sampling at a lower temperature compared to human sources providing similar content

It's still probabilistic from a huge dataset while the human mind is not.

It does not matter really - stiffness does lower up to T=0.7, then platoes; even at high temperatures tics/slop-patterns are still there.

That's pretty model-specific, for example DeepSeek of the v3/R1 era would already start losing coherence occasionally at t=0.7 with no other samplers

I played in the early days and quit to have a life, came back for a bit to check it out.

Basically the community is people checking it out, bots, and the terminally addicted multi boxing 30 accounts.


To be honest I’ll ask a model to specifically think of edge cases but I won’t expect any model to do the edge cases of its own volition

The whole point is that we’re TikToking the software industry and no one will know what to tell the ai to get the software written in a way that works. It’s in principle the same as idiocracy where no one knows how to make burrito covers anymore because shit won’t grow

Maybe we need to start cryofreezing some of our best engineers as a precaution. Or even just a bunch of average ones.

Imagine being Guido van Rossum typing away and then the cryo-freeze squad (new ICE) comes knocking.

Maybe a super smart AI will take over the part of telling the smaller AI's what to do.

Humanity can then relax in floating chairs drinking soft drinks and consuming AI generated entertainment.


You can train and run small models on an old gpu. That’s what I’m doing now at, well, much older than 17. Does it produce a useful model? No. Not even remotely.

However, I do learn stuff about models that takes it from “magic” to “useful tool I understand the limitations of.”

Do I do it for that reason? No not really, I’ve never had luck learning something because it would be good for my career. I do it because at my core I’m a bored teenager who wants to make the computer do cool shit.


Am I the only one who just downloads directly from LM Studio and just runs the server there? It’s trivial.


You're not the only one, I'm running Qwen3.8-27B in LM Studio and it seems to be going great. Was very easy to set up.


Well they decide what a token is. So they can do less superfluous things and backfill with a weaker model.


Okay but is that a relatable sound bite that will make someone click a link? No.


On the one hand we have Claude doing my work for me, on the other hand it is laced with unoriginality, on the gripping hand everyone will get all references.

All witty stuff disappearing and turning into LLM mush is kind of the maximal expression of capitalism for capitalism’s sake. We don’t care if things are novel or interesting, we’re producing “content” to fulfill an imaginary content demand that we can feed to make money.

The best way to feed it is to create stuff that’s just barely enough degrees of quality from a viagra spam email that it can pass for original thought.

—- —- —- Em dashes for emotiveness.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: