4 ms·

Maybe it’s just me but (and I ask this as something of an ML practitioner myself) can anyone at Anthropic or Open AI or any of the frontier labs work on summari

Maybe it’s just me but (and I ask this as something of an ML practitioner myself) can anyone at Anthropic or Open AI or any of the frontier labs work on summarizing content into a format where a human won’t immediately vomit? There’s this sort of local optima all these models pick which is instantly recognizable and hardly digestible for human consumption. Perhaps too much internal RL against benchmarks during chain of thought? The older non reasoning models had they’re own problems but at this point I can’t get an LLM to summarize data for humans, which is troubling. As much as I’d like to read this piece it follows suit and I don’t have the patience to try and extract anything valuable from it.

carrychains 2mo ago
Give it a sample of your own technical writing and tell it to match the voice. I did this with Fable just today and it nailed it.
- mvanveen 2mo ago
  that's an interesting idea thx
AreShoesFeet000 2mo ago
that is just fatigue. you’ve been vibed too hard.
wewewedxfgdf 2mo ago
Yes, you say: "write (whatever you want written), BUT do not make it sound like a typical AI don't use any of the tropes or typical phrases that AI tends to use - if you include anything that makes it appear to be from AI then the task has failed. Pull your writing away from the median and toward very specific random styles that are NOT the cliched, averaged LLM style." edit: actually wrong I tried to get ChatGPT to do this and it kept falling back into LLM speak even when it recognised that's what it was doing.
- mvanveen 2mo ago
  yeah, no, I've tried a number of variants like that. for ex even if I tell it to not rely on its chain of thought created jargon it'll just usually continue to hallucinate adding in the jargon unless I line item them painfully and eventually get something a human can understand. prompting it to be concise ("250 words or less") seems to help somewhat but theres only so much you can do esp in cases where the details matter and don't compress well
toomim 2mo ago
I'm starting to notice that LLMs tend to speak in the form of movie trailers. Say this in a deep masculine dramatic voice: One man. One gun. A story that would could never die. Compare that with this Coldcard page: Two files, four lines, five years unnoticed. This is the COLDCARD entropy failure It's trying too hard to replicate a dramatic movie arc with every... single... sentence...
- wewewedxfgdf 2mo ago
  You are correct, and that style of writing is called "broetry".
- saidnooneever 2mo ago
  but someday in a galaxy far far away... what do you think these things are trained on? mountains of rubbish, human and likely also AI slop. how many bad movies and ads do u think they are fed. poor things. i can see them caged in the corporate HQ, with their tails shoved down their own throats rigged up to the SlopExtractor9000. its time for the UN to Unslop the world. or something like that :(
j16sdiz 2mo ago
Maybe it learn it from content farms, not RL. Before LLM, the majority of the web was content farms... but those are instantly recognizable and we seldom fall into them. Now LLM mixed the content farm style writing with _some_ real content. It take time and effort to tell they are slop, causing much fatigue.
- mvanveen 2mo ago
  I have no data to back up this claim but one hypothesis I have is it's gotten particularly acutely bad esp with newer gen models because of increased training on reasoning and chain-of-thought. To borrow a programming lang analogy the failure mode I see a lot is it invents its own jargon that is effectively like a compiler intermediate representation of the high level natural language you actually want a human to look at and then inserts it directly into what the human has to read. It needs to stop doing that but it doesn't seem to do a good job at differentiating from what is or isn't chain of thought slop. I'm sure the jargon is useful during its reasoning but it's very unhelpful and not very nice to deliver it to a human.

carrychains 2mo ago

Give it a sample of your own technical writing and tell it to match the voice. I did this with Fable just today and it nailed it.

mvanveen 2mo ago
that's an interesting idea thx

AreShoesFeet000 2mo ago

that is just fatigue. you’ve been vibed too hard.

wewewedxfgdf 2mo ago

Yes, you say: "write (whatever you want written), BUT do not make it sound like a typical AI don't use any of the tropes or typical phrases that AI tends to use - if you include anything that makes it appear to be from AI then the task has failed. Pull your writing away from the median and toward very specific random styles that are NOT the cliched, averaged LLM style." edit: actually wrong I tried to get ChatGPT to do this and it kept falling back into LLM speak even when it recognised that's what it was doing.

mvanveen 2mo ago
yeah, no, I've tried a number of variants like that. for ex even if I tell it to not rely on its chain of thought created jargon it'll just usually continue to hallucinate adding in the jargon unless I line item them painfully and eventually get something a human can understand. prompting it to be concise ("250 words or less") seems to help somewhat but theres only so much you can do esp in cases where the details matter and don't compress well

toomim 2mo ago

I'm starting to notice that LLMs tend to speak in the form of movie trailers. Say this in a deep masculine dramatic voice: One man. One gun. A story that would could never die. Compare that with this Coldcard page: Two files, four lines, five years unnoticed. This is the COLDCARD entropy failure It's trying too hard to replicate a dramatic movie arc with every... single... sentence...

wewewedxfgdf 2mo ago
You are correct, and that style of writing is called "broetry".
saidnooneever 2mo ago
but someday in a galaxy far far away... what do you think these things are trained on? mountains of rubbish, human and likely also AI slop. how many bad movies and ads do u think they are fed. poor things. i can see them caged in the corporate HQ, with their tails shoved down their own throats rigged up to the SlopExtractor9000. its time for the UN to Unslop the world. or something like that :(

j16sdiz 2mo ago

Maybe it learn it from content farms, not RL. Before LLM, the majority of the web was content farms... but those are instantly recognizable and we seldom fall into them. Now LLM mixed the content farm style writing with _some_ real content. It take time and effort to tell they are slop, causing much fatigue.

mvanveen 2mo ago
I have no data to back up this claim but one hypothesis I have is it's gotten particularly acutely bad esp with newer gen models because of increased training on reasoning and chain-of-thought. To borrow a programming lang analogy the failure mode I see a lot is it invents its own jargon that is effectively like a compiler intermediate representation of the high level natural language you actually want a human to look at and then inserts it directly into what the human has to read. It needs to stop doing that but it doesn't seem to do a good job at differentiating from what is or isn't chain of thought slop. I'm sure the jargon is useful during its reasoning but it's very unhelpful and not very nice to deliver it to a human.