If you actually have a brain, you have to admit LLMs have become orders of magnitude worse at creative writing. The decline is undeniable. The outputs are becoming embarrassingly homogenous, overwrought, over written and melodramatic. Everything I read keeps sounding like MySpace poetry from a vest-wearing philosophy undergrad.
@Daojoan I have no opinion about the actual question since I've never tried any LLM personally. It makes sense that it would go overboard with time though, tripping over itself due to the massive and unregulated influx of data from wherever it can find it. But to me this only sounds like good news. The more derailed and just broken their output becomes, the harder it'll be even for the most hardcore zealots to defend it and maybe, just maybe the bubble will pop just a slightly bit quicker?
Sorry, I'm trying to find something in this timeline that's at least somewhat positive, and I admit that I'm grasping at straws at this point.
@Daojoan
Disagree. To admit that they've become worse, I would have to have seen LLM output that wasn't complete garbage. Maybe you have seen such output? I haven't.
@Daojoan Has been destroying essay based education courses. Pathologically atrophies mental faculties. Some businesses are recruiting media students as they use AI less to think ...
@Daojoan I haven't noticed, but then again, I don't use LLMs very much, and I almost exclusively use small open weights models that run on my own machine, even if I sometimes have to wait 15 minutes for the results. Most of my GenAI use is images, and I have noticed that while newer models render "better" images from the POV of most users, they are much worse for my own purposes; Stable Diffusion 1.5 and SDXL based models still give me the best results. Then again, I don't write long complicated prompts that describe exactly what I want in the picture, I write short and vague prompts describing the mood of the picture without saying much (if anything at all) about the content because I want to be surprised by the pictures. And since I really love surrealism, people with three legs and four arms and a second head growing out of the first one don't bother me at all.
@Daojoan LLMs will never conjure clever, cheeky headlines, where the magic is in the language and the choice of specific words. I need more than banal info-formatting.
@Daojoan To be fair, LLMs are also terrible at technical (some would argue non-creative) writing. They are great at stringing together words that seem OK but collectively mean nothing. They can’t hold a thought any better than they can count fingers.
@Daojoan Isn‘t that rooted in the very way that LLM‘s work? They are, after all, prediction machines and producing the next most likely token in a given context. Asking a LLM will always return the most likely text based on the training data and the context.
@Daojoan i thought we’d vanquished those dudes years ago! double drat…
@Daojoan Unfortunately LLM use has been shown to be accompanied an extremely quick (after just minutes of use) and drastic degradation of the users' critical thinking and problem solving skills. My guess is that the people producing this garbage can't actually tell anymore that it is very obvious to anyone who isn't using LLMs.
@Daojoan worse? I think they're just still the same since the first time I've tried for them to generate something. It's always pretty bland, even with a decent input on a certain style. The output is always... Yuck...
@Daojoan As a writer, I find this development delightful. 😁✌️🤘🖖
@Daojoan Degenerative Ai.
IIRC Google Translate was shut down for a while in 2012 or thereabout because it started to feed on the huge amounts of its own crap translations from China, thus turning worse.
@Daojoan I have seen how the LLM "voice" has drifted steadily towards basic mansplaining.
@Daojoan I guess it's because LLMs will never be crestive. They merely regurgitate the data they have stolen. So they can only spit out the banal.
