Zos_Kia

I’ve only had issues with fitgirl repacks i think there’s an optimisation they use for low RAM machines that doesn’t play well with proton

Interestingly the pendulum is now swinging the other way. If you look at next.js for example, server generated multi page applications are back on the menu!

I’d place it right around when angular started gaining traction. That’s when it became common to serve just one page and have all the navigation happen in JavaScript.

If I understand these things correctly, the context window only affects how much text the model can “keep in mind” at any one time. It should not affect task performance outside of this factor.

Yeh, i did some looking up in the meantime and indeed you’re gonna have a context size issue. That’s why it’s only summarizing the last few thousand characters of the text, that’s the size of its attention.

There are some models fine-tuned to 8K tokens context window, some even to 16K like this Mistral brew. If you have a GPU with 8G of VRAM you should be able to run it, using one of the quantized versions (Q4 or Q5 should be fine). Summarizing should still be reasonably good.

If 16k isn’t enough for you then that’s probably not something you can perform locally. However you can still run a larger model privately in the cloud. Hugging face for example allows you to rent GPUs by the minute and run inference on them, it should just net you a few dollars. As far as i know this approach should still be compatible with Open WebUI.

There are not that many use cases where fine tuning a local model will yield significantly better task performance.

My advice would be to choose a model with a large context window and just throw in the prompt the whole text you want summarized (which is basically what a rag would do anyway).

And certainly not as spooky as spectrography

Alternative interpretation cause i find i18n extremely boring and hate the indirection it adds to a code base : you’re telling me I can start making an app without this hassle, and it will only cost me a 2Kloc PR some time in the future. That’s a totally manageable price to pay and makes the early dev experience much better (which can have a lot of impact on momentum).

If you like to write, I find that story boarding with stable diffusion is definitely an improvement. The quality of the images is what it is, but they can help you map out scenes and locations, and spot visual details and cues to include in your writing.

I legit don’t get it. Is it about the US ? I mostly speak about France in my political comments so i’m not sure where they are going with that.

oh yeah silly me. That’s definitely not creepy 👍

hey you’re the creepy guy who reads comment history before replying to a conversation aren’t you ?

Am i getting downvoted ? It says 3 upvotes / 0 downvotes on my end.

What’s “eagleflavoured” ?

No i’m saying comedy (as in writing your jokes for you) is not something you should expect from language models. As a general rule, there is no tool that will make you a good writer, only (potentially) tools that can help you do more with your qualities as a writer. But it will never be funnier or more talented than you are.

That’s why i personally experiment with writing tools. Writing standup is one thing, but imagine you’re writing a sitcom or any form of serialized work. That’s a lot of fucking work and obviously if you’re starting out you can’t exactly afford to pay for assistant writers to do the menial labour that comes with it. Language models can come in handy in that scenario, but again you can’t expect them to be the genius in the room if you want a good show you have to bring the good ideas and the funnies. It’s a power tool and power tools don’t draw the plans for the house they just grind where you need grinding.

More like “you’re trying to paint with a hammer AND you’re holding it wrong”

a true conversationalist lmao you’re doing great buddy

Define “good at writing”. Good comedy is very difficult to attain and none of the models are anywhere near it, including the more recent ones.

I’ve been experimenting on creative writing tools with a bunch of writer friends, and the setup described in this paper is notoriously shit. I mean they come up to ChatGPT on v3.5 (or Bard lmao) and expect it to write comedy ? Jeez talk about setting yourself up for failure. That’s like walking up to a junior screenwriter and yelling “GIVE ME A JOKE” to them. I don’t understand why people keep repeating that mistake, they design experiments where they expect the model to be the source of creativity but that’s just stupid.

If you want to get output that is not entirely mediocre, you need something like a Dramatron architecture where you decouple various task (fleshing out characters, outlining at the episode level, outlining at the scene level, writing dialogues etc…) and maintain internal memory of what is being worked on. It is non-trivial to setup but it gets there sometimes - even the authors of this paper recognize that this would have probably produced better results. You also need a user able to provide good ideas that the model can work with, you can’t expect the good creative stuff to come from the robot.

Instinctively i’d say you have to treat the model like your own junior writer, and how do you make a junior writer useful ? By teaching them to “yes, and” in a writing room with better writers (in this case, the user). In that context, with a good experienced user at the helm, it can definitely bring value. Nothing groundbreaking but i can see how a refined version of this could help, notably with consistency, story beats, pacing, the boring stuff. GPTs are better critics than they are writers anyway.

That being said i never really pursued “pure comedy” on LLMs as it sounds like a lost battle. In my mind it’s kind of like tickling : if a machine pokes your ribs you don’t get the tickles, that only works when a human does it. I doubt they can fix that in the short or mid term.

holy shit you’re right i don’t know where i got the idea that it was the same format