Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Longwriter – Increase llama3.1 output to 10k words

github.com

Post reply on HN
Print view View original

21–30 of 30 posts

Re: Longwriter – Increase llama3.1 output to 10k words

#21

Post by ben_w »

mmaunder wrote: post #7

What the difference between this and using chat history to concatenate outputs and prompting with something like “Now write the next section” repeatedly? I’ve done that with NotebookLM and it’ll write a complete fictional story based on sources, for example.

My experience is that by default it's like that game where you fold some paper, draw just up to the fold, pass the paper to the next player and they continue, and then you unfold it and see what came out: https://i.kym-cdn.com/photos/images/facebook/001/782/166/960...

With editing or lots of regeneration it can still work, of course.

(That said, I last tried this with the original ChatGPT…)

Re: Longwriter – Increase llama3.1 output to 10k words

#22

Post by roenxi »

elfelf12 wrote: post #14
vessenes wrote: post #6

The sample output is interesting - it has highly suggestive chapter titles which read like pretty normal story beats. It seems like it's guiding itself on these, then able to chunk out longer form writing per chapter. For what it's worth, the writing is .. bland. In the way that only an LLMs writing can be -- relatively grammatically sound, and totally soulless. I will never think of the love story of Elizabeth and T…

Is it a copyright problem or a capitalist problem or why do we only get nerfed dumb chatbots? Would be interesting to really try hard and create a llm that can write novels in the style of an author. And skip the chat functionality!

Copying writers is probably a copyright thing. But the experience with generative AI for images was that, at least for the early models, it was good to put things like "masterpiece, highest quality" in the prompt. The model biased towards average rather than trying to maintain a high standard. The more general problem here could easily be that people haven't figured out how to prompt interesting writing from an LLM yet.

Although my personal theory would be that LLMs are just writing how someone without an ego or firsthand knowledge would write - it has a bunch of different angles it could take but has no particular reference to draw on to determine which is true. Great human writers are often cataloguing their extra-literary experiences. How is ChatGPT supposed to be inspired by a beautiful sunset to capture it in a way that has never been done before? It is capable of the writing part, but the inspiration part is a lot harder for it.

Re: Longwriter – Increase llama3.1 output to 10k words

#23

Post by numpad0 »

elfelf12 wrote: post #14
vessenes wrote: post #6

The sample output is interesting - it has highly suggestive chapter titles which read like pretty normal story beats. It seems like it's guiding itself on these, then able to chunk out longer form writing per chapter. For what it's worth, the writing is .. bland. In the way that only an LLMs writing can be -- relatively grammatically sound, and totally soulless. I will never think of the love story of Elizabeth and T…

Is it a copyright problem or a capitalist problem or why do we only get nerfed dumb chatbots? Would be interesting to really try hard and create a llm that can write novels in the style of an author. And skip the chat functionality!

LLM can continue anything, chat is simply what worked best so far. Outputs being bland and soulless, and lacking in global structure, if I may add, is just architectural. There's nothing behind that.

Re: Longwriter – Increase llama3.1 output to 10k words

#24

Post by MintsJohn »

vessenes wrote: post #6

The sample output is interesting - it has highly suggestive chapter titles which read like pretty normal story beats. It seems like it's guiding itself on these, then able to chunk out longer form writing per chapter. For what it's worth, the writing is .. bland. In the way that only an LLMs writing can be -- relatively grammatically sound, and totally soulless. I will never think of the love story of Elizabeth and T…

Interesting notion, I notice the same with image models, less stylistic more blandness on the latest generation. Only MJ seems to have style as a feature.

Re: Longwriter – Increase llama3.1 output to 10k words

#25

Post by 6510 »

roenxi wrote: post #22
elfelf12 wrote: post #14

Earlier quoted context omitted.

Is it a copyright problem or a capitalist problem or why do we only get nerfed dumb chatbots? Would be interesting to really try hard and create a llm that can write novels in the style of an author. And skip the chat functionality!

Copying writers is probably a copyright thing. But the experience with generative AI for images was that, at least for the early models, it was good to put things like "masterpiece, highest quality" in the prompt. The model biased towards average rather than trying to maintain a high standard. The more general problem here could easily be that people haven't figured out how to prompt interesting writing from an LLM y…

There is nothing between the lines.

Re: Longwriter – Increase llama3.1 output to 10k words

#29

Post by Netcob »

hshshshsvsv wrote: post #28

Will this ever be able to generate anything meaningful or useful?

We'll probably see a lot of this:

Alice writes short prompt -> Alice's LLM generates oversized slop -> Bob's LLM generates summary -> Bob skims the summary.

Some information may be lost in this process due to hallucinations and errors, but Nvidia and OpenAI will make a lot of money causing and fixing problems.

Re: Longwriter – Increase llama3.1 output to 10k words

#30

Post by theturtle32 »

roenxi wrote: post #22
elfelf12 wrote: post #14

Earlier quoted context omitted.

Is it a copyright problem or a capitalist problem or why do we only get nerfed dumb chatbots? Would be interesting to really try hard and create a llm that can write novels in the style of an author. And skip the chat functionality!

Copying writers is probably a copyright thing. But the experience with generative AI for images was that, at least for the early models, it was good to put things like "masterpiece, highest quality" in the prompt. The model biased towards average rather than trying to maintain a high standard. The more general problem here could easily be that people haven't figured out how to prompt interesting writing from an LLM y…

The notion that generating something "in the style of" a human creator is a violation of copyright is categorically false. Copyright is only infringed when a work is substantially copied. Generating something new but with a similar feel is fair game. That OUGHT to be (but, obnoxiously, seems not to be) universally uncontroversial.

Human creators might bristle and find it distasteful to have works automatically generated in a style they spent a long time honing, but it is most certainly not a violation of copyright.

Post reply on HN