this post was submitted on 24 May 2024
255 points (98.9% liked)
Technology
59135 readers
2532 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
From my experience with Llama models, this is great!
Not all training info is about answers to instructive queries. Most of this kind of data will likely be used for cultural and emotional alignment.
At present, open source Llama models have a rather prevalent prudish bias. I hope European data can help overcome this bias. I can easily defeat the filtering part of alignment, that is not what I am referring to here. There is a bias baked into the entire training corpus that is much more difficult to address and retain nuance when it comes to creative writing.
I'm writing a hard science fiction universe and find it difficult to overcome many of the present cultural biases based on character descriptions. I'm working in a novel writing space with a mix of concepts that no one else has worked with before. With all of my constraints in place, the model struggles to overcome things like a default of submissive behavior in women. Creating a complex and strong willed female character is difficult because I'm fighting too many constraints for the model to fit into attention. If the model trains on a more egalitarian corpus, I would struggle far less in this specific area. It is key to understand that nothing inside a model exists independently. Everything is related in complex ways. So this edge case has far more relevance than it may at first seem. I'm talking about a window into an abstract problem that has far reaching consequences.
People also seem to misunderstand that model inference works both ways. The model is always trying to infer what you know, what it should know, and this is very important to understand: it is inferring what you do not know, and what it should not know. If you do not tell it all of these things, it will make assumptions, likely bad ones, because you should know what I just told you if you're smart. If you do not tell it these aspects, it is likely assuming you're average against the training corpus. What do you think of the intelligence of the average person? The model needs to be trained on what not to say, and when not to say it, along with the enormous range of unrecognized inner conflicts and biases we all have under the surface of our conscious thoughts.
This is why it might be a good thing to get European sources. Just some things to think about.
If the social biases of the model put a hard limit on your ability to write a good woman character, I question how much it's really you that's "writing" the story. I'm not against using LLMs in writing, but it's a tool, not a creative partner. They can be useful for brainstorming and as a sounding-board for ideas (potentially even editing), but imo you need to write the actual prose yourself to claim you're writing something.
I use them to explore personalities unlike my own while roleplaying around the topic of interest. I write like I am the main character with my friends that I know well around me. I've roleplayed the scenarios many times before I write the story itself. I'm creating large complex scenarios using much larger models than most people play with, and pushing the limits of model attention in that effort. The model is basically helping me understand points of view and functional thought processes that I suck at while I'm writing the constraints and scenarios. It also corrects my grammar and polishes my draft iteratively.