Sensitive content
I've been fiddling around a lot with a local copy of DeepSeek-r1. I've come across a lot of limitations and things it can't do, but that's not what I'm here for today.
There's something LLMs are surprisingly good at. Workshopping stories. That is to say, formal criticism, editting, and so on.
Why this, retrospectively, makes sense. It's an actual pure language task. The system doesn't need to know very much about the world, but about how words come together. Hallucinations, if they ever arise, are not a significant problem, since they are obvious within the secondary world of the story being worked on.
In this sense, they can be used like a sort of more potent thesaurus, (give me three ways to say this).
Other useful things. Finding dialogue beats (all the cinematographic-style frowning, blinking, etc). That's especially hard for me as someone who can't see body language.
Refining snappy dialogue, keeping to the voice of a character. This is surprisingly successful. It really helps detecting dialogue that doesn't work and, if you're willing to let it write for you, it can come up with some wicked lines.
Structural criticism. Pacing, conflict, plot holes, uniqueness of character voices. Finding cliches.
Finding problems. Why is it a bad idea if my character does x? Does it make sense psychologically for my character to do y?
The most important aspect is, you don't have to agree with it, but it's still useful. It gives very fast feedback onr your text, obviously not as good as a real editor, but better than you can get rubberducking. Occasionally it will say stupid stuff, and you ignore it and move on.
I realise for many people using LLMs in their creative project is anathema. I understand and respect that. For me, it's not like that. Writers have been using tools forever: dictionaries, thesaurus, concordances, reference works... Why is it suddenly bad if the answer comes out of this tool instead?
Anyway, the immediate feedback turns out to be a huge thing for me. I'm able to write a lot faster by getting the illusion (and I am well aware it is an illusion) that someone has read my text.
Dire Threat to the People
in reply to modulux • • •Sensitive content
It's bad because the tool itself is bad.
Not bad, not-functional, or bad, poor results, but bad, ethically and environmentally.
You can't build an LLM without massive amounts of compute power, and while it's possible to *run* one with less, that still drives the demand for more, bigger, faster models. There's continual hand-waving about it not being a problem forever, but the data centers keep getting built and the power demands keep increasing and the "solution" doesn't seem to be getting any closer.
And then there's the issue of the data sources and models, not just the ethical problems of training on copyrighted data or personal information or private conversation, but also that the resulting models are typically bigoted and racist (which isn't surprising considering the source data.)
And finally, the result for you as an individual might be better writing - but if used widely, the result is stagnation.
modulux
in reply to Dire Threat to the People • • •Sensitive content
I'm familiar with the arguments. I find them somewhat unconvincing. But since you took the time to write them down, I'll reply in a bit more detail:
The tool itself is bad: bad for what? That's the thing. I'm not opposed to making judgements of this kind when warranted, but they should be made about people or specific uses of tools, rather than tools themselves. This is not to say tools are neutral either, a tool can have many more bad or problematic uses than good, and then it needs to be either regulated or prohibited, but going straight to calling a tool bad is more moralism than I can comfortably operate with.
You clarify bad in which sense (which is good). The ethical claims are unconvincing to me, and I'll explain why. The environmental claims are a very real issue and it does worry me. Thing is, we have a planetary challenge to decarbonise our energy system. This is not going away if all LLMs got shut down tomorrow. There are also significant pressures on improving the energy efficiency of these systems. But no, I don't have a good answer to the environmental argument. Do we need a moratorium until we reach certain percentage of non-carbon energy in the mix? I have no idea. I could support something like that.
There are tons of things you can't build without massive computing power and energy outputs. This is not per se an argument for not building them. It may be an argument for not building them now, though. I'd point out the same line of argument applied to computers in general was just as valid in the early days. They took massive amounts of materials and energy to do computations not much faster than humans could. But you can't get here without passing through there. Research can't skip to the final, efficient, functional stages of things.
Data sources and models: I do not find training with copyrighted data an ethical problem. It may be a legal problem (though the EU data mining exception might cover it, for instance) but law isn't ethics. I've never been favourable to copyright and artificial scarcity regimes enforced by state power and I'm not going to start now.
Model bias: absolutely, this is a problem. As you point out, it's a problem that's directly inherited from the data sources. One has to be mindful about this, and it is possible to nudge models into taking this into account. But the same logic for not using a tool like this applies for not using the sources it drinks from: dictionaries, thesaurus, encyclopedias, old (and not so old) books, random websites, talking to some random person... If you generalise this argument it sounds completely implausible, so in order to make it one has to distinguish why it applies to some tools and not others. I don't think this is impossible, but I haven't seen it done.
The last claim is relevant. Maybe this is a tool that helps me but its general use would be harmful. I could believe this, though I think I'd need more evidence that people won't make better use of such tools as they understand the limitations. If true, it still leaves me with a quandry though, because me not using it isn't going to make a difference in that regard.