llms

I use a large language model (LLM) sometimes to write code, but I never use one for writing. All words on this website are my own.

When I use one to write a feature, I use it like a search tool: I explore the possibilities and best practices, determine good organizational structure, define what functions should be made, and implement. Very often, I’ll scrap what an LLM has suggested. Sometimes, I’ll rewrite it myself, and other times, I’ll critique each block of generated code and ask myself and the LLM if it’s needed.

I prefer atomic, simple features. Many of my lua filters are like this. They’re short, and probably couldn’t be written much of any other way. I could also drop any or all of them, and have a fully functioning website.

on em dashes #

I like em and en dashes. When I write one, I press Ctrl+Shift+u, then 2014 or 2013, then Enter for an em or en dash, respectively.

my beliefs about LLMs #

I hold many seemingly contradictory beliefs about LLMs. I agree with LLM haters about nearly every critique, yet I find LLMs surprisingly useful for learning about niche interests, which many of my interests are. Here are some different beliefs I have about LLMs.

• LLMs are addictive. Use one too much, and it’ll be the first thing you reach for when solving a problem.
• LLMs are very often wrong.
• LLMs replace culture and socialization.
• LLMs affect how we think. Using them has affected my thinking in both good and bad ways, but I think it’s mostly bad: reliance and laziness.
• LLMs modify code in mostly additive, complexity–increasing ways. See The Tower Keeps Rising (warning: written by the founder of an LLM company). The author doesn’t say that’s a bad thing, but I do.
• LLMs are very bad for security of the user.
• LLMs are very, very bad for the privacy of the user.
• LLMs are probably very, very, very bad for the security of everyone who uses computers which run on LLM-generated infrastructure.
• LLMs harm the environment. My understanding is that it’s not just training, but inference which causes the harm.
• LLMs aren’t trained on stolen work. You can’t steal something that can be copied. But I it’s the obligation of those who copy to share their modifications. Therefore…
• LLMs should be released with open weights, open source code, and open data.
• I don’t use the term AI, unless speaking about the marketing term and cultural phenomenon. We shouldn’t agree to tech companies’ framing. A convolutional neural network that helps radiologists identify cancer is nothing like ChatGPT, and grouping them together in our speech is giving OpenAI free marketing.
• Using an LLM helps me accomplish tasks that I would have difficulty doing by myself.
• Using an LLM helps me learn me about the work I’m doing by doing it, rather than reading about it. Rather than learning the general case, and then applying it to my use case, I learn in the context of my use-case, and learn to apply it generally.
• Using an LLM helps me when the documentation on an interest I have is poor, which is all the time.

I think using LLMs is one of the least principled things I do. It reminds me of someone who strongly opposes factory farms, but still buys from them since “there’s nothing wrong with meat in principle”. I can imagine a platonic ideal LLM, that runs locally and efficiently, and for some reason I feel that gives me permission to keep using one, even while it’s far from that ideal. Which is ironic, because I wouldn’t apply that logic to animal agriculture or buying fast fashion or picking a job.

LLMs are easy and fun. They help me realize some of my dreams, this website being one of them. But their presence in our world also causes a lot of harm. I hope the harm can be mitigated. I hope open source models win in the end. I hope that a group working in the commons will develop a small model, with low knowledge and high skill, which can run on a simple laptop. I also hope that this hope I’m feeling doesn’t cause me to look away from the harms that the LLM industry causes today.

Things I’ve heard, which I’m thinking about #

Someone on mastodon said that all LLMs are privacy invasive. Not just the closed ones run by sketchy silicon valley companies, but locally-run open weights models too. My reflex is to disagree, but I feel like they’re appealing to something which I don’t understand. My best guess is that since people have the right to be forgotten, and LLMs remember facts about people on the internet so long as those models exist, you can’t respect the right to be forgotten unless you delete models trained on those people.

“There will never be an AI tool that is truly private unless it hasn’t trained on nonconsensual data.” This I agree with, strongly, but I worry this logic is being applied too broadly. People should be able use the internet in a way that doesn’t leave a public footprint. Private social media pages and group chats should never be used to train an LLM. But commercial content doesn’t have the right to privacy like people do. If something’s out there but it’s paywalled or copyrighted, it’s by definition not private. The right to make money on your output on your terms is not the same thing as the right to privacy. Information should be free. Copying isn’t stealing.