But What Is It *Good* For?

Last updated: Mon Feb 10 2025

Today I’d like to talk about AI again. Feel free to come back next time if that doesn’t interest you 🙂

‘Leibniz’s Calculating Machine’ from Theatrum arithmetico-geometricum (1727) | Public Domain Image Archive

Here is, roughly, my ethical standpoint:

• LLM output is fair use. I buy the argument that LLM output is transformative and de minimis — any given set of output tokens has a barely-discernible relationship to the content the LLM was trained on, unless the LLM is carefully prompted to output verbatim content. As a writer and programmer, LLM output doesn’t feel competitive — hopefully you’re reading this because you want to hear from me! Ethically, I don’t feel that I’m “stealing” from any one particular writer or from culture as a whole when I use LLM output.
• ‌LLM training is sketchy, but we need a stronger public domain. Although the output of LLMs feels fair use, the input to LLMs is more complex. LLMs were likely trained on massive piles of pirated books, the legality and ethics of which are unclear at best. I’m waiting for the courts to decide whether that was legal and what an appropriate punishment would be, while still using LLM products due to the above bullet point about fair use. That said, I’ve always been a proponent of a stronger public domain, which would make training “vegan” models vastly easier. When or if a vegan model comes into widespread use, I’ll switch.
• I’m concerned about the environmental cost of LLM usage, but not that concerned. As Simon Willison points out, training the largest Llama 3 model cost about the same amount of energy as a single digit number of fully-loaded jumbo jets from New York to London. Inference costs are dropping, thanks to scaling laws. The multi-billion-dollar arms race to build out data centers will eventually crash, and in the meantime, the big tech companies building out data centers will likely end up relying on renewable energy. Everything productive requires energy — that’s the definition of energy!
• Cultural evolution may save us. I don’t feel ethically responsible for societally-harmful usage of LLMs, like generating slop or cheating on tests, any more than I feel responsible for knife murders because I use a chef’s knife. However, those uses are genuinely concerning, and it’s not clear that alignment research will ever manage to prevent them. I understand the urge to ban the technology to make those impossible. That said, setting aside whether a wholesale ban is even logistically feasible, it may not even be advisable. Culture changes fast, and we may find that within a generation a new set of cultural norms and technologies develops to adapt to these problematic usages while maintaining productive usages. That requires, however, that we continue to explore the technology.
• I’m agnostic on LLM consciousness. I don’t think LLMs are meaningfully conscious yet, but as an illusionist about consciousness, I find it frustrating when people handwave away the possibility of LLM consciousness by calling them “stochastic parrots” or referencing the Chinese room argument. That implies strong philosophical positions! These folks would benefit from reading “If Materialism Is True, the United States Is Probably Conscious”.
• I’ll avoid using image or video generators. They run the risk of cannibalizing paid creative work for working artists in a way that feels less true for LLMs, which are closer to general reasoning engines. However, I’m not going to boycott a movie because the animators generated buildings for a background shot or boycott a game because the technical artists used generative AI to produce sprites — whether these tools have a productive, non-disruptive place in a working artist’s toolkit is an open question I don’t feel qualified to judge. In any case, I’ll avoid using these tools in the future, though I’ll keep up the examples of Stable Diffusion output I’ve included in newsletters before.
• I’ll continue to avoid Perplexity, for reasons I’ve previously outlined. I don’t trust them to behave ethically in this morally-complex space.

‘Sinclair Lewis at the Wheel’ from The World’s Work (1921) | Public Domain Image Archive

But what are these models good for? I would consider myself a “technical skeptic” per this useful taxonomy, a la Simon Willison (you’ll note I cite him often). Although we should remain skeptical of these tools, they are genuinely useful, not just for writing AI slop and cheating on tests! I almost exclusively use Anthropic’s Claude 3.5 Sonnet — I gravitate towards Anthropic’s products for aesthetic reasons, and industry insiders I know argue that Sonnet is the best general-purpose LLM available today. Here are a few of my uses:

• Solving tip-of-the-tongue syndrome: Sometimes I’ve forgotten a word or phrase that I want to use as a synonym or reference. The current generation of LLMs are spectacularly good at figuring out what word I intend. For instance, this morning I wanted to make a joke about Nesquik, but I forgot what it was called. I typed out a message asking about “chocolate mix that’s popular in China,” and after first suggesting chocolate mixed with dark soy sauce, Claude correctly identified Nesquik as what I had in mind. This can also be useful for figuring out terms you don’t know, like “what’s the term for a charcuterie expert?” (a charcutier).
• Expanding acronyms and slang: Claude is (surprisingly?) good at figuring out acronyms and slang. I was recently reading an article that mentioned a KVM, and I couldn’t figure out what it meant in that context, even by checking Google. Claude, meanwhile, happily informed me that it’s short for IP-based Keyboard, Video, Mouse and is a network device that lets you remotely control your computer.
• Programming: This could mean something fancy like Cursor, or it could mean repeatedly asking Claude to output code. In any case, even for small projects I’m technically capable of writing myself, using an LLM can move a project from a whole-weekend endeavor to a relaxing 20-minute task. Claude is also useful for asking simple questions about how CSS or JavaScript works.
• First-pass copyediting: An LLM is nowhere near as good as a human copyeditor — I wouldn’t rely on it for anything professionally published. But for a blog post or newsletter, Claude is more than capable of catching sentences that are incomplete or confusing. I could catch these myself or ask a friend to proofread, but asking Claude takes seconds.
• Brainstorming: An old tip from Simon Willison is to ask an LLM to brainstorm ideas, but provide a few dozen, not just one or two. Although its ideas will be boring, you’ll sometimes find a few gems that inspire you.
• Media recommendations: Since Claude is tuned to output the statistical average of the internet, it’s good at giving middle-of-the-road recommendations for media. If you want to know the essential Decembrists albums, or hear about the five surrealist novels that “everybody has to read”, Claude serves better than trying to parse a Wikipedia page or find an appropriate Pitchfork article.
• Writing equations: I include mathematical equations in my spaced-repetition flashcards app, but it uses MathJax to render, which requires painfully writing out the equations in LaTeX. Given a hand-drawn picture of the equation, Claude can output the appropriate LaTeX in a copy-pastable box.
• Parsing and converting simple data: If you have a messy, unstructured blob of data, an LLM can extract data from it and output it in a usable format like CSV.

It’s important to keep in mind that LLMs are not perfect. They are decent at all these use cases, but you have to pay attention to the output. As you use LLMs more, you’ll gain mechanical sympathy¹ — you’ll “feel” when they’re starting to hallucinate or when their output doesn’t make sense. LLMs are a tool — they’re not a substitute for thought.