My take on AI has been fairly straightforward: some of it is genuinely useful, some of it is interesting but overhyped, and some of it is dogshit. There are also ethical concerns on top of the basic functional concerns: every major LLM has been trained on stolen data, the energy costs of training LLMs remain terrible, and the major companies in the AI space are openly supportive of a fascist regime that murders children. So I use LLMs sparsely; I refuse to give money to the companies that offer them; and when OpenAI or Anthropic start ad-subsidizing their free models, I'll stubbornly go to the effort of finding an approach to consulting LLMs that keeps money out of their hands. 

If you are staunchly zero-tolerance on LLMs, or if you feel the use of LLMs in any capacity instantly makes you complicit in the horrible goings-on of the world circa 2026, I completely understand and am cheering you on. I also apologize, both for my occasional use of LLMs and for the ways that this piece is going to focus on them as a technology, rather than on how they fit into modern times more broadly. 

The interesting question to me about LLMs is ultimately: how are they going to matter? Where are they going to add genuine, lasting, sustainable value to people's lives, in ways that are—by my standards—not completely shit? In other words, what is its potential to make the world (and people's lives) legitimately better?

I'm finding that my answers to that question follow predictable patterns. My hunch is that, over time, these will be the places where we see sustained interest in LLM use, and how we'll eventually come to understand the technology. The other, shittier uses won't go away overnight, but I think we'll see decreased interest in certain areas, less evangelizing hype, and fewer people succumbing to the irritating madness that we have so much of these days.

Let's go over those territories, and why they matter.

I have no interest in LLM summaries of things. If I'm engaging with a thing, I want to engage with the thing—not reduce it to regurgitated pulp that I can digest in a slurry. What I do have interest in is LLMs' ability to scan videos, not to summarize, but to locate: to help me find specific points within 30+ hours of footage that I genuinely want to find.

"What LLMs are technically good at" is a different question from "how LLMs' proficiencies can be applied." LLMs are good at parsing a tremendous amount of data very rapidly, full stop; they are not perfect at it, but that just means you shouldn't rely on them in situations where perfect accuracy matters. They are unreliable reference tools, because you need information to be accurate when you're looking it up. They're capable of summarizing texts and videos, but I don't find those summaries particularly useful.

Location, on the other hand, is incredibly useful. There has never been as useful a tool for hunting through copious amounts of data for specific items that are imprecisely defined. Sure, you can locate an exact phrase in a body of text, as long as you spell it correctly; you cannot, on the other hand, use any other search tool to find a scene or a moment that you can only vaguely describe. The tool that you would need to accomplish that is... a large-language model. Now they exist and are publicly available, and that's neat!

Is it a life-changer? Not really. But it can save minutes or even hours of your day, and that's really nice.

Relatedly, we have:

LLMs are great at scanning code directories. They're also great at identifying exactly why something that I think ought to be working is not, in fact, working. Programming languages are precise enough that LLMs can make sense of how they're operating; if I'm struggling to make sense of a bug, an LLM can often tell me exactly what's going wrong and how to fix it.

Is it perfect? Again, not really. But that's fine: if it catches 80% of my mistakes with a single query, that's already hours saved. And if its analysis is wrong, I'm no worse off than I would have been if I'd hunted down an answer on Stack Overflow that didn't help. What's more, it understands context enough to refine its answers if I want it to help me hunt.

As a programmer, finicky bug hunting is the least satisfying part of my job. I've often joked that my salary isn't there for the things I can design or build: it's there to keep me from storming out of an office and becoming a mountain man every time I spend three hours staring blearily through the same 200 lines of code, trying to figure out what's gone wrong. I am okay with LLMs taking that part of my job away. And I'm sure that there are other dreary parts of other jobs that people would love to make go away too.

I love math in the abstract. I also love the kinds of wizardry that math will let you do, as long as you understand it well enough to create something nifty. But I do not, generally speaking, derive a lot of satisfaction out of spending hours trying to calculate a bunch of different vectors in order to arrive at an ultraprecise result.

This is more-or-less the only kind of "GenAI" that I've touched. It's confined to a specific, limited need; it handles the underlying mathematics for a process that I still need to articulate precisely. If it's off—and it is sometimes off—I often find that I need to fix it by hand, because the LLM struggles to understand what's going wrong. But that's fine: if it does 90% of the work, I'll happily pick up the 10% at the end.

It's been noted that LLMs excel at building what we used to call applets: tiny little bits of code that serve a single purpose and work acceptably well. Do they write maintainable code? Not really—so you want to use them only for little bits and bobs that you won't need to touch again. Tiny helper functions? Probably fine. Actual architecture? Maybe not.

I've heard some developers say that they find LLMs useful for scaffolding new projects, for a similar reason. "Generate the basic architecture for a service or plugin" can give you a nice base template to start from, so you're spending your time less on technical bureaucracy and more on the actual detailed logic of a thing.

I think it's telling that LLMs seem to work best with little intricacies or with big vaguenesses. Again, it comes down to what these technologies actually are: they are approximators, great at translating a precise concept into mathematical terms or at constructing the loose shape of a bigger thing. This doesn't replace a programmer's job, in the same sense that importing other people's code libraries doesn't replace their job either. A part of the job is understanding when to use a certain tool: when you want to use an existing library or framework rather than writing your own; when an LLM will work more quickly and effectively than you would. 

That kind of executive call is similar, in some ways, to creative decisions: it asks you to do something that humans are very good at doing and that computers fundamentally can't do, no matter how hard they try to approximate it. And programming itself is an interesting blend of executive strategizing and creative choice: the ambiguity between the two opens up spaces where LLMs perform quite well, but it highlights that LLMs themselves need to be viewed through both executive and creative lenses. When can it create what you need it to create? When is using it the best possible choice you could make? The agency rests on the person using it; the more deeply they understand how LLMs work, the more easily they'll be able to use them appropriately and sanely.

During the Passover seder, four types of question-askers get described. There's the wise child, the wicked child, the simple child, and—last but not least—the child who doesn't know how to ask.

Expertise boils down to knowing what questions to ask. And one of the daunting things about trying new things is that you don't understand them well enough to even begin asking questions: if you know how to get started, you're in the clear, but what happens when you don't even know how to start?

I like trying new things. But I've never been good at that "learning how to start" bit. I stall out; I get frustratingly inarticulate; I can't seem to figure out how to think about this new territory at all, for all that I learn very quickly once I begin. Neither computers nor books, historically, have been particularly useful for me: I can't phrase my questions precisely enough for a search engine to help me figure myself out, and books and their vastnesses rarely click with me at the start. (Once I get going, I'll devour a dozen books on a subject in a week, but they're not very useful for that first big push.)

LLMs are weirdly effective at helping with all this. They're the first tool that can take my vague, confused hand-waving and help me translate it into something more precise. Partly that's because they usually try to organize my vaguenesses into precise terms, but they have a far more useful ability: they can categorize my imprecisions themselves, and ask me questions about those ambiguities that help me narrow my thoughts into something more specific.

Once I hit that point, I don't usually need an LLM anymore. They've served their purpose: they've helped me out of my own vagueness. At that point, I can use more technically precise tools to figure out what I'm doing; I can form searches and track down books and hone in on what, exactly, I'm trying to do. But that initial vagueness has been a stumbling block for me on a lot of projects in a lot of different ways, and having a tool that can help me work through that is extraordinarily neat.

That's the appeal, I suspect, for people who ask an LLM to help them plan vacations somewhere. It's not that they want a machine to build their itinerary for them: it's that they don't know how to start constructing an itinerary, and have finally been given a tool that will work through that process for them. To me, that's a deeply unappealing prospect... but that's because I know how to plan a vacation already. People who don't have that skill don't always know how to go about developing it within the kind of timeframe that they're devoting to figuring out their travel plans.

But we don't have to be content with the way that people are currently using LLMs. There's nothing stopping us from developing a tool that teaches people how to plan vacations, using LLMs to pinpoint the places where they struggle and walking them through the process until they know how to do it themselves. For that matter, that's true of a lot of territory where people use LLMs to achieve unsatisfying results. It comes back to that difference between summarizing and locating: are LLMs filling in a gap with regurgitations and vague approximations, or are they helping us find a needle in a haystack? Are they helping us figure out difficult things, or are they trying (and mostly failing) to do those difficult things for us?

What's fascinating, and consternating, is that the better tools that LLMs could be used to power already exist within the current LLMs. If you use an LLM correctly, if you approach it for the right reasons, if you ask it for the right things, they can be tremendously powerful. But they don't teach you how to use them effectively. Figuring that out is a human task, not an AI one. And the people who are currently evangelizing LLMs are pushing the worst and bullshittiest uses of them, in part because their deepest usefulness is more limited in scope, less megalomaniacal, and feels genuinely like a tool or aid than like the revolution that the bullshitters are trying to promise.

I want to differentiate between LLMs as a technology and LLMs as an industry. The current industry is horrifically bad; the culture that it has spawned is horrendous too. It's territory for bullshitters and grifters and arrogant idiots, many of whom have unfortunately received tremendous sums of money from people who are arrogant, idiotic grifters themselves. The technology, on the other hand, is very interesting—and there is still an open question as to what form that technology will take. The chatbot format has captured the zeitgeist, and aspects of it will likely remain, but nothing says that that's what LLMs have to look like or how they ought to act. And the interfaces that we design for them will influence our relationships to them, which will in turn yield brand-new kinds of LLM-powered technologies that we're not ready to imagine yet. (In part because the people currently fueling the visions of an AI future are, by and large, dim-witted and awful.)

In their present form, LLMs can be genuinely very useful, especially within the kinds of domains I've articulated above. (I suspect that that "locate versus summarize" hair-splitting will eventually become one of the major ways that we think about where LLMs have genuine utility, and where they become unreliable or unappealing.) To the extent that you can feel ethically okay with using LLMs—and as I said at the stop, I completely understand if that extent is "none whatsoever, ever"—it's worth trying to learn where they serve a genuine purpose. Eventually, I think, the "correct" uses of LLMs will win out; whether that happens in two years or twenty, I couldn't say. But it's neat that, by their nature, you can explore those uses already: they'll operate as the kind of tool you want them to be. And until we learn how to offer LLMs in the forms of better tools to begin with, it'll be a useful skill to understand how to operate LLMs as those kinds of tools here and now.