I recently turned 40. As one does after a major round-number birthday, I've been ruminating. And as I settle into what statistically must be the second half of my life, I'm thinking carefully about the things that actually matter to me, the things that are worth my time and energy.
Cliché I know, but I'm going to savor it for a moment. Ruminating almost always merits my time and energy. I have never regretted a good ponder.
Saving time is supposed to be the primary value proposition of AI. Let the machines do the slog work so you can focus on what matters. Although "what matters" to the purveyors of the technology seems wildly contrary to what matters to me.
I do not want to let the machine plan my travel. Find deals, sure, but that service exists without city blocks of GPUs. Booking my flights, lodging, rentals, etc.—I want to make those choices. And even if I let the model try to do that work for me, I'd have to manually check the work anyway to confirm nothing dangerously stupid occurred. And so the time I was supposed to save is now spent checking the model's work, and likely redoing some of it myself.
Repeat that experience for party planning, gift giving, writing complicated correspondence, and you have the bulk of the use cases for generative AI pitched by its creators—outside of software development, anyway. We'll get to that soon. And these quotidian tasks are what users seem to most use LLMs for.
Of course there are certain tasks that the tool can automate with reasonable success. Almost all of them are low-stakes. The truly important ones, the ones with value and impact, cannot yet be entrusted to the models. Perhaps not ever.
Imagine leaving your tax filing to AI. What level of performance should users require to allow the models to handle their money? I'm talking about normal users here, not the zealots with seemingly endless disposable income to burn on LLM misadventures.
I do notice a pattern to the use cases proposed by AI companies: they mostly all involve care, empathy, for and with others. To plan a trip with a partner or family, you must consider what they'd enjoy doing. You might even (😱) talk to them about their desires.
This is emotional labor. I don't think it's a coincidence that the overwhelmingly male tech industry has deemed this the work to surrender to machines. The dream of the personal assistant/wife/mother-as-a-service is alive and well in the terminally techbrained. To them, care and empathy are inconveniences to be automated away. Such a deranged set of priorities only makes sense in a world where business output is the highest moral good.
I firmly believe that such a world, whether it exists in reality or solely in the minds of capitalists, must be burned to the ground. Life is not for work—not that kind of work, anyway.
And this dehumanization of effort appears to be the extent of the vision for this "transformational" technology in the here and now. Again, beyond software development, the use cases are almost ridiculous in the disparity between energy required and benefit conferred. And even if the busywork of life were automated away, what is the intended use of the additional time? It's plainly not leisure or time with family, since we've seen exactly no reduction in work hours (layoffs notwithstanding) due to generative models. Somehow, whatever time is saved by AI seems destined for more work, more "productivity."
But not more compensation. Perish the thought.
What, then, is the point? What have we won with this automation?
"Ah but it will get better," you say. "Just look at coding."
Believe me, I have. In fact, studying the use of these tools for software creation is a good portion of my day job.
It is undeniable that the models have improved in the realm of code generation. But we must resist the software engineer's fallacy that their domain is so rarefied, so complex, that facility with it necessarily translates to facility in all other fields. Conversely, I contend that the highly structured and predictable nature of programming languages makes the field uniquely suited for LLM application. If they can do anything well, it should be code. But even the latest OpenAI model scores in the low to mid 60% range on benchmark tests. Now granted, no human being could do better across those entire benchmarks, but if the objective here is to remove the need for all experts, we're a long way off.
The objective here is to do things better than a human, and by "better" I do not mean "faster." We do not require the models' speed to build quality software. It is unclear whether the attendant costs of that speed allow us to maintain secure, high-quality software if LLMs are the primary means of production.
Are the models' solutions more elegant? No; the models produce highly verbose solutions that are difficult to maintain and often suffer from "feature creep," adding bloat and complexity to the output.
Are the models' solutions more creative? Indeed, they can't be, since their very architecture will regress to the mean of their training data when generating responses.
Are the models' solutions more secure? Weirdly no—despite advances in finding vulnerable patterns in some types of code, models seem dreadfully locked in to the insecure patterns they've learned from example code. Perhaps they can't be blamed ("I learned it by watching you!") but again, this tool is supposed to exceed our capacity, not repeat our mistakes.
Ah, but I'm getting hung up on whether the tool works. I don't care whether it works right now; I care why we want it to.
We're going to leave aside any claims of humanity-reshaping inventions or discoveries. That is not why ChatGPT is a product, nor why your boss insists on everyone using these tools. The utopian promise justifies the sprawl, but the sprawl is an objective unto itself. It's entirely possible that Sam Altman and Dario Amodei really believe they're building God. In the meantime, they also noticed—and so did everyone around them—that people keep giving them money to try to make the employee replacement machine.
What is it, really, that we're trying to get done so much faster? And why is it good that it's fast? If the answer boils down to "number go up," it's a fairly intellectually bankrupt exercise. Most of the numbers going up have nothing to do with the prosperity of humanity, save for a handful of elites who were already doing better than they had any right to be.
And frankly, it makes the craft so much less enjoyable. That matters, in my opinion—especially if corporate revenues are not the end-all.
The models are faster at finding vulnerabilities in software. Occasionally (but not nearly as much as is hyped), they produce exploits for those vulnerabilities. Also occasionally, they write patches for them. I don't actually know that finding all these faster is a net good. I know that since the technology now exists, there is a sort of escalatory proliferation happening. The attackers will use models to find 0-days; therefore, defenders must use them to defend against them. Seems like the problem and the solution are the same thing, which tends to be a hustle.
The blessing of speed, even without LLMs, is not without risk. The recent TeamPCP supply chain attacks showed just how dangerous speed can be. Not only the speed of the attacks, but the automated CI/CD pipelines that enabled their spread. Introduce to this world the uncertainty, the chaos, of LLMs, and what have we gained? And for whom are we running so fast?
The old engineer's adage still applies: good, fast, cheap—you can have two. And so does the Navy saying: slow is smooth; smooth is fast. Deliberate effort saves time in the long run. I don't see LLMs changing that.
So, a humble recommendation: slow down. Wherever you can, slow down.