Artificial Assistants or Price-Dumping Competitors?

Much of what is written below I’ve already voiced in various debates; now I’ve decided to gather it all in one place so I have something to link to when the occasion arises.

My bubble has split into two roughly equal halves: starry-eyed clowns using Claude in autonomous mode to churn out useless “apps” nobody asked for—and stubborn pessimists judging language model capabilities based on the state of mid-2020. The former love to proclaim: “I don’t write code anymore; I focus purely on the creative side, strategy, and architecture.” The latter somehow deny the obvious and won’t even trust a model to stack JSON neatly.

From outside my bubble come the screams of so-called “artists” who accuse models of plagiarism and lament lost income due to price-dumping. Let me start with them, because this part is much simpler. Fearing that models will steal jobs from actual professionals is laughable. I use generation extensively—I even tried making a video clip recently. Over the past year, all my libraries got clean, handsome logos, clear documentation, and a full suite of supporting texts. I used to try hiring professional designers for logos for my (open-source, I should emphasize) libraries. They want money (which is fine, I was ready to pay), but also—they don’t want (or aren’t able) to delve into what a specific library actually does, or why eye-searing yellow or a toddler’s crayon sketch won’t work for me. I belong to the category of clients who know exactly what they need, but lack the skill to execute it in color and form. I don’t give a damn about “creativity” or deep, arcane allusions. People aren’t going to stare at my library’s logo for hours; it’s not the Mona Lisa. In short, a model that renders a distinct, recognizable image in three iterations is three hundred times more reliable and convenient than a part-time “creative spirit.”

I have a complicated relationship with music: I can’t sing a melody pitch-perfectly, but I always know how it should sound. This is my lyrics, after all. And I care about them enough that I won’t just ask some casual acquaintance to strum a few chords for a couple hundred bucks “the way they see it.” I'm not even sure Max Martin would satisfy me in this context, but there’s no way to test that—I can’t afford Martin anyway. The model, on the other hand, will do it exactly the way I need it to, and it'll touch up my vocals in the spots where a voice crack overpowered my inner Cohen. And it’s still going to be my voice.

For English lyrics (and before that, for a few poem cycles in Russian written specifically for female vocals), I was fortunate enough to find actual human beings who resonated with the material and worked with me—for free, and until I was happy with the result. For, shall we say, important to me poems in Russian, I was handed the opportunity (a crutch—if you will—for a limping pseudo-musician) to do everything myself, without leaving the room, and without making a mistake. I've been writing poetry for over thirty-five years, but the chance to record actual songs set to those texts came only recently. I make use of this opportunity with deep gratitude.

There is, of course, a catch: if I didn’t know precisely what output I wanted to achieve, using models would be completely pointless. You can’t open a chat window, say “Write a brilliant love poem,” and expect anything remotely digestible. Finishing a stubborn quatrain in an existing poem (preferably in English)—that is entirely doable, and if it needs edits, they'll be minor. The operator of the model still needs a sense of rhythm and an ear for language—two qualities that generative-model-based assistants will never replace.

Would all these vocal artists and musicians work with me on a semi-pro-bono basis because they share my views on open source and accessible music? Well, so far I've found two musicians who truly connected with my lyrics, and one artist. I am more productive than what the world has to offer me as a supporting cast.

Now, as for programming.

Both categories of people mentioned above always get the exact same question from me: “Show me the code.” Either the code you blissfully generated as a “strategic architect,” or the code you churn out by the ton yourself without any assistants. And somehow neither camp is capable of producing even a single lousy open-source library. I've been saying for a long time that if a developer has no open-source code, they're a mid-level engineer at best, on any scale. It is impossible to genuinely reach senior level without committing a mountain of code to the core language, the standard library, or colleagues' libraries. Brash amateurs who know a single language within a single paradigm and demand to be addressed as “Señor”—they’re impostors and novices; let’em go screw themselves.

A model is incapable of solving even a moderately complex problem that requires actual decision-making. A so-called “reasoning” model is just generative brute force with pruning. If you want to use the tool productively, read the manual (in this case, figure out what Reinforcement Learning and Chain of Thoughts actually are, why a model has no “memory” whatsoever and never can, and what precisely a model is doing while “thinking”.) So, if you can’t formulate a task such that there isn’t a single bifurcation point (conditional statement) in the middle, you won’t be able to use models effectively. Worse yet, you won’t be able to trust the output. Being able to frame tasks like that requires extensive experience both in writing code and in—well—task formulation itself. “Make me Uber, but better” only works during the UI polish phase. Business logic is still on the human, and if they aren’t tirelessly writing code themselves, how would they even know how to frame the task correctly?

Personally, generative models have helped me become significantly more productive. Many of my libraries from the past year would never have seen the light of day without AI assistants, and my verses wouldn’t have turned into songs that actually evoke emotions in people. I have plenty of code I can show without hesitation. I can also point to music albums I'm not ashamed of.

All of this became possible thanks to generative models, and for that, I am deeply grateful to them. At the same time, I cannot refrain from a warning: if you need anything other than a dumb executor to carry out grunt work strictly following your detailed specification, a model is not what you're looking for.

It is pointless to consult a model, check hypotheses with it, or ask it to find a solution. Not just pointless; it’s strictly dangerous. Delegating domain expertise to an assistant is sheer madness (or sabotage.) Entrusting a problem’s solution to a model is like cheating on an exam: it might work out, or it might blow up in your face. And those are all the simple rules for enjoying work with what is fraudulently dubbed “artificial intelligence.”