Programmer questions whether AI is actually useful after months of testing

The author is a software developer who began experimenting with AI tools in mid-2023, starting with Claude. Initially, the results were encouraging enough that he integrated the tool into his workflow on hobby projects. By early 2024, having subscribed to Claude Pro, he gained access to "agentic" AI—systems that can use tools autonomously—via Claude Code. At that point, he describes being "enthralled" and experiencing a sense of "accelerated productivity" that "felt like a drug." He used AI to generate entire projects at once, and the initial experience was impressive.

However, once the generated code was in place and the software appeared to work, the appeal faded. The core problem: he did not have a mental model for the code in front of him. Whenever a bug appeared or a new feature was needed, his understanding was reset to where it had been before he started prompting. He could not make meaningful changes without first building a comprehensive understanding of the code. The author acknowledges his own working style—building mental models incrementally through the act of writing code, statement by statement—is fundamentally at odds with receiving pre-built systems from AI. He lacks the discipline (or believes he lacks the capacity) to design large systems entirely in his head before touching the keyboard.

Over time, a clear trend emerged: he used AI less and less. He abandoned "set and forget" vibe coding in favor of an active, guided approach where he directed the AI more closely. Finding that still unsatisfying, he pivoted again to AI-assisted ideation—having conversations with Claude to build mental models, then writing the actual code himself, as he always had. On whether this worked, his answer is candid: he does not know for sure, precisely because AI is unreliable. He cites an example of a software component he designed with AI that he never used because it was far more complex than necessary, or not needed at all. Conversely, he has gained genuine insights by rubber-ducking problems with Claude, leading to real project progress. But even in successful cases, he is left with the lingering doubt that he probably could have finished without AI at all.

The author is not alone in this experience. He observes it is common in discussions wherever people debate their own AI use, and he raises a provocative possibility: that LLMs may never produce an obviously indispensable "killer app," and that this technology may be fundamentally overhyped. He compares the current AI fervor to the Metaverse bubble, which seemed plausible for a brief window—the idea that people would spend their waking hours in VR headsets. The technology itself was impressive (he owned and enjoyed Quest headsets), but the core proposition was, in retrospect, absurd. He poses a rhetorical question: what if LLMs follow the same arc, and in a decade we look back astonished that anyone believed intellectual work could be automated? What if developers, engineers, and writers conclude, as he has, that the technology is not useful enough to bother with? What if large organizations, which spend the most on tokens, decide the return on investment is inadequate?

The article presents a detailed first-person account of a common tension in current AI adoption: the gap between the ease of generating output and the difficulty of understanding or extending it. The author's experience reflects a deeper concern about how AI is used in knowledge work. He notes that his mental model is built incrementally through the act of writing code, not through abstract design beforehand—a workflow that tool-generated code fundamentally disrupts. Rather than accelerating development, AI-generated code may actually slow him down by forcing him to reverse-engineer his own tools.

The author draws an explicit parallel to the Metaverse narrative, arguing that the same hype cycle may be playing out with LLMs. The Metaverse seemed plausible during a brief window of enthusiasm but is now widely recognized as impractical and oversold. He suggests that LLMs may follow a similar arc: a period of intense belief in productivity and automation, followed by a reckoning in which organizations, developers, and writers conclude the technology is not worth the investment. This is grounded not in conspiracy but in repeated observation—he notes his skepticism is "not unique" and "not new," citing widespread frustration he observes in developer communities.

No comments yet. Be the first to share your thoughts!