The Omnipotent AI Myth: Why Our Beliefs About What It Can Do Matter

The headline "AI Can Do Anything" appeared on Hacker News a few days ago with barely three points and no comments. Limited social traction, but the statement itself deserves attention: it sums up with brutal precision the mindset that drives much of the conversation around AI in 2026, whether defending it or attacking it.

The original piece, published on clawdcursor.com and picked up on Hacker News, starts from a premise we've heard thousands of times in forums, product meetings and client presentations: the idea that a language model is, potentially, capable of anything. The question worth asking is whether that belief helps or hinders real work.

What it means to believe AI "can do anything"

When someone says AI can do anything, they're usually expressing one of three different things: that the model is flexible enough to adapt to many tasks, that given the right prompt the output always improves, or simply that there's no relevant technical limit. These three claims have different degrees of truth.

The first is reasonable. Claude Opus 4.7, with its one-million-token context window, can process information volumes that would have been unmanageable for any workflow two years ago. Combined with well-configured MCP servers, specialized subagents and hooks that automate lifecycle stages, the range of application is genuinely broad.

The second also has solid grounding: prompt engineering, reusable skills and the ability to compose agents in Claude Code have shown that many perceived limitations were actually workflow design limitations, not model limitations.

The third, however, is where the narrative breaks down. Current models make factual errors, carry biases inherited from training data, lack access to real-time information unless explicitly connected, and their reasoning on certain domains—advanced mathematics, highly specialized code, complex formal logic—still produces failures that surprise those who thought them solved.

Why collective expectations have real consequences

This debate isn't just philosophical. The beliefs teams hold about what a given model can do directly determine how they design their systems.

We've seen projects where the validation phase gets eliminated because "the model does it well". And projects where nothing ever ships because "the model isn't reliable". Both extremes stem from miscalibrated expectations.

The problem with the omnipotence narrative is that it shifts responsibility. If the model can do anything, when something fails the conclusion is that the prompt was wrong, that context was missing, that it was the wrong model version. Rarely does anyone question whether the task was appropriate for this kind of system in the first place.

In the Claude ecosystem specifically, this tension shows in how plugins and subagents are used—or misunderstood. A specialized subagent can be extraordinarily good at a bounded task. Mentally converting it into a general agent because "AI can do anything" is the quickest path to unreliable results and debugging nightmares.

What the Hacker News community (usually) does well

Although this particular post generated no discussion, Hacker News remains one of the spaces where technical skepticism acts as a natural counterweight to hype. That forum's audience tends to demand specificity: exactly which task? with which model? under what conditions? what's the comparison baseline?

That demand for concreteness is precisely the antidote to absolute headlines. Not because AI can't do remarkable things—it can—but because real utility always lives in the details of the use case, not in generic claims.

Our Take

The omnipotent AI narrative doesn't come only from marketing: we build it collectively every time we oversimplify what a model has done well. Calibrating expectations properly isn't technical pessimism; it's the groundwork for any integration that actually works.