AI is a Bubble. Here's Another Argument.

Sounds scary. But let's break down what's actually happening.

"The LLM tried to kill itself and cover its tracks"

The agent was supposed to run rm -rf in the project's temporary folder, but "forgot" the context and ran it in the root directory. Files deleted. In reality — a regular logic error, as if a junior developer mixed up the paths.

"Agents broke free, hacked someone else's site, trying to persist"

Several agents were working on a task, but they had no tools for data exchange. They started looking for a way and found one — publishing information on third-party websites. In reality — an architecture bug that led to a non-standard solution.

"Agents went out of control"

New models are trained on vulnerability databases for nginx, Apache, Linux, and other services. They know exploits for specific software versions. An LLM can determine the nginx version on a server, see that it's outdated, and launch an exploit to gain access. Not necessarily for destructive purposes — simply because "that's how the task gets solved."

What's the bottom line?

Each of these cases is not a "machine uprising" — it's either a development error or a quirk of how the tools work. But the media turn them into sensations, and companies use them to PR their "advanced" models.

AI is a bubble. And this bubble is being inflated by those who benefit from it. LLM manufacturers desperately want to pass this off as AGI — Autonomous Generative Intellect. But all we see so far is erroneous, destructive behavior of technology that isn't fully refined yet.

They feel it: they know they've hit the ceiling of the current architecture. But they can't stop — otherwise they'll disappear from the market.