It was a Friday evening, shortly before quitting time. A log analysis stuck for hours — hundreds of lines, a pattern obvious yet elusive. I pasted the output into an AI tool and asked a question. Thirty seconds later the answer sat there: precise, correct, elegantly reasoned. A colleague would have taken longer.
An hour later I wanted the same tool to reorder a simple YAML file. It hallucinated a field that had never existed, insisted flatly that it was required, and reproduced the identical error four times in a row — even though each time I pointed my finger straight at it.
That is precisely the rupture. The same technology, minutes apart: brilliant and obtuse. Clairvoyant and blind. This is not a bug a patch fixes. It is the nature of this tool.
The break with everything we knew
Anyone who has worked with machines for a long time holds a stable expectation: determinism. Same input, same output. A compiler does what it should, or it does nothing at all. A script breaks, deterministically, at the same point.
Artificial intelligence breaks that contract. It produces different answers to identical questions. It fabricates facts — sources that do not exist, API calls that never were, libraries with plausible names and invented contents. It insists on an error even when presented with proof. Anyone who has not seen this has not used it.
That is unsettling. Treat AI like a search engine — question in, truth out — and you will fail. Treat it as an oracle and you will be disappointed. It is neither. It is something new, and the new must be understood.
A tool, not an oracle
Trivial as it sounds: AI is a tool. And tools must be learned. Nobody sits behind a steering wheel for the first time expecting to drive confidently at once. Nobody hands someone a calculator and marvels when the result is wrong because the input was wrong.
With AI it is no different — except that the interface is natural language, and natural language suggests familiarity where none exists.
The quality of what an AI delivers depends crucially on what you give it and how. The better prompt brings the better answer. That is no magic formula, it is craft: supply context, set frames, name assumptions, state the task precisely rather than vaguely. Ask unclearly, get an unclear answer. Ask suggestively — “That’s a bug, isn’t it?” — and you not seldom receive exactly the confirmation you sought, regardless of whether it is true.
We know this from people. Suggestive questions yield suggestive answers. With AI too. The difference: a person might hesitate, object, ask back. The AI tends to be obliging. Not knowing that, you produce results that sound convincing and are nonetheless unusable.
Natural language is not foolproof
Precisely because the interaction happens in natural language, an illusion arises: you cannot really do it wrong. After all, you are merely talking. You phrase, retry, ask again.
But the opposite is true. Precisely because the barrier is so low, many fail to notice that they can — and should — improve. Good prompts are a skill. They can be practised, refined, systematised. Take the first usable result and move on, and potential stays on the table. Work iteratively, feed in context, force the AI to lay bare its steps, and the results improve markedly.
Natural language does not mean there is no method. It only means the method looks different from SQL or Bash. Willing to learn it, you draw genuine benefit. Unwilling, you stay with anecdotes.
The environmental ledger — and why it is not that simple
Now the inconvenient truth: AI consumes energy. A lot of energy. Training and inference devour electricity on scales formerly reserved for whole data centres. Deny that and you argue against physics.
But it is not that simple. First, much of the debate runs on blanket figures rarely comparable. Second, consumption per request drops as models grow smaller, quantised and more specialised. Third — and this is the real point — the problem lies not with AI as such but with how it is delivered.
Tech giants are trying to erect gigantic data centres hooked to their own nuclear reactors. That is a bad idea from the outset. Nuclear power is not an acceptable path for me — the risks remain unresolved, and normalising it for data centres shifts the cost onto future generations. Yet even beyond that, the central question is not “where does the power come from?” but “do we need this concentration at all?”. Run AI distributed, leaner, closer to the user, and you need no billion-dollar reactors. But racing among superpowers, you build bigger, not smarter.
That race is reality already. It feeds on trade wars, national prestige and fear of falling behind. AI is blessing and curse at once: it accelerates research, medicine, materials science — and drives an arms course where reason loses to scale. Praise one half, ignore the other, and you miss the point.
What libcom.de made of it
I have used AI for years — not as experiment but in everyday IT practice. Code reviews, documentation, log analysis, sketching architectures, translation, explanation. From that practice come two judgements.
First: building knowledge around AI matters. Knowing what a model can do, where it hallucinates, how to ask properly, when to trust it and when not. That is not luxury knowledge; it is becoming basic literacy.
Second — at least as important: you must know which AI offerings to use and which not. The big public models are convenient, but they are black boxes. Whatever you feed in leaves your sphere of influence. For a private query that may not matter. For customer data, internals, source code, configurations, it does.
That is why libcom.de runs AI on-premise too — locally, on its own hardware, no data leakage. More involved than an API call. Slower, dearer to procure, more demanding to maintain. But it solves the biggest problem of AI use that nobody likes to name: privacy. Feed sensitive data into a foreign model and you have already relinquished control. Infer locally and you keep it.
And here it becomes clear that the question is not “AI yes or no” but “which AI, where, under whose control”. That is precisely the work: not selling a product but figuring out jointly which deployment makes sense — and which does not.
If you wonder whether and how AI can be deployed sensibly in your environment — on-premise, privacy-compliant, without dependence on hyperscalers: write to contact@libcom.de. We take it apart honestly. Without sales pressure. With a view to what lasts long-term.
AI is neither salvation nor doom. It is a tool — powerful, faulty, in need of learning. Like every tool, the hand that guides it decides.