Large Language Models (LLMs) cookbook

The model counted wrong, and it sounded completely sure.

A model answered a question about something recent as if it already happened, and it hadn't.

I asked our own AI assistant about a feature in our product, and it described one that has never existed in convincing, confident detail.

My long prompt has the right fact in it, and the model still misses the one buried in the middle.

Why did my prompt get cut off, or my bill run higher than expected, when I never checked the token count first?

I asked it a plain question, and it answered with three more questions of its own.

Temperature zero, same prompt, two different answers.

I want to use a new model, and I have no idea what it was trained on or where it breaks.

I asked something completely legitimate, and the model refused like I'd asked for something dangerous.

Three paragraphs in, the model latches onto one sentence and just keeps saying it.

Is a model with more total parameters really the bigger, pricier one?

I default to the biggest model out of habit, without ever checking whether a cheaper one sitting right next to it would honestly do the job just as well.

Same default temperature, a brainstorm and a fact lookup: neither result landed.

Brilliant in the language everyone tests it in. Silent, then wrong, in the one your team actually uses.

Every vendor calls it an agent now. Not all of them mean it.