Learn

You can run a complete AI stack on hardware you control. Whether you should depends on what an hour of maintenance is worth to you.

The main thing between you and a working local setup is picking a model that fits your memory. Get that wrong and you will conclude local models are useless.

Fine-tuning teaches behaviour. It does not teach facts reliably, and confusing those two is where the money goes.

Nothing is composed. The image emerges everywhere at once, which is exactly why counting and text placement go wrong.

Most teams reach for a dedicated vector store where a table in the database they already run would do. Here is where the threshold sits.

A citation has a recognisable shape, and producing that shape is easy. A three-step workflow that survives deadline pressure.

The loop takes twenty lines. Error compounding, absent memory, no stopping condition and excess permissions are where the work actually is.

Output tokens cost up to five times input. A support assistant at 10,000 conversations a day runs $2,700 or $9,000 a month depending only on model choice.

Classifiers measure how ordinary writing looks, which is not the same as who wrote it. Watermarking is better and still proves processing, not authorship.