What the current generation of models actually makes buildable — each one paired with the shift it represents and the concrete case that proves it.
These models are not text-only — they take images, video and audio directly. A tenant moves out and sends twenty photos of the property: the script returns a structured damage report, scratched flooring and holes in the wall each rated for severity. The same call turns an unlabelled product photo into a title, SEO copy and keywords.
Regex-era extraction broke the moment a layout shifted — find the word "Total", take the next five characters, hope for the best. The modern approach inverts it: hand over the document and impose the answer shape instead. The model reads a creased invoice photo the way a person would and fills the schema, so a supplier redesign no longer takes the pipeline down.
Instead of returning prose, the model is given permission to call functions in your code. Ask which tenant fell behind on rent most often this year: it writes the query, your script runs it against the real database, hands back the row, and the model turns that into a sentence a human can read. The same wiring lets it book, cancel or notify.
Old automation was a decision tree — if the customer says "refund", ask for the order number, otherwise fall through. Rigid, and blind to anything unplanned. The modern shape stops encoding the path and encodes the reasoning instead: define the goal, hand over the tools, and let the model choose the sequence and adapt when a step fails.
Writing work that would take a team months, run overnight. Five thousand customers in a spreadsheet, each with a purchase history: the script loops, and every one gets a genuinely individual email — their name, their last order, an offer that actually fits. Batch processing sends hundreds of these in parallel rather than one after another.
Not every problem is a language problem. Feed in five years of sales history — weather, weekday, product, price — and automated model search trains and compares dozens of candidates for you. What comes out predicts next month's sales with a stated confidence, without a single algorithm written by hand.