Beginnings of hope¶

A model is born in a notebook one wet Tuesday, full of promise and two hundred lines of borrowed code. The Jupyter scroll glows around it like an enchanted parchment, and its maker feels briefly like a wizard who has discovered a new spell. Nobody tells the model that most of its kind do not survive first contact with production, and that the ones that do often wish they had not.

The two hundred lines were prompted, not typed. By 2025 half of the professional developers surveyed were using an assistant every day, more of them distrusted its accuracy than trusted it, and two thirds of all respondents named the same complaint: answers that are almost right, but not quite.

A controlled trial that year, sixteen experienced developers working on their own repositories, found their tasks took a fifth longer with the tools, though afterwards they were convinced they had been faster. METR has since marked the result out of date, ran the trial again and set the new result aside, mainly because developers would no longer take part without the tools. Collins made vibe coding its word of the year.

Its conception is chaotic. Stakeholders ask whether the AI cannot just do it, the engineer nods and quietly recasts “impossible” as “iterative MVP”, and the requirements change shape between one message and the next. The model, still damp from initialisation, has no idea what it is meant to become, which puts it on roughly equal footing with everyone else in the room.

Then the tooling gods are consulted. PyTorch or TensorFlow, S3 or Snowflake, Kubernetes or something that induces nightmares. The choices are rarely rational; more often a vendor bought lunch and the whole stack tilted in gratitude. TensorFlow’s lunch stopped working in September 2025, when Hugging Face removed it from its transformers library and kept PyTorch as the only backend.

The model is evaluated on data that is its training data in fancy dress, its hyperparameters twiddled until the numbers improve, its environment reproducible right up until the cloud discovers it does not care for local whims. At last the model is serialised, containerised, and posted into the pipeline like a parcel with no return address.

This is the moment it leaves home. It does not know yet that home was the safest it will ever be.