How AI models work

An AI model is a general-purpose reader and writer. Hand it a spec section to summarize, two details to compare, an RFI to draft a response to, or a week of meeting minutes to sort into open items, and it can do the job even though nobody built it for that specific task.

A model differs from the other software on a project in one way that matters. The same input does not guarantee the same output.

Deterministic vs. probabilistic​

A quantity formula in a spreadsheet, a schedule field in your authoring software, or a Bluebeam measurement gives the same output for the same input every time, because someone wrote the rule.

A model is probabilistic. Given the same sheet and the same question, it may take a different path each time. A model predicts the most likely next piece of text from two sources:

1. What it was trained on: a large body of text up to a cutoff date
2. What it is given now: the prompt

Ask "What is the specified roof membrane warranty? Answer with the number of years only." and two things can follow:

1. The same question can get a differently worded answer tomorrow
2. The model may ignore part of the instruction, such as "the number of years only"

Instead of expecting identical answers, compare the evidence behind them. The rest of this course builds on that habit.

Open figure

Choosing a model​

Models differ in reasoning ability, speed, cost, and strengths. A fast, cheap model is fine for "what does this note say." A slower, more expensive one can work through a fifty-sheet set without losing the thread. For construction documents, current models are capable, and new ones arrive most months.

In Nomic the picker defaults to Auto: the current default model at Medium effort. Turn Auto off to choose a specific model: Sonnet 5.5 for everyday work, Sonnet 5.5 Extra High for exhaustive reviews, plus Opus 5.5, Opus 4.8, and Haiku. Gemini Flash appears when the organization enables it, and earlier versions sit under More models. Default to Sonnet 5.5. Use Opus for the hardest reasoning and Haiku for quick exploration. See Models & Pricing.

In Nomic, the model is fixed for a session. Effort, which is how long the model reasons before acting, can change within one. More in lesson 3.

Modalities​

Models read more than text. They also read PDFs and images, such as a drawing set, a site photo, or a screenshot from your authoring software. Some tools add voice.

On a project, that looks like:

1. Text to describe the task and set the scope
2. A PDF sheet or a photo of an existing condition, with a question about what conflicts with the drawings
3. A structured file, such as an IFC model or a DWG, queried for elements, layers, and attributes instead of pixels

The Nomic agent does all three. It reads parsed text, sends rendered pages to a vision model when it needs the geometry, and queries BIM and CAD files directly. The steps the agent shows in the conversation tell you which one it used.

Open figure

Next: where models fail, and how to spot it.

Next: Hallucination and limitations