TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 17—19 · FACTUAL · DURABLE
One model, many jobs
Two tools that make the same mistake most likely share one model.
The idea
A foundation model is trained once, on a very broad diet, then adapted to many jobs: chat, code, captioning, drafting. Many products rest on one of a few such model families, with an interface on top and often other parts — retrieval, OCR, a vision encoder, filters, tools. So tools that look unrelated can share strengths, manners and blind spots from one family. But a product can also be a system of several models, and its behaviour cannot be read off the headline model alone. A small model for one narrow task is different: cheaper, narrower, easier to test.
Why it matters
When two tools make the same mistake, the likely reason is that they share a model, not that the mistake is true. If you know which model sits under a tool, you know which other tools will make the same errors.
See it in the studio
A "brief assistant" and a "spec writer" from two different companies both confuse plinth with skirting. That is not a coincidence. One model family under two products, one gap.
Watch for this
Taking a second tool's agreement as confirmation. If both rest on the same model, you have asked one source twice.
Try it
Pick three AI tools you use. Find out which model is underneath each; it is usually one search away. Draw the map: how many model families, how many products — and which products turned out to be several models working together?
Prove it
Explain why many different-looking tools can share the same weakness, say why a product's behaviour can still differ from its headline model's, and tell a foundation model apart from a small tool built for one narrow task.
How it works
The term was coined in 2021 to name the shift from many task-specific models to a few general ones, adapted by prompting, fine-tuning or retrieval. Adaptation changes behaviour at the edges. The core habits come from pre-training. FINE-TUNE and RETRIEVE FIRST on this strand are the two main ways a foundation model is fitted to a studio.
What this idea builds on
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.