TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Foundations · AGE 15—17 · FACTUAL · DURABLE

Training and inference are different times

Training set the weights before you opened the app, and your chat never changes them.

The idea

A model has two separate times. Training is when its weights were set: a long run over a huge pile of data, finished before you ever opened the app. Inference is the moment you ask it something: the weights are read, an answer comes out, and the weights are exactly as they were. Your conversation does not change them. It cannot. So when the tool gets your project right on Monday and wrong on Tuesday, it did not forget. It never learned. Training happens once, elsewhere. Using happens now, with you.

Why it matters

Most of the false hopes and the false fears about these tools come from mixing the two times up. "It will learn our studio's style from our chats" is a hope from the wrong time. "It is learning from me right now" is a fear from the wrong time.

See it in the studio

You spend an hour correcting a chatbot's idea of a Mangalore tile roof. Next week, a new chat, and it is wrong again in the same way. You taught nobody. The weights were set months ago in a data centre. What you typed stayed in that one conversation.

Watch for this

The line "your data is used to improve our models" in a provider's terms. That is real, and it is not the same as learning from you now. The provider may keep your chats and use them in a future training run, weeks or months later, for a future model. That is a privacy question about where your words go, not a sign that this model is learning. Some products train on consumer chats unless you opt out; business and API accounts often do not by default. Read the setting for the product you use.

Try it

Tell a chatbot a made-up fact about yourself — say, that your studio is on the fourth floor. In the same chat, ask about it: it will remember, because the fact is still in front of it. Now open a brand-new chat with memory turned off and ask again. It has no idea. Nothing you said reached the weights.

Prove it

Explain, in two sentences, why correcting a model today does not make it right tomorrow, and say what a provider's "we may train on your data" line actually means.

How it works

Training adjusts the weights, a little at a time, over billions of examples, until the model's guesses match its data well. That run costs machines, electricity and weeks. When it ends, the weights are frozen and copied to the servers that answer questions. Inference is one pass through those frozen weights per token. Nothing in that pass writes anything back. The things that seem like learning — the tool remembering your site, picking up your phrasing, knowing your last upload — are context and memory features added by the product, and FOUR PLACES on this map sorts them out. Fine-tuning is real learning, but it is a separate training run that a developer sets up on purpose, not something a conversation does; FINE-TUNE on the Generative Mechanics strand is about that. And the provider's right to keep and train on your chats is the third thing again: a future training run, on a future model, governed by your settings and the company's terms. TWO CLOCKS, beside this one, follows the two times to the cut-off date; TWO BILLS on Ethics & Provenance follows them to the energy bill.

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.

Age grows from 11 at the centre to 22 at the edge, and six sectors show the learning strands. Tab into the map and the arrow keys step from idea to idea, following the links where there is one. Enter opens the idea under the cursor, and E reads out its links and the reason recorded on each. Press slash for Search, question mark for the full key list, and Escape to leave. Open Ideas for the complete readable list, including what each idea builds on and what it opens up.

LOGIKA · RBDS AI LAB INDIA
ON-RAMP · AGE 11 · FIRST ENCOUNTERS, NOT GATES — IDEAS · — DEPENDENCIES
DONE
OPENS NEXT
SOLID — STANDS ON