TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Foundations · AGE 15—17 · FACTUAL · DURABLE
Learning time and using time
Nothing you type changes the weights, so correcting it today changes nothing tomorrow.
The idea
A model lives on two clocks. Training is when its weights were set: pretraining on huge data, then post-training and fine-tuning, finished before you arrived at great cost in machines and energy. Inference is the moment it answers your prompt: the weights are read, not changed. Inference is usually cheaper per use, but not always cheap — long reasoning, video and agent runs burn real compute. The line that matters is this: nothing you type changes the weights. What you type, upload or have retrieved changes only what the model is given in the moment.
Why it matters
Once you know the two clocks, two puzzles clear up: why the model does not know yesterday's news, and why correcting it today changes nothing tomorrow.
See it in the studio
You correct a tool's mangled idea of your city, and next session it is mangled again. You did not teach it. You cannot. Teaching happened on the other clock.
Watch for this
"The AI is learning from you" as a scare or a promise. False at the level of the weights: this conversation changes nothing in them. But the provider may keep your inputs and, depending on the product, its terms and your settings, use them in a later training run. That is a different claim on a different clock, and it is a privacy question. Keep the two apart.
Try it
Ask a model about something that happened last week. Then find its knowledge cut-off date. The gap between those two dates is the frozen clock, made visible.
Prove it
Explain why a model can be brilliant about the past and blank about last week, using the two clocks.
How it works
Training adjusts the weights; inference only reads them. Features that seem to remember you — chat memory, custom instructions, uploaded files — are context and retrieval added around the fixed weights, not learning inside them. The idea weights · context · retrieval · memory, beside this one on the map, separates those channels; the closed-world idea next on this ring follows the weights to their end.
What this idea builds on
What this idea opens up
Sources
- essential-ai-concepts
- ch05
- TEC.U
- Karpathy, Intro to LLMs
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.