12 min · Beginner
How Generative AI Works, Without the Math
You will see how a reply is predicted from your prompt and the chat, why a long thread can crowd out the task, and why the model sees an image only when you provide it.
Learning objectives
- Explain, without math, that the model predicts a plausible next piece of the reply.
- Define training and context, and say when a chat does not remember you.
- Treat an upload as part of the prompt, and ignore creativity controls you do not see.
01
Why this matters
If you imagine a tiny person reading your mind, you will expect memory and eyesight the product does not have. A plain picture of prediction makes the failures less mysterious. You can then fix the chat instead of blaming yourself.
02
Concept
Picture a busy cook who has practiced thousands of meals. When you order, they do not run a laboratory. They reach for patterns that usually fit and plate something plausible. A generative model is similar only in that thin way. Training — earlier learning on huge amounts of text and other data. The cook picture is a start, and it breaks as soon as you expect taste, shopping, or a look inside your cupboard.
Here is the correction. The model predicts a plausible next piece of the reply given your prompt and the conversation so far. It is not opening your files and it is not keeping a diary of you. It does not remember you across chats unless a product memory feature is on, and those features change, so check the current screen. A new chat starts without your story unless you paste the parts that matter.
Instructions you place at the start of a chat shape everything after them. A long messy chat can crowd out the original task because the working set is finite. Context — the working set of instructions and conversation the model can see right now. It is not permanent memory. When side topics fill the thread, start a fresh chat and paste only the brief that still matters.
Some products let you push the model to be more literal or more varied; if you do not see that control, ignore it. Multimodal means an image can be part of the prompt, not that the model sees your screen unless you upload it. Model names change. In 2026 public names included 3.5 Flash and live voice models. Read the model picker instead of memorizing a brand, and check the current screen.
03
Real-world example
Priya's cake ratio was never stored
Priya runs a small bakery in Leicester. On Monday she tells a chat her lemon-cake butter ratio and asks for a prep list. On Thursday she opens a new chat and types continue the cake plan. The reply invents a ratio she never uses. She thinks the model forgot her. It did not forget, because a new chat does not keep a private memory of her kitchen unless a memory feature is on, and she should check the current screen before trusting one. She pastes a three-line card: lemon cake, her real ratio, twelve servings, no nuts. The next prep list matches that card. The card did the work. The model's memory did not. She keeps the card in her recipe notebook, not only inside a thread.
04
Gemini demonstration
A realistic exchange. Open Gemini beside this page and try a version with your own facts.
You
I explained my project at the start of this long chat, and now the answers ignore it. Did the model forget me, or is something else going on?
Gemini
This thread can crowd itself out. I predict a plausible next piece of the reply from the prompt and the conversation I can see right now. That working set is the context. It is not permanent memory, and a long messy chat can push the original brief aside. I also do not remember you across chats unless a product memory feature is on. Those features change, so check the current screen before you rely on one. A practical fix is a fresh chat. Paste a short card at the top: the goal, the reader, the constraints, and what done looks like. Then ask for the next step only. Do not paste secrets into that card. Instructions at the start shape what follows better than a pile of side questions.
You
There is a photo of our chalkboard menu on my laptop screen. Can you read the prices from there without me uploading anything?
Gemini
I cannot see your laptop screen just because the chat is open. Multimodal means an image can be part of the prompt when you upload it, not that I watch your desk. If you are allowed to share that photo, upload the file and tell me the job: read the item names, list any price that is hard to see, and do not invent a number if a digit is blurry. If the board shows private staff notes or a supplier cost you should not share, do not upload it. Describe the menu in your own words instead. I also will not memorize a model brand for you. Read the model picker on your screen, and check the current screen, because names such as 3.5 Flash can change.
05
Prompt breakdown
The prompt
I am starting a fresh chat. Use only this brief. Role: assistant to a bakery owner. Context: lemon cake for twelve, no nuts, plain language, do not invent prices. Task: write a prep list for one person working alone. Constraints: at most twelve steps, no secret recipes beyond what I paste next, mark any assumption. Output: a numbered prep list plus three questions you still need me to answer. My notes: [paste only non-secret notes].
- A fresh chat plus a brief at the top keeps the real task in context.
- Role, context, task, constraints, and output tell the model what to predict toward.
- The ban on invented prices stops a fluent guess from looking like a fact.
- The three questions show what the brief still lacks, so you can answer them next.
06
Common mistakes
Usually
Continue from yesterday in this brand-new chat. You already know my ratios.
Better
Here is the brief again: lemon cake, twelve servings, no nuts. Use only these notes.
Usually
Look at my screen and fix the slide I have open behind the browser.
Better
I will upload the slide image. Read the title and list typos. Do not invent numbers you cannot see.
Usually
Keep this huge thread going. The original goal is in here somewhere.
Better
Start a fresh chat and paste the original goal, the reader, and the output I still want.
07
Practice exercise
Reset with a brief card
Done when: You can point to one answer that drifted and one answer that stayed on the brief because the context was clean.
08
Challenge
Prove what the model cannot see
Pick a harmless image you have the right to share, such as a photo of a handwritten shopping list you wrote, or skip the upload and use a short description instead. In one chat, ask Gemini to read the image without uploading it and notice the limit. Then upload it, or paste your description, and ask for a grouped list plus any item it cannot read clearly. Finally, open a fresh chat and ask it to recall the list without pasting anything. Write one sentence on what context actually included. Do not use a private document, an ID card, or anyone else's photo.
Hint: The third chat should fail to recall the list unless a memory feature you checked is on, and even then you still verify.
09
Key takeaways
- The model predicts a plausible next piece of the reply from your prompt and the conversation it can see.
- Training is earlier learning on huge amounts of data, not a memory of you.
- Context is the working set right now, not permanent memory. A long chat can crowd out the original task.
- An image counts only if you provide it. If you do not see a creativity control, ignore it and read the model picker.
10