weeklyAI
TertulIA · courses
Create · lesson 1 of 10next lesson →
C01 · The Machine That Continues
You can read the twenty seven lines behind any chat window, say what the model does, and name the slot that decides who a message is aimed at.
Ask TertulIA
Ask me about this lesson, or what to learn next.
Conversations are saved for as long as weeklyAI exists, to improve the publication. Answers come in the language you write in.
The lesson
Short cut · 15:31
Trailer · 1:34
Are you ready for the challenge?
In this half I took the box apart one layer at a time and what was left was twenty seven lines, one slot and somebody's key. Now I'd like you to catch one in the open. Not in a chat window where you asked it something, but out in your week, where a machine wrote to you and did not say so. Find the message, then ask the two questions of it: what went into the slot, and whose key paid for the call. You will not always be able to answer the second one. Noticing that you cannot is most of the point.
These are situations, not queries. Change one thing and it's yours.
- Take a confident paragraph from any assistant on a topic you know well enough to grade, and find the one detail that is wrong. Then check where the world wrote least about your topic, and see whether the error landed there. It usually did.
- The next email or message that arrives written for you personally, with your name and something it knows about you, decide what the spreadsheet row behind it must have contained. Then ask what your neighbour's version said, and notice you have no way to find out.
- Ask an assistant something about your own trade that only somebody in it would know, twice, in two separate windows. Keep both answers. The distance between them is the weighted dice, and it is the cheapest demonstration in this lesson.
- Anything that told you a fact this week and could not tell you where it got it. A summary, a recommendation, a description of a place you have never been.
Whichever you choose, verify it the way the lesson taught: against something the machine could not have written. Not against a second machine, which will agree with the first and prove nothing.
When you're done, tell us what you pursued and tell us what you got.
Keep your questions as you go; this page reserves a place for them, and what you ask here will shape where the course grows next.
What you work with
- crear_texto_en.ipynbipynb
- crear_texto_es.ipynbipynb
- motor_texto_en.ipynbipynb
- motor_texto_es.ipynbipynb
- teoria_llm_en.ipynbipynb
- teoria_llm_es.ipynbipynb
Field Card · C01 · The Machine That Continues
The distinction
A language model does not know anything. It continues text, one next word at a time, rolling weighted dice. Fluent is not true, and confidence is a cheap style.
In practice
- Verify names, dates and quotations against a source the machine did not write. A second machine agreeing with the first is not corroboration.
- When an answer sounds current, ask whether the frozen model answered or the search bolted on top of it.
- Read the model's name like a medicine label: family, size, version, and whether the word preview sits inside it.
- Faced with any message written for you alone, ask who filled the slot.
The question to keep asking
Who filled the slot, and who holds the key?
The fine print
No knob makes the machine honest. It only changes how much the machine deliberates, and careful is not the same as correct. The same twenty-seven lines write a poem for one person or a letter for fifty thousand.
Glossary · C01 · The Machine That Continues
- Language model
- A machine that continues text, one next word at a time. It does not look up a stored answer or consult a warehouse of facts; it takes what has been written so far and proposes what comes next. The poem was not waiting on any shelf; it is made in the act of continuing.
- Weighted dice
- What the machine rolls at every single word: a list of candidates for what comes next, each carrying a weight, and a pick among them rather than always the heaviest. The weights came from reading more text than a person could read in a hundred lifetimes. So the same words handed over twice give two different answers, and that is the design working, not failing.
- The last day
- Training has one, and the model is frozen at it the way a photograph of a room is frozen, not even aware that today exists. Nothing inside is counting forward while you sleep. A fact that quietly stopped being true comes out wearing the same fluent coat as one that is still true.
- Grounding
- What the panel calls the switch that lets the machine go and fetch from the web before it continues. The internet is a tool bolted on from the outside, not something the model contains. Knowing whether the tool or the model answered you changes how much of that answer you are entitled to believe.
- Confabulation
- The most important idea in the lesson: the machine has no truth check anywhere inside it, so a confident falsehood costs it exactly what a confident fact costs, which is nothing. It is not lying, because lying requires knowing better. It fails where the weights are thin, which is usually where the world wrote least.
- The knob
- The control for how much the machine deliberates before it answers, named in the program with a word rather than a number. It does not measure randomness; it measures deliberation. No setting of any AI turns confidence into accuracy.
- The slot
- The gap between two quiet quotation marks where the ask goes in, with a second, quieter slot that says who the model should be before anything is asked. That gap is where your life enters when someone holding the key aims at you. Every argument about AI you will hear is an argument about who fills it.
- Mass personalization
- Fifty thousand letters, each one written for exactly one reader, out of a spreadsheet, a short script, five dollars of calls and an hour of running. A population of one is not part of a public anymore, because the neighbors cannot compare notes when no two of them received the same thing.
- Narrowcasting
- A message no neighbor sees, and so one that cannot be argued with in public. A rumor used to have to survive two people comparing what they had heard; a claim told to one person at a time never has to survive daylight. Nothing about it is new; only the cost of doing it has collapsed.
References · C01 · The Machine That Continues
This lesson cites no published literature: its source is the service's own documentation, Google AI Studio and the google-genai library, as the lesson's packet names and uses them.
Next: C02 · Run The Function Yourself. Learn AI at TertulIA