Concept 01 · 8 min
It doesn’t understand you. It predicts the next word. Extremely well.
Scroll to start
Scene 1
You already use one
Your phone suggests three words while you type. Pick the one you’d choose.
Your meeting is confirmed for next
Your phone learned from your messages. The assistant you use does the same, with far more text: billions of pages.
Scene 2
One word at a time
Sarah starts an email to Mr. Miller and lets the assistant finish it. Scroll slowly: each step adds one word.
An animated fan of candidate words follows. The same steps are listed in the table below it.
Keep scrolling: each scroll adds one word.
Dear Mr. Miller, thank you for your letter, which I have read with great interest.
| Step | Sentence so far | Most likely next words | Picked |
|---|---|---|---|
| 1 | Sentence so far…thank you for your | Most likely next wordsletter18%email6.6%time6.6%kind4.3%interest3.5% | Pickedletter |
| 2 | Sentence so far…you for your letter | Most likely next words.27%of20%and10%dated6.6%,5.2% | Picked, |
| 3 | Sentence so far…you for your letter, | Most likely next wordsand31%which18%in8.3%but6.6%dated3.4% | Pickedwhich |
| 4 | Sentence so far…for your letter, which | Most likely next wordsI42%was7.7%has6.4%we6%arrived3.9% | PickedI |
| 5 | Sentence so far…your letter, which I | Most likely next wordsreceived27%have21%read9%am5.3%will4.5% | Pickedhave |
| 6 | Sentence so far…letter, which I have | Most likely next wordsread40%received11%just8.1%carefully4.8%been4.4% | Pickedread |
| 7 | Sentence so far…which I have read | Most likely next wordswith44%carefully18%and11%very5.9%several3.4% | Pickedwith |
| 8 | Sentence so far…I have read with | Most likely next wordsgreat36%interest30%much12%pleasure4.2%the2.1% | Pickedgreat |
| 9 | Sentence so far…have read with great | Most likely next wordsinterest83%pleasure9.2%care2.1%appreciation1.1%attention<1% | Pickedinterest |
| 10 | Sentence so far…read with great interest | Most likely next words.69%.↵↵14%and9.5%,2.9%.↵<1% | Picked. |
Dear Mr. Miller, thank you for your letter, which I have read with great interest.
| Step | Sentence so far | Most likely next words | Picked |
|---|---|---|---|
| 1 | Sentence so far…thank you for your | Most likely next wordsletter18%email6.6%time6.6%kind4.3%interest3.5% | Pickedletter |
| 2 | Sentence so far…you for your letter | Most likely next words.27%of20%and10%dated6.6%,5.2% | Picked, |
| 3 | Sentence so far…you for your letter, | Most likely next wordsand31%which18%in8.3%but6.6%dated3.4% | Pickedwhich |
| 4 | Sentence so far…for your letter, which | Most likely next wordsI42%was7.7%has6.4%we6%arrived3.9% | PickedI |
| 5 | Sentence so far…your letter, which I | Most likely next wordsreceived27%have21%read9%am5.3%will4.5% | Pickedhave |
| 6 | Sentence so far…letter, which I have | Most likely next wordsread40%received11%just8.1%carefully4.8%been4.4% | Pickedread |
| 7 | Sentence so far…which I have read | Most likely next wordswith44%carefully18%and11%very5.9%several3.4% | Pickedwith |
| 8 | Sentence so far…I have read with | Most likely next wordsgreat36%interest30%much12%pleasure4.2%the2.1% | Pickedgreat |
| 9 | Sentence so far…have read with great | Most likely next wordsinterest83%pleasure9.2%care2.1%appreciation1.1%attention<1% | Pickedinterest |
| 10 | Sentence so far…read with great interest | Most likely next words.69%.↵↵14%and9.5%,2.9%.↵<1% | Picked. |
Nobody wrote this sentence in advance: each word was picked because it’s likely after the ones before. That’s a language model. LLM: large language model.
One precision: it doesn’t always write whole words. See how it cut this sentence:
Dear Mr. Miller, thank you for your letter, which I have read with great interest.
Scene 3
Turn up the adventure
The assistant favors the most likely next word, but it doesn’t have to take it. A setting makes it more or less adventurous. Try the slider.
Challenges
- Get three different openings that all make sense.
- Bonus: make it go off the rails.
StartDear Mr. Miller, thank you for your…
Start with Predictable, then try the others.
Write three versions at this level
Your three versions will appear here.
Most assistants hide this setting, called temperature. “Precise” modes lower it; “creative” modes raise it. Same model, more or less adventurous.
Scene 4
It continues. It doesn’t check.
Sarah starts a sentence about her client’s company.
Miller Joinery was founded in
What will the next word be?
That’s the catch: it produces what sounds right. When it matters whether it’s true, see the Hallucinations page.
Scene 5
Notes before the answer
Your assistant may have a “think longer” button, often called the reasoning level. Here’s what it changes. Pick a request, then choose a level.
Start with Low, then try the others.
It answers straight away.
Ask it
Its notes and its answer will appear here.
At a higher level, the assistant first writes notes, often out of sight, then uses them to answer. More notes: often a better answer, but slower and pricier.
Hype vs reality
- What the hype says
“It understands what you write.”
What actually happensIt computes which words usually follow yours. That’s enough to sound smart, not to be right. Even a “reasoning” mode only writes notes first.
- What the hype says
“It searches the internet for answers.”
What actually happensOn its own, it searches nothing: it writes from patterns learned during training, frozen at a date. Search is something we add around it.
- What the hype says
“It has a giant database of facts.”
What actually happensIt stores no table of facts, only billions of numbers tuned to predict words. Frequent facts come out right; rare ones come out plausible.
Under the hood
Everything on this page fits in these lines.
def next_word(text, temperature): scores = model(text) # one score for each possible next word probs = softmax(scores / temperature) # scores become probabilities return random_choice(probs) # draw one, weighted by its probabilitytext = "Dear Mr. Miller, thank you for your"while not text.endswith("."): text += next_word(text, temperature=0.8) # add it, then predict againTraining is the expensive part: billions of pages to tune the scores, which is what makes the model large. Writing is just this loop; a reasoning level only makes its first phase longer: notes, then the answer, written the same way. On this page, the five settings were temperatures 0, 0.3, 0.8, 1.1 and 1.8, recorded from an open model so you can replay them.
Some languages, like French, take more tokens than English: models read more English.
In 3 sentences
- A language model writes one word at a time, each time picking a likely next one.
- It learned these likelihoods from an enormous amount of text: it predicts, it doesn’t look things up.
- Settings like temperature and the reasoning level change how it writes, not what it learned, and fluent never means true.
Did this make sense?
0/9 concepts unmagicked
Go to the challenge- 01LLMAvailable · How can it write so well without understanding? · Available
- 02ContextAvailable · Why does it forget what I said earlier? · Available
- 03PromptAvailable · Why are my results mediocre? · Available
- 04HallucinationsAvailable · Why does it make things up? When can I trust it? · Available
- 05RAGAvailable · How do I make it use our internal documents? · Available
- 06ToolsAvailable · How can it act, not just talk? · Available
- 07AgentAvailable · What is an agent, concretely? · Available
- 08SkillsAvailable · How do we specialize it without retraining? · Available
- 09Limits & safetyAvailable · What must I never trust it with? · Available
Solve the sandbox challenge to unmagic this concept.
Next step
Why does it forget what I said earlier?
A goldfish with a small notepad
Use this journey with your teamsRecorded with Ministral 3 (3B), Oct 3, 2026. Low, Medium and High are three instructions we gave this small model.