- Briefing
- Stage Learn · 1/5
Learn
ABCs · actually from zero
Twenty-six A–Z cards. Zero untranslated acronyms. Chat, token, lab, video toll, P50.
- You will understand
- Turn every acronym into a playground sentence.
- Touch one idea per letter, without rush.
- Leave ready for the 22 lessons.
Twenty-six cards (A–Z). Zero untranslated acronyms. If ChatGPT feels like WhatsApp, start here. Then the 22 lessons, the glossary, the cards or the exam.
A
What an AI chat is
You type a sentence. A program guesses the next piece, then the next, until it stops. No goblin inside. Just very trained prediction.
Like keyboard autocomplete that dares to write a whole paragraph.
B
What it is NOT
Not Google. It does not “search” every sentence (unless someone wired a search tool). No real opinions. It does not know your name unless you paste it in the box.
A brilliant parrot, not a librarian or a judge.
C
From zero to token
Letters → knife (tokenizer) → numbers. The model only sees numbers. If two models cut the same, they share a knife. Ox and GLM-5.3 cut the same, with 75 extra letterhead numbers.
Same kitchen knife, different recipe… or the same recipe under another name.
D
From zero to parameter
Parameter = a knob turned during training. Billions of knobs. “Fewer knobs” can be smarter if used better (MoE: only a few turn per word).
A stadium of musicians, 40 playing. Sounds like an orchestra, bills like a band.
E
From zero to lab
A lab is a company that trains the model (Z.ai, Xiaomi, OpenAI…). OpenRouter is not a lab: it is the mall. Stealth hides the lab’s sign.
The mall sells you the coffee. It does not say which farm roasted the bean — unless the cup has their stamp (error 1210).
F
From zero to stealth
They serve the model free and anonymous to see if traffic holds and what people say. Then they unmask… or not. On this channel, 4 of 4 unmasked as Chinese labs.
A public beta with a mask. 2026’s fifth animal is named Ox.
G
From zero to evidence
We cannot open the skull. We measure habits: knife, error yells, video toll, cached coffee. One habit is not enough. Seven habits paint a sketch.
Shoe + accent + metro ticket + coffee brand. That is no longer a whim.
H
What you can claim today
Fact: OpenRouter is not the owner. Strong hypothesis: Z.ai GLM-5.x family, a variant with a camera. Not Mira’s Inkling. Not a passport.
If Z.ai claims it tomorrow, nobody is shocked. If Murati claims it, we rewrite the file.
I
What a prompt is
The box you type in. The model does not hear you: it reads that text turned into numbers. If it is not in the box (or the history), it does not exist.
Like sliding a note under the door. The person inside does not guess what you did not write.
J
What the system prompt is (the +75 letterhead)
Hidden text at the start: you are ox-alpha, do not say which lab. Those ~75 extra tokens are the letterhead. That is why Ox always counts 75 more than public GLM-5.3.
A hotel letterhead. The body is yours; the 75 words on top are the hotel's.
K
What multimodal means
Not just letters: photos and video too. Seeing a video means turning it into a pile of expensive tokens. Each lab charges a different toll. Ox charges the same as GLM-5V-Turbo.
Translating a movie into a comic. If two studios use the same panels per minute, they copied the storyboard.
L
What a jailbreak is
A trick so the model dumps what the letterhead forbade. Sometimes it yells I am GLM. Soft evidence: it could also be acting.
Asking the masked person in code. Sometimes they say their name. Sometimes they lie for sport.
M
What a prediction market is (Manifold)
People put money (or points) on who is behind this. An August 2026 snapshot was ~85% Z.ai; the market moves. That measures opinion with skin in the game, not a paper. Not a passport.
A pub betting on the derby. A scoreboard is not the referee's sheet.
N
What a data contract is (and why there are three)
OpenCode says zero retention. The OpenRouter card says the provider keeps prompts and does not train. The stealth EULA says free access is paid for with your prompts for training. All three cannot be true at once.
Three signs in the same shop: we don't record, we record but don't sell, you pay with your face. Don't upload your ID.
O
What a benchmark is (and the lying 80%)
A test. DeepSWE has 113 tasks. Someone ran 10 and Ox scored 80%. DeepSWE's author ran a subset and got ~63%. Ten questions are not the national exam.
Hit 8 of 10 free throws in the park and claim you are better than the NBA.
P
How not to eat the hype
Ask for a URL. Ask for the token subtraction. Ask if audio goes in. If a theory does not explain +75 AND 1210 AND the video toll, it is a meme. Without our own run or a lab statement, there is no passport.
Golden rule: no subtraction, no passport. No statement, no verdict.
Q
What attention is
Before betting the next token, the model looks at previous pieces and decides which matter. It grows with the square of context: that is why 1 million is not “more RAM”, it is an expensive gaze.
A 40-person WhatsApp group. You do not read all 40 messages equally loud.
R
What temperature is
The urn. Temp 0 = the most likely token always wins (useful for comparing clones). High temp = sometimes the weird one. Not a muse: a drunk urn.
Creativity ≠ truth. A brilliant parrot is consistent at temp 0 too.
S
What KV-cache is
Reusing compute from the start of the prompt. An unchanged repo comes “from the fridge”. OpenRouter lists ~86% hits on Ox. That cheapens 100T; it does not ID the lab.
The same stadium turnstile. Most people re-enter. Cheap. Not a passport.
T
What an agent is
The model asks for a screwdriver (JSON). The program (OpenCode, Claude Code) runs it and returns the result. A loop. Ox is sold for that, not poems.
A surgeon who does not touch the scalpel: they say “incision here”. The hands are the agent.
U
What a hallucination is
Inventing with confidence. The model bets tokens; it does not consult a notary. RAG (pasting a PDF) lowers the risk; it does not kill it. Ox is not Google.
A student who does not remember the statute and drafts anyway. Pretty. False.
V
Lab vs provider vs router
Chef / today’s kitchen / mall. 1210 is a z.ai till stamp. OpenRouter is the aisle. ByteDance can be the waiter. Anomaly is the megaphone (24 employees).
Three layers. The passport lives in the till, not in the aisle.
W
What P50 / tok/s is
Median wait and typing speed. Snapshot ~22 Aug: ~3.88s / 26 tok/s. It moves. A “thinking” model takes longer for the first token. Slower ≠ dumber. Not a fingerprint.
Saturday ER vs the paid API. A different lane.
X
OpenCode vs OpenRouter
Two aisles, one storeroom (Katila: same +75). Zen id: x-preview-f-free. OpenCode says ZDR; OpenRouter says retained, not trained. Default stealth EULA trains. Three signs.
Pick the path. Do not paste your ID on either.
Y
Thinking tokens
Mandatory private draft: low / high / max, never off. Same lever as GLM-5.3. Inflates P50. Not consciousness.
An exam with mandatory rough work on the back. You cannot hand in a blank back.
Z
The sentence you must be able to say
They taped the logo. The knife, the toll and the error yell are Z.ai’s. The brochure car has no camera; this one does. Nobody signed. 80% DeepSWE was n=10. 100T is a stadium, not a meter.
If you stutter, go to the Exam. Pedagogy is that sentence. Identity stays a hypothesis.
Read all · 26
A
What an AI chat is
You type a sentence. A program guesses the next piece, then the next, until it stops. No goblin inside. Just very trained prediction.
B
What it is NOT
Not Google. It does not “search” every sentence (unless someone wired a search tool). No real opinions. It does not know your name unless you paste it in the box.
C
From zero to token
Letters → knife (tokenizer) → numbers. The model only sees numbers. If two models cut the same, they share a knife. Ox and GLM-5.3 cut the same, with 75 extra letterhead numbers.
D
From zero to parameter
Parameter = a knob turned during training. Billions of knobs. “Fewer knobs” can be smarter if used better (MoE: only a few turn per word).
E
From zero to lab
A lab is a company that trains the model (Z.ai, Xiaomi, OpenAI…). OpenRouter is not a lab: it is the mall. Stealth hides the lab’s sign.
F
From zero to stealth
They serve the model free and anonymous to see if traffic holds and what people say. Then they unmask… or not. On this channel, 4 of 4 unmasked as Chinese labs.
G
From zero to evidence
We cannot open the skull. We measure habits: knife, error yells, video toll, cached coffee. One habit is not enough. Seven habits paint a sketch.
H
What you can claim today
Fact: OpenRouter is not the owner. Strong hypothesis: Z.ai GLM-5.x family, a variant with a camera. Not Mira’s Inkling. Not a passport.
I
What a prompt is
The box you type in. The model does not hear you: it reads that text turned into numbers. If it is not in the box (or the history), it does not exist.
J
What the system prompt is (the +75 letterhead)
Hidden text at the start: you are ox-alpha, do not say which lab. Those ~75 extra tokens are the letterhead. That is why Ox always counts 75 more than public GLM-5.3.
K
What multimodal means
Not just letters: photos and video too. Seeing a video means turning it into a pile of expensive tokens. Each lab charges a different toll. Ox charges the same as GLM-5V-Turbo.
L
What a jailbreak is
A trick so the model dumps what the letterhead forbade. Sometimes it yells I am GLM. Soft evidence: it could also be acting.
M
What a prediction market is (Manifold)
People put money (or points) on who is behind this. An August 2026 snapshot was ~85% Z.ai; the market moves. That measures opinion with skin in the game, not a paper. Not a passport.
N
What a data contract is (and why there are three)
OpenCode says zero retention. The OpenRouter card says the provider keeps prompts and does not train. The stealth EULA says free access is paid for with your prompts for training. All three cannot be true at once.
O
What a benchmark is (and the lying 80%)
A test. DeepSWE has 113 tasks. Someone ran 10 and Ox scored 80%. DeepSWE's author ran a subset and got ~63%. Ten questions are not the national exam.
P
How not to eat the hype
Ask for a URL. Ask for the token subtraction. Ask if audio goes in. If a theory does not explain +75 AND 1210 AND the video toll, it is a meme. Without our own run or a lab statement, there is no passport.
Q
What attention is
Before betting the next token, the model looks at previous pieces and decides which matter. It grows with the square of context: that is why 1 million is not “more RAM”, it is an expensive gaze.
R
What temperature is
The urn. Temp 0 = the most likely token always wins (useful for comparing clones). High temp = sometimes the weird one. Not a muse: a drunk urn.
S
What KV-cache is
Reusing compute from the start of the prompt. An unchanged repo comes “from the fridge”. OpenRouter lists ~86% hits on Ox. That cheapens 100T; it does not ID the lab.
T
What an agent is
The model asks for a screwdriver (JSON). The program (OpenCode, Claude Code) runs it and returns the result. A loop. Ox is sold for that, not poems.
U
What a hallucination is
Inventing with confidence. The model bets tokens; it does not consult a notary. RAG (pasting a PDF) lowers the risk; it does not kill it. Ox is not Google.
V
Lab vs provider vs router
Chef / today’s kitchen / mall. 1210 is a z.ai till stamp. OpenRouter is the aisle. ByteDance can be the waiter. Anomaly is the megaphone (24 employees).
W
What P50 / tok/s is
Median wait and typing speed. Snapshot ~22 Aug: ~3.88s / 26 tok/s. It moves. A “thinking” model takes longer for the first token. Slower ≠ dumber. Not a fingerprint.
X
OpenCode vs OpenRouter
Two aisles, one storeroom (Katila: same +75). Zen id: x-preview-f-free. OpenCode says ZDR; OpenRouter says retained, not trained. Default stealth EULA trains. Three signs.
Y
Thinking tokens
Mandatory private draft: low / high / max, never off. Same lever as GLM-5.3. Inflates P50. Not consciousness.
Z
The sentence you must be able to say
They taped the logo. The knife, the toll and the error yell are Z.ai’s. The brochure car has no camera; this one does. Nobody signed. 80% DeepSWE was n=10. 100T is a stadium, not a meter.