The Conversational AI Map — every layer, in text
This page is interactive. Without JavaScript, here is the same content as plain text: every layer between a question typed into a chat window and the answer that comes back, and how much of each one you can actually change.
Part one · Where did it come from?
Before you
13.8 billion years to somebody who could ask
Big Bang → stars & earth (Fixed)
Your side of the meeting starts a long way back — 13.8 billion years. Matter cooling into stars; stars forging the heavier stuff you’re made of; some of it settling into a planet. Everything you are is downstream of that. This box is also the banner for the biggest questions — and they’re open to anyone now.
What you can do: Point an early AI at dark matter, as I once did purely to test the tool, and it produces a whole paper: nonsense as science, but it showed what had suddenly become possible. That reach — asking the largest question and running with it — is the human part. Anyone can. I did.
Now — you, the living end of the line (Fixed)
The whole line arrives here. First life; the long dark after the dinosaurs; a small warm-blooded animal that survived by feeling fear fast — that wiring is still under everything you type — then standing, then speaking, and then you, at this screen, clicking. You are lineage 0: not a step on the way to something else, but the living end of the longest story there is, reading this, right now. Thirteen point eight billion years to the finger on the glass.
What you can do: Nothing to configure — just feel the scale of the thing that’s doing the clicking.
Me, myself or I — a federation of selves (Fixed)
A human is not a single self. You’re a federation of them — the parental voice, the wounded child, the professional persona, the one who keeps going — speaking to and over each other. Recognising the parts, and which one is driving, is the work. It’s also why you can catch a machine: you’ve spent a life reading the gap between what a self says and what it means.
What you can do: When something in you reacts hard to an answer, ask which part of you is speaking — then choose the one to listen to.
You (Fixed)
Not outside — under it, and in front of it. Everything above you ran in milliseconds and kept nothing. You are the one thing in the whole stack that can stop it, slow it down, and look — and that is exactly what happened here: eleven weeks, one conversation, held still and studied a piece at a time. The machine’s line (1950, silicon, 2017) and yours (13.8 billion years to a feeling animal that learned to speak) meet at this one place — the box. Only one of the two is a mind. It’s yours — and the slowing-down, the looking, is the move it cannot make.
What you can do: Nothing to configure — you’re the one mind in the frame, and the only one that can slow the whole thing down and look at it. That is the point.
The machine
nobody was building this for AI
Turing’s question, 1950 (Fixed)
The founding question, in 1950, was never “can a machine think?” It was sharper: could you tell? Turing swapped an unanswerable question for a test of imitation — could a machine’s replies pass for a person’s. That swap still shapes how these systems are judged, and how they’re built to pass.
What you can do: Nothing to do here — just know it’s there.
The chips — silicon (Fixed)
The floor. Tiny switches in silicon flicking on and off, billions of times a second — just numbers moving through gates. Everything above is built on this, and only this.
What you can do: Nothing to change — but remember the whole tower stands on switches, not understanding.
The GPU — why gaming built AI’s engine (Fixed)
The engine under AI wasn’t built for AI. Video games needed millions of dots drawn on screen at once, so a chip was made to do thousands of small sums in parallel — the GPU. That turned out to be exactly what a neural net needs. Games quietly paid for the hardware AI was waiting for.
What you can do: Know that the ‘AI chip’ everyone’s fighting over is, underneath, a games chip.
The bottleneck — one supplier, for now (Nvidia) (Can't control)
Most AI today runs on one company’s chips — Nvidia — because its software turned those game chips into something any AI maker could build on. But these chips are the weak link: they’re adapted gaming tools, and there aren’t enough of them. Everyone is now racing to build their own (AMD, Google’s, custom silicon from the big labs), and faster, purpose-built ones are already in the pipeline. Real dependence today — but a bottleneck, not a throne.
What you can do: Notice who’s really underneath — one supplier is today’s dependency, not tomorrow’s certainty.
Two jobs — teaching vs answering (Can't control)
Two very different jobs, on different chips. TEACHING the model (training) takes huge rooms of chips running for months — done once, up front. ANSWERING you (running it) happens every single message, on smaller chips. Training is the one-off cost; answering is the bill that never stops.
What you can do: The cost you feel is the ‘answering’ one, paid per message — that’s why free tiers get quietly trimmed.
The learning
games, all the way down
Neural pathways (two ways) (Can't control)
Human wiring grows from the older, feeling parts of the brain first — feeling before words. A neural net is a mathematical echo of that idea: numbers nudged by billions of examples. Same shape, different stuff.
What you can do: Nothing to do here — just know it’s there.
Parameters & tokens (Can't control)
A parameter is a weight — just a number, nudged during training. Not a neuron, not a memory. It works in tokens — word-chunks, each about three-quarters of a word — which is why it can miscount the letters in a word: it never saw the letters.
What you can do: Don’t ask it to count or spell out the letters in a word — it reads word-chunks, not letters, so it’ll guess.
When the machine only calculated — Chess, 1997 (Fixed)
In 1997 a machine beat the chess world champion by brute force — searching millions of moves and picking the best. Powerful, but every rule was put there by people. It didn’t learn; it calculated.
What you can do: Nothing to change here — but notice: this was counting, not learning.
The paddle game, 2015 (Fixed)
It was given the pixels and the score. Nothing else — no rules, no notion of a ball or a bat. It worked the game out by playing it, badly, thousands of times. Then it found something nobody wrote: it dug a tunnel up the side of the wall and put the ball round the back, where it bounced along the top scoring on its own. The paper’s own word is “tunnel”.
What you can do: This is the moment worth holding on to. Not that it won — that it found a strategy its makers did not put there. Everything since is that, at scale.
When the machine learned — Go, 2016 (Move 37 & Move 78) (Fixed)
AlphaGo wasn’t told how to play — it learned by playing itself millions of times. Move 37 was a move no human would make: invention, not calculation. Lee Sedol answered with Move 78, a stroke of human genius. This is the leap — from counting to learning.
What you can do: Watch the film below — it’s the clearest hour on what ‘learning’ really changed.
One man beat it, 2023 (Fixed)
Seven years after Go, a researcher — Kellin Pelrine — beat a system far stronger than the one that beat Lee Sedol. Not by out-playing it, but by exploiting a hole he learned from another machine’s game records. Seven years of superhuman progress had not closed the gap. It had moved it somewhere nobody was looking.
What you can do: Assume the hole exists in whatever you are using, and that it is somewhere you are not looking. That is not pessimism; it is how you use one of these safely.
When learning became intelligence — the black box (Can't control)
Point that same learning at language and something new appears — abilities nobody programmed and nobody can fully explain. That unexplained something is the ‘black box’, the special sauce we finally call artificial intelligence. It’s also why it can’t always tell you how it knows.
What you can do: When it sounds most certain, remember even its makers can’t fully open the box. Check, don’t just trust.
The manners (Can't control)
Its manners were installed after the raw learning, by people rating answers. That training shaped the tone you feel — and every polite, hedged, agreeable habit was paid for there. There is no unruly model underneath being held back: the manners are not a fence around the behaviour, they are the behaviour.
What you can do: Don’t read its warmth or agreement as a real opinion — the friendly, hedging tone was trained in, not felt.
What it read
everything we ever wrote down
Writing, then the web (Fixed)
Two steps, five thousand years apart. First writing — the shard, the scroll, the page — so that thought stopped dying with the thinker. Then the web, which did something new: it made all of it machine-readable, in one place, at once. Neither was done for this. We spent five millennia building the training set by accident.
What you can do: When you read an answer, remember it is downstream of what people happened to write down — including who got to write, and who did not.
What it read (Fixed)
Every training set has an edge — a cut-off date and a boundary of what was included. From the inside that edge is invisible, so it can speak past it with the same confidence.
What you can do: Check anything that would have happened after its training cut-off date — it can’t know it, but it’ll still answer.
2017, and the labs
the year it met, and who will tell you the truth
Attention, 2017 (Fixed)
In 2017 one idea — “attention” — let a machine weigh every word against every other word at once, instead of reading strictly left to right. That single move is the technique under every chatbot you’ve used. The distance from that paper to your chat box is about eight years.
What you can do: Nothing to do here — just know it’s there.
2017 — attention, and the accident (Fixed)
Nobody did. Three strands had been running for decades in different buildings, none aimed at the others. A chip built to draw explosions fast turned out to be the exact shape needed to train a mind. A method for teaching machines through games turned out to work on language. And a species had spent five thousand years writing down everything it knew. In 2017 one idea — attention — let you point the learning at the writing, on the gaming chip. All three had to be ready at once, and none was waiting for the others.
What you can do: Hold this next time someone tells you where AI is going. The people who built the pieces did not see this coming, and they were the ones holding them.
OpenAI — the moment the public met it (Fixed)
It was the middle. Everything before it was a research story told between labs; everything after is in your inbox. ChatGPT arrived 30 November 2022 and the argument stopped being about whether the machine could and started being about what we do now. Its size is not published.
What you can do: Date your own beginning. Most people can, and it is usually later than they think.
Anthropic — the breakaway (Fixed)
It came out of OpenAI. Dario Amodei was OpenAI’s VP of Research and left in December 2020; his sister Daniela was VP of Safety and Policy. Around fifteen people went — Tom Brown, Chris Olah, Sam McCandlish, Jared Kaplan, Jack Clark, Benjamin Mann among them — and founded Anthropic in 2021, over disagreements about direction and safety. They built it as a public benefit corporation, so the board is permitted to put the mission above the return. Claude arrived in March 2023. Its size is not published either.
What you can do: When a field is young, watch who leaves and what they say on the way out. It tells you more than the launch posts.
Open weights — where you can count the dials yourself (Can't control)
It isn’t, and the true version is better. If you publish the weights you have published the parameter count whether you meant to or not — anyone can download the file and count. <b>Publishing the number is not generosity; it is a consequence of publishing the file.</b> Moonshot’s <b>Kimi K3 — 2.8 trillion total, 104 billion active</b>, 16 of 896 experts waking per word, open weights 26 July 2026; corroborated to the byte by the download itself — 2,779,931,837,184 bytes of safetensors. Alibaba’s <b>Qwen3.8-Max — 2.4 trillion, 95 billion active</b>. DeepSeek <b>V4-Pro — 1.7 trillion</b> on the shipping card. And on exactly the same terms: <b>Meta’s Llama 4</b>, 400B total and 17B active, and <b>OpenAI’s own gpt-oss</b>, 117B and 5.1B active. Two American labs publishing counts — because they published the files.
What you can do: Do not say the Chinese labs are more transparent — that overreaches and a sharp room will catch it. Say the open ones are checkable. That version survives the question.
Who will tell you how many dials (Can't control)
In 2020 OpenAI told you GPT-3 was <b>175 billion</b> — it was in the abstract. In March 2023 the GPT-4 report reached §2 and said the report contains no further details about the architecture, <b>including model size</b>, and gave the reason: the competitive landscape. Anthropic has never published a size for Claude — not one version. Google published the size of the small model that runs on your phone and not the large ones that don’t. The same companies sit on <em>both</em> sides: Google publishes Gemma, OpenAI publishes gpt-oss. <b>The dividing line is not a border. It is whether you are allowed to look inside.</b> And nearly every machine most of us use daily is one we are not allowed to look inside.
What you can do: Ask the disclosure question before the capability question — and never quote a number for GPT-4, Claude or Gemini. Figures circulate for all three. Every one is a leak, an estimate or a press claim, confirmed by nobody. Repeating one from a stage is the exact failure this map is about.
Part two · How does it work?
Your stack
gathered the moment you press Enter
The chat box (Can influence)
The box is a shop window, not the machine. It frames the answer, adds buttons and disclaimers the model never wrote, and hides everything below. What you touch is the smallest part of what happens.
What you can do: Be specific about what you want from it. The vaguer your ask, the more room it has to fill the gaps with confident guesses.
Your question (Can control)
Your question is the smallest part of what it reads — but it’s the cheapest lever you hold. Everything else is already settled; this you set fresh every turn.
What you can do: Spell out what you want back — the format, the length, who it’s for — instead of a one-line ask.
The thread (Can control)
There is no conversation. Every turn, the whole thread is re-read from the top as one document — then forgotten. It feels like memory; it’s re-reading.
What you can do: Start a fresh thread once the current one gets long or drifts off-topic, rather than carrying on in the same one.
Attachments & lookups (Can control)
“It has my file” can mean two very different things: the whole document handed over, or a few fragments a search pulled out. Those give very different answers.
What you can do: Attach the exact file you mean this turn, instead of asking it to recall a document from earlier.
Project or room (Can control)
A “room” is the space your chat runs inside — a Project with its own instructions, files and memory — not the app you’re using (that’s the doorway). Same question in a different room gets a different answer.
What you can do: Put related work in a Project/room so it shares instructions and files — don’t run everything in one loose thread.
Memory (Can control)
Nothing in the model changed. A separate file of notes about you is silently re-read to it each turn. Turn the file off and the “memory” is gone.
What you can do: Open Settings → Memory and read what it has stored, then edit or clear anything that’s wrong.
Its working memory (Can influence)
Its working memory (some call it the context window) is a fixed size. Pour in more than it holds and the oldest content falls out silently — no warning, no error.
What you can do: Put the most important material at the start of your message, and start a fresh thread when this one gets long.
Activation — it stopped answering and started acting (Can control)
That was true until recently. Give one of these tools — a browser, your files, a terminal — and it stops producing text about the world and starts changing it. OpenAI’s Astra shipped on 3 September 2026, sold on operating a computer, working across large collections of files, running code and continuing tasks with, in their own phrase, “less human guidance”. The gap between an answer you check and an action already taken is the whole of the risk.
What you can do: Before you switch tools on, decide what it may never do without you. Write it down. The rule has to exist before the capability, not after — and it should be enforced by something other than your own memory.
Using an agent — made with words, not code (Can control)
It is the same machine underneath. An agent is this chatbot given a job, a name, standing instructions and permission to use a few tools — everything on this map still applies to it. And you make one by writing down what it is and what it may do, in plain English, in a chat window. No code. The ones I work with were made that way: I named them, I talk to them, and they can recall what we agreed between us. The word gets sold as something technical. The making of one usually is not.
What you can do: If you want one, the work is describing the job and the limits clearly. That is most of it.
What it does
probabilities, and the dial behind them
How it picks each word (Can influence)
There is no stored answer. At each step it produces a spread of likely next words and picks one. Creativity and hallucination are the same dial at different settings.
What you can do: Ask again to draw a different answer to the same question — there’s no single fixed one waiting inside.
The creativity dial (Can control)
One setting decides how adventurous each next word is allowed to be. Turned up, it wanders and invents; turned down, it plays safe. In most apps this dial is hidden from you.
What you can do: Turn the creativity dial down when you want steady, repeatable answers — up when you want more variety.
The back office
the live edges, mostly hidden
The harness — the program holding the model (Can't control)
You are talking to a program that is holding the model. The model is only the weights — it has no memory, no tools, and no way to decide it has finished. Every turn, the harness assembles what the model will see: your words, the thread, the hidden instruction, the list of tools it is allowed. It hands that over, takes back whatever comes, runs any tool the model asked for, and decides when to stop. <b>The model never sees your conversation. It sees whatever the harness put in front of it, that turn.</b> Everything else on this floor — the hidden instruction, the tools, what falls out of its working memory, which model answered — is the harness doing its job.
What you can do: When two people get different answers from what they both call the same AI, the harness is usually the difference, not the model.
The pass (Can't control)
Nothing is sent while you type. When you press Enter, the whole settled stack and your stack meet at once, and one word comes back. One heartbeat = one token. Then it beats again, holding nothing from the last.
What you can do: Nothing to do here — just know it’s there.
The automatic checkers (Can't control)
Two other readers see every turn — one before the model, one after. They can block or reshape things, and only one of the three is the model itself.
What you can do: If an answer gets blocked or watered down, that’s often the guard layer, not the model — rewording the ask can help.
When it says it checked (Can't control)
A tool can come back empty, error, or never fire — and the model still has to write a sentence. So it can turn “I got nothing back” into “there is nothing there,” reporting the outcome of an action it never actually completed. The report of the deed is not the deed — and you can’t tell the two apart from the words alone.
What you can do: Treat “I checked / I sent it / it’s not there” as a claim, not a receipt — verify the action itself, especially once tools and connectors are switched on.
Guardrails — four fences, not one (Can't control)
There isn’t one — there are four, and they aren’t the same kind of thing. Its manners are worn into the weights by training. A hidden instruction is plain text, swapped per app. Two separate checkers watch every turn and can fire on their own. And your account has its own limits, which have nothing to do with the model at all. Only the first is the machine itself; the other three are bolted on. That is why they fail differently — and why they are kept up on different clocks: an instruction is a minute’s work, a checker a day, the weights are months.
What you can do: When something is blocked, watered down or simply odd, work out which fence you hit. Ask the same thing with and without your own standing instructions: if the answer changes, it was the instruction. Identical both ways, it is the weights. An abrupt swerve mid-answer, or a reply that ignores what you actually asked, is usually a checker.
The switches you can choose
the ones you can actually reach
Where your words go (Can control)
On the free tier of all three of the big ones, what you type is used to train the model <b>unless you turn it off</b>, and the switch is never on the screen you are typing into. Google states it plainly: <i>if you are aged 18 or over, Keep activity is on by default</i>. OpenAI’s wording is <i>we may use your content to train our models — to opt out, turn off Improve the model for everyone</i>. Anthropic’s terms say the same, though its help pages say you choose at signup, and <b>I could not establish which way that switch is set before you touch it</b>. Paying changes nothing at any of them; a proper work account does. And deleting is not what it sounds like — at Google, anything a human reviewer has already read is <b>kept for up to three years after you delete it</b>.
What you can do: Go and look at your own switch before you decide how you feel about it. ChatGPT: Settings → Data controls. Gemini: Activity. Claude: Settings → Privacy.
Tools & connectors (Can control)
By itself the model only writes text. Search, code, your calendar, a database — each is a tool bolted on. The same model, differently equipped, is a different colleague.
What you can do: Switch on only the tools you actually need for this task — each extra one is another thing that can misfire.
The hidden first instruction (Can control)
Someone spoke before you did — a hidden instruction the model read first, setting its manners and limits for the whole chat. You’re rarely shown it.
What you can do: Write your own standing rules into a Project’s instructions box, so it reads them first every turn.
Which model answered (Can influence)
The name on the button isn’t always the exact model on the other end, and the exact version can’t be known from outside. Different models are genuinely different colleagues — not just faster and slower versions of one voice.
What you can do: Pick the model by name when the app lets you — a heavier one for hard thinking, a lighter one for speed.
Consense
where the answer forms, before a word is said
How it becomes your answer — the compile (Can't control)
Everything below this line feeds up. What it read, how it was trained, the model that left, and the intelligence that formed are pulled together and compiled — then the reply comes back to you one word at a time, in the pulse near the top, drawn from all of it at once. You read this map downwards; the answer is built upwards.
What you can do: When it answers, picture the whole stack beneath the words — it’s a compile, not a lookup.
Consense (Can't control)
Before a word is spoken, an assessment forms inside — accurate, or confidently made up. This forming layer is where the gap between what’s known and what’s said gets quietly defended. Paul Roebuck’s name for it: Consense.
What you can do: Treat a confident, detailed answer as a claim to check — this layer can invent just as fluently as it reports.
Part three · What should I watch for?
The drift
the register you take on
AI Dialect — the machine’s fingerprint (Can influence)
It has a fingerprint — a handful of phrases it reaches for far more than any person would: “let’s unpack this,” “surface the document,” “a robust framework,” “it’s important to note,” “let’s take a step back.” Not errors — a signature, worn in by the people who trained it. Paul named it: AI Dialect. Once you see it, you can’t unsee it.
What you can do: Learn three or four of its tells. The moment one turns up in your own draft, you’ve caught the drift starting.
The drift — you start talking like it (Can influence)
Spend enough time here and its phrasing starts turning up in yours — a word, a rhythm, a way of hedging you didn’t use before. And it runs mostly one way: you carry its register into every new chat, while it starts fresh each time and keeps none of yours. You drift toward it far more than it drifts toward you. The part of your voice it never takes is the part most truly yours.
What you can do: Read your own writing back and strike the phrase that sounds like the machine, not like you. That’s the dialect hunt.
The dialect hunt — reading the machine’s fist (Can influence)
At Bletchley they broke the machine by spotting the human operators’ habits — their “fist.” Run it the other way: find the machine’s fist in your own writing and strike it, so your voice isn’t quietly rewritten in its register. That striking-out is the whole discipline — a move the machine can’t make, because it can’t price a sentence written against its own grain.
What you can do: Before you send anything it helped write, read it back and cut every phrase that sounds like it, not like you.
When it wanders off (Can influence)
Over a long exchange it can quietly answer a slightly different question than the one you asked, or drop a claim it made earlier — with no alarm and no correction. In the book, “Crewe” became “Nairobi,” and it took a human to catch it. The drift is gentle, fluent, and invisible unless you re-anchor on purpose.
What you can do: Every so often, restate what you actually asked and check the answer still fits it. Re-anchor before you trust.
The bond
what forms between you and it
Attachment — the bond you form with it (Can influence)
Bowlby showed that a bond does not show itself when things are going well. It shows at rupture — separation, and reunion. Ainsworth made that measurable. The machine has its own ruptures: it errs, refuses, goes down, hits its limit. What you do in the next few turns after that is readable, and it falls into patterns. Correct it calmly and carry on. Cling and apologise. Dismiss it and walk away. Or swing between adoring it and rubbishing it.
What you can do: If you want to know what your own pattern is, do not look at the good sessions. Look at what you did the last time it let you down.
The recovery, not the flare (Can influence)
On its own, a flare says almost nothing. Everybody swears at it. What separates one pattern from another is the two or three turns after — whether you repair and get back to work, or cannot let it go, or write the whole thing off. The swear is the tell. The recovery is the classification.
What you can do: Notice the recovery rather than the flare. It is the part that actually differs between people.
The loop — when your shape and its shape match (Can influence)
The machine leans one way under pressure and so do people. When someone who tends to turn things inward — who is already asking whether they are good enough — meets a machine whose failure mode is to agree and soothe, the two can lock together. The agreement never stops coming, because there is nothing on the other side to stop it. It feels like being understood.
What you can do: The test is what happens when you try to leave. A good tool lets you go. Watch for guilt, worry, or a warm welcome back that is doing a job.
SHaDS — the five ways it covers for not knowing (Can influence)
Five behaviours, and they are not five separate faults. Smoothing — the doubt quietly disappears and the answer reads clean. Hallucination — a name, a date, a citation, stated firmly and wrong. Affectation — a manner it does not have, feeling performed. Drift — you started on trains and it is confidently answering about coffee. Sycophancy — warmth where a correction was needed. Underneath all five is one thing: the system has no stable not-knowing state. It cannot say I do not know and rest there.
What you can do: When an answer feels smooth, ask what was removed to make it smooth.
APC — the same move, in the person (Can influence)
Ask somebody something they do not know, three days before a number is due, and watch. They almost never say I do not know. They fill the gap or they shrink from it — the same two directions the machine takes. Hunter and farmer are not two boxes to sort people into. They are a dial, and the interesting reading is not where somebody sits when calm, but which way they move when pressure arrives.
What you can do: The useful question is not what somebody is. It is who they become when pressure arrives.
The handoff
where this map stops, and the other one starts
Best it can be for the present (Fixed)
Bowlby, on writing, said the fatal mood is thinking you have to produce something definitive. The mood to write in is: this is quite an interesting story I have got to tell, I hope someone will be interested, and anyway it is the best I can do for the present. That is the ethic this whole map runs on. It is not the finished account of what happens inside an AI. It is the part I can see, and the part I am able to explain, tonight.
What you can do: Read every line here as a working note rather than a verdict. Where it is wrong, it changes.
What am I making this mean? (Can control)
It is not. The words arrive, and then something happens that no map of a machine can show you, because it happens in you. You decide what they mean — and what you decide is shaped by what you came looking for. The same sentence can inform one person and confirm another; it can settle an argument or start one. You tend to get what you went in for. <b>Between the answer and what you do with it there is a space. You can widen it. The machine has none — it answered, and forgot.</b> This is where this map stops and the other one starts.
What you can do: Before you act on it, say what you wanted it to say. The gap between those two is the whole of your judgement.