Five things that explain most of what you have seen AI do, and then three routes through
what we have published. No prior knowledge is assumed.
Every reading time below is measured from the actual length of the piece, not estimated.
The articles are published by My Digital Sovereignty Limited across its two sites, so some
links change domain — same author, same infrastructure.
All of it may be quoted or republished with attribution. There is nothing to sign up for.
Start here · about six minutes
Five things that explain most of what you have seen AI do
Not a summary of the field. Five mechanisms, each of which accounts for something you have
probably already watched one of these systems do.
1. It predicts the next word. It does not know things.
A language model is trained to continue text. Given what has come so far, it works out
what is most likely to come next, then does that again, and again. That is the whole
mechanism.
Nothing in that process checks whether the result is true. It checks whether the sentence
looks like the sentences it learned from. Fluency and accuracy are produced by different
things, which is why you can get a beautifully written paragraph citing a court case that
does not exist. The model was not lying. It was doing the only thing it does, and a
plausible-looking citation is exactly what should come next in a paragraph like that.
The practical consequence: confidence tells you nothing. It is equally confident when it
is right and when it is inventing.
2. It learned from somebody's writing. Whose?
Try this. Ask it something detailed about New York. Then ask it something equally
detailed about the town you grew up in. The difference between those two answers is the
training data.
These systems learned from an enormous quantity of text, and that text is not evenly
spread. Places, people and subjects that are heavily written about in English are
represented richly. Everywhere else is thin, and where it is thin the model fills the gap
with whatever pattern fits. It does not announce the difference.
So the question "what does it know about this?" is really "how much did people write
about this, in a language and a forum the model was trained on?" For a New Zealand
subject, that is often the whole story.
3. Whoever owns the model sets its manners.
Ask the same contentious question of two different systems and you will often get two
different answers, sometimes in different registers. That is not the models disagreeing
about the facts. Someone decided what each one will discuss, what it will decline, how
cautious it should sound, and whose complaints it should avoid attracting.
Those are ordinary product decisions, made by the company that trained the model, in its
own country, under its own law and its own commercial interest. There is nothing
conspiratorial about it. But it means a sentence beginning "AI says…" is doing a lot of
hidden work. What is actually being reported is what one company has configured its
product to say.
It follows that who owns the thing is not a side issue. It is part of the answer you get.
4. It forgets because it has to.
There is a hard limit on how much text one of these systems can hold in view at once.
Everything counts towards it: your instructions, the conversation so far, any document
you pasted, and its own replies. When the limit is reached, the oldest material falls out
of view.
This is why a long conversation degrades. It contradicts something you agreed half an
hour ago, or forgets a constraint you were careful to set at the start. It is not being
evasive. That part is simply no longer in front of it.
The practical move is to put what matters in the message you are sending now, rather than
trusting that something you said thirty messages ago still counts.
5. Some of it you can check. Know which.
Claims from these systems come in three kinds, and they need different treatment. There
are things you can verify in under a minute — a date, a quotation, a citation, a
statistic. Check those, every time, because they are exactly where invention hides. There
are things you cannot verify but can attribute: who said this, where, when. And there are
things that are simply the model's pattern, with nobody behind them at all.
The dangerous kind is the middle one dressed as the first: confident, specific,
plausible, and attached to no source you can follow.
It is a fair question to ask of anything you read, including this site. Some of what we
publish carries a sealed record — the text hashed, signed, and stamped with a date by an
outside authority — so you can confirm for yourself when it was written and that it has
not been altered since. There is more on that below, including how to check one.
The shape of it · about four minutes
What sits on what
People say “AI” for the whole of this, which makes it hard to ask a useful question about
any part of it. It is a stack. Each layer rests on the one below, and each has a different
owner, in a different country, answerable to different law.
Most public argument is about the middle: which model is best, whether it will take your
job. The layer that decides what actually happens to you is the one directly above it.
Agentic infrastructure is where it is settled what an agent may do, on whose
behalf it acts, what it is allowed to reach, and what evidence it leaves behind. A model
that can draft an email is a curiosity. A model wired to your calendar, your bank and your
filing, with standing permission to act, is a different proposition, and the difference is
not in the model at all.
That layer is where the next wave of products will be launched from, and it is the least
examined part of the stack. It is also the part a community, a business or a country can
actually own — you are unlikely to build a frontier model, but you can absolutely decide
what agents may do in your own affairs, and keep the record of what they did.
A worked example, from Christchurch
Ben Reid of Memia
built Sensorium over eighteen months, working with AI coding agents. He calls it an
exocortex: it reads thousands of technology articles a week and gives him back what matters
for his newsletter. He did not train a model. He assembled the layer above one, and got a
capability that did not previously exist for a research outfit of his size.
That is the shape of the next few years for most organisations. Not building the engine.
Building the thing that decides where the engine is pointed, and keeping the receipts.
The other bet: models that know how things work
Everything above describes systems that predict the next word. There is a serious argument
that this is a dead end for anything requiring judgement about the physical world, and its
best-known advocate is Yann LeCun, who spent a decade as Meta’s chief AI scientist.
His alternative is a world model: rather than learning what text usually
follows other text, the system learns how an environment behaves, so it can predict what
would happen if it did something — before doing it. The technical approach is called
JEPA, and it predicts at the level of abstract representation rather than generating raw
output. A system like that can plan, because it can imagine consequences. A next-word
predictor cannot, which is why it will confidently describe a physical impossibility.
LeCun left Meta in December 2025 to found AMI Labs on this bet. It raised a billion dollars
of seed funding at a three-and-a-half billion dollar valuation, which tells you how
seriously the argument is taken. Worth noticing where: Paris, not California.
The most substantial challenge to the American approach is being mounted from Europe, and
nothing about the next decade is settled.
Machines that sense, rather than read
Everything so far learned from text and pictures scraped off the internet. A separate line of
work starts from the other end: give a machine actual senses and let it learn from what it
encounters. This is where the phrase embodied AI belongs, and the current effort goes
under the ungainly name vision-language-action models — systems that take
an instruction in ordinary words, look at the room, and move something. The unsolved parts are
exactly what you would guess: turning a sentence into a safe sequence of actions, fusing senses
that disagree with each other, and knowing when the machine is not sure enough to proceed.
The sensing hardware is changing under it. An event camera throws the frame
away. Rather than photographing the whole scene sixty times a second, each pixel reports only
when the light falling on it changes — which is roughly what your own retina does. The
result sees in near-darkness, resolves movement at microsecond speed, and produces a fraction
of the data. Prophesee, Sony and iniVation all ship them, and they are arriving in cars,
headsets and robots now rather than eventually.
There is a reason to care beyond the engineering. A sensor that emits a trickle of change
instead of a torrent of video is one you can process on the device, in the room, without
shipping the footage to somebody else’s computer. The economics that pushed perception
into the cloud are not fixed, and some of what made surveillance cheap is being undone by
people optimising for battery life.
Where it is going · about three minutes
What is coming, and what we build on
Models are getting smaller, and stranger
Two shifts matter more than the headline race. The first is small models.
A great deal of what people use these systems for does not need the largest model ever
built, and a specialised model of modest size, trained for one community or one trade, will
often beat a general giant at that community’s work. Small enough matters because small
enough runs on hardware you own.
The second is a different way of generating text altogether. Everything described earlier
writes left to right, one word at a time, which cannot be hurried. Diffusion
language models instead start with masked noise and refine the whole passage at
once, the way image generators do. LLaDA came out of Renmin University, Mercury out of
Inception Labs, which reports over a thousand tokens a second on standard datacentre hardware
— several times what conventional models of similar capability manage. Inception argue the
gain is not only speed: because the model is not restricted to what it has already written,
it can structure an answer as a whole. If that holds, the cost of
running a capable model on your own machine falls by an order of magnitude, and a good deal
of what is currently rented becomes ownable.
Situated language: the thing we build
Our own work sits at that junction. A situated language layer is a model
that has been placed somewhere: it knows a particular community’s vocabulary, its
history and its obligations, and it runs on that community’s own infrastructure rather
than answering from somebody else’s data centre. Not a general assistant with a local
accent — a smaller model that is accountable to a specific group of people.
Placing the model is the easy half. The harder half is that the rules a community sets have
to hold at every layer of the stack above and below it, or they are decoration. Almost
everything published under the heading of AI values is a document: a policy, a charter, a
paragraph in a system prompt asking a model to behave. A model can be talked out of a
paragraph. It cannot be talked out of a condition in the code.
So the same rule substrate runs through the whole of what we deploy, and it bites in three
different places.
While the software is being written. The AI agents that help build this
platform work under hooks that refuse actions rather than advise against them. Material
classified as culturally sensitive cannot be altered by an agent at all without a
single-use authorisation that only a person can issue — and that authorisation
deliberately cannot be granted for a whole session, only one action at a time.
When it ships. A deployment is refused, not warned about, if a page
carries prohibited terminology, if long-form writing has not been through its editorial
pass, if translations are incomplete, or if a check finds internal material reachable from
a public response.
When it runs. The deliberative tool will not average participants into
a number. Where a summary cannot be written without losing a disagreement, it returns a
refusal instead of a summary. That is a post-condition in the code, so it holds whether or
not anyone is watching, and it holds against us as much as anyone.
None of that is remarkable engineering. What is unusual is where the rules sit: not in a
document describing intended behaviour, but in the layer that decides whether an action
completes. The point of building it this way is that our own good intentions are not load
bearing — and neither is our continued existence.
The word gets used loosely. Underneath it there are four separate questions, and they have
different answers.
Sovereign data — whose information it is, who may hold it, and under
whose law it sits when it is at rest.
Sovereign records — whether the account of what happened belongs to
the people it concerns, or to the vendor whose software recorded it.
Sovereign databases — whether one body’s material is genuinely
separated from another’s, in the machine rather than in the terms of service.
Sealed records and timestamps — whether a claim about when something
existed can be checked years later by a stranger, without trusting anyone who holds it.
The last of those is the one that survives when the others fail, and it is the subject of
the longest piece on this site.
Pick by what you are interested in rather than by what you do. Each route names a short
form — the first two steps — if that is all the time you have.
If you have a decision in front of you rather than a question, there is a fourth way out:
Which
kind of AI suits the job? — six questions about your situation, and a prompt you can put
to whatever you are considering. It takes about three minutes and needs no account.
How it works, and why it can be confidently wrong
Take this if you keep meeting claims about AI that you cannot judge.
Writing about AI? Routes one and two together come to about an hour of core reading, and
everything here may be quoted or republished with attribution.
Optional coda: what an alternative looks like in production. Also published in te reo Māori — that translation is provisional and awaiting review.
40 min
Quoting and republishing
Everything linked above is published under CC BY 4.0. Quote it, republish it, translate it,
build on it — with attribution. The urls are stable and each piece carries its date.
Two of the pieces carry a sealed record: the text is hashed, signed under a key we hold
ourselves, and stamped with a date by an outside authority with no stake in the outcome.
You can confirm the text has not changed since that date without asking us for anything.
This guide is a starting order, not a syllabus. If it turns out to be useful, three places
carry it forward.
Ask the Glossary — the button at the corner of every page. If a term
in any of this is unfamiliar, ask it there and you will get a plain definition without
leaving what you are reading. 316 entries, English and te reo Māori. It is quicker than
asking us, and it works at midnight.
Field Notes
— shorter pieces on what is happening as it happens.
What's New
— everything recently published, in order.
Media enquiries
— if the Glossary has not answered it and you are writing something, ask us directly.
For the research itself — papers, threat models, open questions — the
guide to the research
sets out three routes into it.
Where this page relies on work that is not ours — world models, embodied AI, event sensors,
diffusion models, Sensorium — every source is listed at
Sources for the reading guide,
marked according to whether it is the party themselves, peer literature, or reporting.
