LLMs Unplugged

Does ‘Unplugged’ still matter in the age of AI?

Everyday life is increasingly regulated by complex technologies that most people neither understand nor believe they can do much to influence.
Albert Bandura, 2001

what we get told about AI

it understands you—it’s basically a mind

it’s just autocomplete—it’s basically a parlour trick

it’s hard to have an opinion
on a machine you’ve never seen
the inside of

a page of token pairs — a word, and the word before it

the algorithm

I call out the last token we wrote (and its colour)
hands up if it’s the boxed token on your page
I pick a hand at random — you read out the token beside it
I write it down, and we go again

Walk it once, slowly. Reading it off the screen is fine.

Step 2: one hand each. More than one match? Go with the first one you spot.

Step 3 is the one to labour, briefly: I pick at random — not whoever’s fastest, because that samples the quickest reader in the room rather than the text. Random is what makes the room behave like the model. Don’t explain why yet; they’ll see it.

Then the seed: someone near the front, sheet in hand, reads out any one pair, boxed word first. It matters that it came off a sheet rather than out of your head — the room supplies the first two words as well as every word after them. If their pair is an odd one (a full stop, a name that appears once), take it anyway: a thin first round is a real property of the model, not a failure. It can’t dead-end on you, so you never need a second volunteer.

4:00

Shift-W, write the volunteer’s two words, call out the second of them, go. Four minutes, 12–15 tokens, write BIG for the back of the room. No commentary between hands going up and a word landing on the board.

Narrate two things, once each, without stopping the flow. The hands are the distribution: call a common word and a forest goes up; call a rare one and you get two. Nobody here has been taught weighted sampling and the room is doing it anyway — the number of hands is how often that continuation happens in the text. (Pedant’s footnote, only if a pedant asks: someone holding several matches for one context raises one hand, which flattens the busiest words slightly.) And the choosing is a coin toss: when five hands are up, the four I don’t pick were all real continuations. Different pick, different sentence. Same reason the “regenerate” button gives a different answer.

Hands will go up every round. Every distinct token appears as a boxed word somewhere, so there is no word you can call that has nothing following it — don’t plan a “running dry” beat. The one way the room goes quiet is an empty seat: that sheet went home with someone, and its pairs are simply gone. Say exactly that, and reseed. The loud seeds for the corpus are on the brief (page 1 of the printed set); avoid restarting from a full stop, or the first thing the room does is argue about punctuation.

Stop while they still want more, around 15 tokens. Read the sentence back.

we just ran a language model

what role did you play?

each word on your sheet → a token

the pairs on your page → the parameters of the model

the hands in the air → the probability distribution over next tokens

my random pick → sampling

the sentence I wrote → generation

the sentence we wrote

is not in the book

between us, we have the whole book

1,962 words · 240 distinct · 1,961 token pairs dealt round the room

no single sheet has the whole picture

the same model in grid form

lookatme!lookatme!lookatmenow!

Token	`look`	`at`	`me`	`!`	`now`
`look`		\|\|\|
`at`			\|\|\|
`me`				\|\|	\|
`!`	\|\|
`now`				\|

three things Claude* does that we didn’t

it looks at more than one word of context

it can say what it never saw—cat teaches it about dog

it’s been taught to answer, not just continue

The asterisk is yours to explain, in one line: Claude here means any frontier model — ChatGPT, Gemini, Llama. Nothing on this slide is specific to Anthropic’s; I just need a name to point at.

Say each in a sentence, resist explaining any of them; each one has a full module on the website and none of them fits in this talk.

One word of context is why our sentence wandered — when I called the, nobody could tell whether we’d just written “sat in the” or “hit the”. Claude looks back over the whole conversation.

Our model rates any pair it never saw as flatly impossible, however sensible. A real model spreads what it learns across billions of numbers instead of separate cells, so what it learns about cat reaches dog.

And ours only ever continues text. Ask it a question and it continues the question — that’s not it refusing, answering just isn’t a thing it does. Turning that into something that answers you is another round of training, on conversations.

None of these changed the loop. Look at the context, weigh the options, pick one, write it down.

context how much it looks at before each guess

1 token 1 token101001 thousand10 thousand100 thousand1 million tokens

training text how much the model reads

1,962 tokens 1,962 tokens10 thousand100 thousand1 million10 million100 million1 billion10 billion100 billion1 trillion10 trillion15 trillion tokens

model size how many learned numbers it contains

57,600 grid cells 57,600 grid cells100 thousand1 million10 million100 million1 billion10 billion100 billion~1 trillion parameters

one axis, in powers of ten · three independent dials · different kinds of model

Three independent scales grow. The first is how much the model looks at before each guess. Ours looked at exactly one token — which is why our sentence wandered — and Claude looks back over up to a million tokens, the whole conversation and then some.

The second is how much text went past the counter. Ours read 1,962 tokens; fifteen trillion is the published figure for a recent open model’s training run. Every one of those tokens went past a counter, the way the book went past yours.

The third is how many learned numbers the model contains. Ours had a 240 by 240 grid: 57,600 places a count could live. A trillion parameters is the right order of magnitude for a frontier model, and the honest answer for the closed ones is “about that, and nobody outside will tell you”.

The bars have separate scales because these quantities can be changed independently. And the endpoints are different kinds of model: grid cells on the left, neural-network parameters on the right. This is a scale comparison, not a claim that one is a compressed version of the other.

much more context

much more training text

many more learned numbers

a different kind of model

still just words in → words out

back to Bandura

most people neither understand nor believe they can do much to influence

a room full of people just did, slowly, the thing you’re told needs a data centre

it doesn’t make you an expert

it makes you harder to bullsh*t

llmsunplugged.org/talks/unplugged-in-the-age-of-ai/

LLMs Unplugged#

Does ‘Unplugged’ still matter in the age of AI?#

what we get told about AI#

the algorithm#

what role did you play?#

between us, we have the whole book#

the same model in grid form#

three things Claude* does that we didn’t#

back to Bandura#