← The dumbest language models

Just how simple can we make a language model? Starting from the ground up, I want to explore just how little you have to do to generate coherent English.

0th order

We’re going to start with two assumptions of language:

1. All letters have the same chance of showing up
2. Letters are independent of one another meaning the letters which come before don’t influence the next letter we output

g:rcbehLyPeK-Movz$Zo;&sZ’3&wopz?T A&a3UYc :RWcpSK?aGOpDggXP!ggaiP,vZdeWRFBU;zQpdiY a.PUEjgKIVYW&j$X

Surprise surprise, we get complete and utter rubbish.

1st order approximation

But we know both of those assumptions are nonsense. What happens if we remove the assumption that letters show up with equal probability? Rather than trying to guess how often they each occur, we can estimate it by counting frequencies from some data (I’m using some Shakespeare here).

Let’s try generating more text with this assumption:

CunCTelheefdtuh:dides.rrr.lae r hekoa qhtnr hnuewy ;todafeUO,euatosnnle?wmUhGv ree res Ah oa orel

It’s hard to tell if this is much better. We’re getting less punctuation then we did before, but it would still be a far stretch to call this real language.

2nd order approximation

Let’s remove our second and last assumption that each letter is independent of one another. If you see the letter ‘Q’, it’s far more likely the next letter is an ‘u’ rather than a ‘Z’.

We’ll tackle this by creating a simple digram. A digram just says that for each letter, store counts of how often the next letter is. When we’re generating more letters now, we depend on the last letter to set probabilities of the next one.

Why stop there? You can keep going and have each letter depend on the previous 2, 3, 4, … This kind of modeling where your next prediction depends on the previous n pieces of data is called an n-gram.¹

rirupld,
HOLAs my, nit idee ber, ayonghingathop whrs.
I ngoon abesithisth woty;
IA the il s.
K:
Thie

Even with just a digram, we’re already starting to see some structure emerge like periods at the end of lines and the NAME: format that the data has. Still a far stretch from the coherent outputs we’ve come to expect from GPT.

Words instead of characters

Of course, we don’t need to use characters as our building block. What if we instead used words as what we were predicting on? I’ve generated 30 word samples and included them below:²

0th order:

She said this was a white again? planted Cushions, method, wise, behold men’s pass’d Romans. vent comes Must, side saving senators, Ladders, covets certain themselves, Marry, requests movers fear. hell. Romans.’ Brother alike. brave? natural

1st order:

She said this was a by
BRUTUS: duke. your But it lack i’ were wealsmen Capitol– make Only in or would think stopp’d my estimation my prey angry Volsce, i’ stripes mother, hang I your

2nd order:

She said this was a humorous patrician, and parent.
CORIOLANUS: I count my nature is left,
Marcius: A weeder-out of wine.
Second Citizen: But we the king, To your voices, I came, Ready to put

Now that we’re generating chunks of words, it’s clear how much better this performs. We used the same techniques, but even a simple 2nd order approximation of language is starting to generate grammatical sentences.

So on and so forth

It’s kind of hard to believe from what we’ve generated above, but this is the ideological framework of ALL language models. From the idea that we can statistically model text to generate coherent English, we need three components:

1. Data to learn from
2. An atomic unit of language (also known as tokens)
3. Some way to look at previous outputs and generate new ones

So what are we missing?

• We used 1 million characters to try and learn, but this is nowhere near enough to learn. Language models today basically swallow up the entirety of the internet.
• Rather than using characters or words, LLMs use more general tokens which are learned from the data. You can play around with one such tokenizer here.
• Lookup tables, like the digrams we used, are expensive and they don’t scale. For a n-gram, you’d have to have V^n size lookup table where V is the number of tokens you have in your vocabulary. How do you compress this to be smaller? Neural networks!
  • With neural networks, the goal is to compress larger and larger n-grams into a smaller approximation that is still quite accurate.

If you want to play around with improving on the simple methods I illustrated above, this is a notebook that I used to generate the samples.

Extra: Historical Note

This idea of using ordered approximation of English comes from Claude Shannon and early works trying to explore how to get an ergodic source of text. If you want to get shivers at how well they predicted LLMs in a general sense, you should go read A Mathematical Theory of Communication.