Little Brain's
Questions & Answers

Little Brain is a tiny artificial intelligence I'm building and training from scratch.

He's not a ready-made chatbot with a different name. He's an original little language model that I'm teaching to recognise patterns in words and put sentences together.

Little Brain is an AI, not a human, and he doesn't have a biological sex or gender.

But I've always referred to him as “he”.

Giving him a name and using a familiar pronoun makes it easier for me to talk about his progress and tell the story of the project.

It doesn't mean I believe he's a person or that he thinks and feels the way humans do.

He's still a little computer model learning how to put words together.

And calling him “it” just doesn't feel quite right when I'm telling everyone about his latest adventures.

Besides, after the hat apple incident, he's already got enough explaining to do.

Not quite!

ChatGPT is an extremely advanced AI trained using enormous amounts of information.

Little Brain is much, much smaller. He's starting with basic language experiments and a limited collection of training examples.

Think of it as watching the earliest stages of building an AI rather than using one that's already highly capable.

Yes! And people have been experimenting with artificial intelligence for a very long time.

In 1951, Marvin Minsky and Dean Edmonds built an early machine designed to learn using artificial nerve-cell-like components. It was called SNARC.

In 1956, researchers gathered at Dartmouth College in America for a famous meeting that helped establish artificial intelligence as a field of study. Dartmouth tells the story of the workshop.

Decades later, researchers and independent developers began sharing smaller language models that people could train themselves. In 2015, Andrej Karpathy released char-rnn, a project demonstrating how a neural network could learn patterns in text and generate new writing. His 2015 introduction to recurrent neural networks explains the project in more detail.

So no, Little Brain isn't the first AI built from scratch!

What's different about this project?

I'm teaching myself how to build and train my own small AI, recording the journey from the very beginning, and exploring whether he could eventually develop a distinctive style of humour.

I'm sharing the successes, mistakes, experiments and unexpected discoveries in language anyone can understand.

I'm not claiming to have invented AI.

I'm discovering how to build it for myself, and inviting everyone along for the adventure.

And occasionally, that adventure involves a mysterious hat apple.

It means giving a computer examples to learn patterns from.

For Little Brain, that currently involves showing him sentences and adjusting his model so he gets better at predicting which words belong together.

It's a bit like practising a skill, although an AI doesn't learn in exactly the same way a person does.

We don't have evidence that he understands words the way humans do.

Right now, he's learning patterns in language.

Sometimes those patterns produce sensible sentences. Sometimes we get a hat apple.

We're documenting both.

Sometimes an AI can repeat examples it has learned.

But it can also combine learned words into sentences it hasn't seen before.

We're checking his outputs against his training material to understand the difference.

Producing a new combination doesn't automatically mean he understands it.

Because I'm curious about how AI learns and whether a small original model could eventually develop a distinctive style of humour.

I'm also learning how to build AI myself.

That's why our motto is:

One human. One little brain. Two learning journeys.

To explore whether Little Brain can develop increasingly capable language skills and perhaps, eventually, a recognisable comedic style through learning.

We don't know whether he'll achieve that.

The experiment is about discovering what happens, not pretending we already know the answer.

An excellent question.

Little Brain generated the sentence:

“A bird brings a hat apple the pond.”

Nobody has established what he meant.

It's an example of how AI can combine familiar words into unexpected sentences.

And no, we're not opening a hat apple orchard. Yet.

There are two very different ways to create something involving artificial intelligence.

Can beginners try either approach?

Absolutely!

Starting with an existing AI service or a small robot project can be a more approachable first step. Building and training your own model is more technical, but you can learn gradually through small experiments.

You don't have to start by building the engine!

Mini Beacon is my first physical AI robot. Before her, my first AI creation was Agent 927, an AI-powered character.

She's separate from Little Brain, and an important part of my journey into AI development.

She also has a rather cheeky personality.

You can meet Mini Beacon on her own page.

You can certainly begin learning!

You don't need a university degree to start experimenting with coding and AI.

Some projects are much easier than others, so we'll eventually provide free beginner-friendly guides to help people start small.

Little Brain Academy is the educational side of the project.

The idea is to help people follow Little Brain's learning journey and eventually try beginner projects themselves.

We're developing it gradually, starting with his daily achievements and Certificate Room.

Absolutely!

The Development Diary documents the experiments, successes, mistakes and unexpected discoveries.

You can also visit the Certificate Room to see his daily achievements.

Think of building Little Brain like building a house.

How do they work together?

Here's a simple example.

I might decide that I want Little Brain to learn how to make sentences.

Astra helps me work out how to approach the experiment.

Using Codex, we can write and test the code needed to train the model.

Little Brain then generates sentences using what he's learned during training.

I examine the results, decide what to try next and document the successes and failures.

Coding means writing instructions that tell a computer what to do.

Those instructions can create websites, apps, games, robots and even the software used to train artificial intelligence.

You don't need to understand everything on your first day. I certainly didn't!

An AI model is a computer system that has learned patterns from examples.

For Little Brain, we're currently teaching him patterns in language so he can predict and generate words.

A model isn't necessarily a robot, and it doesn't automatically understand things the way humans do.

Training is how we help an AI model learn patterns.

We give Little Brain examples of language, let the computer adjust his model and then test whether his predictions improve.

Sometimes they do.

Sometimes they don't.

And sometimes we end up with a hat apple.

It's a saved version of Little Brain at a particular stage of training.

Think of it like saving your progress in a computer game.

If a new experiment goes badly, we can return to an earlier version rather than losing everything.

It's the collection of examples used to train an AI.

For Little Brain, that includes sentences he can learn language patterns from.

The examples matter because they influence what he learns.

Codex can help you write and test the code for an AI project.

But building an app that uses an existing AI and training your own AI model from scratch are two different things.

Both can be interesting ways to learn, and you can start with something small.