And the next thing we did, just as I said, different environments necessitate different

kind of agents.

We looked at five, no, eight kinds of agents, kind of a non-learning agents, and then you

can add learning to all of those.

This is just a broad way of classifying agents.

Of course, the agent designs are much finer.

There's just much finer taxonomy, but this already gives me kind of a way of talking

about agents and what they might need and so on.

We're going to progress over this year down this line.

Good.

Remember, you can ask questions.

Well, almost.

We don't have the cube for some reason.

That's fine.

When you have a question, I'll try to repeat and remind me

of it, please.

So the first kind of agent we looked at

is a very, very simple thing.

It has a sensor which tells it how the world is now.

It does not have any memory.

Everything it sees is its world.

It reacts to that given some kind of condition.

Action rules.

And then the next preset rolls in, it reacts to that.

Very simple.

Bacteria level.

Simple agent can do simple stuff.

And for some things, you don't need more.

For a thermostat, like the one on that heating system there,

it actually measures the temperature in the environment.

If it's too high, it closes down.

If it's too low, it opens up.

And it even does that in a continuous world.

But let's say we have a thermostat that looks once

in a second, then that would be something

that would be exactly this.

Boy, how nice.

Very simple.

It does what it's supposed to do.

It doesn't do so optimally, surprisingly enough.

And that's a current research topic.

Namely, if the thermostat were a little bit more clever,

it would notice that between, say, 1205 and 1215,

the doors are open.

And so it should, before that, get lots of hot water

so that it has a lot of power.

And then between, say, 1215 and 1 o'clock,

we're adding heating capacity of, I don't know,

you're maybe 30 right now times 200 watts.

That's quite a lot of heating power.