Tech and AI: 6. What is AI?

This BBC podcast is supported by ads outside the UK.

Suffs!

The new musical has made Tony award-winning history on Broadway.

We demand to be home!

Winner, best score!

We demand to be seen!

Winner, best book!

We demand to be quality!

It's a theatrical masterpiece that's thrilling, inspiring, dazzlingly entertaining, and unquestionably the most emotionally stirring musical this season.

Suffs!

Playing the Orpheum Theater, October 22nd through November 9th.

Tickets at BroadwaySF.com.

BBC Sounds, music, radio, podcasts.

Welcome to Understand Tech and AI, the podcast that takes you back to basics to explain, explore, unpick, and demystify the technology that's becoming part of our everyday lives.

I'm Spencer Kelly from BBC Click and you can find all of these episodes on BBC Sounds.

Something important has happened in just the last year.

The world suddenly seems to be talking about artificial intelligence.

AI.

Now, I think that's mainly due to the arrival of something called ChatGPT, which is an AI that you can talk to, that you can have a conversation with, and my goodness, does it sound human?

Suddenly, we're all talking about whether it's smart enough to take our jobs or even take over the world.

Well, in these next five episodes, we're going to talk about all of that stuff.

And the first thing to note is that AI has actually been around for longer than you think.

Many, many years ago, I did a degree in computer science and I studied artificial intelligence.

And that was way back in 19...

So it's not new, it's just that up until recently, it's kept its head down, quietly changing our world behind the scenes.

But now the chat is out of the bag.

Now, I've got someone here to help us understand what AI is and what AI isn't.

It's Dr.

Michael Pound, Associate Professor in Computer Vision at the University of Nottingham.

Mike, hi.

Hello.

Can I just start with a frustration of mine?

Whenever I see a news story about AI, you usually get either a picture of the Terminator next to it, or you know that white humanoid robot that's reaching out westfully to touch something, or maybe it's reading a newspaper.

I find that really frustrating because it really confuses the issue.

It suggests that AI is either killer robots or AI will only do human-type things.

And that's just not the case, is it?

Not at all.

And in fact, that's a frustration of mine as well.

I think that we're a long way off Terminator.

I'm not particularly worried about this.

AI has been around for decades.

Can you give us some examples of where AI has been working?

Yeah, so analysing images is a very common use for AI.

So that's what actually I do, so computer vision.

We've had cameras, CCTV, other footage for many, many years.

Since that time, we've basically been trying to train computers to do a better job or a quicker job than us on finding out what's going on in images.

So let's try and explain it in a nutshell then.

Artificial intelligence, how do you train a computer to recognize cats?

What you would do is you'd use something called machine learning.

So AI actually is a sort of catch-all term that encompasses a lot of different fields, but machine learning is perhaps the one we're most accustomed to.

And what it really comes down to is instead of directly programming a machine to do something by writing software, what you do instead is you give it lots and lots of examples of cats or things that are not cats.

And over time, it builds up an understanding of what those two kinds of things look like.

And what I find really interesting is we don't know what it is doing, how it is doing that.

All it is doing is forming connections, a bit like in the brain, isn't it?

It's forming connections that over time get better at recognizing cats.

But if we were to open this imaginary black box and look at the kind of wiring that is inside, this is all metaphorical, of course, we would have no idea why something is connected to something else and why this bit is not connected to that bit.

Yeah, part of understanding these networks is something that bothers people.

You know, if you're going to use these networks in medical imaging, we'd like to be able to understand what it is they're doing because doctors can explain themselves and that's quite important for them and for the patients.

Understanding machine learning is important, but actually it is very, very difficult with these big networks.

So to summarise at this point then, artificial intelligence programs itself to do the job that you're training it for in a way that a human computer programmer wouldn't or couldn't.

That's right.

So if we're defining artificial intelligence as broadly machine learning, I don't want to program up every single possible rule to solve some task.

What I'd like to do is present this machine with some data and have it program its own rules.

And I'm also passionate about helping people to understand the limitations of AI.

So it's also true, isn't it, that once you've trained an AI to do a job, like identify cats or not cats, it can only do that job.

It can't then also say, that's a dog.

No, it can't.

It might be ever so slightly better at that than something that's not trained at all, but basically it won't do anything you don't ask it to do explicitly.

I think it's also important to understand that when you're feeding these pictures of cats or whatever into the AI, they've got to be labeled as cats.

And that is something that humans do before it gets chucked into this empty AI brain, yeah?

That's right.

We're going to take tens of thousands of images of cats, right?

That's a lot of data to annotate and go cat, not cat, cat, not cat.

And then we've done that, we've spent ages doing it, and then we just show these examples over and over again until it eventually learns.

Who is labeling those pictures for the AI?

Well, sometimes it's a poor scientist like me

who wants to get a result and publish a paper, and so they'll just sit there and plow through their own images just to get the job done.

There are lots of companies actually that do this.

You can go to pay per image or per data set and say, I need you to go through these 10,000 images and label all the cats for me and I'll pay you this much money.

Are we also unknowingly training AI?

I am thinking of the capture images that you get on certain websites when they ask you to prove you're not a robot by clicking on all the pictures of traffic lights.

Is that us training AI?

Capture is a very good example.

It's a clever way of verifying that someone's not a bot, but also getting just a little bit of annotation done.

And you can imagine if you multiply that by millions of people over a period of time, you actually now have really impressive data sets that you can train on.

And when it is pictures of roads or cars or traffic lights that we're being asked to identify, I wonder what type of artificial intelligence we might be training.

I can only hypothesize, but possibly something to do with roads and driving.

Mike, this is really fascinating stuff.

And I want to come back to you in a minute or two, if I may.

But first, let's take a slightly longer view and get a bit of the history of AI from Dr.

James Sumner, our resident tech historian from the University of Manchester.

It was in 1956 that American researchers settled on the name artificial intelligence for what was emerging as a new area within the still very young field of computer science.

The projects they were working on mostly aimed to simulate specific kinds of tasks that were normally done by human thought.

For instance, translating from one language to another or recognizing objects from the way they look.

Objects are put under the camera and an image is fed into the computer system.

Light reflected from the object is turned into an electrical pattern on the surface of the chip.

Their priorities were driven very much by funding.

AI software was complex and needed the fastest and most expensive computers, beyond the budget of almost any funding agency, but not the United States Department of Defense.

In Washington, President Truman went to Congress.

From now on, he announced the United States would contain the advance of communism anywhere on the globe.

This, at last, was the official declaration of the Cold War.

The Cold War presented problems it seemed AI could tackle.

Piles of documents in Russian waiting to be translated into English, or how to instantly and reliably tell a tank from a tractor.

Early hopes for these projects were very high, but progress faltered and funding was scaled back drastically in the 1970s.

The most popular approach, which focused on feeding the system detailed knowledge of the world and a set of logical rules to deal with it, what's sometimes now called good old-fashioned AI, turned out to be something of a dead end.

AI survived by reinventing itself, pursuing other approaches that could learn how the world works by interacting with it, sowing the seeds for today's deep learning learning approaches, which you'll hear more about in a moment.

Dr.

James Sumner, thank you.

Now, Mike, James just mentioned deep learning.

How does that differ from machine learning?

The rise of AI in the media and in the news in the last decade has been deep learning specifically.

So what we're doing is this extraction of knowledge from an image, for example, finding the edges, finding the corners, combining them into features like eyes and mouths, and then combining those into features like faces and animals.

It sounds more like a human way of learning because we learn to identify a cat because it's normally got two ears, two eyes and we already know what an ear and an eye is.

Is that the sort of thing that you're talking about?

Yeah, we take simple features and we combine them into an idea of something more complex and actually that is what happens in deep learning just in a slightly different way.

So we've talked a lot about cats.

We need to talk about the elephant in the room now, which is chat GPT.

More generally something called generative AI.

This is AI that can generate new content, pictures, music, videos, and importantly conversations.

It sounds very human and I've had some really fantastic conversations with this thing.

Why is it so convincing?

What it's actually doing is in essence a predictive text.

Like the ones you get on your phone.

Just like the ones you get on your phone, just much, much bigger.

And so we have a situation where you've given it some prompts.

So you've said, I'd like you to write a story involving these characters.

Those are now written down.

and something it can read.

And so it reads that sentence and then this passes through all the layers of its network and it starts to predict what the next likely word of that sentence would be and so maybe you start a book with the word the I'm not good at story writing they all start with the for me so it says okay there's a 90% chance that the next word is the and then it says okay there's a 80% chance that the word after that is story and so on and so forth and it just predicts one word at a time I suppose what's most spectacular about this is the fact that just by predicting one word at a time, you can still get all this incredible text out.

Because you'd think if you were just writing down one word at a time, it would soon degenerate into nonsense.

But it doesn't seem to.

It seems to be able to go for quite a long time producing really, really good text.

It's really interesting thinking about something like ChatGPT just as predictive text, because that, I think, helps to explain why sometimes it just comes up with complete rubbish.

It's not looking for facts.

It's looking for the statistically most likely word that comes next.

Yeah.

And often the statistically most likely word is the one that we want to see.

Over time, this ChatGPT and similar networks have been trained to give us what we want to read.

And that doesn't necessarily have to be true.

It's basically like a really well-versed sycophant.

Exactly right.

If you ask it to write a Wikipedia page for me, it's very complimentary and most of it's untrue.

I have an example here that I think demonstrates that ChatGPT doesn't actually understand what it's saying.

I asked who holds the world record for crossing the English Channel on foot.

And it said the world record for crossing the English Channel on foot is held by Christoph von Druch from Germany, who completed the crossing in eight hours and 34 minutes on September the 15th, 2018.

Amazing.

Now, there was a guy of that name who swam the English Channel, but not on that day and not in that time.

And of course, he didn't walk it.

There's no concept that a body of water cannot be traversed, for the most part, on foot.

No, that's right.

But it also has introduced elements of truth.

The person does exist.

They are associated with the English Channel.

but not in the right way.

And so it's kind of this half-truth that makes it even more difficult to spot sometimes.

Would you say then that artificial intelligence is not really intelligent?

It doesn't actually understand what it's doing.

What people are really wondering is, is it like a person?

And the answer is definitely no.

So I think there is intelligence there, but there's also intelligence in ants.

We think ants are amazing, but we don't necessarily think they're quite the same as us.

Famously, last year, there was a Google software engineer who had a conversation with Google's chatbot, Lambda, and the chatbot started to talk about feelings of loneliness.

It might even have said, don't switch me off.

And apparently it convinced the software engineer that it was not only intelligent, but conscious.

It had feelings.

It was a real kind of human-like mind inside there.

Are we just throwing that concept in the bin right now, aren't we?

Politely, yes.

I think

the issue there is that if you ask an artificial intelligence like ChatGPT, if it's sentient, it's going to give you an answer based on predictive text.

So basically an answer it thinks you want to read.

And the answer may well be yes, because that's an interesting thing for us to talk about and we can delve into that topic more.

And so, I think that it's still just predictive text, nothing has changed.

It's just that prediction has output the word yes this time instead of no.

And so, I think that's a better way to look at it.

Mike, thank you so much for that.

This is so fascinating.

Would you mind coming back for the next episode?

Because I would like you to help us understand where AI is working already in more detail.

Absolutely, it would be my pleasure.

So, what have we learned so far?

Well, AI is an excellent imitator, but it's not conscious, it doesn't have feelings, and it often can't tell the difference between fact and fiction.

But if you train it to do specific tasks, which it can practice millions and millions of times, it can get better than any human.

I like to think that AI is a really slow learner.

It just does it really quickly.

Bye for now.

Dreaming of buying your first car or new home?

Knowing your FICO score is the first step to making it real.

With My FICO, you can check your score for free and it won't hurt your credit.

You'll get your FICO score, full credit reports, and real-time alerts all in one simple app.

Your credit score is more than just numbers, it's the key to building the future you've been working toward.

Visit myfico.com/slash free or download the MyFICO app and take the mystery out of your FICO score.

Com.