Is it too late to contain the spread of AI? As experts debate the development of truly super intelligent machines, the next generation of machine intelligence is already creating answers to questions we can only guess at. 

Hosts: Matt Armitage & Richard Bradbury 

Produced: Richard Bradbury for BFM89.9

Episode Sources: 

www.descript.com

https://www.youtube.com/watch?v=bHIhgxav9LY&t=585s

https://www.newscientist.com/article/mg25233651-900-2021-in-review-ai-firm-deepmind-solves-human-protein-structures/

https://www.theguardian.com/technology/2022/jan/09/are-we-witnessing-the-dawn-of-post-theory-science

Listen:

The music that made this episode: https://open.spotify.com/playlist/5o85xTNsPrQ2PWH8n2RanP?si=211f46a9cef846e9

Follow us:

www.kulturpop.com

https://www.instagram.com/kulturpop/ 

https://twitter.com/kulturmatt

Cyber attackers.

Aren't a priority until your company is number one on their head list.

Vulnerabilities can be ignored until one makes your business.

The front page news, secure your business digital future with silk on

cyber threat by Telefonica tech, the complete cybersecurity service with

round the clock monitoring and detection.

Get serious about cyber security.

Find out how only at Cellcom business.

This is a download from BFM 89.9.

The business station.

Eight to 9.9, the business station.

My name is rich, Brad Bri.

And, uh, this is Matt explained, is it too late to contain the spread of

AI as experts debate the development of truly superintelligent machines

and the role that they should take in our societies, algorithms are already

controlling and influencing our presence.

The next generation of machines is poised to take scientific discovery

beyond the point where we can do anything, but accept its conclusions.

But all those predictions as diet as they sound AI.

One of the things we come back to most often on this show

was 20, 21, a big year for AI.

Hey,

rich.

It's a difficult question to answer.

Not because there hadn't been huge advances in the field, but because of the

way we think about it, we tend to think of AI as this standalone discipline, which.

But it's also part of so much of our daily lives to one of the biggest smartphone

developments in recent years has been embedding AI into our devices, which

means that for all kinds of services, from photo processing to digital assistance, to

voice transcription, the devices no longer have to send the files to the cloud to

process you don't need a data connection.

It's all done on your own.

But that isn't something you really notice unless you notice the speed

jumping processing it provides.

And that's quite common with a lot of AI applications.

They aren't, things we're aware of.

They happen in the background.

We see the big stories about the AI that is teaching itself to design or program.

Or beating chess masters.

What we don't notice are the chat bots that take our initial

queries and seamlessly pass them onto a human operator.

If our request exceeds its program parameters and the numerous other

incremental advances that blur the lines between machine and human.

Has there been something that's particularly notable for you?

Of course it's long been my dream to create a synthetic version of

myself that can exist in the cloud.

My ultimate aim would be to do these shows with no human input whatsoever.

This year.

I've tried out a few of those AI story and text generators with mixed results.

A lot of the time, the text they generated was too generic and repetitive.

It didn't really save time or effort as to be useful.

They had to be extensively replaced or rewritten in terms of

my voice and personality though, the results have been better.

It helps that my personality is rather mechanical at the best of times.

Some listeners may remember that rich and I did an AI voice test.

Last year, we uploaded our voice prints to a service called a script, and we

recorded a segment of the show using those synthetic versions of our voices.

Their app has made enormous developments over the last 12 months.

In fact, What you are listening to now.

And what you've been listening to for the past few minutes has been

machine generated from texts that the frankly inferior human version

of Matt has forced me to read.

Okay, this is me again, by the way, uh, the mat, rather than the, uh, the map

bot or rather, this is actually me for the first time today, which may be a

disappointment for those of you who were enjoying that robot voice doppelganger.

Uh, so I sent Richard A.

Short clip using this new voice model.

I think a few weeks before Christmas.

Yeah.

Uh, and when you sent the, I, I was in the studio, I'm doing a live show,

uh, and I, I was between talk sets and I, I played it whilst Christine

was in the show at the studio.

Uh, and she was like, are you speaking to Matt right now?

That's how scary it was.

I really, I didn't realize that Christine actually thought it was me.

Um, but no, I mean, that is impressive.

And, and the same with the same with, you know, elements of the, uh,

The voice that you've just heard.

Some of it sounds very natural, very realistic, and other points

you can tell it it's treated.

But on our last episode, I think somewhere back in December, uh, we

talked about blockchain technologies in terms of the way that they've become

this enormous hype machine, you know, Linked to wild financial speculation.

And that has been one of the more dominant tech stories, I think over

the past 12 months, uh, coupled of course, with the, the misdirection,

I guess, of the metaverse both of which as the map bought helpfully

points out, have rather drowned out.

What's been happening in terms of artificial intelligence is

especially as we got to the latter part of the year where we started

to see these stories coming out.

Well, once again, predicting all of the dangers of super smart AI.

So anyone wanting to go into sort of more detail about that and with the speaker a

lot more learned and informed than I am.

Uh, I direct them to the BBC's rife lectures.

I think I've said that right.

Brief or rife, R E I T H, which you can find online there.

Podcast stream, but they're also on the BBC website this year, Stewart

Russell, who is a professor of computer science and the founder of the center

for human compatible, artificial intelligence at the university of

California, Barkley academics are terrible at coming up with names for that.

Yeah, right?

Yeah.

He gives four lectures around the theme of living with AI.

He talks about AI and the fears that we have, uh, he talks about it and the.

Influencing economic activity, but he also talks about its increasing

military and policing roles.

And I think the final one is about how we can actually find ways

to coexist with these systems.

Uh, and, and Ukraine, we, we, we still have nothing.

I don't know if it's so much that we have nothing to fear.

It's just about how we shape those decisions.

You know what AI does, what we allow it to become.

So it's also a bit like those metaverse conversations that we had last year

at the moment, there is no metaverse, uh, we can have those discussions.

We can try and define it, but companies like Facebook, sorry.

Metta are jumping into that conversation.

Now they're telling us.

What they want it to be.

So we need to have those conversations around, I guess,

big AI for want of a better term.

What we aren't having, uh, are those conversations about the, uh, the

algorithms and how those dumb AI systems are sweeping through our

world and the uses that companies and governments are actually putting them to.

And as I just demonstrated with.

My voice clips, it's perfectly possible to create a reasonably convincing deep, fake

voice and direct it to say what you want, because all I did was type those words in,

but are there enough safeguards in terms of what those systems can actually do?

Um, are you referring to the, uh, Alexa challenge story?

Well, yeah, that one's specifically.

So in case you haven't heard of this story, it was a report from

the BBC in late December that a 10 year old child had been at home.

She'd been doing some physical education challenges on YouTube because you know,

it's Britain, the weather was bad, so she couldn't go and play outside.

When she asked her Amazon echo speaker to give her another

challenge, it told her to try pulling.

An electric socket, the plugs slightly out of the wall and

touch a coin to one of the prongs.

Now in case.

I need to say it.

That is a potentially lethal activity.

Nobody should ever try that.

It can give you a massive electric shock.

And it was a trend that surfaced briefly on social media and was

quickly removed about a year ago.

But somehow the girl's Alexa assistant reference those reports about the

challenge and her echo speaker.

Happily told her to try it.

Thankfully, her mother was with her at the time.

And her mother also added in the report that her daughter is too

sensible to do anything that daft.

And of course, Amazon quickly patched the issue once they were made aware.

But that's kind of the problem it's that the users have to tell the owner

of the technology in this case, Amazon, that something has gone wrong with the

information that the AI is actually.

Isn't that normal though.

I mean, we buy something, it doesn't work.

And we tell the retailer or manufacturer that there's something wrong with it.

Um, can't we put that down to something like a faulted device or a system error.

Well, I think that's really the crux here because the AI didn't malfunction,

there was nothing faulty about it.

What the algorithm lacked was the ability to contextualize it.

You know, it was asked and it was answered.

It found a challenge, it delivered it.

So as far as the Alexa was concerned task done.

Right.

So as I said, the AI didn't malfunction.

It wasn't faulty.

It was that ability to contextualize.

But what was lucky was that this happened to a relatively sensible young girl

and there was the double protection that her mother was actually present.

What

we're seeing is that AI is still very much an

experimental and developmental.

Well, yes, to the machine.

The experience with the girl is an exercise, a driving assistance software.

For example, that causes an accident, uh, causes harm, or even death to

the cars, occupants or road users will learn from that experience so

that it hopefully won't be repeated.

You know, there are all sorts of ways to train artificial intelligence.

Adversarial networks where two or more versions of the

software challenge, each other.

That's one, uh, neuron networks where no implicit instructions are given.

That's another very often.

And in the real world, at least AI learns from its experiences

and interactions with us.

And it's interesting because one of the backdrops to the last 12

months has been the development of new vaccines for COVID, which many

people are unwilling to take before.

Of the fact that they are new and to their minds, new equates to

experimental and experimental.

In their minds means not safe yet.

We embrace these truly experimental machine tools, these artificial

intelligence and put them to work in our everyday lives.

But what safety tests have they been put through what government bodies,

what organizations have approved these machines as being fit for purpose?

So, um, is there a frame of reference off kilter then when it

comes to that?

Well, I think that's where the vaccine analogy is actually quite useful.

If you flip the narrative a little, it would be easy to frame artificial

intelligence in terms of a global pandemic in the Terminator movies.

We believe that Skynet is a threat because it's self aware, but really

that's not the biggest problem.

Skynet is a threat because it's everywhere.

It's.

Every system, it is an infection.

And that's the situation that we approach today.

These little bits of code are everywhere.

Uh, having said that I don't believe AI is a pandemic.

I'm just trying to make that point.

You know, these are complex machines and the majority of

people don't understand that.

In any real or kind of coherent way.

And because of that lack of understanding, we often miss the point that though they

are assisting us, they're incapable of understanding us machine intelligence

can be guided by human responses.

You know, it can be guided by how many people react in a certain way to a

certain set of actions or circumstances.

Something we'll talk about more after the break, but it doesn't

ascribe an immoral weight to it.

You know, it looks at murder as being bad because one of its parameters

says it is, or because it finds it's relatively uncommon in terms of frequency.

Yet these are still the tools that we allow to sit.

Unsupervised in a smart speaker in a kid's bedroom, but these machines

are choosing what news we see.

They guide our purchasing decisions.

They help to manage investment portfolios.

Yet we haven't the slightest idea why these very stupid and limited pieces

of code actually do what they do.

After the break artificial intelligence and the end of all things you

tune into mansplained here on BFM 89.9.

The business station.

Stay tuned to tech top brought to you by Cellcom the stairs.

Be financially minded BFM 89.9.

Tech-Talk brought to you by Cellcom business.

Uh,

BFM 89.9.

The business station.

My name is rich bribery.

Welcome back.

Mansplained if 2021 was a year that we started to reconsider our relationship

with AI could 2020 to be the year that AI makes knowledge irrelevant.

Well, you know, one of the best things about this digital world is the

amount of knowledge that we all have.

You know, with a few clicks, we can find out pretty much.

Anything, you know, you can find out how to drill a hole in

a world without electrocuting yourself or flooding your home.

That probably tells you a lot about me.

Uh, you can find out why cold fusion just won't work and you can research how

aerosol particles disperse by watching some Australians drop paint field exercise

balls off a giant tower onto a giant ax.

But one of the drawbacks is that we have developed this

tendency to treat information.

A bit like a pub quiz.

We can reel off these impressive lists of facts without necessarily

having developed a corresponding sense of understanding to accompany them.

These shows certainly wouldn't be possible without the internet.

Well, as it's the first show of the year, um, Take that as any kind of

insult and it's true and not just for me, but I imagine BFM would need an enormous

staff, producers and researchers, and fact checkers without the internet.

You know, it's hard to think back to that old world of just a couple of

decades ago, where you had to go to libraries, you have to order books,

you had to painstakingly troll through.

Physical copies of newspapers and old publications, just to find something

that might make, you know, a sentence or two or a quick aside, w which isn't

to say that you can't find in-depth information

online?

No, of course not.

You know, the ability to publish all kinds of research online has enabled us

to create this vast commons of knowledge.

And sure, not all of it is free.

You might find papers and journals behind paywalls, but there is usually a synopsis.

And very often you'll find that there's someone somewhere who's done an

explainer video or written an article.

I mean, I don't know how many of our listeners follow the YouTube

channels, like, uh, very Tassian where you'll find really complex physics

and maths and computational science.

So.

AI explained, but done in really engaging ways, including using rented helicopters.

You know, one of the great mysteries of the modern age is how electricity works by

which, I mean, how does the power that's generated, you know, dozens of kilometers

away, turn on lights and make devices.

In our homes and I'm not shown to stay, uh, that when I watched a very Tassian

video about electricity, there was lots of stuff in there that I didn't

know, particularly when it comes to the idea of why as being stuffed with

electricity, they aren't, it's about the magnetic, the electromagnetic fields.

They create.

So we have this weird combination of being able to learn the answers

without knowing the, working this out or having that deeper understanding

of what the answer actually means.

And that takes me to my favorite Hitchhiker's guide analogy,

the supercomputer deep thought.

Which determines that the answer to life, the universe and

everything is 42, of course, which is a completely useless answer.

So it designs another far more complex computer to determine

what the question actually is.

Okay.

Kind of the stage that we're getting to with AI right now.

Yeah.

We're using it to compute more and more complex ideas to run simulations

and experiments at speeds, that human researchers simply can't.

And we saw that with the development of the COVID vaccines

with AI being used to do.

Potential candidates, which shortcut a lot of the trial and error work that using

traditional methods could have led us to spend weeks or months or even years doing.

And we're also seeing it in one of 20, 20 ones, most remarkable AI breakthroughs.

And again, you know, Lincoln.

You might have missed it.

A deep minds, alpha fold network, which mapped all the proteins in the human body.

Something human researchers have spent decades doing, and they managed to

identify less than 20% of those profiles.

Those proteins, uh, alpha fold, I think brought it up to about 98.5%.

So the hope is that, um, machines that are four-fold will enable us to develop

personalized medicines, treatments that bind to specific parts of molecules

to, to quote the new scientist.

But we largely in the dark as to why machines like alpha

fault, make the predictions.

They do a to quote a guardian article by.

Laura Spinney called, are we witnessing the Dawn of posts, theory science.

You can't lift a curtain and peer into the mechanism because there

isn't a set of workings out that you can extract and look at.

Is this

partly a data?

Of course, I mean, otherwise, you know, why talk about it on a tech show,

but, um, you know, machine learning can, uh, both generate an output, you

know, vast tranches of data and that changes theories or at least the shape

of the theories at a conceptual level.

Uh, Laura Spinney quotes, Tom Griffiths, a psychologist at Princeton

university who says, you know, these theories, uh, that make sense.

When you have huge amounts of data, look very different from the ones that make

sense when you only have small amounts.

So you're back to that deep thought problem.

The issue isn't the answer.

The machines can give us the answer in ways that are easy for

us to digest, but we no longer truly understand the question.

I

mean, it's always a dangerous question to ask though.

Does it matter?

Well on one level, no, you know, perhaps we're getting to

that stage in our development as a species where we don't need to know

the questions behind the answers.

Uh, science fiction is full of stories about machine dependent species that

suddenly adrift when that machinery fails and they realize, you know, they don't

really have any basic life skills anymore.

You know, they don't know how the lights work.

They don't know.

I mean, You know, they simply can't do anything, but at the same time,

you can argue that as a species, we can only progress so far.

If we insist on controlling all of the processes by which we get information,

you know, we, we can't really talk about the superintelligence of the future

at this point, at least in terms of.

Abilities, but we can talk about machines.

We have now, especially in terms of the data that we feed it.

And that data is often a weakness.

We've talked so often on the show about how data sets and AI can lead

to unusual and unpleasant outcomes.

And there was that example of course, of the coin and the plug

in the first half of the show,

those larger issues of distortions as a result of, uh, bias

datasets.

Certainly, you know, we've seen racial profiling and bias being baked into data.

AI concluding that it's better to hire men in senior positions because most

senior positions are occupied by men.

Uh, in medicine, this is particularly important because women are non-European.

Under represented in a lot of medical tests.

And there's the classic case of medications being developed for

women being tested entirely on men, because it was thought that

menstruation would affect the testing.

Are we seeing.

Same course to limit how involved AI becomes in the

process of, uh, discovering the,

I mean, I don't think so.

As we've noted before, uh, AI, as it currently stands is usually an assistive

tool rather than a truly independent one.

And what's exciting in some researchers is the ability of the machines.

Demonstrate when the theories that we have are too simple.

When we talk about machines, looking at things in different ways, you

know, it's often said with the implicit assumption, that our way,

the human way is the right way.

And there are often limitations in terms of the way that we formulate ideas and

theories, uh, quoting Tom Griffiths again, uh, his team did a piece of research

on risky behavior, taking the new.

The Princeton team used predicted the outcomes that people would

choose in various scenarios with quite a lot of accuracy.

But it also helped highlight where the initial theory broke down or where

it was insufficient, which was the point where we become overwhelmed by

possibilities and switch to simpler or competing strategies to make decisions.

So these larger data sets allow the machine to come up with much more

complex strings of possibility.

You know, we tend to look at things in terms of if this, then that.

But we can only hold those in our heads up to a certain point.

The machine can calculate these strings almost indefinitely and

produce likely outcomes for situations that we can't even conceive.

Is that enough for as though.

Do you think humans can, can live in a world of correlations

rather than knowledge?

Well, that certainly is the bigger question.

You know, we can improve an enlarge datasets to overcome

some of those implicit biases.

We may be able to accept that.

Certainly in terms of medicines, it's more important to have the.

But to know how we reach them, but whether it's enough for us just to have

the answers, that's difficult to answer, you know, that's a much more metaphysical

issue, human history, human development.

It's about that search for information.

It's part of that obsession that people like Elon Musk have with pursuing the

question of whether humankind can live and settle and thrive on other planets.

You know, and turning that into a reality.

It's why we have all those concepts in mathematics that most of us

thought were useless and we'd never use in our daily lives.

So perhaps we can develop machines.

Better at informing us how they arrived at those answers or machines that are

better at framing them in ways that are familiar to us and more comfortable.

But there's also the argument that humans are needed for that spark of

creativity and for their intuition.

But machines like DeepMind have started to make their own intuitive

leaps machines are also creating.

Writing stories, poems and articles, making videos.

They're not always coherent, but it is going to be interesting to see

the DeepMind equivalent of the apple falling on Newton's head and finding

out where that process takes us.

Um, Laura spinny.

Picasso, who said, I think in the 1960s, computers are useless.

They can only give you answers.

And while there is a truth to that, that I think people like Douglas

Adams and of course deep thought would appreciate we can apply the ingenuity

and creativity that characterize humans to those machine derived answers.

Fantastic.

Thanks for that, Matt.

Thanks, Richard.

Uh, it's been a while since I've been on actually.

Yeah, it

has now.

Um, you can find me on Instagram and Twitter at culture mat.

You can also head over to culture, pop.com for transcripts of these

shows and information about culture pop and it's consulting services.

If you missed any part of this show, don't forget.

You can download the podcast wherever you normally download it from.

We recommend the BFM app it's available from the apple app store.

I'm rich, Brad Bri for BFM 89.9.

Tank top was brought to you by SOCOM business.

Get serious about cybersecurity and secure your businesses.

Digital future at dot dot com.my thank you for listening to this podcast.

To find more great interviews, go to BFM dot mine or find us on iTunes BFM 89.9.