The engaging dialogue with Robin Osborne reveals the profound ways artificial intelligence (AI) is reshaping human relationships and interactions.

Osborne, a veteran in the tech field, shares his early passion for computers and programming, leading to his influential roles at companies like ASOS and Just Eat.

He recounts how the rise of machine learning has revolutionised customer service, particularly through the development of chatbots that manage repetitive inquiries, allowing human agents to focus on more complex issues.

This narrative sets the stage for a broader exploration of AI's impact on society, emphasising the need for technology to enhance human connectivity rather than diminish it.

Takeaways:

• The impact of AI on enhancing human connectivity is a crucial discussion today.
• AI technologies can efficiently handle mundane tasks, freeing humans for more complex interactions.
• Generative AI can inspire creativity by offering ideas beyond the usual constraints.
• Ethical concerns arise when AI imitates personalities without explicit consent from the individuals.
• AI's role in customer service can significantly improve user satisfaction by deflecting repetitive inquiries.
• The future of AI should focus on improving quality of life rather than just cost-cutting.

Links referenced in this episode:

• robinosborne.co.uk
• hachyderm.io/@rposbo

Companies mentioned in this episode:

• asos
• Microsoft
• Amazon
• Google
• Just Eat

Hi, how you doing?

Good.

Thank you for coming on.

This is the next show on relationships with AI.

Really appreciate you coming down and this is an opportunity just to hear about AI and relationships and how that impacts the fabric of society.

Also just enhancing human connectivity as we talked about before.

I wanted to just have the opportunity to or really thankful for the opportunity of having you come down here and also just to sort of hear about your take on this particular subject and particularly from the angle of relationships and human connectivity.

So firstly, because many people might not know who you are.

Would love to.

Robin, this is.

Your name is Robin Osborne and could you tell us a little bit about yourself and your background in this particular space?

So hello, I'm Robin Osborne.

I met you when I was studying computer science way back when, a while ago.

Talk about that.

Little while ago.

Always been into kind of computers, programming a spectrum for when I was like 6 years old and through that kind of that thread just continued computer science and then through big e commerce companies and then AI started kind of coming into the this sector.

I say about maybe nine, ten years ago it really started to pick up where you had these big players, the big companies, the Microsofts, Amazons and Googles.

And then I got really into that, seeing what was possible, seeing what was available and kind of sharing that with the community, with the tech community.

And it's just gone on leaps and bounds from there.

It's just quite incredible.

And I'm just blown away every time we see the next thing that's there.

Like, I don't know, there were things I discovered maybe yesterday that I didn't know about.

And I'm just incredibly enthusiastic and will share with you and say this is amazing or this is scary or both.

So could you just give a bit of a flavor about sort of the different companies that you've worked with and sort of the different types of things that you've done and what your area of expertise is now.

So yeah, let's go back.

So in the earliest days, asos, I would say that third engineer at asos but didn't buy shares, which is why I still need a job.

The other two seem to be doing quite well and even quite early on there was a concept of recommendations engines which would be based on I guess what you had bought and what we think you might like next.

So we're going to predict that.

And that was more kind of machine learning and kind of data analytics, I guess.

And the first time that it really blew me away was when there was the Microsoft tooling their Kind of machine learning algorithms.

Someone did a demonstration where they would take historic orders that had five items in them and it would train them on.

When you buy these four items, the fifth item is X whatever it is.

So it could be.

It was trained on orders with five items and then it was shown orders with four items and said predict what the next item should be based on all of this order history.

And it would predict that very accurately.

Based on orders that actually were five items.

It was told that there were four.

What should they get next?

And it would predict it.

And just the idea that, well, wait a minute, now we just show them the sixth item and the seventh item and the eighth item.

These are items that people will actually want because we've trained it on this person's purchase behavior.

So you've got that kind of recommendations, algorithms that previously were not that intelligent, not that complex.

And then they're based on like the machine learning.

And that was like Asus.

That was a good maybe 15 years ago.

And then.

And that was within Microsoft then moved on to just eat.

So food tech, not E commerce.

But you end up having lots of customer interaction problems with an order and the idea that it costs the company a significant amount of money.

Like a not a not insignificant amount of money.

I guess every time a customer gets handed over to a support agent there is a cost to do that.

And it tends to be quite mundane things that's kind of a waste of their time.

They should be used for their.

The more kind of human side of things, the more kind of creative elements, the more it's more than just a late order.

It's.

You've got a question.

There is something that needs a human.

So we created the Just Eat chatbot which was the first time using Microsoft's language understanding and intelligence service or Luis.

And that was way before like ChatGPT was around.

So this is the early, early days where you could train a brain with a whole load of text.

Where you said this text has this intent and here are some identifying things within this piece of text.

So we trained it on all of these live chat transcriptions.

And you could say right out of these thousands tens of thousands of live chat transcripts from customer support agents when the first sentence is this, the intent was late order.

The thing that looks like XYZ is an order number so you can identify those when you have a conversation with a chatbot.

So then that meant that any conversation with just eat around my order is late was deflected from humans handled by brain computer brain pre chat GPT.

So there's the intelligence there was pretty limited and saving money and it was a case of.

That was the majority of the kind of the questions.

It was like my order is late.

My order was missing something.

Being able to deflect those.

And that was Microsoft's tool.

But I could geek out about that so much because they had all the stuff around kind of translation, text summarization, image recognition, voice recognition, like speaker identification, like me kind of talking now it could kind of tell Robin is talking, Yabo is talking.

Just by us giving like 5 seconds of audio or something like that was just incredible.

The capabilities there and being able to do like live translation to another language with speeches, speech bubbles or something with.

Sorry, with subtitles, which I just thought was incredible.

And that was Microsoft then.

Oh man, I'm going to go off on tangents now.

There ended up creating a.

So my wife used to work as well for Asus and she came up with a hackathon idea where you could take a photo of someone's clothes and it would figure out very quickly what they're wearing accurately and give recommendations and say you can buy that thing over here so you could be on the street and go oh, I like that outfit.

Click and they go oh, you can buy it over here.

And then we ended up building that using all these other tech that was so accurate.

It wasn't just a case of lady sitting chair, it was like sparkly tops, zip and we can buy that over here.

Getting that kind of affiliate purchase thing.

So that was all that image recognition and video recognition and video summarization.

So much, so much stuff, such a wealth knowledge.

It's just amazing and mind boggling to sort of hear you talk about those things.

It's interesting.

I love wanting to go back to the point about sort of how your experience with Just E and using Lewis, just that whole process of early stage AI and how that impacts in a particular way the business.

So human connectivity.

You mentioned about sort of the importance of how do you shave off all of the mundane repetitive tasks and leave the human to interact in the more qualitative decision makings that would help enhance the business.

That's a good way of putting it.

The qualitative decision making.

Yeah, yeah.

Because the.

So being able to.

The interaction for the user was from the Justy app and it was text.

So pretty basic.

Not even kind of talking to a AI.

It was typing that message would hit the.

The chatbot underneath.

It was called Bot Framework.

It would kind of look at what they've just Written look it up in the Language Understanding service to go, do you know what this means?

Is there an intent that you can tell me?

And if it says it's a late order, I've got a 90% confidence that's to do with late order.

And it looks like they've given me this order number which matches the one that you know about because they're logged into the app.

So with a high degree of confidence you can say, are you asking about this order?

And also while you're doing that, I can go and check in the background the status of that order.

Because Just Eat will get updates whether it's still being cooked, whether it's been picked up, whether it's in transit, how far away it is.

So it can immediately give an update and go, are you asking about this order?

That's 10 minutes away.

And at that point it's like, oh yeah, do you still need to talk to someone?

No, that's okay.

Done, moved.

And then if that didn't answer their query, the layer underneath that, this was incredible.

At the time it was called the Q and A maker.

The idea that you could take a list of kind of frequently asked questions, just your standard kind of FAQs, your returns policy, your whatever, your refund policy, and feed it to the a more simple version of the Language Understanding service.

And it would create very easily create its own little chatbot based on your FAQs.

The techies didn't even build this.

This was built by a business analyst in Just E who just took their FAQs, gave it to the service and then you didn't even have to ask the questions accurately based on kind of the frequent, the FAQ kind of headings of what is your refund policy?

It was like, my food is terrible.

Can I get my money back?

And it will go, okay, that's not a late order question.

That's not a.

Something's missing from my order according to the Language Understanding service.

I don't know what that means.

Go down to the next layer.

The next layer is this FAQ kind of catch all.

And it's like, oh, that looks like it might be a refund policy.

Okay, here's our refund policy.

Does that answer your question?

If not, okay, I can't help.

We need some human to help out here because this is beyond our basic understanding.

But it was all kind of based on facts, data, real things, not kind of hallucinated paragraphs of response.

This is all quite factual and helpful and got very quick.

And it meant that when you have A company with a limited number of customer support agents or a limited number of software licenses and you can't answer all these questions that are coming in at peak time, particularly when Just eat Sponsored the X Factor so it's Sat.

Friday, Saturdays, audience numbers kind of went up and users of Justin went up as the ads went out.

And then customer support agents, it's the same number.

So when a whole load of requests came in, they'd be waiting and their satisfaction level just goes down and down and it impacts the company.

So being able to just deflect say 80% of those because people are asking where their order is, leaves the customer support agents to deal with the more relevant customers complaints.

Amazing.

What things have you seen and how have you.

Have you used AI?

I mean it's moved on just so, so much from this basic.

Here is a spreadsheet.

Give me some fuzzy matching around the words in a spreadsheet to what is used within engineering.

So not civil engineering, but computer programming.

You've got the concept of the kind of the coding copilot and it's literally called that.

It's the GitHub copilot that is embedded in programming environments.

So what would normally happen?

Most coders don't know everything.

Surprise.

I'm sorry.

And a lot of the time we'll just go to Google and then Stack Overflow and then these other various sites where someone else has had the problem and you're now looking it up, getting it copy pasting shouldn't copy paste, but you know, and then putting it in solves the problem.

Now all of that knowledge is kind of embedded within your development environment and you can just ask questions, go, how do I do this?

What happens there?

Not even asking questions.

But as you're typing the autocomplete kind of intellisense is there to guess what you're probably trying to do based on this vast amount of information that exists in vast numbers of code repositories that has been absorbed into a large language model and then pushed out into Microsoft's GitHub Go pilot because they now own GitHub.

GitHub is like the biggest code repository.

And the idea that you can start typing, give it some, just some comments on a page, write in English, I would like this.

This is what I'm trying to do.

Hit one button and it just kind of goes and spits out kind of boilerplate.

But possibly working code.

You're not quite sure where it came from, what it's based on, but.

And I guess the problem there is if you're say a junior engineer.

You might not know the quality of what you've just been given, like looks right.

I might just take that and go with it.

And then this code that you might not fully understand has made it into a production environment and now you've got to support it and you've got to maintain it.

But maybe you didn't quite fully understand it because you didn't write it, it was just generated for you.

Whereas the good relationship in this scenario is a more kind of experienced engineer where you just not quite sure about something in a specific area, just get it to help a bit and go, oh yes, that looks right.

Hold on, I need to just change that a little bit.

It's just helping you out.

It is that, that little co pilot, that wingman, it's just there giving you the assistance that you need in engineering, in computer programming.

So I suppose sort of just using that example of like the junior engineer versus the more experienced one, that whole thing of again, it's about the expertise and about sort of how can you use generative AI as a specific tool to assist you in making producing the best type of work so you could have the best commercial outcomes?

Yeah.

And I, in the sectors that I've worked in, I haven't seen a valid use for generative AI yet, which is kind of interesting.

But large language models are the ones that are giving you all this text output and it's just outside of the coding environment, actually.

It can get incredibly confused even within coding, but outside of say the coding boundaries, just in generally talking with a large language model, like a chatgpt or whatever, the hallucinations, I'm pretty confident the next word in this sentence should be this.

I'm going to put it there whether it makes sense or not.

And you can read paragraphs of stuff and think.

Yeah, that kind of makes sense.

Yeah.

And it's complete lies, completely made up.

There's no, not necessarily based on fact, based on whatever data was ingested.

And there's no kind of quality metric of the data that it ingested.

You don't know where it came from.

You can't validate that in a lot of the bigger large language models.

You've got some more open source ones.

I think it's called mistral or something.

It's like a big French one where it's open source and you can say this is where the data came from and it's quite open and you can, you can understand it.

That one's getting quite big now.

I think that's kind of the odd one.

And the generative AI thing of I've got a whole load of this type of information, this type of data.

Give me more that is similar to that.

I haven't seen that in say, the coding or kind of E commerce getting.

Getting much use out of that.

I have seen some examples around kind of product development where it will just go off and give you completely random clothing ideas, which means you kind of escape boundaries of human possibility.

Your thoughts, I guess, like thinking.

I won't even consider that because actually, how would I make that.

That's going to be really difficult.

Generative AI can just come out with completely random things, which at least can help you overcome the friction of developing in your product and kind of go, oh, right, yeah, no, that.

I mean, that one's impossible.

But I could do something similar to that.

And the idea that it's giving you that nudge again, it's kind of pulling you by the hand, going, no, no, no, look, come on, this way.

The co pilot, the wingman.

I need an idea and it's overcoming the friction.

So from my perspective of being more AI inquisitive, like, so you'd say that that is a really good example of how it can be used in the space of creativity and that's how it enhances humanness.

So that, that way it can just help pioneer ideas that perhaps you might.

Not have the time to, or, or even you've got.

Everyone tends to have their.

Maybe not everyone, but a lot of people have the kind of boundaries around what they perceive to be possible, what they, what they'll even consider when given a particular challenge or task, and they'll just think about things that are within that boundary of kind of experience, of knowledge.

And maybe the generative AI doesn't and will give them ideas outside that then they can reach for and it will help them go a different path.

That's super interesting.

One of the things that we're kind of really keen to know about as well is sort of where do you see AI being used in the next five to 10 years?

How do you see it in the context of like, relationship and connection with enhancing human connectivity?

So going way back to like the original chatbot.

I don't know if you ever played around with this one called Eliza.

I just had my previous guest mentioned about Eliza and apparently you can still have a go.

Yeah.

And I mean written in the 60s, passed the Turing Test in late 80s, which is where a human is interacting with Something they don't know whether it's a person or a computer.

And they have to kind of say, I thought this was a human, I thought this was a computer.

Eliza passed that Test in the 80s.

And the surprising thing is it actually beat Chat GPT 3.5 in the past few years.

Chat GPT 4 is now like way ahead and the other ones are way ahead.

But.

And the thing is, it's not AI, it's a case of.

It's like a psychologist is kind of saying, and why do you feel that that's interesting, Tell me more.

There's no kind of commitment, there's nothing.

It's not saying things are completely unrealistic where you go, well, that's a computer.

That was weird.

It's just non committal, just like I, you know.

And how does that make you feel?

All these questions and that.

I remembered messing around with this back in like late 90s and just having kind of a full conversation, seeing how far it could get before it felt or looked like it wasn't a regular conversation, even though it could go off on complete tangents, like I was pretending to be a mob boss or something, having a psychologist trying to get it to help me with, I know, disposal of something.

And that felt like quite an interesting interaction that I can imagine.

Even people now just want to get something off their chests just to talk to a really basic psychologist chatbot, that even then, that same thing from the 60s, it's still quite useful and helpful and can get that emotional reaction, I think.

And now we've got the fact that there are, well, any number this, so many different versions of chatbots, all based on certain ingested data and prompts.

A friend of mine made a kind of Dungeons and dragons Dungeon Master 1 where all it took is you just gave it a prompt to say, you are a Dungeons and Dragon, a dungeon master.

Here's what you can do.

Don't stray from the things that you have created.

Do not include the real world.

Go.

And you could kind of have this conversation.

It says, right, you're in this setting surrounded by woodland, there's a mist, and they're like really quite detailed descriptions of places.

And what do you do?

Like, okay, I'm gonna.

And you weren't given like, go left, go up, go down.

You could describe, you could say, well, what time of day is it?

Oh, is this nighttime?

The moon is in the sky.

And oh, wow.

Having this really interesting interaction, not like emotional unnecessarily, but it's the same that you would have with A, I guess a more junior dungeon master playing Dungeons and Dragons, like more advanced people will give you so much more, I think.

But you could have, like a full game, just spend an afternoon having a chat with this thing that has been told what the bounds of conversation are, the bounds of its world, and try not to stray out of it.

And there are any money, you can choose to have conversations with large language models and AIs that are supposed to represent celebrities.

People have kind of trained them on that.

Celebrities, things that they've said, things that they've written, even their voice.

And it could feel like you're having that conversation with that person.

I don't know what that could be.

Quite an interesting one, but bit bizarre to then if were you ever to meet that person, would you feel like you know them because you've been talking to them for so long, but you haven't.

You've never done that.

So raises two questions to me.

Sort of.

First thing is, like, when you've talked about the creation of Yusuf Eliza and just how.

And the Dungeons and Dragons option that you create, that your friend created, that helps enhance community and that's one way that AI can be used positively to enhance human connectivity.

And just where there's a shared sort of global experience, because obviously that game can be played internationally.

And then the other side, though, is you talk about interactions with, you know, with somebody famous, you've never ever met them, but you're having deep chats with them.

Do you think there's a question about the ethics about behind that and.

Absolutely.

And how.

How does that play out in that space?

I mean, even so, with the conversations with chatbots, has the person that they're imitating given any kind of.

What's the word?

Permission to use their likeness?

Whether it's.

I mean, a likeness is, I think we know is kind of like the look and the voice of someone, but what about their personality?

Like, how could you say, don't use my personality as a likeness?

But if that is being generated based on things that you've said and you've done fed into a AI that's processed them, that's going to be really difficult.

And at what point do you know who you're talking to?

There's the tech that recently came out from ByteDance, the TikTok company, loopy, where you could take a flat image and some audio and just that single image will now animate pretty accurately, pretty realistically.

A single image with some audio.

Previously it would be, the mouth might move a bit and you go, well, that's just a weird moving mouth of a photo.

But now it kind of understands the tone of the audio.

When there's a break for them to breathe in the audio, the picture will inhale and its shoulders will raise, its head will move side to side, it will give eye contact, it will look off to the side as it's thinking.

Eyebrows will move, they'll be blinking, there are lips smacking, which I found really interesting when someone in the audio just kind of pauses and does a.

Now that's being animated.

It's just incredible and it's so realistic.

There was like just a photo of Leonardo DiCaprio singing in Chinese and you just, wow, it's gone from a photo.

And his cheeks are just like perfectly animated to form those sounds.

There's a statue, a marble statue that is, that is now singing or chatting.

That looks incredibly real.

And that kind of, if you can do that from a photo, at what point?

It just really worries me that people who don't really know that much about AI will find these in the wild and go, you know, on Facebook or whatever, which is now basically full of this kind of stuff and believe it that it's real or that that person actually said that thing.

Because it's so easy now to just go, oh, I'm going to take some AI generated audio that has been built from audio snippets that this famous person has, has made.

Because of course their voice is going to be out there in the public domain.

I can create a audio generating model from their voice.

I can just type whatever I want.

It will say it.

And now we'll have this picture that will look like them saying that thing.

Wow.

And it's just really weird.

That's crazy.

But I mean also quite fantastical about this whole new sort of area that we're moving into.

One final question that we're looking at is what do you see?

How do you see AI being used in, in the future and how it can.

So I, I like this one looking, looking at this one from the kind of, the different viewpoints.

The idea that, and unfortunately I think this is the one, the, the viewpoint that we're stuck in at the moment is looking at AI through the capitalism lens and how can I save money?

How can I get rid of people?

How can I do this cheaper?

Not really caring about the people that are now unemployed or now can't find work because it's just being automated.

It's so much easier and cheaper and that's quite a depressing place.

And I would hope that that's not where we'll end up because there's kind of the, the socialism lens of well, if we can take this work away from people and that means we can now provide universal basic income, then people don't need to work.

They can if they want to, but that work that had to be done for a company is now being done by a computer.

So if we can get to that kind of, that side of things where people get to choose what they want to do as opposed to doing what, doing something that they hate because they have to do it.

AI could be that thing that helps take jobs that people don't like, don't enjoy, but still allows people to live a reasonably good life without being too worried.

That's a very interesting sort of look at how society can progress as we've seen through the different ages and the different sort of technological advanced advances through time that I just have really enjoyed like our discussion.

It's been really, really good just to sort of like think about all these different issues, these ways that in which human connection can be enhanced or it can be fractured.

Both certainly in the workspace and also within the friendship and community building space.

So just as we sort of come to a close I want to say thank you so much for coming and being part of the show and sharing your knowledge and just.

And your sort of fabulous insight into this new whole world.

As I said I'm learning all the time so it's a good, it's.

I'm sort of sat looking at like absorbing all this information.

Where might people find you on your.

On the socials if people want to connect?

Well, I bailed from X a while ago so I'm either my website robinosborne.co UK it's where easier to find me.

I'm on Mastodon as it's hard to say it now it's rposbo@hackydomeio but yeah, just go to robinosborne.co uk.

You'll find me there.

Brilliant.

Well, thank you so much Robin.

Really appreciate you coming down and thank you for being on the show.

Thank you very much.

Yeah, I really enjoyed it.