WBD025 Audio Transcription

Threats From Artificial Intelligence with Toby Simpson from Fetch.ai

Interview date: Friday 13th July

Note: the following is a transcription of my interview with Toby Simpson. I have reviewed the transcription but if you find any mistakes, please feel free to email me. You can listen to the original recording here.

In this interview, I talk with Toby Simpson, the co-founder and CTO of Fetch.ai. We discuss the world they are building for autonomous economic agents and broader opportunities and threats from artificial intelligence.

“I don’t like calling it artificial intelligence as it implies it is not real. At some point, we will create this digital intelligence, and it will think, communicate and exist. I think that is an exciting thing and can be something which transforms our lives in all sorts of positive ways.”

— Toby Simpson

Interview Transcription

Peter McCormack: Good morning, Toby.

Toby Simpson: Good morning.

Peter McCormack: This is very nice to meet you here, because usually I have to travel a long way and I've had a 40-minute journey this morning, which is very nice.  So, nice to meet you here in Cambridge.  Can you just tell me your name, your background and let everyone know who you are and what you do here at Fetch?

Toby Simpson: So, I'm Toby Simpson.  I am Co-founder and CTO at Fetch and I originally come from a computer games background.

Peter McCormack: And how did Fetch come to be?

Toby Simpson: Well, Fetch is something that one way or the other we've been talking about, myself and the other co-founders for, goodness, probably about the last decade.  And as part of the work I was doing in the past trying to build, well I wanted to build The Matrix for my living room, trying to figure out just how to make a virtual world of such magnitude that you could put all this stuff in it.  Actually, we were talking about whether or not if you could build one really big enough, then maybe you could put representatives of everything in it, representative data, people, things, hardware and you might actually end up with something that was a new way in which value could be exchanged.

But we couldn't figure out how to get the scale, because with conventional virtual world techniques, you don't trust the clients but you do trust the servers and there's a scaling point there.  But actually along comes decentralised ledger technology and we suddenly thought, "Wow, goodness me.  Now we don't have to worry about this, because now we can distribute the load of creating this world across every single computer on the network", and that suddenly freed us up to think about putting billions, if not tens or hundreds of billions, of digital entities in this kind of space.

Peter McCormack: That's quite interesting though, because one of the big problems across the blockchain space is scaling; a lot of people talk about scaling issues and perhaps people are adding blockchains to projects when they don't need it.  But for you, it's actually a solution.

Toby Simpson: Yes, that's exactly right.  We didn't start thinking about this problem as a blockchain or a decentralised ledger problem.  This was something where that was one of the key parts that allowed us to get the scalability that we wanted, because ultimately when you've got these digital representatives of all their stuff, not only are you going to need a very large number of them, but you're going to need a very, very large world for them to live in.  This just allowed us to create that world in a decentralised fashion where you never quite know who you can or who you can't trust.

Peter McCormack: So, I think a good starting point would be getting to the nuts and bolts of Fetch, let's understand exactly what it is.  I've been on the website, I've read parts of your whitepaper, it's quite complicated, but there's a number of components within it in terms of marketplaces and your open economic framework.  So, in layman's terms, can you try and explain to us what Fetch is and what it does and what its application will be in the real world?

Toby Simpson: Yeah, I like to think of Fetch as the ultimate dating agency for value providers.  Fetch is about connecting people that have something with people who need that something or who might need that something as effectively as possible.  It does this by being this decentralised digital world in which useful economic activity can take place.  That activity is done by these digital entities or representatives that we call autonomous economic agents. 

Their view on the world is provided by our open economic framework.  That is effectively the senses like eyes, ears, touch; it's the agent's window on the world.  But what's really interesting about that is of course in the real world, we all have to see the same space; in a digital world there's no such restriction.  What Fetch does is it restructures that world in real time to ensure that what you see in front of you is ideal for you, and that's one of the really interesting aspects of this.

Peter McCormack: I'm starting to get a picture of it, but I think for most people it would be good to maybe hear an actual real-world example of an application and how and when someone might start or want to use Fetch.

Toby Simpson: Yeah, so a phrase that I've heard around recently is "economy of things" as opposed to the internet of things where all the little individual things are able to trade and do stuff on their own behalf.  I like to think of Fetch as the world in which the things in the economy of things live; it's all about solving problems with a large number of moving parts. 

Examples of problems with a large number of moving parts is anything involving mobility at all, supply chains, energy and all sorts of other aspects where increasingly there's more and more bits and pieces involved; but actually, it's really tough to try and manage those things from the top down, because there's only so many individual bits and pieces that a centralised entity can look at and manage and understand. 

The way we approach this is what we ideally want is for these problems to be able to solve themselves from the bottom up by breathing life into those individual moving parts, and letting them work it out between them.  Now, this is something that's connected with my past in artificial life, the idea that you could create a large population of things and they would work together.

Peter McCormack: Did you create some kind of game, like a Sims-type game or something?

Toby Simpson: Yeah, it was called Creatures.

Peter McCormack: I think I've heard of them.

Toby Simpson: Creatures was interesting because we created a biologically-inspired reinforcement learning agent that was made up of things like chemical receptors, emitters, reactions, neural dynamics, and somehow you stick all these things together and you had a creature that could adapt to its own world no matter what you threw at it.  That was really interesting, because that was a great example of where you didn't need to understand or program the really complicated behaviour in order to get it.  I liked that self-organisation, the fact that the computer was doing the hard work. 

I have to say it was probably the first time and quite possibly the last and only time where we actually used natural selection to refine a product for release.

Peter McCormack: Let's look at, say, the energy market you've mentioned there.  Did I maybe read on your website or somewhere that you could have the ability to move between suppliers instantaneously as and when needed?  Can you give me an example of how it may work?

Toby Simpson: I think one of the things we're going to see a lot more of is people are going to stick these big storage devices in their houses; there's these Tesla, I can't remember what they call them; power packs, or whatever they are, and they're interesting when you've got renewables, because of course, if you've got some solar panels, then it picks all that power up during the day and then stores it so you can use it back through the night. 

Now, of course, if large numbers of people are doing that in an area, then it would be great to be able to trade between each other.  That's the kind of thing that Fetch enables, because each one of those things, be it power generation or power storage, can have its own autonomous economic agent and it can look out for itself to try and deliver its value and get as much as it can for supplying what it does.  That's quite interesting, because then these devices can work together.

What we're seeing in the energy market, and this is good for suppliers as well as users and I remember the old, I don't know what it's called now, but we used to call it Economy 7 where you either had cheap-rate power or you didn't, and that meant that you'd try and kick off your washing machine at a certain time of night in order to get the cheapest-rate power that you could.  But that's a very digital thing; that's either cheap or isn't.  The reality is actually the cost of power varies constantly throughout the day; so wouldn't it be nice, and this has been done and worked on, if you're going to do it in, say, 20-minute segments, because then that tends to incentivise people to change their usage patterns, because you've got an economic incentive to work that stuff out. 

For example, you don't care if your electric car charges between 1.00am and 1.30am, a little bit more between 3.00am and 3.45am and then a touch more between 4.00am and 4.30am, so long as it's ready when you're leaving the house at 6.00am.  It may well be then, as a result of all these individual autonomic economic agents talking to each other and interacting with each other, they can get the best possible deal either with local and mini suppliers, like the person with the Tesla pack or something like that just down the road, or with their energy supplier.

Peter McCormack: It sounds like these autonomous agents and the fact that a lot of this is going to be built using machine learning and AI, that these will become very efficient markets and almost perfect economies in distribution of value.

Toby Simpson: That's very much what we believe will happen, because they can trade with each other.  They're not agents that only act as a result of input from somebody else; they're agents that can make decisions on their own behalf.

Peter McCormack: That brings a couple of questions to mind for me.  I was trying to think of this specifically, say, for the energy market and perhaps I have an energy pack at home and energy supplier, but I have the ability to move between suppliers as and when required.  It feels like there's a potential here that the competition within the energy market and the energy suppliers will be at the production level and no longer at the marketing and sales end of the market, because the consumers will be buying from an efficient market which will automatically make the decisions for them; therefore, they have no need to be buying or sold to.  Therefore, the energy suppliers compete on the production level.

Toby Simpson: Yes, and is that a bad thing?  It's a different thing.

Peter McCormack: It's a different thing, yeah.  I'm not saying it's a bad thing.

Toby Simpson: And there may be other aspects that they would compete with each other as well in that how that's all segmented up and how all these processes work, and so there's plenty of opportunities for that to change.  But I think it will change.

Peter McCormack: It's a change I see in market dynamics in that in certain sectors, and we already have it now with price comparison engines, that you can go out and find things at the lowest possible price; it's still a manual task.  Like if I want to find a flight, there's three or four different sites I can use.  If you take this interview of what you've said here, it just changed the market dynamics of the business; they no longer appear to need to have sales and marketing teams in that everybody will purchase from the open market automatically.

Toby Simpson: I think there's definitely going to be a change in that area for absolute sure, because it is more efficient, it's better for the users and it turns everything inside out.  Since you mentioned organising flights and bits and pieces like that, we had a bit of a horror story a few weeks ago when we were trying to get to Amsterdam for a conference.  Having sat there and waited for them to try and find out what was wrong with our aeroplane for three hours, they eventually canned the flight. 

What was interesting about that was suddenly there was 100-odd people who all had the same problem; they all needed to get to Amsterdam and the vast majority of them needed to get there that day, which is why they had taken an early flight, to buy themselves some time.  But because we'd been sitting there waiting for ages and then we had to wait again for a bus to take us back to the terminal building, we were running very short of time to do stuff. 

But the effort has to come down to the human being in those cases: we all had to get on the phone, we had to make calls, we had to try and compete for internet connections on a saturated wi-fi and saturated mobile phone connections in order to solve this.  Now, eventually we did but it involved quite a journey, well I don't know quite a journey, but into London and back out again to get to a completely different airport only to discover that that flight was then delayed by an hour, but we did make it!

But it's a great example where we had to make all of the effort.  It was interesting that the people sitting in front of us on the plane and behind us were also people who'd come off the cancelled flight; others had gone to different airports under different conditions, and you wonder how they got on.  But what you really want in a situation like that is for representatives of you, agents that represent you and your preferences to go out there and sort it out for you and then deliver you a solution that works, because they're going to be able to do it far quicker, they're going to be able to incorporate more things.  We only incorporated the obvious, we checked the obvious airports, the obvious flights and we found one that seemed okay and that was the one we went with.

Peter McCormack: A similar problem that I might face if England win tomorrow, trying to get to Russia for the weekend.

Toby Simpson: That is a great example of a very similar problem.

Peter McCormack: It's funny you should say that, because actually it is a problem; it came up the other day after England won.  There were lots of tickets available, but people didn't know how they could get to Russia, there were multiple flights, they could fly via one country or another.  So, I guess what you're saying here is that the manual task we do of using these comparison engines and searching through all the data will be done by an agent and just prepare us options to go here or there.

Toby Simpson: Absolutely and one of the things that's interesting about the kind of digital world that Fetch provides is, I've often been told that it's not the data that we do use that's interesting, it's the data that we don't, either because we don't know it's there or it's just too much hassle to look for, or the cost to deliver it exceeds its value. 

Now, when you can attach a digital representative to any piece of data, then it brings all that stuff into play because it's virtually zero cost to put that data into an environment, and when you've got an environment like Fetch that's constantly restructuring itself to put people in contact with each other that are likely to need what each other has, then you've got something very interesting.  When we eventually arrived at Amsterdam at nearly midnight, we were curious as to whether or not there would have been a train option.  Do you know what?  It turns out there is.  I didn't know this, but you can take a direct train to Amsterdam from London.

Peter McCormack: Yes.

Toby Simpson: I didn't know that.

Peter McCormack: It wouldn't have crossed my mind.

Toby Simpson: No.  So, I immediately thought, well obviously I have to go via Paris, right?

Peter McCormack: Yeah.

Toby Simpson: But no you don't and actually, that doesn't take nearly as long as I thought it would take.

Peter McCormack: I guess so because in your mind you're thinking there's an ocean, because that's the way our brains work.

Toby Simpson: Yeah, exactly.

Peter McCormack: Or we've got to go south to go north.

Toby Simpson: Yeah, so it just doesn't seem right and yet it exists as a possibility.  In fact, the time cost of that is not outrageous, but we never considered it because we didn't know about it; it was just, "Oh, I wonder if we could have done that?"  That's the problem when you don't have all this stuff in play; it relies on you as the individual to do all the work.  This is getting worse for us as time goes on; there's more and more things to consider, there's more and more moving parts and how are you supposed to hold all that in your head?  Are you going to miss something that was actually the optimal solution?

Peter McCormack: Yeah.

Toby Simpson: Yeah, quite possibly.

Peter McCormack: It's interesting, because I'm always trying to relate it back to how I would use something like this.  I travel a lot with doing the podcast and I try and do multiple interviews in a country at the same time to group them together, and I'm looking at flights and internal flights, but also I'm looking at the cost.  If I have to go to Austin, Washington and New York in one trip, should I fly to Washington and get a connection to Austin or should I go to New York first, and I end up creating a spreadsheet.  The reason I do it is because if you plan this correctly, the cost difference could be up to £1,000 on a trip, and also you can save yourself hours being stuck in airports. 

I'm trying to picture if everything connects, I'm assuming I could get up one morning and say to Alexa, "Alexa, I need to go to Russia this weekend for the England game.  I'd like my cost to be below this and I would like to be there by this time.  Email me the results", Alexa would activate the agent and the agent would do the work?

Toby Simpson: We very much believe that will be the case.

Peter McCormack: Okay, fantastic.

Toby Simpson: That ultimately you'll ask your digital representative to go out there and solve these problems for you.

Peter McCormack: How far away are we from this being a reality?

Toby Simpson: The good news is that the Fetch technology works.  We've been very, very quiet about it for quite some time until we were sure that the innovations that we had actually behaved in the way we expected them to.  So, we have a private test network of this operating internally and over the coming months, we're going to be releasing some of this code for people and we're looking at a public test network this year, probably late summer, early autumn. 

That's very exciting because it means that people will be able to download this stuff, they will be able to set up a node and they'll be able to create agents.  We know that we need to make it very, very easy for people to create these autonomous economic agents and benefit from the Fetch world in order to make it a little bit of a no-brainer to do this, because it's about utilisation of value.  We all walk around on a continuous basis wasting a huge amount of value; your mobile phones are full of really interesting sensors.  One of the things we were talking about recently was that on a busy London street, if everybody puts their phone away at the same time, the chances are it's just started raining. 

That's actually really interesting information and for me, if I'm walking from the river up to King's Cross, it may be the difference between whether or not I get on the Tube at Holborn or not, where I normally wouldn't because I'd walk it; it's good for me.  That kind of local immediate information that's available is only interesting if I know it's available, and that's why Fetch is there because it delivers that to you and ensures that you know what you need to know, rather than having to wait around for it or go looking for it.

Peter McCormack: So, over the next year, two, three years, I guess the development and our exposure to these agents is going to accelerate and probably happen a lot quicker than we expect.  Doing my research for this, what I realise is we're already interacting with a lot of AI, which we don't realise, but you will see you expect an acceleration of this; how far away do you think I am from being able to ask Alexa to find me something and be emailed the results?

Toby Simpson: I think that stuff's going to happen really, really quickly.  I think it'll catch us by surprise.  One of the things that's really interesting about things like the global network is, I remember writing back in the early 2000s that I thought that there'll come a time very soon when you won't know where your computer ends and the global network begins.  Now I know that's the case, because my children don't get it when they can't watch their videos on an airplane and I say, "Well, you don't have a connection to the internet" and that makes no sense to them, because as far as they're concerned, these iPads are always connected to all of the collective knowledge of the human race all the time and they don't understand why that wouldn't be available. 

So, to them the global network is everything; it's through all these devices on a continuous basis and they don't understand when it's not available.

Peter McCormack: It's funny you should say that.  My son is absolutely useless at being on time when I collect him from school and I'm always phoning him.  I said to him, "When I was your age, you had to arrive on time when you would meet somebody".  If you were to meet them at 4.30pm you pretty much had to be there, and people arrived on time; there was no alternative to not being on time for things because we didn't have phones.

Toby Simpson: Because you'd never see each other if you weren't on time because no one would ever manage to get together, so you can't do that thing where you just send a little text message of whatever saying, "I'm going to be ten minutes late, sorry about that".  It's a change to the way that we conduct our lives, but it's happened really fast.  I was only theorising about this in the early 2000s and you look at what we have now and it's not substantially further on than that timewise. 

So, you can see that with something like these agents, which takes the hard work off the human being, which is neat, and gets the computer to do all of this, it changes the way in which we interface with everything that we come up against in our daily lives.  It's curious how we much we put up with in just about everything we do.  And it still astonishes me, for example, that you go to the supermarket, you put everything in the trolley and then you take everything out and then you put everything back in again.  Really, what's that all about?

Peter McCormack: Yeah.

Toby Simpson: Transport, just generally; getting anywhere these days is really complicated.  We take it for granted, so we don't see it as complicated because we know how to do all this stuff, that you need to check this out, then this one, then that one and then you do need to do this, you need to double check on that one.  It's all these different things that you have to assemble, but we shouldn't have to do that.  Life's too short to worry about assembling all these bits and pieces just to do something simple.

Peter McCormack: I use a website called Momondo for flight search, and occasionally I check Kayak and Skyscanner and usually Momondo's the best for me.  Do you envisage a scenario where Momondo will create their agent within your ecosystem which does a more automated process for me; or who do you imagine building these agents?

Toby Simpson: That's a very interesting question, it's an interesting question because some of these services effectively act as middlemen, as intermediaries, whose job it is to connect somebody who has something with somebody who wants it.  Now, in a decentralised ledger world where you've effectively got self-service trust, any individual can help themselves to trust in that they can see that all the parties are who they say they are. 

Then you've got something really interesting because that changes the role of those intermediaries potentially quite significantly, because the agents will do all of this themselves, the individual entities will be out there making all of these happen.  So, particularly in the short term, there is a great need for people who know how to do this stuff and people who built all the algorithms, particularly all the really clever ones that connect people to things, to be represented in these spaces. 

So yes, they could; they could create their autonomous economic agents that represent the travel options that they have, they could let them loose in the world knowing that Fetch will reorganise itself so that anybody who wants something that they've got is connected.

Peter McCormack: What is the role of Fetch; is Fetch essentially a marketplace?

Toby Simpson: No, Fetch is a world.

Peter McCormack: Okay, it's a world, but isn't the world a marketplace?

Toby Simpson: I guess it's the ultimate marketplace in the sense that it's one that's constantly investigating new market intersections.  This is one of the reasons why we call it a smart ledger that underpins all this, because we're using all of the computing power of the network to build these predictive models and build this trust information so that agents can very effectively find each other, but also so that ones that shouldn't be connecting don't get connected.  That's really important, because if you've got a world with 10 or 100 billion autonomous economic agents running around representing all manners of things, then that's a lot of stuff and that needs to be sorted in a way that you only see what's important to you. 

That's the key thing that Fetch is doing; it's organising all of that stuff and not just on one dimension either, not just on a geographical dimension.  It's doing it on a number of other ones relating to enabling different kinds of economic activity.  For example, junction points, which exist throughout mobility, you're very interested in when a road junction occurs or where there's a railway station or something like that.  Those are all things that it's a really interesting dimension to sort a world by, because it enables all sorts of interesting solutions, particularly when you combine it with other ones. 

So, that's what Fetch is doing; it's providing this world, it is a fully decentralised world and there's no magical superpowers that are held by any one of the nodes over any of the other ones, so it's truly freeing that information, the collective intelligence that it builds about all this to all of the users.

Peter McCormack: What is the risk of malicious agents being built and being loose within the world, and what can you guys do to police that?

Toby Simpson: Policing, these things need to be self-policing; you can't have a central entity that says, "Oh, you're a bad agent, you're a good one, you're a bad one", because that just doesn't work.  Also, it doesn't scale for that matter.  You need a mechanism by which all of this happens.  What was clever, of course, about Bitcoin and blockchain was it was more profitable to be honest than it was to be dishonest.  So, why on Earth would you try and rig transactions when the reality is you could have done much better just sitting there chugging away and mining or doing other things?

Peter McCormack: Let's say misinformation.

Toby Simpson: Misinformation, well you're always going to get bad actors in any scenario like this.  The trick is how do you ensure that the damage that they cause is limited, and one of the things that's very important for us is the trust information that the network builds about agents and nodes.  Because, if you start behaving differently or outside of what you would normally do, then that affects your trust values. 

Certainly if you'd set yourself up with a whole load of fake transactions, just two or three bad agents travelling between each other to try and build up some history and then you go out there and try and do bad things, then you would have very low trust values until you build that up.  What you build up over a period of time, you can lose very quickly by behaving poorly.  So, that kind of stuff is really important, and we've done an enormous amount of economic modelling on all of this to try and see what happens. 

Of course, we will spend an extraordinary amount of time trying to figure out how to attack our own network in many different ways.  What happens when we create a million agents whose job it is to mess around with everybody else's stuff and just dump them in the world at random points?  How does the network react to that, how do the trust values correct all of that and does that affect people's ability to get things done?

So, we know the importance of all of this, and we know the importance of making sure that that is sorted at a protocol level and at a network level so that everybody is heavily incentivised, because it is better for them to be honest.  Then you recognise there'll be people who have a motivation which is completely different, whose goals are to achieve something that you really don't want.  You build the mechanisms in to prevent that where you can.

Peter McCormack: Can you talk me through in terms of the technology, what are the layers of tech that exist here?  Obviously, you have your smart ledger.

Toby Simpson: Yes.

Peter McCormack: Let's talk a bit about that actually.  How does your ledger compare to ledgers that we are aware of already in the blockchain space, because if you're enabling billions of transactions instantaneously, it probably isn't a blockchain ledger similar to Bitcoin?

Toby Simpson: It's fair to say that it is, yes.  There are some similarities, we're all drawing from the ideas of decentralised ledger technologies and we're all trying to create or build a system where you can establish that global truth when you never quite know which individual parties you can trust; so, reaching consensus on that truth and an ability to add things to an ongoing chain of stuff.  But yes, we had to do it differently. 

We needed to paralyse blockchain effectively so that we could get a much larger number of transactions through the system.  We combine that with a bunch of other things in order to increase those volumes even further.  Actually, this is where we started on the journey to useful proof of work, which is what we call it.

Peter McCormack: Okay.

Toby Simpson: You've got Bitcoin network and Ethereum network and you look at the huge amount of computing power that has and you think, "Wow, look at all that computing power.  You're talking about supercomputers over supercomputers over supercomputers.  Wouldn't it to be cool to be able to do something useful with that for the benefit of the network and the users of the network?"  For us originally, this was like a giant Tetris game really where you're trying to figure out which transactions don't collide with each other and don't share resources, because you can execute those in parallel. 

It turns out that when you're arranging a large number of potential transactions over, say, 1,000 individual lanes, then that's computationally non-trivial.  That's where we started on the road to useful proof of work, because we thought wouldn't it be nice if we could reach consensus on the network but do so as part of optimising those transactions where individual nodes were trying to figure out effectively how to arrange this puzzle piece so that all the lanes were occupied as best as possible, like a giant Tetris game, and then submit those blocks saying, "I've done well, look at me.  I think this should be the block". 

Then we've got those blocks then go on to a directed acyclic graph, because we can scale all the different results, and there is a mechanism by which somebody is an elected leader and they pick a block to go on to the global state based on the fees, because it gets split between the person who solved the puzzle and the leader.  Then the next leader is the person who achieved the best occupancy and so it goes.  And all built into the protocol is a mechanism where if the elected leader doesn't choose quickly enough, it falls to the second one and the third one and so on.  And that person is heavily incentivised to do it, or someone else will, because there's fees to collect by doing it. 

That was great, because that gave us one massive level of scalability, but also the computationally challenging exercise of arranging all of that became part of reaching consensus, which was neat.  Then we thought, "Well, why stop there?  Wouldn't it be nice if all this other stuff that we wanted to do, like trust information and predictions, we could also do as part of useful proof of work, because then we would be harnessing that computer?"  

There's a lot of people talking about it's a huge waste of energy and will possibly go down in history as the biggest waste of energy of all time.  You think, "Well, if we're doing something useful with it, then it's not anymore".  Then it's really cool, because then we've got the globe's largest supercomputer and we're doing stuff with it.  Trust is really important on the scalability thing, because trust for us, one of the things we've got on a probabilistic level is the probability that any given transaction that hasn't yet made it to the global state will make it to the global state. 

So, if you're able to look at, say, a transaction that's pending there for the weather information just down the road for a tenth of a cent, and that's got a 96% chance of making it to the global state, "Yeah, you know what?  I'll let that one fly, I'll take it", knowing that maybe 100 transactions down the line one of them might bail, but generally it means that you can make very, very quick decisions safe in the knowledge that the network is providing you with information that is accurate.

Peter McCormack: How does the incentive mechanism work within Fetch?  Will you have your own reward mechanism?  Will you have your own currency and token?

Toby Simpson: This is a tokenised --

Peter McCormack: Economy.

Toby Simpson: -- network and it needs to be.  It needs to be, because some of these transactions are very, very low value transactions.  If you're strolling down the road and you want to know the local temperature every 100 metres from somebody's weather station in their garden, you're not going to be paying big bucks for that.  So, you need to be able to get the transaction fees low enough so that you can charge these very low token amounts so that all this stuff comes into play.  It's also tokenised because it's a global network; you need to be able to move all these things around.  That is the method by which it makes most sense. 

Of course that way, that can be part of the incentive mechanism which operates on many levels, I might add, because if you're operating a Fetch node, one of the things that you're doing is providing services to agents.  Because some of those services involve more computing power than others, such as delivering trust information or allowing agents to walk around the network or see and gather information about a big space, then that is something that will cost the agents, not an enormous amount, but it scales very quickly if you've got large numbers of agents there.  Plus of course, you've got the rewards for solving these puzzle pieces because of all the transaction fees that eventually get split up between people.

So, there's multiple ways in which this stuff works and one of the problems we wanted to solve with this is to ensure that it's not a winner-takes-all scenario.  Winner-takes-all is dangerous, because that's where you lose a lot of computing power and waste it, because you're in that situation where 1,000 people do a lot of computing power, one of them wins and 999 of them effectively did nothing.  We didn't want that to be the case, because it's not useful work unless everything that everybody does is useful. 

Of course, if you can solve that problem, then it means that you, anybody can operate a node, even if they've got a relatively low-powered computer, because as a result of performing any kind of computation on that network, they're going to be rewarded for doing so.  That's one of the things that encourages a larger number of nodes on the network, which also happens in Fetch because sometimes people want to place nodes geographically in space in order to attach more agents that might want to be in an area, like around a large city or somewhere like that.

Peter McCormack: So, Fetch will have its own token?

Toby Simpson: Yes.

Peter McCormack: Which will distribute value, but to extract value the token therefore will need to exist on exchanges for people to buy and sell out?

Toby Simpson: Yes, there will need to be a mechanism for that otherwise you wouldn't be able to get the tokens to start with.

Peter McCormack: Of course.

Toby Simpson: And you wouldn't be able to take them out of the space.

Peter McCormack: So, one of the areas I've struggled with understanding is token economies and token metrics and crypto economics and I've read a lot about it.  Help me understand from your side whether this is a complicating factor or something that you guys are pretty clear on.  People are going to be rewarded in a local token, but obviously we're aware in this crypto economy there's lots of speculating, we have market-makers, we have volatile prices.  Do these present any risks to you with trying to price value and trying to distribute value in the system?

Toby Simpson: It does, it makes it a lot more complicated on the surface in the same way that you wouldn't buy a cup of coffee with a Google share; you have this issue in tokenised space.  If people are, for whatever reason, holding on to tokens, either for future utility value or for any other reason, that potentially takes them out of play in the economy and you want people to be using these things.  And particularly in somewhere like Fetch where there's an enormous amount of these things going on, what you can't and don't want is for the value of these to change substantially over a transaction.  So, you need to have a method for dealing with that. 

Now, we do have a method of dealing with that because we've separated high-velocity and low-velocity tokens and we've actually got a very interesting paper.  We're going to be publishing a lot more information about this over the next month that talks about all of this in a great deal of detail.

Peter McCormack: Great, okay.

Toby Simpson: Effectively in the same way with Ethereum and gas pricing, you need to have solutions to these problems, but when you're doing large numbers of rapid transactions, the difference is that if you put it into something which is more stable for the duration of that transaction, you need to be able to reverse it back out again afterwards.  

So, there are some additional complications and also that works for us as a direct minor reward for performing processing on the network.  So yes, we did need to have a solution to that problem, because Fetch is a world and we want people to use it, we want to make sure that it can be used and that some of the results and some of the things that we see in some of these tokenised economies aren't going to grind everything to a halt by removing the tokens from use and from circulation.

Peter McCormack: Yeah, it's an area I've really struggled with recently, the more I've investigated and the more I've read about it.  What I struggled with, I struggled on a couple of levels; firstly, that we're creating new complicated business models for businesses to plug into, based on token value and token prices which are volatile, when their business probably operates in a fiat model in any scenario.

Toby Simpson: Yes.

Peter McCormack: Therefore, the value of what they're exchanging can change quite wildly.  So, I struggle with that; and I struggle to understand why we haven't therefore seen a bigger adoption of stablecoin or stable tokens within these models to eliminate that risk.

Toby Simpson: I think we will see things like that happening differently.  One of the things that's different about Fetch is it isn't just a mechanism of changing or exchanging some kind of token with each other; it is a world in which all this stuff takes place.  The fact that there is a token is necessary.  We didn't set out to create a new, for want of a better phrase, cryptocurrency or something like that.

Peter McCormack: Of course.

Toby Simpson: We set out to create a world where agents could have their own lives and hopes and dreams and can get out there and do stuff.

Peter McCormack: It feels like a lot of these token economies though are now realising they have to have some form of mechanism to restrict supply, whether it's some form of staking or some form to build value in their token to give the token something worth investing in, which also then sometimes feels it's maybe a distraction from what they're actually trying to create.

Toby Simpson: Maybe it's a side point, I don't know, but one of the things that is interesting is the accessibility of all this stuff.

Peter McCormack: That's true as well.

Toby Simpson: I was highly amused that firstly, as a point, I doubt they created Ethereums or CryptoKitties; but secondly, when it did, suddenly there was a whole generation of young people who ended up with Ethereum wallets as a result of that.  They all went through the pain of realising that actually, this stuff isn't nearly as accessible as you might think and the learning curve is extremely steep; it's like snowboarding from that perspective.  You spend three days falling on yourself in different ways and getting an increasingly large number of cuts and bruises, and then suddenly you get it.  But getting to that point is really difficult, and one of the things that we're going to see in these tokenised economies is that people will be less aware of the fact they're even there.

Peter McCormack: I think that's a necessity.

Toby Simpson: Yeah.

Peter McCormack: I gave a comparison the other day to somebody where, when I was on holiday in Thailand, you know when you're away you have the exchange rate in your head and you're constantly converting?

Toby Simpson: Yes.

Peter McCormack: I had the Thai baht conversion wrong and when I came home, I'd spent around £800 more than I thought just because I hadn't had the conversion rate in my head.  Currently, you have to deal with the token, the native token, and understand the value and the exchange.  It's almost like if you went to a shopping centre and every shop you went into, they had a different currency you had to deal with; it's overcomplicated for the majority of people.

And also, with something like MyEtherWallet, if you want to invest in an ICO, you have to understand the GWEI and gas and all this stuff that even me, as somebody who has worked in the space for quite a long time, finds also very complicated.  It feels like we're missing a UX layer between the user and their use of fiat and the token, which has made everything rather complicated.

Toby Simpson: We are most definitely missing that layer and it's not a very welcoming space without it.  I'm still amazed that so many of these wallets, when you download them, the first thing that happens where you try and run is it says, "This is an unsigned application.  Are you really sure?"

Peter McCormack: Yeah.

Toby Simpson: Really?  Is it?  Should I go any further or should I not?  Then, as you said, you've got to understand you need all this extra stuff there to pay for the transaction fees, so you need your gas fees and all these other unhelpful abbreviations for things that don't make sense until you understand enough about it to do that.

Peter McCormack: Right.

Toby Simpson: With Fetch, of course, we can create and we imagine these interface agents that their job is to connect one aspect of one world to the Fetch world and vice-versa.  This is interesting, because effectively you can disguise the fact that that's taken place.  It's necessary in order for the Fetch world to do its magic, but actually we can ease the burden of performing jobs on it.  And actually you want to do that, you really do want to do that, because if the barrier to entry to do anything on Fetch is that you have to jump though 100 flaming hoops to set X, Y and Z up before you can even start, then most people aren't going to bother.

Peter McCormack: Nope.

Toby Simpson: With Fetch, where we're talking about creating tens of billions of autonomous economic agents.  If every single person who is creating agents and developing them has to go through all that as well, they aren't going to bother either.

Peter McCormack: Yeah.

Toby Simpson: This is part of this space growing up.  I guess we've been involved with it for long enough to think that it's always been there, but it hasn't.  It's brand new, it's still got that new-car smell to it, and we're still investigating all the different applications and things we can build with it; it's just that we've forgotten, in some cases, to bring everybody along for the journey.

Peter McCormack: I used to have just a very simple web agency; we built websites and various things.  Every employee who joined, I always would buy them one book by Steve Krug called Don't Make Me Think.

Toby Simpson: Yes, yeah.  The user interface book of the gods, that one, isn't it?

Peter McCormack: Yeah, and it was just to get everyone in the mindset that everything we create, it has to be very simple to use.  I used to have some amazing developers, brilliant programmers who could build great websites or applications very quickly, but they would always have to work alongside two other people.  You would always need a usability designer to crate the interface for them to work with, but you would also need some form of account manager, business director who would be thinking commercially as well, just trying to think of the commercial application of it.  With two of the three, you would usually find problems but with the three working together, you would get something commercial that was easy to use and brilliantly built. 

I feel in the crypto world, we've got an awful lot of the building, very little of the user interface design and a lot of people are retrospectively fitting the commercial models and reinventing the commercial models.

Toby Simpson: Which I think is going to be the case, isn't it? I mean, let's look back to the worldwide web back in 1994/95.

Peter McCormack: Of course.

Toby Simpson: It's exactly the same, I think anyway.  Two of the funniest books on my bookshelf are the Internet Yellow Pages 1994 and the Internet Yellow Pages 1995.  One is about a couple of centimetres thick; the other one is about five or six centimetres thick, and they never made another one.  But in that space, we were all investigating, doing a lot of development, were doing a lot of ideas, were creating a lot of things; we weren't necessarily taking everybody along for the journey at that point because we were still poking around. 

All of those, as you say, the other two parties that need to be there in order for stuff to work well for people, and that didn't come along until later.  As a software developer myself, I can tell you that programmers shouldn't build user interfaces.  I built some terrible ones because I thought, "Oh buttons are cool, I'll have some of those" and before you know it, you've got tons of these things.  Then somebody else looks at it and goes, "What on Earth is that?"  You say, "Well, all these things are important".  Are they?

Peter McCormack: Yeah, are they?

Toby Simpson: And they aren't.  I guess this is one of the things that Apple understood with things like the iPod.

Peter McCormack: If you look at the Apple TV remote compared to any other remote.

Toby Simpson: And I do. I've got my Apple TV remote and I look at that and I think, "That is actually all I need".

Peter McCormack: Yeah.

Toby Simpson: Then I look at my actual remote for the TV and I thought, "What were you thinking?"

Peter McCormack: Yeah, I know.

Toby Simpson: My kids can use the Apple TV, they can't use the television or the PVR because it's way too complicated, unnecessarily so.

Peter McCormack: Yeah, I completely agree.

Toby Simpson: Because they've literally got a button for every single operation you could do.  That's not even how human minds work; if your house was organised like that, then you'd have a menu of everything you could do in your house, in your living room, and you'd never find anything.  How would you ever go to the toilet again?  It's ridiculous to be presented with all of those things.  Human beings are terrible at this.  We really aren't very good, particularly if we're involved in it, because when you know something, it's very hard to understand what it's like not to know it.  I did a talk recently called The First Rule of Blockchain Club which I suggested --

Peter McCormack: You don't need a blockchain?

Toby Simpson: No, The First Rule of Blockchain Club is not to be able to explain blockchain to anyone.  This seems true, because people who know it find it very difficult to explain it to other people in the same way that people who are in the space, full stop, find it very difficult to explain what's going  on, and that will change.

Peter McCormack: I thought the first rule of blockchain is you don't need a blockchain!

Toby Simpson: In a lot of cases, that's probably the case, right; particularly people who went through this in the 1990s.  You take your company name, you slap dotcom on the end and you've tripled your value.

Peter McCormack: Yeah!

Toby Simpson: But in and amongst all of that, and in the late 1990s when that was taking place, there were some incredible innovations going on and things that are still with us today, having fundamentally changed the way we go around our daily lives.

Peter McCormack: But you couldn't have built Fetch without a blockchain.

Toby Simpson: We couldn't have built Fetch without decentralised ledger technology.

Peter McCormack: Without decentralised ledger technology.

Toby Simpson: Yeah, exactly because that gave us the scalability.  I was used to building worlds in a client server basis where eventually there's a pinch point of the server.  The fact that I could create these worlds in the early 2000s with tens of thousands of agents, so with today's computing power it may be a million or two and that's not enough; a million or two isn't enough to represent very much at all; we need billions, and that's why we needed that kind of technology.

Peter McCormack: What challenges does that bring to you?  One of the biggest challenges with decentralised ledgers is essentially they're immutable.  Does that present you guys with any issues if there's any conflict, any misinformation in there?  Are there any challenges?

Toby Simpson: That depends on what kind of misinformation we're talking about, because ultimately what is on the ledger is the global truth.

Peter McCormack: Right.

Toby Simpson: If that's not necessarily what people wanted to be there, at least you've managed to reach consensus on that, and that is storing the results of all of the value exchanges that are taking place between agents.  Things like the predictive models and people's data, now that's not stored on the ledger; that's represented by the agents.  What it is that's learned from the data that's going by on that and delivered as the predictive models and everything else to the agents, that's one of those things that the information you built a year ago is worthless, so why on Earth would you store it on a ledger?  That's just an insane waste of space; so you don't. 

You've effectively got a wave of predictions that's been built and distributed across the network between nodes all the time that provides you with that trust and those predictive models.  In the meanwhile, all of the value exchanges and some other information is being stored on the ledger.  But the ledger isn't a database for people's stuff; the ledger is there to provide integrity and to ensure that you can reach a consensus on all of these things and that you do have a global truth that anybody on the network is able to verify is in fact the case.  That's what it's there for; it's like the underpinning that makes the whole thing work.

Peter McCormack: We've done quite a bit of time already and obviously a lot of my research was the AI side of things, so I'd like to move on to that and ask you some questions about that.  I've got wider AI questions I want to ask, but could you explain to me with Fetch what is the deep learning component, the AI component, and where is it you're adding value with that?

Toby Simpson: Well, we knew what we wanted to achieve and what we wanted to achieve was to put an agent with something in touch with an agent that wants it or who might want it, and we wanted to minimise the distance and the hassle between the two. 

Now, there are a number of machine-learning and AI techniques that you can use to do these things.  They're called embeddings and this is where you take all of the actions that an agent has done over a period of time, and you effectively pass that through a neural network called perennial network.  That does a dimensional reduction, and you end up with something called semantic hash.  This is unlike a cryptographic hash where there's no resemblance at all between what goes in and what goes out.  With these, there is.  Effectively you're creating a space in which you can put all of these things and the ones that are near each other tend to be related to each other.

This is a really interesting field and a lot of people have been doing some pretty amazing stuff. Actually, it helps us in our daily lives in all sorts of interesting ways to try and relate what should be near to something.  Now, we can use this to figure out what kind of agents are likely to want to be with other agents.  You get mistakes with dimensional reduction; a bit of a weird example, I guess, but if you take a sphere and you collapse it down to a circle, then London and Sydney are suddenly a lot closer than they should be.  Now, that's a mistake. 

As it happens with Fetch, we happen to know whether or not any of the results that we're delivering work because if they are delivered and value exchanges take place, they were probably good.  If they were delivered and no value exchanges took place, then they probably weren't that good at all.  So, you can effectively perform reinforcement learning on all of those things to get rid of the ones that aren't working and to reinforce the ones that are.  This also allows an exploration of new ones incidentally which is quite interesting, because one of the things we wanted with Fetch is for new intersections between marketplaces that weren't obvious to be explored by the network.  So, that's one aspect of where all of this happens. 

Also, we use Bayesian networks for the trust, the probabilistic layer that I was talking about where, as you build information as transactions are moved around the network before they get organised in one of those puzzle pieces we were talking about, like the use of the proof of work and put on the network, they have been signed by every node that says, "Oh yeah, you know what?  I'd include this if I could, I like the look of it".  Then you can run that through Bayesian networks and you can end up with a trust response which you can deliver to people.  So, those are two aspects of where we're using this. 

But that is very much just the beginning from our perspective because the VM that allows us to do those things allows us to do all sorts of other, more general machine learning things and that's stuff that we may create, but that's also stuff that users may create.  That's super-duper interesting, because we've got this huge supercomputer that's capable of performing these things and a lot of machine learning tasks that can be distributed and the ability to do so.  So, agents can package up these problems, submit them to the network and then have them done. 

We're also interested in building collaborative models.  Collaborative models are where a number of people can contribute to a model, but don't have to give away the data that they use to contribute it.  That's cool because if you've got two or three companies with some proprietary information that know that they would all benefit if they were contributing that to the same model, but don't want to give away their data, suddenly that's possible and that's also very interesting.

Peter McCormack: To somebody who's not hugely experienced with AI, is there a significant difference with doing the aspect which is dedicated to machine learning and improving data and then machines being able to actually think and operate independently or is it all one sphere?  And, therefore, I think what I'm getting at, I obviously want to start talking about the risks with AI.  If you're creating a world which is inhabited by artificial intelligence, are there unknown unknown risks?  I guess you wouldn't know because they're unknown unknown, but you see where I'm getting at?  Is there a chance where it can start to create a life of its own that you potentially can lose control of?

Toby Simpson: First things first, would that be a bad thing? 

Peter McCormack: Depends how far it goes.

Toby Simpson: Human beings are remarkably insecure and paranoid and this is just my own warped personal view on this.  I think that we imagine, for some reason, that any form of thinking machine that we create would somehow be saddled with the same 3.7 billion years of evolutionary baggage that we've ended up with.

Peter McCormack: It depends how we model them.  One of the things I've been doing in my research is that the AI seems to be modelled based on human thinking and is that really the great starting point?

Toby Simpson: Quite possibly not, but in a lot of cases it actually isn't, and I call this the thinking gap, which is the gap between where we think we are with thinking machines just generally and where we really are.  Actually, we're a long way away from a true digital intelligence right now.

Peter McCormack: Are we?

Toby Simpson: I think so.  A lot of what we're doing is very, very powerful narrow AI.  In fact, a lot of what we perceive as being machine learning and AI out there is actually just statistical analysis and very large data sets.  One of the things we've got now is an enormous amount of computing power and an enormous amount of data and some very, very refined and very clever algorithms that we've created for analysing that.  That's not the same as something that can think.

Peter McCormack: But there is money and brain power going into developing computer-based neural networks?

Toby Simpson: Absolutely, but we've been planning around with neural networks one way or the other for quite some time.  They are not really the same neural networks that exist inside our heads.

Peter McCormack: Of course, yeah.

Toby Simpson: There's still a big gap from the way in which biology does these things and the way in which we do these things inside computers.  That gap will need to be closed quite significantly for us to see any progress on that front.

Peter McCormack: The first film I ever went to the cinema, where I wasn't old enough and I went along with my friends to get in, was Terminator 2.  I was 13, I think; I think was 13.  But that, The Matrix, there isn't a good AI story from history.

Toby Simpson: Yeah, history's loaded with good AI stories.

Peter McCormack: I remember all the bad ones!

Toby Simpson: But then that's a human trait, isn't it?

Peter McCormack: Of course.

Toby Simpson: Terminator wouldn't have been much fun if he'd have been a real good guy that had come back in history to make some nice little corrections so that human beings were happier and live longer lives.

Peter McCormack: No, of course.  We might not get Skynet but very interestingly last week, in preparation for this, I binge-watched every Black Mirror episode.

Toby Simpson: Oh!

Peter McCormack: Through all of that, a number of the stories seemed to be a lot closer to where we could be in terms of very basic AI helping us make decisions or guiding us in making decisions.  I think a very interesting one was the Hang the DJ episode, the dating one where it said who you should be with and how long for.

Toby Simpson: I remember that one and yeah, that is quite interesting. 

Peter McCormack: Do you think we've moved closer with that?

Toby Simpson: What's clever about Black Mirror is it takes one of these narrow examples and it takes it to an extreme, but not so much of an extreme as it becomes implausible.

Peter McCormack: Yes, of course, yeah.

Toby Simpson: It's just enough to dance on that line where you think, "Oh, goodness me.  Yeah, right.  Hmm, this is a bit creepy, this is a bit scary and I'm not sure I like this".  There was a weird black and white one with these little four --

Peter McCormack: Metalhead?

Toby Simpson: Yes.

Peter McCormack: That's probably my favourite episode.

Toby Simpson: That was really spooky from that perspective.

Peter McCormack: Yes.

Toby Simpson: Although quite why she didn't bury the creature once she'd managed to run it out of power is anybody's guess but, you know, different thing.

Peter McCormack: Yeah.

Toby Simpson: That's what's very clever about it, because it makes us think and it's very important that we do.  I'm not suggesting for one moment that we all sit down and go, "Ah, it's going to be fine.  Let's just carry on rolling and see where we end up".  We are going to have a go at creating this and, as human beings, I'm absolutely convinced that we will create a digital intelligence.  I don't like calling them artificial intelligences because that implies it's not real.

Peter McCormack: Yeah, that's a fair point.

Toby Simpson: I think at some point, we will create this digital intelligence and it will think and it will communicate and it will exist.  Will it be like us?  No, it won't; it'll be different, and I think that's actually an exciting thing.  I think that can be at least something that transforms our lives in all sorts of wonderfully positive ways, because we'll have a different kind of intelligence to interact with.  It's very easy for us to get worried about how that might negatively impact our lives, but then that's also a natural human thing; we get worried about these things until we're in it. 

I guess some of my concerns would be, I've noticed that we are busy training ourselves to be rude to computers.  So take Siri, for example, which is not the greatest of personal assistants when it comes to understanding what you're asking, so if you ask a very polite question, "Excuse me, Siri, would it be possible for you to play my favourite track?  You know that… etc?  Thank you" then it will misinterpret half of the things that you've said to be polite as part of the query, and in the end you get used to the fact that you've got to stop saying "please", you've got to stop saying "thank you"; you've got be direct and you've got to remove all the niceness from the way in which you communicate, otherwise it won't work.

Peter McCormack: But that's a very British problem.

Toby Simpson: Particularly, I know and it infuriates me because I want to be nice!  That's where we're busy training people to be rude to computers and there'll come a point when digital intelligence is reached and what kind of position will we be in? 

Just since we're having a talk about how it could all go terribly wrong, I think that one of the quotes I've used in a presentation, which funnily enough was entitled No, It's Not Going to Kill Us All, was in fact something along the lines of, "The first computer to be able to pass the Turing test may choose not to".  That gave me a bit of a shiver down my spine when I read that, I thought, "Oh yeah, that might be a very good reason why it might want to just not bother letting us know". 

But then I'm an optimist with these things; I see computers helping us out in our daily lives in all sorts of ways.  When was the last time you had to get an A-Z out or unfold a map in the passenger seat of a car to try and figure out where you were?  All these digital assistants are taking all these tasks that used to be complex and painful and awkward, and they are removing that awkwardness from our lives.  Some might say, "Oh, yeah, but we're losing these skills".  It's, "Yeah, but when was the last time you had to go out into the woods and kill an animal, skin it and cook your own food to survive?"

Peter McCormack: No, I agree.

Toby Simpson: Us human beings, we're moving on and we're losing some of those old skills, but that's not a bad thing because we're able to do different things and our lives get better as a result.  So, narrow AI, which is what we're seeing right now, is transforming our lives in many ways and I think that's wonderful.  Strong AI, that's digital intelligence, that is coming.  I don't think it's as close as people think and there'll be all sorts of little hops along the way while we try and figure out techniques that will allow it to happen, but we will get there and I think that's probably the closest I'll ever get to interacting with an alien intelligence during my life, and I hope to see it.  I think that will be really quite something.

Peter McCormack: I agree to some extent, but I do have some further questions because last night I watched Do You Trust This Computer? the documentary about AI which Elon Musk showed to his staff?

Toby Simpson: Right.

Peter McCormack: Which I know comes from a specific angle, but it closes on the quote, and I've written it down here, "The pursuit of artificial intelligence is a multi-billion-dollar industry with almost no regulation".  Elon Musk said it was a threat.  I think I've got it here that Stephen Hawking said it could end mankind.  There's a lot of intelligent people who see the threat, but at the same time you have governments who always have an arms race, who whilst they'll see the threat they will want to ensure they're ahead of enemy governments. 

You have companies such as Facebook, Amazon, Google who won't want to be left behind, who will therefore pursue the development of AI.  It feels like without any form of regulation, that it could be a risk to mankind eventually and it also feels that even if you did regulate it, AI is one of those things that could be developed rogue, like a rogue weapon anyway.

Toby Simpson: Yes, and this is pretty much every technological leap that we've made since we started ploughing fields; regulation and control has been behind the reality.  Regulation and updating data rules and things like that to create crime, so digital crimes and all that stuff, that was lagging behind the actual technology by a decent margin. 

We see it now in the blockchain space where you've got these new ways of fundraising, these token generation events and things like that; regulation is catching up, but it's behind and we'll see it with AI.  I don't personally subscribe to the Elon and Stephen, "We're all going to die, let's all be terrified by it" playbook.  I think that regulation for this kind of stuff is going to come, but that's not going to stop people from trying to develop it.  I think the benefits to mankind from creating a digital intelligence far, far outweigh the negative uses of these things.  But there will be negative uses.

Peter McCormack: Of course, of course.  I agree.

Toby Simpson: I don't think there's a piece of technology that we haven't chosen to misuse in some way, but on the other hand, we wouldn't get rid of that technology because most of it is beneficial to us.

Peter McCormack: No, of course, and the documentary covered an example where an AI computer can read a 1,000 lung X-rays for cancer and spot things that the human eye can't, that would take a doctor an hour, two hours each, which obviously in the field of healthcare is hugely beneficial.  But at the same time, it did talk about the potential to unintentionally create dangerous things.  I'm not sure if you saw the AI Twitter account that was created that went rogue that was racist?

Toby Simpson: Yeah, that was very well gamed by the people who gamed it into doing that!

Peter McCormack: Of course.

Toby Simpson: That's because it's an algorithm and it's not a very smart one, in the same way that IBM's rather famous chess-playing computer actually doesn't know how to play chess; it's just a lot of very good algorithms, a lot of very good data and a lot of processing power.  You can't stop it halfway through and say, "Do you know what?  I'm bored of chess, let's play Monopoly.  In fact, let's not bother with Monopoly, let's just have a chat about what we're going to see at the weekend, do at the weekend".  You can't do that, because there's no context; it really doesn't understand what it's doing.

That's why things like speech recognition are so awful in normal human environments.  As a human being, you and I can be at a party, a very loud party, and we can have a conversation with each other and we can isolate that one single conversation in and amongst every other conversation that's going on without making any mistakes.  If we miss any words, we'll fill them in in our mind from the context alone and that's because we understand what we're doing, whereas your deep learning algorithms, they don't really understand what they're doing, so they miss the context. 

One of my favourite little pictures is in Cambridge actually where the Google Street car was doing its pictures and it went past a field with some cows in it and it blurred out one of the faces of the cows to protect its privacy!  It didn't with all the other ones, but for this one it thought it was human.  No matter how much you throw at it, you're always going to get those mistakes, because it doesn't really understand what a face is in the same way that we do. 

I call it "chairness"; you walk into a room and you will be able to sit on chairs you've never seen in your life.  Even if they're not chairs, you can see a log on the ground, "Yeah, I can sit on that".  You can see a table; you evaluate how strong it is.  You go, "Yeah, I can sit on that", because you understand the concept of "chairness" and you're able to do that without fail, because you've learned it over your life through mistakes and success and failures.  It's an unsupervised learning; you've done it by yourself, you haven't had to have anybody else say, "That is a chair, you can't sit on that.  You can sit on this".  That's amazing and trying to train a computer to do that using the techniques we've got right now; you can get close, but you can't get perfect.

Peter McCormack: Aren't we thinking on too small a timeframe here?  I mentioned the example of my son where I'd have to arrive and meet somebody when I was his age which was 25 years ago, and in that time, we've gone from essentially no real mobile technology to having a supercomputer in our pocket where I can FaceTime somebody on the other side of the planet.

Toby Simpson: Isn't that wonderful?

Peter McCormack: Yes, it's wonderful and it's amazing and if you're thinking in terms of AI now, 25 years might be a short timeframe, but 200 years?  Isn't there always going to be some kind of risk if we create computers that think, that they will think in a different way than we expect?

Toby Simpson: Yeah, and we talk about we won't understand what's going on inside them.  Yeah, but we don't understand what's going inside ourselves.

Peter McCormack: Of course, yeah, but we're a risk to ourselves.

Toby Simpson: We are, and you walk up to a human being and you ask them something; you rely on them to be honest in order to get the answer.  With a computer that's intelligent, you're going to ask it a question and you're going to have to figure it all out from the answer.  Maybe we won't know what's going on inside it in the same way that we don't really know what's going on inside an individual's mind.  Again, I'm not particularly alarmed by that in any way, shape or form because we're creating something with its own goals, hopes, dreams, life and its ability to make decisions on its own behalf.

Peter McCormack: I guess the risk I see is that, if you create AI with human-style thinking, human-ability thinking but computational power to act and move quicker and make decisions across a global network, there are potentially unintended consequences.

Toby Simpson: Yes, there are potentially unintended consequences from that kind of thing, and I think it is right that people are taking this seriously.  I think it is right that people are talking about all of these issues and how we might protect ourselves about this, even though we're, in my view, still quite a distance away from getting it.  The fact that we're actually thinking about it significantly in advance under the basis that it will happen is a good thing; it's good for us to discuss these things, it's good for us to think about all the worst-case scenarios and we're really good at thinking of worst-case scenarios.  We may not get them all, but human beings, when it comes to imagining how something's going to go really, really badly wrong, we're right up there.

Peter McCormack: Cinema usually does it for us.

Toby Simpson: Exactly, indeed.  We're exposed to a lot of it and we're creative as well.

Peter McCormack: Of course, yeah.

Toby Simpson: One of the things that human beings do, every time you're walking down a street, you're constantly imagining different futures.

Peter McCormack: Sorry just to jump in, but I think one thing Black Mirror did very well was to extract empathy for digital beings in that there was the Star Trek episode where they were just digital representations of themselves; there was the episode where you could create a version of yourself which became your assistant and she didn't listen to the instructions, so he put her to sleep for six weeks, which was seconds for him, but she lived in an inner torment for that six weeks inside a container on her own.  So, I think in the episode Black Museum, we had the guy who was a digital representation of himself constantly sitting in the electric chair.

Toby Simpson: I remember that one.

Peter McCormack: But they extracted empathy for digital beings.  Do you envisage a scenario in the future whereby this will become a consideration?

Toby Simpson: That we'll make friends with digital things, we'll have relationships with digital things?

Peter McCormack: Yes.

Toby Simpson: Bound to happen, isn't it? 

Peter McCormack: Equal rights for digital?

Toby Simpson: Why not?  If they're a true intelligence, why wouldn't we do that?  It would be wrong not to, wouldn't it, because otherwise you'd be recreating slavery inside machines; that doesn't seem right.  If you've got something that's got its own hopes and dreams, its own ability to think and make decisions on its own behalf and it's conscious in the way that we would imagine that to be, then I think that's a natural extension of all of that.  I don't see that as being unreasonable. 

I think that us humans will be alarmed about it as we always are, and then there'll be a generation of people that will grow up taking it for granted and they'll have no worries about it at all.  We see that now; I see it now.  I'm getting older and I look at some of these new things and I think, "Really?"

Peter McCormack: How old are you?

Toby Simpson: I'm 48.

Peter McCormack: A bit older than me. 

Toby Simpson: I look at these things and I get worried about them, oh yeah, indeed.  Like AI, we're worrying about it, right.  And the internet, we're all worried about that; but actually, my kids, they've grown up with it and they take it for granted.  It's really no mystery to them and they expect it to be there.  They're not alarmed by it at all and I think that as they get older, they'll grow up with decentralised ledger technologies.  For them, this will all be normal in every way.  It will be inclusive; it allows them to do things and exchange value with people without having to get a third party's permission.  No matter where you are in the world or what situation you're in, you can take part in the global economy.  That's a really good thing.

Peter McCormack: Yeah.

Toby Simpson: I think that could help in a number of areas where currently people are excluded from the global economy.  All these things will happen, and they'll be the new generations of people who grow up taking it all for granted.  I suspect that there'll be a time when we're all saying, "Oh dear, AI …" but it'll be fine.

Peter McCormack: Before we finish, there's just a couple of other consequences I just want to get your opinion on, because this is absolutely fascinating; two specifically.  One thought came to mind, and I tweeted it out on the way: we are educating children and sending them to university to train for careers that AI will likely wipe out.  How does education catch up with the future?

Toby Simpson: Yeah, it's a tough one, isn't it?  Actually, it depends on what scale you look at.  If there's one thing that we know ever since we started creating things and inventing things, the scope of human jobs has changed.  That's just a natural result; it hasn't wiped out human jobs completely, it's changed them.  There's a transitionary period where some potential employment opportunities disappear and new ones appear. 

You think right now and who would have thought data scientists and AI machine-learning engineers and all of these things relating to space travel and quantum this and quantum that, these are jobs which simply didn't exist and now they do.  People were worried at one point in history about what would happen to everything related to horses and carriages, etc, when the horseless carriage came along, but actually all these new jobs were created in supporting and servicing all of that and putting down the infrastructure for them; this is just the way it is.

Education is very dear to my heart.  I personally believe that no civilisation ever went wrong when it improved the quality of education that it gives its citizens.  I think education is an enabler, it's a great equaliser; it gives people opportunities to do things.  When you provide people with a good education, you provide them with the ability to adapt, particularly if that is part of that process. 

We live in a world now where the idea that you would leave school, go into a job and stay in it for your entire life is frankly laughable; it just won't happen.  You're likely to change career several times and probably a couple of them will be to things that don't even exist today, so how on Earth do you prepare for that?  Well, you don't, but human beings are adaptable and none of us are born knowing how to use a computer or drive a car or get on an airplane.  We learn that stuff and that's amazing, it's an incredible human trait. 

So, the ability to adapt to all of these changes, I think we're really good at it.  It may not seem like it sometimes, but we are, at embracing these new technologies and looking for new opportunities and enjoying all these wonderful new careers and new opportunities that appear.

Peter McCormack: What about the risk of concentration of power and wealth?  One of the things that has come with advancements of technology is arguable monopolies within specific sectors.  There's an argument that in an automated driverless world, that Uber concentrates the revenue that originally would have been distributed amongst taxi drivers to a very small group of people.  Do we risk with AI seeing an even greater concentration of both power and money?

Toby Simpson: That's a very interesting question and I would argue that this is one way a decentralised ledger technology plays a key part, because it distributes all of that to all of the users without the centralised entities being there with this self-service trust.  If you think about what it does, it provides trust, it allows somebody who wants to go somewhere to be connected with someone who is able to take them somewhere. 

But in a potential future world, just hypothetically speaking, where you as an individual could verify that trust yourself without that third party being there, then wouldn't that be interesting because that's changed the role of the intermediary.  As you were saying, where a lot of that wealth that would have gone to the individual drivers get concentrated with them, that's suddenly not the case; it is a value exchange that takes place directly between the person who wanted the transport and the person who provided the transport. 

There is a paper trail of trust there on the ledger, so you can verify that this person is who they say they are, that the vehicle is correct because you can check all the digital signatures.  In a world like that, well that's different and that's the kind of thing that knocks down all these centralised silos and distributes that information amongst the users of the network to allow everybody to gain from that stuff.

So, I think that this sort of concentration of wealth and power and information was bound to happen as all this data increased, because the cost of actually doing anything with it and having the resources to doing anything with it was so astronomically high that it was reserved for the people who were able to raise the money in order to be able to do it.  But that's changing and decentralised ledger technology is one of those things that's changing it; it's inclusive, it's open to everybody if they can just figure out how to install the component parts, but we'll work on that. 

I think that's a good thing for society; it puts a lot more control over your stuff with you; how that information is released and who it's released to becomes your problem and that's nice because it's yours.

Peter McCormack: This has been an absolutely fascinating interview.  I've really, really enjoyed it.  I think I could talk to you for hours about this stuff, probably over a good bottle of wine as well.

Toby Simpson: Definitely.  Let's do that over a good bottle of wine!

Peter McCormack: Just to close out, because I think it's quite a long interview, can you close out just by telling everyone what's coming up for Fetch now and for yourself, how people can follow the project, keep in touch and who you want to hear from?

Toby Simpson: We want to hear from anybody and everyone who can create something for Fetch.  We're providing the world but we're not developing the agents; that's up to individuals to figure out.  We're providing the collective intelligence that enables the individual intelligences around the outside to do cool things.  We've got lots of ideas and we'll be creating a whole bunch of these to make it easier for people to do this and roll their own agents on mobile apps and things like that.  But that's nothing in comparison to what people are going to come up with on a network like this, so we're dead excited. 

People can visit our website; they can join us on Telegram.  All the information for all of that is on our website and they can get involved and ask us questions; we really want to be able to do this.  We're going to be setting up developer and community sites as well, and we're going to be providing people with the code and the support and everything else as well as the examples to be able to build stuff.  That's all going to be happening this year.  So yeah, come and help us put together the Fetch world.

Peter McCormack: I'll share all the links out in the show notes.  Thank you so much for your time.

Toby Simpson: Thank you very much.