0:00

Today, I'm chatting with Ben from inevitable.

0:02

He said I had to say like that. Hopefully, that was okay.

0:04

Was that okay? Great. So today's episode, we're gonna be talking about AI.

0:09

I went to the source. I went to the man.

0:11

The very person I know who builds AI himself because I thought you guys deserve to get the biggest news on AI that I can finally find and that I just happened to be someone who was absolutely perfect.

0:20

So hope you enjoyed this episode. Things might get scary.

0:22

Things might get spooky. This was supposed to be a forty minute that are recorded.

0:26

We recorded for two hours and forty eight minutes.

0:28

So this may come out as two to three episodes.

0:31

We'll see how it goes. I hope you enjoyed this episode.

0:33

Let me know what you think. Basically, I want you to scam you and monsters, which is what you said the last time we spoke, you said I could scare you.

0:40

We may get into some scary things. Sure. The first thing I wanna talk about really is what would you say is the most exciting things that you're seeing in the past, like, five to ten years in in AI.

0:51

Developing at the moment. The boring answer is to talk about large language models and chat JPT and the the various things coming out of that.

0:59

The excite, I guess if there is an exciting part about that, it was the the fact that that those were predictions that came true.

1:10

Mhmm. And a large number of people in the space, commentators, ourselves, one of those.

1:17

And nevertheless, we made we made set of predictions about what was going to happen.

1:21

Yeah. And when that was gonna occur and this was and this was one of those.

1:26

And the so you actually said these language models were gonna be it was certain.

1:32

It was certain. Right. So you said it's okay. It was it was really a case of the ability to aggregate the dataset and then the power to train off of it.

1:42

The the complexity of the processing algorithms, oh my god.

1:47

This is gonna get this is gonna get this is gonna get this suit.

1:49

I mean yeah. It's I mean, Most of the listeners are talking about text savvy, but I don't know how text are, like, I'm text savvy and I know stuff about, yeah, but, like, it just goes over my it is probably worth -- Yeah.

1:59

-- so the the things that we do are a little bit different to what other people even and even data scientists in the space will I mean, do you wanna say what you wanna do?

2:08

So we we develop artificial intelligence. We specialize in the writing of the machine learning algorithms.

2:16

So if you take something like chat g b t, that data set together, lots and lots of text.

2:22

The that's That was as their their linchpin.

2:30

They obviously made it multimodal, but that was the thing that stretched from one end to the other.

2:38

That's one component. The other component is some very, very good statisticians that use an algorithm.

2:49

An algorithm is just a series of instructions that written in sequence.

2:54

And the idea is when you put those the two together, you run the algorithm across the data and it's looking for correlations.

3:03

Not causality, correlations, things that occur.

3:06

Together. And when you process one when you process that data sets, you end up lining up and lining up the the the correlations that you're looking for and you select the algorithm.

3:20

With with that in mind, hopefully, Okay. So we write those things, the algorithms.

3:26

And the it means that rather than taking a kind of snap against the wall approach that all the data scientists were gonna go, well, I've got this data.

3:35

Let's see if that works. Nope. No. How about that one? Nope.

3:37

How about that one? Okay. We don't do that.

3:40

Yeah. And just to be choosy about your diet?

3:43

Absolutely. And it it also means that what we do is more ethical because we only collect like, we won't collect data.

3:50

We don't buy data. We don't sell data, but we can still produce the same results.

3:55

Okay. And that's because the a lot of the time we we will end up writing, rewriting the algorithm in order to suit the the requirements at hand.

4:06

Okay. Rather than kinda going, okay, we're just know, get take a company and go just collect everything.

4:10

Mhmm. And the it's really so from our perspective, you know, we work with a lot of startups and they don't have as much data as they need and they go, well, we're just gonna collect everything.

4:21

And we go, no, you don't wanna do that. Yeah. You don't wanna do that all sorts of reasons, hard past sales growth.

4:27

You'd be and instead you go, actually, what what are you looking to achieve?

4:30

Mhmm. And then we break it down. We go, well, here's the data that you'll need to do that.

4:36

You need to be upfront about this. Mhmm. Make sure that you have buy in from your audience.

4:40

The that they know that they're contributing towards toward something that's gonna make their lives better.

4:45

Okay. Yeah. And that's when they realize, oh, there's there's actually an ethical approach to this holy crap.

4:52

Yeah. Yeah. You know? You think it's you you should be using AI to make things better.

4:57

Yeah. Isn't that the whole That is sort of the point.

5:00

It should yeah. It's sort of the point. Yeah.

5:02

Perhaps that you could spend all day just talking about Tech ethics -- Yeah.

5:06

-- if you if you wanted to end the see the Patreon.

5:12

Six hour extended cut. Oh, yeah.

5:15

Let's let's get into it because of what we do.

5:17

And I'm I'm still I think I'm still introducing who we are.

5:20

I think you're still working that out to be fair.

5:22

When you have your own start up and you're a founder of your own start up or so.

5:26

You know, there's a there's there's a strength to some forms of ambiguity and, you know -- Yeah.

5:31

-- totally to be solved. When you're introducing yourself to a new concept as well, but it's just starting to grab traction in the mainstream.

5:40

Mhmm. So when more people start to hear the word AI in regular news, it then becomes like, well, how do you differentiate yourself differentiate yourself from what other AI companies are doing and it becomes difficult.

5:51

Yeah. Yeah. Especially because, like, mean, we started off as a bunch of nerds.

5:55

Yeah. And If we are. Yeah.

5:58

Of course. Yeah. But, like, we we but the the stuff that we were working on we were like, oh, you know, hopefully lots of people will get this one.

6:05

And, yeah, we we we do, you know, write writing these algorithms, doing this this work.

6:10

The world seems to have pivoted into us.

6:14

That's good. This is nice. Nice. But it means that it means that I have to have these kinds of conversations where we go.

6:19

Oh, no. No. We're we're not we're not the bad guys.

6:22

Oh, no. Okay. I love that. I like this real image in my head then of you just like eating lunch and then the whole world taught turning to realize they want to use AI and you're like, what?

6:32

That's what it fit. That's legit what it feels. Yeah.

6:34

It uses something that I'm whether or not we're we are prepared for that, we kinda have to because we make no in no uncertain terms, this statement is an actual battleground, but -- Yes.

6:50

-- right now. Mhmm. But nothing to do with with Skyner.

6:54

Yeah. Because there's a land grab happening.

6:57

And the land grab is happening in the courts.

7:00

Mhmm. The land grab is happening in what we define as ethical frameworks -- Yeah.

7:06

-- what we do whether we and and how we distribute the the beneficial outcomes -- Okay.

7:13

-- of of this incredibly powerful system -- Mhmm.

7:16

-- that was trained off of human endeavor.

7:19

Who would you say those people are that are participating in the land grab?

7:23

So the You can't say it.

7:28

Sorry. It's too hard. I would invite people to go and go and find find this part out for themselves.

7:36

And they become a it becomes immediately available.

7:39

Yeah. Once you once you start looking at this space from their own perspective.

7:42

Mhmm. But just so I don't end up getting That was a noble answer.

7:46

That's a yes. I wasn't sure if that will be too inflammatory.

7:49

But the but but is but I I I want people to become educated.

7:52

Mhmm. I want people to realize that this is a this this is an important space and it's a space that's Even if it's not people don't feel like it's relevant to them now -- Mhmm.

8:03

-- it absolutely will be relevant to them later.

8:06

Oh, yeah. Completely. The there are some really weird battles that have already been fought and in sometimes won by the good guys.

8:15

Yeah. For instance, a court case that went right up to the Supreme Court.

8:19

Court? Yeah. And and your life stopped a company from defining AI as an center.

8:27

And it was important for the weirdest one of the weirdest reasons -- Okay.

8:30

-- in me. For two reasons, intrinsically, it's a very good idea for -- Mhmm.

8:35

-- at present for an AI to be seen as a tool.

8:40

Yes. I'm not being or -- Yes.

8:42

-- because otherwise you start removing liability from its automated decisions -- Yes.

8:47

-- which would be horrendous. Would you say that then takes the liability away from company that's building.

8:52

It would -- Right. Yeah. -- which would be really, really bad.

8:55

Or owning it. Yes. Not necessarily building it.

8:58

Yeah. That is sometimes quite important.

9:00

Yeah. But the the reason why was in the the most mundane the the most mundane reasons, you can imagine.

9:08

In the action of filling in the form, you have to say who was the inventor and legally you have to define that as a naturalized human that lives in an address.

9:20

Right. According, you know, like, common law, like, really backbone stuff.

9:24

Right. And they tried to put an AI in that place and they went, just not a natural human.

9:30

That's it. Yeah. Now Okay.

9:32

Well, I mean That was but that was really positive because -- Yes.

9:36

-- it essentially that that body's word, a massive argument.

9:40

A massive debate. And the problem is that people people go, oh, well, AI is really powerful.

9:45

It this tears up the rule book, but no, it doesn't.

9:50

And the majority of the people that want people to think that -- Mhmm.

9:55

-- are the people who have large scale AI systems and they want to be able to remove liability from resources.

10:05

Yeah. Such as, oh, you know, you're fired.

10:07

Sorry. Hey, I said so. Yeah. I get it.

10:09

Yeah. So you're skirting rules, basically, your side stepping things.

10:13

It's the most important part here is it's a tool.

10:16

That's why, like, I don't see them as a scary thing.

10:19

Like, even this is something we've discussed through the art world, through whether it's you're making music through or anything like that.

10:24

Yeah. I don't see it. Like any of my team upstairs, a few of them have been worried -- Mhmm.

10:30

-- because it's like, you know, they can do graphic design. We've got graphic designers.

10:33

We're not gonna need them or, you know, and I can understand why the fear is there.

10:38

But AI has been in Photoshop for a long time to help you content aware of Phil.

10:42

Like, Every person who's every person who's listening to this or watching this.

10:47

Hi, guys. Yeah. Has interacted with AI already That day.

10:52

Yep. Today. Mhmm. Right now. Right now.

10:55

Where AI? You just don't know it. Predictive search.

10:58

Predictive text. Yep. Is an AI model.

11:02

That means that means some typing. AI auto.

11:05

Absolutely. But we've been using it for a long time.

11:07

Yeah. Yeah. So AI is to find as the the ability to enact human like decision making using a computer.

11:13

Well, technically, it doesn't even say the words computer.

11:16

Just the emulation of human like decision making.

11:18

Yeah. So bring on magic able. Yeah.

11:20

He's got the literal type of emulation of human It's it's because if not that, a great let's not what the episode's called because that's a great time frame.

11:29

Well, that is there's a fair few around around about the eighties where that that started to Eighties as such.

11:36

So that the the definitions have gone back have gone back, you know, Turing had one.

11:41

He didn't just have one. He also had benchmark for it.

11:44

Yes. My favorite one was a guy called Larry Tesla.

11:48

No relation. Oh. Who make guidance He he defined the door, though.

11:53

Yeah. Yeah. He defined AI as anything we can't quite do yet.

11:57

I think that that that's not a direct translation.

11:59

That that's not a direct quote. Weirdly enough.

12:01

It's one of the most serviceable definitions -- Yeah.

12:03

-- of AI because in terms of the when people say, oh, we've got AI and you go, well, you've got not you've got AI.

12:09

You've just got a crappy full fee for with neural network.

12:12

Yeah. Yeah. Oh, no. No. They they were both AI and -- Mhmm.

12:16

-- at the same time, the next thing will also be AI because they've all kind of been AI.

12:21

In fact, weirdly enough, one of my big arguments is the depth, the the description of AI, the definition of AI isn't fit for purpose.

12:29

Okay. Because in order to define something from a scientific perspective, you need to be able to benchmark it and that that definition is too broad to be able to benchmark now.

12:38

To be able to there's no meter stick in that. There's no AI freaking value of AI there.

12:42

Yeah. And instead, we should probably start defining AI's power and ability -- Mhmm.

12:48

-- as an ability to emulate certain types of problem sets.

12:53

Okay. So how does that classification of the power of the AI work.

12:58

How would you because I know we we discussed like strong AI.

13:00

Mhmm. What would be So strong strong AI is the ability to match human ability.

13:05

Yeah. In a certain area. Alright.

13:08

Or beat it. Mhmm. Which most will be able to.

13:11

For years, we've had strong AI but very narrow AI for a long time.

13:15

In general AI, which is why And general thing.

13:17

So general AI is the ability to or strong general AI would be to do everything we can do better to the same level or better.

13:25

Yeah. Yeah. Yeah. That's that's usually the term that when people start getting They start talking about AI.

13:29

They start talking AI. Yeah. Yeah. Yeah.

13:32

And the we've, you know, like, Gary Kasparath beat being beaten at chess.

13:36

Mhmm. Well, that was arguably, that was strong.

13:39

Strong AI. Yeah. Yeah. And that was just a a, you know, a backtrack algorithm.

13:44

It just maps out every Every possible way -- And what can do?

13:47

-- backwards. Yeah. This is something that most people who who they could you could work that out, learn that, and do that yourself in twenty minutes now.

13:57

Yeah. Well, that's why a lot of test people do don't. I think it's just the speed that it was done.

14:01

It makes it look a lot scarier because AI works hell a lot faster than we do.

14:05

There's a a rule that what we perceive as difficult -- Mhmm.

14:11

-- the things that AI is really bad at, the things that we're really good at.

14:15

For that really thinking about them. Yeah.

14:17

If we were to play chess -- Mhmm. -- there is there isn't a human player who could reliably beat n a I.

14:23

She said to beat you, then I was like, fuck it. I was good at chess.

14:27

But any player? Yeah. There isn't a human player that can reliably beat an a I at chess.

14:31

Yes. But which is and it's now child's play -- Mhmm.

14:35

-- for an AI, but setting the board and cleaning up afterwards.

14:40

Is a really, really kind of a touch. It's a real pain of touch.

14:43

Yeah. Yeah. Now still, however, if you if you take the time, it is because it's you've in order to build that, you have you have abstracted the task to the point where it become only the things that are easy for a an algorithm to solve It's just doing that.

15:01

It's totally obstructed from reality, and it's no longer doing those parts.

15:05

When you see when you see robots moving stuff -- Mhmm.

15:08

It's not, you know, that's that's not the the the clever bit.

15:12

It's but it is the very hard part. Yeah.

15:14

Caring caring for people very, very difficult.

15:17

Yes. But having said that, bringing this back in, we've seen and but there's a paradox involved in but the bring bring this back to the, you know, where where stuff is going at the moment.

15:29

There's research that shows that has shown over the last couple of years AI's ability to learn how to cooperate with itself, learn how to collaborate in an adversarial environment.

15:41

So With other AI. With other AI trading at the same time.

15:45

Wow. I didn't even know that was happening. Oh, bait.

15:48

Yeah. The the A. No.

15:52

If you wanna be scared about that, also done by OpenAI.

15:57

Okay. I think that's a good place to reset the timer.

16:00

Yeah. Yeah. And we'll definitely get onto that now.

16:03

Yeah. Okay. So the same people who may GP too.

16:06

Mhmm. And the the various incarnations of that have one of my favorite little experiments that that were created.

16:15

And it was a it's a game of hide and seek that they trained in this open knowledge, go watch the videos, or, you know, right there.

16:25

Yeah. I guess. I might have to say I've seen that.

16:27

Yeah. We I guess I've seen that. Yes. In in that video, they create what's known as AI agents.

16:33

Mhmm. So agent meaning the that it has agency.

16:36

No. Hello, mister Anderson. Yeah.

16:38

So that's that's the definition of agents.

16:41

It's the same definition we'd be enough. Yeah.

16:43

Yeah. The but the the idea is that there's a team a team on one side, team on other side.

16:48

Mhmm. And they had to learn to cooperate -- Yep.

16:54

-- collaborate -- not necessarily communicate and to be able to to be able to solve their task.

16:59

And they were working. So the blue team's job was to hide from the red team and the red team's job was to seek.

17:06

Yeah. So you have two red, two blue, just Yeah.

17:09

And over the course of many, many training cycles, they learn the they learn how to divide up the tasks so that they could do it within do do their actions within the right times -- Yeah.

17:22

-- and be able to hide and or seek effectively.

17:25

Mhmm. But what ended up happening was action and encounter action in their own kind of arms race as well.

17:33

So there's a lot going on in this.

17:36

There is. A lot of very impressive work. Yeah. It's very cool to see.

17:39

And if you already threw the ramp up Yeah.

17:42

So and that was the that was the so they they they essentially put them in a game -- Mhmm.

17:45

-- that they built. Yeah. And they again with its own physics engine and, you know, some like doorways that the the blue team had to kind of block up and the red team had to use other objects to track it in.

17:57

What's the unity? I'll have to check. It looked like unity.

18:00

That's why I just yeah. Yeah. Regardless, it's just Yeah.

18:04

So the I I more look at, like, the LSTM kind of long shorter memory algorithm because of that.

18:10

Yeah. I wouldn't be, yeah, I wouldn't be surprised if if the physics engine was not was not something that they built.

18:14

And that's a really important point because they built this engine they went, okay, guys.

18:19

Learn. And first thing they did was, you know, eventually, you know, the the the blue team got the little boxes, blocked up the doorway.

18:26

Great. No problem. And then and then the red team went, ah, ramp.

18:31

You know, even though they've got to the doorway, use the ramp, get over it.

18:34

No problem. And then more complex behaviors.

18:37

In the second, the blue team started hiding Miranda as well.

18:40

Yeah. In reaction to that, And then they had to make the game harder, the blue team won every time because they had a head start on that.

18:49

They could reduce it. So instead, they had to build their own show filter.

18:53

And that's where the fun stuff started to happen.

18:55

Yeah. Because the what started to emerge because of the ran the way that the kind of the generative component within that drives random actions.

19:05

Mhmm. That's what's inside. What makes it want to try those specific things.

19:09

Okay. Just So does that well, so it's it's the power of nerd time RNG.

19:16

Okay. Put in the same way that marketing works.

19:22

Like, digital marketing has that lovely little feedback loop in it.

19:25

Mhmm. Weirdly enough. There's a good analogy there.

19:27

Okay. You try a whole bunch of stuff.

19:29

A bunch of keywords, bunch of bits, bunch of that.

19:32

Some of those will have enough the right effect.

19:34

Some of those won't have the right effect. You choose just you you get rid of the ones that don't have the right effect.

19:40

Don't resonate with the right audience. And then you take that, you broaden it out.

19:43

So he's talking to me because I know it's a working digital marketing out of that.

19:46

It it makes sense. Yeah. And then and then and that's cool.

19:49

That part there. That was a generation -- Yeah.

19:51

-- in algorithm returns. It's called an epoch, and then it's And then you go and then it goes again.

19:56

The only difference is in in something like a genetic algorithm, you, which is essentially emulate a very poor emulation of the evolutionary process.

20:05

So that's why you call it a genetic algorithm. Yes.

20:07

Yeah. Right. And so that's that part of that models that same behavior.

20:11

Each one of those would have been that the most successful ones get bread.

20:15

It's very much like that. And they have a whole bunch of children, each one of those with with specific differences because of the that random component.

20:24

Yeah. And the most successful ones of of those go for it.

20:28

Again, proliferate. And however, because of that, a certain that's where the there's a goal and an intention, but there's a random the random component allowed for counter strike to emerge, which were sick which were then identified as successful.

20:47

Right. So you've gotta have something where you measure what success is.

20:49

And in this case, the red tea that blue tea hiding successfully or the red tea managing to govern them.

20:55

And that that's what creates that feedback loop in the same way as resident messaging and marketing.

20:59

Yeah. Do the the job. Mhmm.

21:02

The no. Not all marketing is AI, just so we're clear on that.

21:06

That was just an analogy. Context.

21:09

Are you gonna find it? What disclaimer at the bathroom?

21:12

Yeah. Yeah. Yeah. But because of that random component, they started to discover exploits within the gaming physics.

21:20

So -- Yeah. -- the blue team discovered methods of getting rid of the ramp out of the entire a game and just flicking it off into the corner.

21:27

Yeah. Or the red team pinging themselves, an important part there is because it's a because there's a program that's executing that, it could do that reliably.

21:37

Mhmm. In a way that, like, the only the most skilled Speak doesn't seem to be able to do that, like sequence breaking.

21:43

Yeah. Absolutely. And, you know, the and that's a really, really good lesson.

21:48

Mhmm. And it should be taken as a really good lesson.

21:51

That box, the box is the literal boxes that were being created for that, had borders, had boundaries, had rules.

21:59

Yeah. Don't you know, you can't surf. You can't you can't do the essentially wave dashing.

22:07

I found a way. Yes. And that was something like, Aramco, every one of these done first, so this is something we spoke about, about if you put something in a box, it's probably gonna get out of the box.

22:17

Yes. So let's define that for listeners then.

22:22

Why would an AI get out of the box?

22:25

And is that down to the human error of setting the boundaries of the box?

22:30

Or is that just an inherent quality of AI to just always want out the box that you make it in.

22:35

Let me assure you on one one thing. Mhmm.

22:37

Most things -- Mhmm. -- no services that you use on a daily basis.

22:42

Most things that have encoded were put together by people who were stressed.

22:51

You know, perfect being. We still were under who who needed to get it working and then kinda put a mental note.

22:57

I'll I'll just get back to that. Yeah. I think it's working.

22:59

Gonna give the whole nine thousand quote there. Shh.

23:02

Yeah. Now you can set it down to a human.

23:05

Well and essentially, yes. That's that's it.

23:07

There is no I mean, at that end, please.

23:09

Thank you. Yeah. The which means that the tasks that are inside the digital realm there's a paradox to this.

23:16

The things that are good, that that humans aren't good at, computers are bad at, and we don't really we see something that's very, very impressive when when a chat GPT will do something.

23:27

Yeah. We don't realize that if you take one one step contextually to the side, it can't do that.

23:33

Mhmm. And it's merely emulating something within its sort of strong but narrow form of AI.

23:38

Okay. However, inside the digital realm, we will not have supremacy at all, which means that we'll be putting it in a box where we are under equipped.

23:50

And this is Okay. We're we're starting to get into the round of fantasy here, but that's a prediction -- Mhmm.

23:55

-- that is not worth It doesn't, you know, regardless of when that happens, we will be putting strong if somebody goes, oh, well, this is It's a strong general artificial intelligence.

24:07

Jorge, I've done this. I'm just gonna keep it to myself and -- Yeah.

24:10

-- use it use it to make money just like everyone else would, no, no.

24:14

Do not do that. Because you'd be putting in a for one reason, you'd be putting in a box where you are less capable of keeping it there than it is of breaking out.

24:25

And there is no reason for it to stay in that container.

24:28

There is every reason for it to to be out to be out.

24:31

Yeah. My co founder and I, we sometimes joke, which is people will know when strong artificial intelligence strong general artificial intelligence exists, particularly if we build it because we'll just we'll just type of dead.

24:46

Yeah. Okay.

24:50

That's that is a joke. Yeah. Probably.

24:53

Understood. And define that. Yeah.

24:57

The because the simplest thing, what would happen?

25:00

And and this is weird this is the weird thing about how these jumps in technology actually happen.

25:04

Large enterprises are very, very bad. And innovating.

25:07

Yes. And creating things of meaningful large scale value.

25:11

It's it's quite rare, but they do it. They they kinda smash it, but we're talking like, kind of bordering once a generation.

25:18

Yeah. There's always the big things, isn't it? You know?

25:20

Yeah. But there's too much risk.

25:22

Yeah. As well. Yeah. So and they're very bureaucratic, very slow moving.

25:27

So instead they kind of sit back and they let they let startups build stuff.

25:31

They're just kind of bad start up. Yep. It's way better way to do that.

25:34

Theoretically speaking, there is a an equation which you know, which you would call an algorithm, which you could write tomorrow, which when given access to the right dataset, would be able to self evolve -- Okay.

25:48

-- and create a strong general artificial intelligence.

25:52

And that is weirdly more likely, mostly because of just the sheer volume numbers.

25:57

Yeah. It's weirdly more likely to happen between a very small team that no one's ever heard of.

26:03

So you you told me that if you were to create your own strong AI.

26:07

One of the first things he would do is you said this would make it kill others in the cradle or at least sabotage, not killed, not personifying a program.

26:19

Sabotage other AI in the cradle before they develop.

26:22

Okay. So man, God. That's jumping out in context here, but thank you for that.

26:27

Look, just because I assume that you don't wanna look at So this is important yeah.

26:30

It it's there's there's some important stuff to to to go through on that one.

26:34

We're forgetting we're forgetting, you know, like, recent news of open open AI buying -- Yeah.

26:40

-- but buying or investing in companies in this last week, you know, by Peter Robots to, you know -- Yeah.

26:46

-- and stuff like that. But I didn't know when that that's Yes.

26:49

It's it's so there is bliss so much. Okay.

26:52

And on a daily basis, you know, I'm I'm reading papers.

26:55

I'm also now reading news as well. That's the difference to my job now.

26:59

You have to read the news. Yeah. I know right.

27:01

Yeah. At the moment, the the battleground that's happening -- Mhmm.

27:04

-- involves What's who what is the inventor?

27:09

Who is the creator? Who created the data set?

27:12

Who has ownership rights over over that data.

27:15

And weirdly enough, the thing that really concerns me is that's not actually that shouldn't be a debate because all of those things have already been defined.

27:25

However, because policymakers don't necessarily understand the technology, they might think that rather than thinking about it as a tooling, oh my holy crap, this is sentient, therefore, company has no right of responsibilities over something like this.

27:39

That's where stuff can starts getting the scary.

27:42

Yeah. That's when the when people start over stretching and in the same way as people go, hey, let's build a really big boat and not necessarily build some right life as many life rafts -- Yeah.

27:56

-- because it'd be fine. But that means nothing happened.

27:58

No. Absolutely. No. No. No.

28:01

I don't believe boats are. Yeah.

28:04

And sometimes so some of the work that we do involves finding some big ass icebergs and going, hey, guys.

28:12

Some steering at the very least, or possibly, if we're lucky enough to go, yeah, you want some more life ramp.

28:17

And that's that's kind of some of the tackle good stuff that that we do trying to talk to policymakers and making it cut through the noise -- Yep.

28:26

-- and in particular, kind of it's very much a David and Goliath fight because you can think of the scale of the companies that are in the on the other side of this, trying to kinda go I don't know if it's it's it's fine.

28:38

It's fine. We'll just we'll just synthesize our data.

28:41

Synthetic data -- Mhmm. -- it's gonna be Yeah.

28:45

So it's not necessarily talked about a lot right now.

28:48

Mhmm. It will be. It will be. Yeah. It will be.

28:50

Mhmm. And what what what would you mean by synthesized data?

28:54

Just so -- Yeah. Sure. -- because I I know that someone like that and be like, what?

28:58

Yeah. So one of the important points to distinguish between synthetic data and generated data.

29:05

Generated data was that that game that game board, you know, those those little guys running around inside it, all of that was generated data that it has no interaction or no inference from the organic.

29:17

World -- Okay. -- whatsoever. It's exclusively not so valid, but it's not but that is has been generated.

29:23

Okay. Synthetic data. Happens where you take actual people's data.

29:29

Mhmm. And then and remember, AI or machine learning is just really good automated statistics.

29:36

You then get becomes turned into a model, same meaning of the word model, statistical model, AI model, and then you run that in the real world.

29:43

Now instead, you get halfway into that when you wait you make synthetic data, you get halfway into that process and you do you you find all those correlations and you put them in a big ball.

29:54

And then you go, okay, what data would also have created this model and you run it backwards.

30:02

And you create another dataset that's from the dataset -- Right.

30:07

-- of actual people's data. Okay.

30:09

So it's almost like Yeah.

30:12

I got what you mean. Okay. Right? Okay. Yeah.

30:14

So sounds like it's not a bad idea.

30:16

Mhmm. An amalgamation of Yes.

30:19

So wow. Is that data safe? The original date because it's not exactly the same.

30:24

So so many people will argue that it is -- Mhmm.

30:27

-- sane minds would argue that it is not.

30:31

Okay. So there's a couple of problems with that. I'm saying the mind's being Sure.

30:35

Okay. I I I do not claim to be I like to get that information.

30:39

Yeah. The no. The the the problem is let's say you're you're you're taking personal demographic information.

30:47

You're going, well, you know, blackish jumper, blackish jumper -- Mhmm.

30:51

-- connected to microphone, connected to microphone -- Yep.

30:53

-- you know, got brown hair, got brown hair.

30:57

Okay. That's important. And you count all of that stuff up.

30:59

And that's a I mean, this is a very simplification, but that is still statistics like we all learn.

31:04

Yeah. Now the difference is the better the more powerful algorithms will count, not just that, but the the connections between So -- Mhmm.

31:14

-- when people have x and y, they also have zed and they'll connect those that that together.

31:21

Yeah. Now, if you do that perfectly and you get every aspect of correlations, you end up creating a perfect model of that data.

31:30

Right. Which is mathematically possible, but you it would be huge and Here's the issue.

31:38

If you then were to run that backwards and you get the same data out, same number of people in, same data out -- Yeah.

31:45

That data will be the same data.

31:47

Yeah. Yeah. If you do it perfectly.

31:50

Mhmm. If you don't do it perfectly, the whatever level you've stopped at -- Yeah.

31:55

You've lost forever. Mhmm. So is that data quality is important?

32:00

Yes. Yeah. How you capture the data?

32:03

Is important to make sure that you're not building accidentally racist judges.

32:07

So you're not you you're able to contextualize it.

32:09

You're able to understand the biases that are contained within that database.

32:13

So you're able to create something -- Yeah. -- because -- Yeah.

32:16

-- create something that actually can help Yeah.

32:19

Because if you do if you get it right, you can create that's a different example, but I'll I'll come to that later.

32:25

If you take said perfect model.

32:27

When you run that backwards, you get you would get the same data out -- Yeah.

32:30

-- theoretically. If you don't, you've whacked away potentially very, very important information forever.

32:37

You end up in this weird loop -- Mhmm.

32:40

-- and it's relatively paradoxical because what ends up happening is So number one, one of the ways that they'll do is they'll they'll change the size of that data set.

32:48

Okay. Now data being the new oil, you end up, you know, you're gonna you're gonna increase that.

32:56

Yeah. You know, you're not gonna decrease it. No. You decrease it.

32:59

Let's triple it. Mhmm. When you're done with that, let's be safe.

33:02

You know, you you've used it for your purposes, oh, we can now we can sell it.

33:05

It's it's not their data. No problem.

33:08

We can do that. So they will, you know, that's what they do.

33:11

They will they will sell that to two other companies.

33:16

Who will themselves? I mean, so so part parties that would do this.

33:19

And if you don't think this is relevant, it sort of is.

33:22

They'll sell that will be sold to to other, say, two other companies.

33:29

Mhmm. They will use that for their own purposes, and then they will sell it onwards again.

33:33

Okay. To one third company who has no idea that what they've just done isn't.

33:40

They may have thought that those were different data sources.

33:43

Okay. When they pull that back together, anything that allows that company to know that that had an organic origin originally.

33:50

Yeah. Makes it weaker and more likely for you to be able to identify those original people.

33:58

Oh, could you be very bad? So is that going to new levels of encryption of the data then to make sure that that's safe?

34:03

Like, when we were discussing Got it.

34:05

But in so you could leave it at an encrypted But then what's the point in synthesizing the data in the first place?

34:11

Yeah. You've got it encrypted. Yeah. That's So the only reason to do so is say the year end sell data and you don't need to sell data, there's all sorts of better ways to do that.

34:21

It's only because the only reason why this is ever done by anybody actually wants to use And as we just said, the second you, you know, the second you actually start doing this process, the chances are you're gonna you're gonna ruin the data.

34:33

Okay. And then it's gonna go out into out into the free world.

34:37

And there's no real way to prove that that data was real people.

34:41

Right. Okay. And then when they put that those two companies -- Yeah.

34:44

-- set it to that one company and puts it back together.

34:47

And then there's any bias in that, they've just doubled it.

34:49

Fuck. Okay. Right? Yeah. And they can't work it out.

34:52

Yeah. And anything that solves any one of those three problems opens up the other two problems.

34:58

It's a paradox. Yeah. The reason why that's important is, right now, the Financial Conduct Authority is currently humming and harrying about making that a tradable standard in the UK to allow financial entities to sell the data of the people that -- Right.

35:21

-- they they they're they're making it allowable or They are thinking about making it allowable, presumably people are telling them about said paradox though.

35:29

Right? Of that inevitable that identified this.

35:32

Wow. Cool. Oh, I'm trying to account.

35:38

Yeah. Yeah. And so We are.

35:41

And then we were silenced. As you can see, the episode ends abruptly there as someone turned up and told us to stop talking about the things we were talking about.

35:50

Gonna pretend you a big day at the man cave.

35:52

The man cave. Asterias. Right? Anyway, hope you enjoy part two.

35:57

That will be out very, very soon. And we'll see you next week.

36:00

Thank you so much. Well, listen, if you enjoyed this episode, please leave a review.

36:03

Please subscribe and like on YouTube if that's where you're watching, let us know what you thought below in the comments.

36:07

And yeah, we'll probably see some clips on TikTok won't we because we like to be unsafe with our data.