Episode 5: Evolving AI solutions in business - Transcript

Available on: 

Soundcloud

Spotify

Apple Podcasts

TRANSCRIPT

Episode 5

Transcript edited for clarity and context

00:00:13:22 – 00:01:24:06

Marina Santilli

Hello and welcome to our fifth episode of the podcast series from UCL Business Big Talks on Big Impacts, UCL Business (or UCLB) is the commercialization company for UCL, and this yearlong podcast series celebrates the company’s 30 years of collaboration and impact.

I’m Marina Santilli, associate director of Physical Sciences and Engineering at UCLB and today I’m looking forward to discussing my topic AI in Business with my guest, Dr. Daniel Hulme. Daniel is founder and CEO of UCL’s very first AI spin out company, Satalia, which was incorporated in 2008, just as he was wrapping up his doctorate.

Satalia has grown impressively in size and impact over these past 15 years and became part of the WPP Group of global companies in 2021. Daniel also holds a position of Chief AI Officer at WPP.

Welcome, Daniel.

Dr Daniel Hulme

Hi Marina.

Marina Santilli

So Satalia has been a significant success story for UCLB, UCL, and in particular the computer science department where you gained your undergrad, master’s and doctorate, all in AI. Can you give us a brief overview of the company and what it does?

00:01:24:21 – 00:02:12:15

Dr Daniel Hulme

Satalia combines two fields in computer science. I guess, the first field is traditional operations, research optimisation, discrete mathematics to solve computationally difficult problems, and also a combined start with machine learning, deep learning, data science, is obviously technologies that are very good in extracting insights from data.  By combining these two types of technologies, you can really move the needle in solving problems for organisations.

Marina Santilli

These are enterprise specific solutions you tend to build.

Dr Daniel Hulme

We tend to build enterprise solutions, innovations that problems have never been solved before. Or if you move the needle significantly, it has a massive impact on the business. And what we try to do then is build assets of repeatable products that then can be licensed and taken to other industries.

Marina Santilli

Right.  So anyway, you and I have recorded many fireside chats over the past years. I’m particularly looking forward to this one, as I know that solving problems in business using AI tools is something that you’ve dedicated your working life to. Also, you’ve witnessed first-hand how attitudes to integrating AI into business have evolved since deep learning came of age around ten years ago. I hope you have plenty of thoughts to share with us on this.

But anyway, before jumping to the current day, let’s go back to the mid-2000s when you first started working on the idea for Satalia. Notably, this was several years before the impressive breakthroughs in convolutional neural nets jumped out of computer science labs to give a step change improvement in image and speech recognition, and certainly before businesses had any idea of the revolution that approach would bring. What was your thinking at that time regarding the extent of the commercial opportunity for AI algorithms?

00:03:05:03 – 00:05:15:24

Dr Daniel Hulme

 

When I joined UCL during my undergraduate in AI back then, and my master’s also in Intelligence, I was fascinated by the idea of building intelligent machines without really thinking about its impact on industry. I started my Ph.D. and tried to model the brain of a bumblebee. Bumblebees have a million brain cells. Bumblebees can do amazing things. They have very, very similar ways of looking at the world as human beings. And if we could understand how bumblebees see the world, then we could use those insights in how human being see the world.

Back then, I also had the opportunity to go to London Business School to do MBA electives. And it was really then, that I saw the opportunity of applying algorithms in industry to make a big difference. From my experience working with industry, there’s a huge amount of inefficiencies and opportunities to make things better. These algorithms that I was interested in during my PhD could really help solve some of these problems.

Actually, at the time I was very interested in neural networks because obviously a bumblebee brain is a neural network. And the question for my PhD was “Can you model a million brain cells in a machine?”  Twenty years ago,  it was impossible.

At that time, I was exploring with ideas about how you could prioritise the training of these networks. One of the bottlenecks in building big brains was the ability to train these known networks quickly. I was actually looking at trying to implement some of these algorithms in an FPGA and CPUs.

It just so happens that that’s what Jeff Hinton was doing in Stanford (I don’t remember where) so they managed to really achieve the prioritization of the training of networks, which now allows us to build very large brains.

But I was a terrible programmer. I was terrible at building software.

I understood that the computational complexity around the training of the neural networks was a problem if we wanted to build big brains. I got very interested in the fields of computational complexity and optimisation of algorithms. That’s where I spotted the opportunity to bring these two types of technologies together.

00:05:16:19 – 00:05:20:06

Marina Santilli

And where did you think those first opportunities might be?

00:05:20:16 – 00:05:58:23

Dr Daniel Hulme

In London Business School, we pitched an idea based on the application of these algorithms and the obvious industries cropped up, you know, the airline industry, any industry that had large scale logistics problems. These supply chain problems are traditionally optimisation problems. Again, if you can move the needle by two, three or four percent, it usually translates to millions, tens of millions, even hundreds of millions of dollars of savings for organisations. It’s a logistics type problems are ripe for optimisation. And again, you can enhance those solutions by using machine learning.

00:05:59:13 – 00:06:14:22

Marina Santilli

And in those days, obviously, there were optimisation algorithms around – how widely were they actually being used in industry, and was it a hard sell to encourage businesses to think about adopting them into their workflows and systems?

00:06:14:23 – 00:07:00:17

Dr Daniel Hulme

I heard a phrase recently, which is, if you’re one step ahead, you’re a pioneer. If you’re two steps

ahead, you’re a martyr. And I think for the past 15 years, I’ve been trying to kind of educate business leaders that the problems they have, are not insight problems. Actually, the problems they have, are decision problems.

It’s very likely it’s one of these computationally difficult decision problems that they’re having to deal with in that business.  Traditionally, human beings who are trying to solve these problems are solving them badly. I mean, I’d love to geek out around the mathematics here, but if you apply the wrong algorithm to solving these problems, it will take longer than the end of the universe to solve them. If you apply AI algorithms, it’ll take milliseconds.

I think the battleground now for organisations, is the realization if they want to win, they need to have the best algorithm to solve these problems.

00:07:01:02 – 00:07:08:13

Marina Santilli

Have you seen that message change, over the past 15 years, quite significantly, or is it still a challenge?

00:07:08:13 – 00:08:38:02

Dr Daniel Hulme

Yeah, I think what was interesting is, in the early days, people weren’t really thinking about computational algorithms because we hadn’t made a huge amount of progress in academia

in solving some of these problems. Neural networks were still nascent and there hadn’t been much research being pioneered in traditional operations research.

So companies at the time were looking at innovations on how can you unlock the creative capacity of people to do more interesting things rather than looking at technology. And then what happened

is that organisations realized to be able to ask questions of their business, to be able to then answer those questions, they needed to bring data together, which was then, the birth of big data. They realized that they needed to bring data together to extract insights for human beings then to go and use those insights to drive the business forwards.

But then the realization was that human beings are actually not so good at making decisions.  What we need to do is use algorithms to replace a lot of what decision making happened with human beings. We’re moving from big data in the 2010s – 2015 to now using algorithms to extract the insights from data.

Now I would argue that’s still not the right investment of time and energy. I think that companies don’t have insight problems, they have decision problems. And over the next five years, organisations are going to realize that they need to be applying optimisation algorithms first.

00:08:39:13 – 00:09:09:00

Marina Santilli

So  as you say, during 2010-2015, things are beginning to change a little bit.

And data science obviously depends on collection of data. So was the collection of data an issue at that time? Was that an initial hurdle for you trying to sell a decision-making product into a company that potentially hadn’t collated much data yet?

00:09:09:01 – 00:11:27:05

Dr Daniel Hulme

Yeah, there was a few hurdles and I think that there’s this kind of handful of things that have come together over the past ten years that now allow us to do really interesting things.

I now look at A.I. not through definitions (and we can talk about some definitions later if you’d like). I now look at AI through the emergence of qualities, of algorithms, of data, of converged to allow us to now do really interesting things in industry.

I think that what happened in the 2010s was that we were able to start to get access to large data. The cloud obviously came, which enable people to then process that data. And then there were some advances in in algorithms in neural networks that allowed us to now do interesting things. And of course, what happened then is, we realized that we don’t have enough skills, people with the talent, to be able to then utilize those things. There was now a push from academia to make sure that we are educating the next generation to utilize these technologies.

Back then, I guess none of this would be called AI. It was called data science or business analytics. I think that there’s still a bit of a misconception around AI. My master’s and Ph.D. were on AI and the current popular but weak definition of AI is getting computers to do things that humans can do and actually humans are bounded; we are limited by our abilities in many ways. So benchmarking machines against humans is a very silly thing to do.

Humans are good at solving problems, of finding patterns in about four dimensions. But computers can solve problems with thousands of moving parts, and they can find patterns in thousands of dimensions. So, using human beings as the benchmark for intelligence is not the right thing to do. I think that’s what people have been doing over the past decade.

There’s a much better definition of AI, actually, that comes from a definition of intelligence, which is

which is goal directed adaptive behaviour. What you want to do is ultimately build systems that can make decisions, learn about whether those decisions are good or bad, about themselves;  so that next time they can make better decisions.

If I’m being totally honest, over the past ten years, we don’t see adaptive systems in production.

That was one of the differentiators for Satalia is that we would architect systems that could safely

adapt themselves in production. I think, only really now that organisations are realizing the importance of building systems that can learn.

Marina Santilli

Okay, so let’s go to that period then mid 2010 – 2015 when Satalia started to achieve its first large business customers. Can you talk us through some of those early adopters and the kind of solutions you were building for them, and the scope of the project you were given?

Because that was exciting times, as I remember.

00:11:49:14 – 00:13:49:15

Dr Daniel Hulme

Yeah, I think that the original idea for Satalia was essentially optimisation as a service, you might call it now, an AI service you would aggregate lots and lots of algorithms that were being pioneered by academics, and industry would be able to tap into those algorithms by building applications on top of them to use the best algorithms in the world. This was before the cloud and before APIs and things like that.

Because Satalia was close to academia, our relationship with UCL, obviously my PhD, organisations saw us (Satalia) is not just doing consultancy, but having access now to technologies that could help them solve problems.

Our first big break, I think around 2014 was with Tesco. Tesco had a large-scale optimisation problem that they needed to solve, which is delivering groceries to their customers. They knew, like a lot of organisations, that that was an optimisation problem.  They knew that it wasn’t the machine

learning problem; that they needed to use leading edge algorithms from academia to solve that. They saw us as a conduit for those algorithms.

When they saw that we could build end to end systems that could adapt themselves with the UIs and all that kind of stuff, they essentially commissioned Satalia to build their entire ‘last mile’ delivery solution, which was really pioneering of Tesco.  They could have easily got a lot of the new solution off the shelf, or they could have other options out there.

But they wanted to build this themselves. They wanted to make sure that they had control of that innovation (that) they could then internalize and develop themselves. So we built for Tesco what they claim to be the best last mile delivery solution in the world. It was phenomenally successful.

Off the back of that, Satalia then built our own last mile delivery solution that we’ve subsequently taken to market.

And over the past ten years we’ve continued to solve problems across the supply chain that move the needle for organisations in warehousing, in middle mile, in supplier confidence prediction, all of this kind of stuff.

00:13:52:17 – 00:14:03:12

Marina Santilli

So your Satalia delivery product still exists; presumably it’s evolved over the last ten years or so.

What are some of the main changes or features that you brought into the product?

00:14:04:09 – 00:15:26:00

Dr Daniel Hulme

Alright, I could really geek out around this.  Optimisation was just one area that could be pioneered in delivery. Figuring out how to route vehicles in an efficient way uses optimisation algorithms, but one of the big innovations was actually being able to calculate very rapidly how long it takes to get from A to B.

If you went to Google Maps now or any kind of routing engine and ask it to give you a thousand A to B routes, it will take several minutes to give you that response.  We needed to solve that problem

in 50 milliseconds. We need to be able to produce a thousand A to B routes in less than 50 milliseconds that are incredibly accurate because those routes that are then used by the optimisation algorithms to decide what the schedule looks like. So we pioneered geospatial routing.

That was another kind of algorithmic advance.

Once you’ve solved that problem the optimisation problem, as I said, there are now ways of using machine learning to extract more value from that schedule, at predicting how long it takes to deliver to a particular customer, for example. I can use machine learning to do that better. I can use machine learning to predict driver behaviour so I can encourage them to drive differently.

There are lots of opportunities to apply different types of algorithms to solve that problem, and that’s what we did with Tesco. We bring all these different technologies together.

Marina Santilli

..and again, that’s pulling in more data then, to deliver more insights

Dr Daniel Hulme

That’s right. And then, I guess, off the back of that, we’ve built our own product and in that case (the) ‘last mile’ delivery that’s now delivering groceries for Woolworths in Australia. It delivers sofas for DFS here in the UK, equipment hire for SAS. We generalize that solution and we’ve taken that to market and we’ve also, as I said, solve problems across the entire supply chain.

I think what’s exciting for me is not just about using these technologies to solve individual problems across the supply chain, but actually to create a digital twin of a company. We’re getting to the point

now where we can represent an entire organisation as a large optimisation problem and the efficiencies that you can get from modelling organisations in its entirety is phenomenal. The next five to ten years is really about building digital twins of entire companies.

 

Marina Santilli

That’s really fascinating. The other leading product that Satalia developed was the workforce product.  That’s again, another optimisation type of solution which you pioneered.  Tell us a little bit about that one and how that evolved for a specific customer’s requirements and then moved on from there?

00:16:43:20 – 00:18:42:21

Dr Daniel Hulme

Yeah, I’m really passionate about workforce. I’ve always been interested in about how would you create an operating system for a company that allows for the right structure to emerge according to the innovation you’ve taken.

I’ll say that again because it’s quite complicated.

Most organisations start out as being a product company or a service company, and you organise yourself or you structure yourself around that offering. And then the other pressure to  diversify your revenues to take into different markets means that you end up trying to be a services company  if you’re a product company, or a product company if you’re a services company. Marina, I’m sure you’ve experienced this with your startups.

I sort of saw that coming at the very beginning of Satalia and what I wanted to do, is operate (or) create a new operating system allowed for the right structure to emerge according to what we’re taking to market; which is why Satalia is good at doing services and we’re really good at also building product and platform.

So workforce allocation, allocating people to work in a fluid liquid way is something I’m very interested in.  We had an opportunity to build a workforce solution for PWC.  They have 5000 auditors in the UK and they wanted to allocate those auditors in a way that maximised their utilization, maximize their career development, continuity for clients and minimise travel time.

It’s a very, very complex problem to solve. Historically, they would have 40 people trying to solve that problem. And the power of algorithms is that you can solve that problem in 4 hours, significantly better than any human being.

Again, this is an example of where you don’t just use optimisation, but you use machine learning. So you allocate people to work. You then use machine learning to see whether they’ve worked well in those teams, whether they’ve developed the skills that they claim to have developed. Using machine learning, let them do that. Then you can better allocate them in the future. That’s where that feedback is really, really important. These technologies, these capabilities, that Satalia has developed over the past 15 years were then really attractive for an acquirer.

00:18:43:23 – 00:19:13:18

Marina Santilli

Absolutely. We’ll get to that in a minute.

So, these are really holistic solutions, aren’t they? And I think it might be worth saying a little bit about the way you built the company. That’s probably important just as well to the acquisition story. You didn’t build the company just as a normal research team with a marketing front-end on it or a product development front-end, you built something that was a little different and that mimicked this idea of holistic solution.

00:19:14:01 – 00:21:03:23

Dr Daniel Hulme

Indeed yes, I spent some time in Silicon Valley working with large companies. Obviously, we’ve been building solutions for companies for over a decade and, you know, I can see some of the difficulties, the bureaucracies, the frictions exist in organisations. I didn’t want that of Satalia.

I’ve been challenging myself almost from the beginning. How can we create a company that enables us to organise ourselves in the most effective way. If we want to be the best innovation company in the world, we need to be able to move quickly, more quickly than anybody else, and that means you need to remove those frictions, those barriers, and enable people, get out the way of people so they can do their job really well. Fortunately, over the past five years, the idea of Scrum and Agile and Design thinking, are kind of supercharged by AI.

Actually, Satalia was about 120 people when we were acquired and we didn’t ever have any fixed managers or fixed hierarchies. People were always free to work where they want, how they want on

whatever they want, even way before COVID. My goal is to not just scale that across 120,000 people at WPP. My goal is actually try to scale up to a planet.

I would like to try and create a platform where everybody in the world has access to the ability to work on what they want, where they want, how they want, and I think actually it might take the edge off some of the impending challenges we have around the impact of AI over the next decade or so.

I’m really passionate about how do we unlock the creative capacity of people by structuring ourselves in in new ways. I think Silicon Valley call this now, liquid democracy or liquid or fluid organisations, but it’s something that we’ve been thinking about for many, many years.

Marina Santilli

What might your next steps be in that direction then?

Dr Daniel Hulme

I think the advances in AI recently has changed my view on what the world might look like over the next 20 years. I thought we might get a bit philosophical here. I thought that we might have 20 years before we build a superintelligence and within that 20 years, how could we get humanity cooperating better as a species?

One step to enabling people to cooperate better as a species is to rethink how we allocate work. But now, because obviously the advances in AI which has caught everybody by surprise.  Me, I’ve been in this field for 20 years. It’s now changed how I’m thinking about things. It might be that we see superintelligence much, much sooner and the economic impact these technologies will have.

Marina Santilli

… which requires some serious thinking about. There’s obviously a lot of commentary on that at the moment and the subject of an entirely different podcast, I think. Hopefully, we’ll have you back to talk about that one time soon.

But, obviously, we can’t really have a conversation in April 2023 without touching on the likes of the generative A.I. models that have sprung out of nowhere in the last six months or so. Taking it back specifically to a product perspective, as opposed to the philosophical angle what are you thinking about how, and if, you would integrate such technologies into the kinds of products Satalia is building.

00:22:32:17 – 00:27:12:19

Dr Daniel Hulme

I would like to think at the forefront of this now. So obviously WPP, who acquired us 18 months ago, produces, a quarter or a third of the world’s content. Generating content is something that’s very, very important to WPP.

The reason why we were acquired were multiple reasons. One is that they wanted to be able to augment our supply chain logistics capabilities with marketing They can take those capabilities to to customers rather than just marketing. They wanted to utilize our knowledge and experience in workforce allocation to enable WPP to operate more fluidly. We’re building internal solutions for WPP, which I’ll talk about a second, because that touches on generative AI. We take AI technologies to clients. So when people think about AI, they don’t just think about Accenture and they think about WPP.

Because of my chief AI officer role, I get that platform to think about the impact that these technologies will have on society in the longer term and how we can use these technologies.

Like “Is it cool? Is it creepy? ”, kind of how we can hold all of ourselves accountable to using these technologies well.

From it, from an internal kind of generative AI perspective, we are building a number of different solutions in this space. My hypothesis is, that over the next several years, everybody will be able to generate any content they want. The question will not be “how do you generate content?”. The question will be what content you generate to get the response that you want.

So, one of the things that we’re building is a concept called “brand brains”. It’s a really interesting concept. If you go and ask MidJourney or Dall-E or whatever ChatGPT to create an image of a cat in space holding a soft drink or ice cream, then it will create a generic picture of that.

But if you’re a creative working on, let’s say, the Coca-Cola account, you need to make sure that that content is aligned with Coca-Cola’s brand guidelines. They understand that it’s not just the soda, but it’s Coca-Cola. It has Coca-Cola’s tone of voice, all that kind of stuff.

What you can do is build brains, build “brand brains” that you can glue on top of these large language models that first of all, enrich the prompt. It expands the prompts and then it also takes the response and recontextualize the response to align with the brand. That’s how we can, first of all, build content that’s much more specific to the brands.

These large language models have kind of inverted how we think about machine learning now. What we do is we build what’s called, I guess, audience brains. So, when you’re marketing content to people, you have a target market and you want to try to understand how does that person, what does that person when they see a piece of content, what do they construct in their mind? They hope to see a cat in space with a soda, but it will elicit nostalgia, excitement that they will create a complex narrative in their mind.

What we want to do is essentially get AI. To replicate that narrative, you will create a different narrative in your mind, Marina than I will.  So how can we get to create a complex narrative?

And then how do we correlate that complex narrative with clicks and sales?

These audience brains are phenomenally powerful and not only that, you also need to build audience brains of cultures or newspapers or political parties or minority groups, because you want to see how they will respond to content; and you’re not offending anybody or you’re not breaking the law. You have brains that represent your target audiences, but also brains that represent groups of people who you care, how they think.

Obviously, you need to understand when you produce content, you show it to those people, you see what how they react, that you learn from it. I think the differentiator for companies like the WPP is not creating content very quickly. It’s figuring out what content to create and to what audience to show it to. And that’s the stuff that we’re working on.

00:22:32:17 – 00:27:12:19

Dr Daniel Hulme

I would like to think at the forefront of this now. So obviously WPP, who acquired us 18 months ago, produces, a quarter or a third of the world’s content. Generating content is something that’s very, very important to WPP.

The reason why we were acquired were multiple reasons. One is that they wanted to be able to augment our supply chain logistics capabilities with marketing They can take those capabilities to

to customers rather than just marketing. They wanted to utilize our knowledge and experience in workforce allocation to enable WPP to operate more fluidly. We’re building internal solutions for WPP, which I’ll talk about a second, because that touches on generative AI. We take AI technologies to clients. So when people think about AI, they don’t just think about Accenture and they think about WPP.

Because of my chief AI officer role, I get that platform to think about the impact that these technologies will have on society in the longer term and how we can use these technologies.

Like “Is it cool? Is it creepy? ”, kind of how we can hold all of ourselves accountable to using these technologies well.

From it, from an internal kind of generative AI perspective, we are building a number of different solutions in this space. My hypothesis is, that over the next several years, everybody will be able to generate any content they want. The question will not be “how do you generate content?”. The question will be what content you generate to get the response that you want.

So, one of the things that we’re building is a concept called “brand brains”. It’s a really interesting concept. If you go and ask MidJourney or Dall-E or whatever ChatGPT to create an image of a cat in space holding a soft drink or ice cream, then it will create a generic picture of that.

But if you’re a creative working on, let’s say, the Coca-Cola account, you need to make sure that that content is aligned with Coca-Cola’s brand guidelines. They understand that it’s not just the soda, but it’s Coca-Cola. It has Coca-Cola’s tone of voice, all that kind of stuff.

What you can do is build brains, build “brand brains” that you can glue on top of these large language models that first of all, enrich the prompt. It expands the prompts and then it also takes the response and recontextualize the response to align with the brand. That’s how we can, first of all, build content that’s much more specific to the brands.

These large language models have kind of inverted how we think about machine learning now. What we do is we build what’s called, I guess, audience brains. So, when you’re marketing content to people, you have a target market and you want to try to understand how does that person, what does that person when they see a piece of content, what do they construct in their mind? They hope to see a cat in space with a soda, but it will elicit nostalgia, excitement that they will create a complex narrative in their mind.

What we want to do is essentially get a high. To replicate that narrative, you will create a different narrative in your mind, Marina than I will.  So how can we get to create a complex narrative?

And then how do we correlate that complex narrative with clicks and sales?

These audience brains are phenomenally powerful and not only that, you also need to build audience brains of cultures or newspapers or political parties or minority groups, because you want to see how they will respond to content; and you’re not offending anybody or you’re not breaking the law. You have brains that represent your target audiences, but also brains that represent groups

of people who you care, how they think.

Obviously, you need to understand when you produce content, you show it to those people, you see what how they react, that you learn from it. I think the differentiator for companies like the WPP is not creating content very quickly. It’s figuring out what content to create and to what audience to show it to. And that’s the stuff that we’re working on.

00:29:33:12 – 00:29:52:07

Marina Santilli

And it’s been, as you say, nearly two years since WPP acquired (Satalia).

00:30:24:00 – 00:30:44:19

Marina Santilli

When they acquired Satalia there where already many other AI-first companies providing consultancy services and products to enterprises. What do you think it was that about Satalia that caused WPP to pick you?

00:30:44:21 – 00:31:57:11

Daniel Hulme

I think, you know, organisations out there, can specialise in one type of AI and I think that because of my background in AI over the past 20 years, I’m very fortunate to have both, machine learning and optimisation operations research. And because I’m a popular speaker on the ethical social safe impact of these technologies, we brought to them kind of like a more complete package both from, you know, capability and logistics, which is what they wanted to expand into in workforce. We did have some experience in generative AI at the time and then, you know, how do we position ourselves to use these technologies. We were a complete package for WPP.

At the time I think there’s quite a few organisations that were wanting to acquire us. But I just felt there are so many synergies with WPP and we’d been working with a few op-cos from WPP on, on clients. And it just been really enjoyable working with creative people, solving difficult problems.

It’s been great; one of the best business decisions of my life.

00:31:57:15 – 00:32:06:00

Marina Santilli

It’s great. Presumably there are still misconceptions in industry as to what AI can achieve. What do you think that current misconception is?

00:32:06:14 – 00:32:29:10

Dr Daniel Hulme

I think we still seeing organisations hire data scientists. The idea is that data scientists are able to extract insights from data then and human beings are able to make better decisions. I can’t say that enough that the human beings are not typically very good at making decisions.

I’m still a massive advocate of companies hiring optimisation people, which they’re now starting to do. Trying to steal some of my people…

(Marina laughs)

00:32:35:02 – 00:34:49:12

Dr Daniel Hulme

I think it is frustrating, but actually more recently (and I’ll try to go with very, very quickly) more recently I’ve been thinking of AI not through definitions but through its application.

I believe that there are six applications of AI that are really making a difference for organisations. Just very quickly: task automation, whilst it gets this kind of a bad reputation in the AI community because it’s very simple algorithms that automate relatively standard tasks. The fact is that it could free a huge amount of cost of human labour and get it to do more interesting things. Task automation is, having a big impact.

Obviously, generative content generation augmenting the creative process. That’s a second category.

The third category is what I call the humanization, i.e. you’re taking a human being and you’re replacing it with something that looks and behaves like a human being. So going back to this idea of audience brains, we’re creating an audience representation of a human being. So replacing a human with a like for like, that’s the third category.

The fourth category is machine learning or extracting insights from data. I think that’s where a lot of mis-investment is currently happening still, although what’s exciting about extracting insights from data is these technologies can help us explain why those insights exist, so that we can learn new things about the world. So that’s where I think there’s opportunities.

The fifth category is complex decision making. So again, traditional optimisation; what is Satalia very strong at and we’re continuing to see a massive impact on business.

And then the final category is the augmentation of your physical self with cybernetics, with robotics, and also, the digital self. The digital you, are in the metaverse with something that’s making decisions on your behalf.

And maybe just to bring this to life a little bit, we actually build digital representations of employees. This might sound a bit creepy and then we then (I have to use a better word) we query or interrogate that digital representation and ask, you know, if we put you on this team, will you thrive.

If you go on this project, what would it help you develop your career?

So it’s going back to this idea of audience brains. Instead of seeing how they respond to marketing content, I can build a brain that represent you in your world of work and understand how to have you better fit into the organisation.

00:34:49:24 – 00:34:51:03

Marina Santilli

That all does sound a little bit creepy Daniel!

00:34:51:03 – 00:35:10:19

Dr Daniel Hulme

It’s all about the intent. It’s about the intended use of these technologies. If you intend to use them to squeeze utilisation out of them or to manipulate them, then people won’t allow you to do it. But we intend to use this to use these technologies to enrich people’s lives, as long as we continue to do that, people will be comfortable.

00:35:11:03 – 00:35:19:15

Marina Santilli

I know you’re one of the good guys now, but obviously the challenges of convincing people that you don’t know is a tough one.

00:35:20:09 – 00:36:48:24

Dr Daniel Hulme

Yeah, exactly. I don’t know what the next ten years are going to look like for us, but I think there’s a lot of concern about these technologies that people have. I think the next ten years is going to be crazy. It will see Cambrian explosion of not just application of generative AI, but all of these other AI that is going to just exponentially accelerate. It’s going to be an exciting decade ahead.

Beyond that, I don’t think anybody really knows what’s going to happen. But what I do know is this : it’s within our gift to create the future, we don’t just sit around and wait for that future to happen.

It’s up to all of us to create that future.

A couple of weeks ago I was frustrated reading LinkedIn, and everybody’s now become an AI philosopher, everybody has an opinion about AGI or superintelligence. I’ve been working with people for many, many decades who’ve been thinking about these things extremely deeply. I kind of realized that actually it’s really good that people are concerning themselves about these things.

It’s really good that now everybody’s philosophising about AI I think what we want to make sure is that if we’re informing policy or, you know, making big decisions based on what’s going on. (All) this chatter means we have the right people at the table. I think it’s great. Everybody’s talking about it.

But one thing I would ask of everybody is, is to be intellectually honest with ourselves and make sure that we have the right people at the table making the right decisions because these things will change the world in incredible ways. And we need to make sure that we use them for the right reasons.

00:36:50:03 – 00:37:10:14

Marina Santilli

The final question …Is it a bit like global warming? Are we a little bit too late coming to the discussion about whether AI is moving too fast, too quickly, or have we still got time?

00:37:10:14 – 00:38:34:23

Dr Daniel Hulme

It’s a question I ask myself every other hour of every day. The answer is I don’t know. When I engage with my peers, people who are working at a company like DeepMind and authors that have been writing on this subject for many years, I think the general concern now is that we will see an AGI happen in the next 4 to 8 years.

I had a private audience with Sam Altman a few years ago and we were talking about the impact that was going to have. Back then he thought we would achieve AGI by the end of this decade. I think now the feeling is that we could start to see AGI happening even sooner.

We don’t know that once we’ve built a machine that’s as intelligent as us, how quickly it will take for that machine to become a million times more intelligent than us. We really don’t know. I think that one of the concerns that we currently have.

00:38:35:16 – 00:38:48:16

Marina Santilli

That’s an interesting thought to finish with, Daniel. As usual, it’s been an absolute pleasure

to talk with you. Before we get too down in the in the philosophical weeds, I think we should probably leave it there. But thank you so much for your time. And I look forward to chatting with you again!

Dr Daniel Hulme

Thank you.