Introduction

So I'm Laura Gemmel and I run Taught by Humans.

About Taught by Humans and Its Mission

We do AI and data confidence and what we mean by that is we try to help everyone have the skills and knowledge that they need to use these tools in work or their everyday life.

Background and Audience

We've been doing this for about five years, doing content, mostly going out and going into companies and actually doing training, working a little bit with universities.

So our target audience is actually working adults.

mostly data people, but we also get interest from engineers and people in marketing and things like that.

The Need for Personalized Education

And for a really long time, I've had this idea that we really need to personalize education and then generative AI happened and it made it a lot easier.

Understanding the Challenges in Data Education

But just to talk you through the problem a little bit.

So who's an engineer in the room, everyone?

Most people could.

So it's hard sometimes to know what to learn, but that's particularly true in data.

There's so many different skills, and most data people aren't, for example, a JavaScript engineer or a Python engineer.

They just need to know data.

And every single job they do requires different skills.

You get turned down from jobs because you don't have the exact experience I think engineers can

relate to that.

But one of the things with data is when you try to go and learn it, it's really overwhelming.

So on Coursera alone, there are 10,000 courses on data and AI.

And if you don't really know what you want to learn, you might end up doing the wrong thing.

And there's all these questions to answer.

Research Insights and Learning Barriers

So I actually did my PhD research in how people learn

data science, AI, and robotics skills.

And 80% of people I spoke to had started learning something, or started trying to learn something, and then been like, oh, I don't actually know which one of these courses to take, and given up.

The Difficulty of Adult Education

And the other thing is, this is my weird little diagram of what a classroom of 20 people looks like.

So this is just in my head.

Education's actually really hard to get right.

So you don't know the person's background.

It's particularly hard for adults, because if you think about educating children,

Up to that point, they have followed a set curriculum, obviously all to a different level, but if I was to start teaching one of you something, I've literally no idea what you know, right?

You could be an expert in one topic, so you could be an expert in Python wanting to learn JavaScript, and that's not the same curriculum as someone who's never coded.

It's actually really hard to get adult education right.

An intermediate education is really hard.

Beginner's education is easy.

You assume someone knows nothing and you just start at the beginning, you explain every single term to them.

For me to do intermediate education I have to decide what you already know and unless I've taught you it, it's quite hard.

And also the end of intermediate education isn't like a particularly defined point.

I'm defining what intermediate means so you might do an intermediate course.

And at the end of it be like, oh, I didn't learn the skills I wanted to learn on this intermediate course.

And education for groups, as you can see from this, we do this thing called designing for the middle.

1So you pick the middle person in the class and then you probably end up with people who are super bored because they already knew this and you have people who are lost because they can't understand it.

Taught by Humans' Approach to Personalization

So what my business does is actually we do like really tailored education and we charge a lot of money for it because it's really hard to tailor education to a group of 20 people particularly if they come from different departments in a company.

So for about five years now, I made this diagram five years ago, so it's now getting brought out.

Vision for Personalized Learning

But I've had this idea for five years that it'd be really great if we could magically just personalise education to people.

So you fill in a quiz, you do something, and we're like, cool, this is what you need to learn, and this is the exact order you need to learn it in, and this is what you'll get at the end.

And I've played around with it for a while.

So if you're familiar with graph databases and stuff, I have this massive one of data skills that I made by hand where I mapped out the individual links between different spreadsheet skills and at what point you should switch to Python and stuff like that.

It's very painful.

It's very slow.

It didn't quite work.

And one of the things is people actually don't often know where they are and where they want to be.

We actually need to tell them that.

Leveraging Generative AI for Education

Generative AI happened and now everything is hyper-personalised.

You're going to have hyper-personalised holiday recommendations, hyper-personalised skincare, hyper-personalised everything.

And I'm trying to apply this to education.

But I don't know that you're engineers, so you probably remember when the super promise of recommender systems happened and everything was going to be personalised by a recommender system.

And if you shopped on Amazon and bought a Hoover, it thought you were obsessed with Hoovers and you wanted to buy like 50 more of them, whereas you actually just needed one Hoover and you didn't need 50 Hoovers.

1So I don't know if it's a little bit overhyped about how well generative AI is going to personalize.

Has anyone actually implemented generative AI systems to do personalization?

One person.

Awesome.

Cool.

Okay.

Development of the Personalized Learning Platform

Initial Prototype and Challenges

So what I did was I had a weekend free, and I was like, right, I'm going to build this prototype from scratch by myself.

I used OpenAI APIs, so not ChatGPT itself, the algorithms that actually par ChatGPT,

You can actually edit them and make them whatever way you want.

I got it to work because I was demoing it on the Wednesday after I started building it, so nothing like a fire underneath you to make you build something.

It worked end to end.

It asked you five questions backwards and forwards, and it spat out some really ugly JSON, and I didn't have time to tidy it up, so I was just showing this on a screen being like, this is your learning journey.

This works great.

It was really slow.

If anything broke, it just broke.

And also, I ran it for two hours and it cost me $47 because I hadn't optimized at all and every single run was costing me $1.

And it only worked on my computer.

So what I built it in was NixJS React using OpenAI.

Iterative Improvements and Vector Databases

I then didn't really talk to anyone for six weeks and literally just coded and got this working end-to-end model that actually spat something usable out.

So as you can see, it's a tiny bit nicer.

It's got like a border and stuff now instead of just being a chat bot.

What I did was I started using vector databases.

Have we heard of vector databases?

Getting some nods.

So for those that don't know, what happens is, very basically, I give it a bit of text and OpenAI gives it

1,536 numbers between negative 1 and 1 or 0 and 1.

And that somehow sums up all the words in the text I gave it.

And using some math, it works out if it's similar to other text.

So what I did was all of the content that I've been creating for five years, I put it in a vector database.

They all got a score.

And it suddenly worked so much better.

So version one worked surprisingly.

I really didn't expect it to work that well.

It worked really well.

This worked amazingly.

So what I actually did was I got you chatted to the chat bot.

We still got the same results.

But instead of all of our content just being fed into the chat bot,

It now ranked it, and it pulled it back like a search system.

So it pulled back the 20 most relevant bits of content, and then it looked through which ones were best, and it put a perfect learning, I'm saying perfect, the perfect learning journey for you.

It was a bit faster, way cheaper, you can see, gone down from one pound per run to four cents per run, which I was very happy with.

It was still a bit unreliable, and it was still one chat bot that did everything.

So it was like a bit slow, and if anything happened, it just kind of freaked out, and you had to start again.

Multi-bot Architecture and User Experience

So what I ended up doing was building this thing that I'm calling a multi-bot architecture.

So at the minute, I actually have, I think it's actually more than five now, I think I might actually have six chatbots that talk to each other.

So they are small chatbots, they all have one,

They all have one goal instead of one big chatbot that did like seven different things.

They all do one thing.

It's like having a worker.

So the first one is get to know you, which I call dotly one.

I've got like weirdly attached to these chatbots as well.

When I run the code, I'm like, well done, dotly one.

That was a great response.

So dotly one is now almost like the customer service representative, the front facing person that if you were coming to talk to us, would do like the needs assessment and be like, okay, what does this person need?

I need to get to know them.

It then writes a search string.

which is sent to the relevant content bot, which I weirdly haven't named.

It's just called the search bot.

And all it does is it gets given a search string, and it looks up our vector database, and it goes, cool, this is the relevant content.

These two bots then both pass.

some information to my learning designer bot.

And I've done the learning designer bot in a really interesting way.

I've actually fed in loads of learning science to it.

So it's actually, well, it's not actually a learning designer, but it's as close to a learning designer as any bot could be, I think.

And it then has a little think and takes its time and then spits out a learning journey, the reasons it made that learning journey for you.

So you actually get to see what it based it on, which actually our users are quite into.

Then you can do your learning journey.

And when you're done, I have a separate assessment bot which looks at what you learn, why you wanted to learn it, and then writes you some questions.

And one of the things that sits behind our entire platform is the skills mapping.

So we kind of map out what you know and what we think you should learn next.

So the assessment bot currently is the only one that writes to the skills mapping.

So based on how your assessment goes, it might update it for you.

And our plans are to add in this learning companion the whole way through.

So anyone old enough to remember Clippy from Microsoft Word?

It deeply divides a room that I stand in at the minute.

So I say it to some people and they look at me blankly and I'm like, God, I'm old.

And then I say it to other people and they're like, of course I know who Clippy is.

And I'm like, okay, that's my goal.

I would like to have Dottley, Clippy, in the corner so that you can talk to it whenever you want to.

And one of the reasons for doing that is it really helps neurodiverse learners.

So I'm dyslexic and I've realized that I skim read everything I do.

And then I'm suddenly like, oh, I read a heading on this topic and I'd quite like to remember what it said.

I've spoken to some friends who have ADHD and they have this thing called global understanding.

So a friend of mine was learning SQL.

heard about what data engineering was, and then went and took a different data engineering course so that she could understand data engineering, and then had forgotten where she was taking the original SQL course.

And then, yeah, it took her three months, but she did eventually do it all.

But what we're hoping for is that our platform, you could actually tell it, OK, I need to understand how the data gets to SQL, and then we can tailor our learning to you.

And yeah, so the way this is built, it's really cool.

So if, for example, a new algorithm that came out that was better, I can actually just slot little bots out and update them instead of having to do the whole architecture.

And I can slot things in and out.

Live Demonstration

So now for a demo.

Showcasing the Platform

I'm going rogue, and I'm doing a completely live demo and not a video.

So let's hope it goes OK.

Everybody cross your fingers.

It's just going to be my luck that, like, OpenAI breaks right now or something.

There'll be, like, an outage.

So this is our platform.

When you're logged in, you can see I'm really into learning.

So I have, like, 20 learning journeys on the go at the minute.

Oh, I'm going to kneel down.

So these are all the learning journeys that I've created before.

But if I scroll down to here...

One thing at the minute is it doesn't remember your other learning journeys.

So that's something I'm working on at the minute.

So it does pretend like you know nothing other than the skills mapping.

So sometimes it forgets your learning journeys.

So you've got .ly.

.ly is trying to get five bits of information out of you.

Topics, motivation.

I should probably type and do it at the same time.

Topics, motivation.

I'm going to say I want to learn Google Sheets, because that's actually the most popular topic at the minute.

It's trying to find out how you learn, what way you want the, it's not going to work.

I don't have any faith in this actually working, even though it's live and on the internet and people use it.

I don't know.

Very much sorry.

It's not a great typing.

One of the things also is quite interesting.

So you can see I'm writing like five words.

When I put this in front of people, they write paragraphs.

And I'm like, why are you talking to the chatbot so much?

So I didn't realize how much information people would want to give to the chatbot.

And one of the things we've actually put into it is if you just say you don't know, it actually can still continue.

And if you say random things to it, it's really on brand for being like, that's wonderful, but what about data and AI?

So it's very interesting.

So I'm going to say I want to learn formulas.

just for speed.

So I think it asks seven questions usually.

Has anyone built anything with generative AI?

You never know exactly how many questions it's going to ask.

So it generally asks seven questions.

Sometimes it's really interested in what you're saying, and it goes off on a weird little tangent.

I'm going to say I like videos and blogs.

And hopefully this should be the last one.

If it's the last one, it'll say one final question.

Oh, it does.

One final question.

Well done.

And Guinness say no.

So we ask at the end if there's anything you'd like to tell it, because that's actually when people tend to tell us about neurodiversity and stuff like that.

We don't want to ask specifically.

But actually, people do tend to tell us.

So this is when it's going off, looking up the vector database.

And then the learning designer bot is going to actually build a learning journey for us.

If it works.

I'm so not confident in this.

I don't know why I'm demoing it and stuff.

Nothing's gonna work.

Oh, it's being very slow.

Hopefully it will work.

Oh, no, it's just the screen.

Come on.

I do have one I can demo from earlier, though, if it doesn't work.

It is.

Yay!

Sorry, I'm very happy.

So we're working on this screen.

We've had some feedback on the UX on it.

But basically, it's decided.

So I think at the minute, we have 100 bits of content on the platform.

And it's decided that these seven are the ones that you should do.

If you click down here, we have a little summary on what you're learning, a learning plan.

So it started with intro to Google Sheets.

Interestingly, we have five intros to Google Sheets.

So the fact it picks one specific one for you is always quite interesting to me.

And then I've said I wanted to learn formulas.

And it's actually just spot all these functions, which is pretty good.

And then it's given pivot tables at the end.

The based on is really basic.

But basically, this is ChatGPT summarizing it.

Or not ChatGPT, .ly.

It's not ChatGPT.

Yeah, so then this is what it looks like, and you can go through, and at the end, it'll write an assessment for you.

So that's my demo.

I'm not going to risk doing the assessment bot because I'll stress out.

But yeah, you can go through and mark it.

And yeah, that's basically what we've got at the minute.

We're working on improving it, particularly the UX side of it.

And we've noticed that people would like to be able to edit their learning journeys.

So just because the chat bot's told you something doesn't mean it's the best.

So we're working on that as well at the minute.

Okay, so that's a great noise.

Lessons Learned and Conclusion

Key Takeaways from the Personalization Project

So if you're thinking about using personalization in a project or something, I think there's some lessons I've learned that I'd like to share with you.

So one of the things is I think these are the five things you need to get it to work.

The first thing is you need a lot of tech expertise.

So to get something like that to work, it was great that I could do

the web development and one of the things that was really important was the data flow.

So how the data moved from one point to another and through the chat box and what the end goal that I was trying to get was.

And also UX is really important for stuff like this.

Like if I just, when I was spitting out JSON, everyone was like, who cares?

What is that?

It doesn't mean anything.

And so actually getting it pretty and having the dots and making it look nice was really important.

The AI side of it is good to know, but it's an API at the end of the day, right?

You can use out of the box if you need to.

If you want to tweak some of the parameters, great.

It works pretty well.

Like we can all admit chat GPT is pretty cool and it works pretty well.

So the algorithms behind it work well.

The bit that made it work better was the architecture.

So pausing for a second, instead of just focusing on your end goals, focusing on

what way we actually architect it.

So thinking about how people are going to use it, thinking about what might happen in the future.

And that's what actually has made this platform work so much better.

So I got it to work end to end.

It was perfectly fine.

Getting this architecture flow has made it so much better.

The things that are non-techie or kind of techie.

We've spoken to 200 people about how they learn, both actually as user research interviews and through my PhD.

And that part of it was really important because I know how people think, and that was why the UX side of it was easier, but I also knew what people needed out of it.

So why the buttons were hidden is because some people don't really give a crap about their learning outcomes.

Other people do.

If you care about learning outcomes, you're going to click on the buttons.

If you don't, you just click Start.

And testing is so important.

So every time I run it, it does something else.

So it's not like normal development where you can literally test outcomes.

You had to write loads of test cases.

I had to pretend to be other people.

I applied a lot of friends in beer and pizza.

to come to my house and break it, and actually try and break it.

And that was the best bit, because I was like, why did you type that in?

But also, I'm not allowed to say that, because you can type whatever you want to into it.

It's a chatbot.

So that was one of the things that I found most difficult, was I expected everyone to use it like me, sensibly, and say short answers.

And that is not how anyone has ever used it that I've put in front of.

I read my friends' responses to it sometimes, and I'm like, why did you say that to the chatbot?

Fine, but why?

So you've got to get as many people involved as possible in testing.

And why I'm really happy to stand up here and tell you the architecture of something I've just built and why I don't think it's a secret is because anyone can build a chatbot.

Anyone can go out there and build a chatbot.

It's actually the content that we're feeding into it and the expertise on how data people learn

and just learning design itself that actually makes our platform work.

So building it on something you've got good data on, good expertise on, is what's actually going to make this work.

If you're looking for investment defensibility on AI products, it's quite hard.

Subject matter expertise, and that's why we're the best, actually seems to work quite well.

One of my friends came here.

Final Thoughts and Contact Information

And that is, so I'm Laura, I run Taught by Humans.

We have LinkedIn and a website if you want to find us.

One of our big, I guess, ethos is that we're actually designing this as education for humans by humans, powered by tech.

So we don't want to replace people's jobs.

We want to help people use these tools to do their jobs better.