TL;DR: I built an open-source, two-hour interactive presentation on AI safety for non-technical audiences. Here’s a recap of what it covers, how it went, and how you can use it yourself.
Epistemic status: personal experience from one session with six close friends, plus impressions from conversations with people inside and outside the field over the past six months.
I've been a machine learning engineer for almost 10 years, working in scale-ups and startups, even as a CTO. And not once during all those years did I really realise how dangerous AI could be. Not until a friend actually showed me.
Six months ago, I quit my job to transition into AI safety, after realising that AI might be one of the most critical dangers we're facing. Today I’m launching my own AI safety non-profit.
But this post is not about me.
It's about something I've been realising over the past six months: most people outside of AI safety simply don't understand the risks of AI. And I think one of the main reasons is that they don't even realise how powerful AI already is. Or at least, that’s my impression when I talk to people around me. Polls in the US show that many people say they are worried about AI, but in my experience, few of them understand why it could actually be dangerous.
When I talk to some of my non-technical friends or family (my stepmother, for example), I sometimes realise that, for them, AI is still basically a chatbot. Something you ask a question to, then copy-paste the answer to help you write an email.
And at the same time, I've spent the past six months discovering how little is actually happening around AI safety, especially politically.
But how could anything happen politically if most people don’t really understand the problem?
Why I built this
Two weeks ago, I was at my first ML4Good bootcamp. I was messaging a friend, explaining what we were discussing there and some of the risks of AI. At the exact same time, he was at a conference with his company, where they were pushing everyone to use AI more and more.
And don't get me wrong, I don't think the two are incompatible. I use AI all the time myself.
But it made me realise how disconnected these two worlds are.
On one side, you have people working in AI safety, who are fully aware of the risks and don't understand why everyone, including politicians, isn't doing more about it.
On the other side, you have regular people, people like I was six months ago, who don’t really understand those risks. (In my case, I wasn’t even aware of them.)
And I think we sometimes forget just how big that gap is. (Again, this is my personal impression, and I’d be interested to hear your thoughts on it.)
What's in the presentation
At the end of the ML4Good bootcamp, we had to work on a personal project. And I decided to build a two-hour, interactive presentation about AI, designed for a non-technical audience.
One of the interactive demos: words placed by meaning, using a real embedding model.
I wanted to trick my non-tech friends into learning about AI safety.
The idea was to invite some friends over, telling them they'd learn how to make better use of AI, understand how it works, connect agents together, and play with some models.
Which was all true. But my main goal was to progressively bring them to the risks of AI. So I structured the presentation in five parts.
First, showing them how powerful AI already is today. Because I think a lot of people still have a 2022 picture of what AI can do. And it's hard to understand the risks of something if you don't even realise what it's capable of.
Then, a bit of AI history and an explanation of how it actually works. The idea was to use live interactions and animations to help them understand the difference between a classical program or classical machine learning on one side, and deep learning and large language models on the other. In the end, we don’t build these models, we train them, so we don’t understand what’s happening inside.
One of the animations: the model learns a rule nobody intended, and nobody notices until it meets the real world.
After that, the idea was to talk about who’s actually building all of this: what’s happening in the AI ecosystem, who’s racing against whom, and why nobody seems willing to slow down. It has the structure of a prisoner’s dilemma.
And finally, the risks. Starting with the ones that are easiest to understand, like misuse, then going through all the systemic effects, and ending with the one that sounds the most like science fiction: loss of control.
Systemic risks covered in the talk.
But by that point, after everything they’ve learned so far, my hope is that it doesn’t sound so much like science fiction anymore, and that concepts like artificial superintelligence (ASI) and the possibility of losing control over it start to seem worth taking seriously, especially when you realise that some of the people who understand these systems best are seriously worried about it.
Finally, I didn’t want people to leave with only things to worry about, so the talk ends with what we can actually do.
How the talk ends: what anyone can do, starting today.
Because yes, there are researchers, organisations, and some political initiatives working on AI safety. But it's nowhere near enough.
And I don't think everyone needs to quit their job and start working in AI safety.
As a normal citizen, even without being actively involved, there's already something we can do. I think it starts with understanding the problem, talking about it with people around us, and making it a political subject.
The same way we expect environmental issues to be part of every political programme, I think AI should be too. Whether it's for geopolitical reasons or because of the existential risks to humanity.
How the first session went
I finished the presentation and invited six friends over to try it out.
And it went really well!
We ordered some pizzas, had a few beers, went through the presentation, skipped one or two subtopics because of time, and then just stayed around discussing it. Asking questions, sharing what they thought, talking about what could happen next.
It was a really chill evening.
And they all told me how much they'd learned, and how little they actually knew about what was happening in AI. And these are engineers, people working in sales, finance, etc. Not exactly people disconnected from technology.
What surprised them most was how predicting the next word, at scale, is enough for a model to generate coherent text, solve complex logical tasks, and even call tools and take actions. Playing first with a base model, which is mostly unhelpful, and then with an assistant model that can genuinely change your daily life, made that jump very concrete for them.
How you can use it
So I decided I wanted to keep giving this talk, and I already have plans for it. But more importantly, I wanted other people to be able to give it too.
That's why I've put the whole presentation online, along with its open-source GitHub repository.
The presentation is interactive: for instance, you can play with an embedding model, a base model, an assistant model and a tokenizer, and there are animations to help explain how things work. People attending can also connect with their phones to access a glossary and play with the models themselves.
I've also written a script explaining the storytelling behind the presentation, with additional notes in presenter mode, so you don't have to figure out how to present everything yourself.
And of course, I used AI to build it. It helped me a lot with the content and the animations, but the storytelling is my own, and I spent hours testing it, checking the sources and reworking it.
And I'd really love for people to use it.
You can give the full two-hour talk, take only the parts you find useful, or fork the repository and make your own version. Maybe for younger people, older people, or with a completely different angle.
If you decide to give the talk, please let me know on Signal! I'd be happy to help you prepare, and I'd love to hear what worked, what didn't, and how people reacted. It would also be great to keep track of how many people we manage to reach.
And if you fork the project, let me know too. If you want to make your version available to others but can't host it yourself, I'm happy to help with that as well.
Six months ago, despite having worked in machine learning for almost a decade, I had no idea about most of these risks.
All it took was a friend taking the time to explain them to me.
I'd love for this presentation to help other people have that same conversation.
Open questions
I'll leave you with a few questions that I'd genuinely love to hear your thoughts on: Do you think AI safety should be everyone's concern, or is it too technical and something that should be left to technical experts and people working in governance? Could there actually be a danger in having everyone involved in the conversation, or do you think it's the only way to pressure AI companies into slowing down?
TL;DR: I built an open-source, two-hour interactive presentation on AI safety for non-technical audiences. Here’s a recap of what it covers, how it went, and how you can use it yourself.
Epistemic status: personal experience from one session with six close friends, plus impressions from conversations with people inside and outside the field over the past six months.
I've been a machine learning engineer for almost 10 years, working in scale-ups and startups, even as a CTO. And not once during all those years did I really realise how dangerous AI could be. Not until a friend actually showed me.
Six months ago, I quit my job to transition into AI safety, after realising that AI might be one of the most critical dangers we're facing. Today I’m launching my own AI safety non-profit.
But this post is not about me.
It's about something I've been realising over the past six months: most people outside of AI safety simply don't understand the risks of AI. And I think one of the main reasons is that they don't even realise how powerful AI already is. Or at least, that’s my impression when I talk to people around me. Polls in the US show that many people say they are worried about AI, but in my experience, few of them understand why it could actually be dangerous.
When I talk to some of my non-technical friends or family (my stepmother, for example), I sometimes realise that, for them, AI is still basically a chatbot. Something you ask a question to, then copy-paste the answer to help you write an email.
And at the same time, I've spent the past six months discovering how little is actually happening around AI safety, especially politically.
But how could anything happen politically if most people don’t really understand the problem?
Why I built this
Two weeks ago, I was at my first ML4Good bootcamp. I was messaging a friend, explaining what we were discussing there and some of the risks of AI. At the exact same time, he was at a conference with his company, where they were pushing everyone to use AI more and more.
And don't get me wrong, I don't think the two are incompatible. I use AI all the time myself.
But it made me realise how disconnected these two worlds are.
On one side, you have people working in AI safety, who are fully aware of the risks and don't understand why everyone, including politicians, isn't doing more about it.
On the other side, you have regular people, people like I was six months ago, who don’t really understand those risks. (In my case, I wasn’t even aware of them.)
And I think we sometimes forget just how big that gap is. (Again, this is my personal impression, and I’d be interested to hear your thoughts on it.)
What's in the presentation
At the end of the ML4Good bootcamp, we had to work on a personal project. And I decided to build a two-hour, interactive presentation about AI, designed for a non-technical audience.
One of the interactive demos: words placed by meaning, using a real embedding model.
I wanted to trick my non-tech friends into learning about AI safety.
The idea was to invite some friends over, telling them they'd learn how to make better use of AI, understand how it works, connect agents together, and play with some models.
Which was all true. But my main goal was to progressively bring them to the risks of AI. So I structured the presentation in five parts.
First, showing them how powerful AI already is today. Because I think a lot of people still have a 2022 picture of what AI can do. And it's hard to understand the risks of something if you don't even realise what it's capable of.
Then, a bit of AI history and an explanation of how it actually works. The idea was to use live interactions and animations to help them understand the difference between a classical program or classical machine learning on one side, and deep learning and large language models on the other. In the end, we don’t build these models, we train them, so we don’t understand what’s happening inside.
One of the animations: the model learns a rule nobody intended, and nobody notices until it meets the real world.
After that, the idea was to talk about who’s actually building all of this: what’s happening in the AI ecosystem, who’s racing against whom, and why nobody seems willing to slow down. It has the structure of a prisoner’s dilemma.
And finally, the risks. Starting with the ones that are easiest to understand, like misuse, then going through all the systemic effects, and ending with the one that sounds the most like science fiction: loss of control.
Systemic risks covered in the talk.
But by that point, after everything they’ve learned so far, my hope is that it doesn’t sound so much like science fiction anymore, and that concepts like artificial superintelligence (ASI) and the possibility of losing control over it start to seem worth taking seriously, especially when you realise that some of the people who understand these systems best are seriously worried about it.
In the talk, I show behaviours already observed when testing current models: they notice when they are being tested, the reasoning they show is not always what drove the answer, some resort to blackmail when threatened with replacement, and some try to save themselves when facing shutdown.
Finally, I didn’t want people to leave with only things to worry about, so the talk ends with what we can actually do.
How the talk ends: what anyone can do, starting today.
Because yes, there are researchers, organisations, and some political initiatives working on AI safety. But it's nowhere near enough.
And I don't think everyone needs to quit their job and start working in AI safety.
As a normal citizen, even without being actively involved, there's already something we can do. I think it starts with understanding the problem, talking about it with people around us, and making it a political subject.
The same way we expect environmental issues to be part of every political programme, I think AI should be too. Whether it's for geopolitical reasons or because of the existential risks to humanity.
How the first session went
I finished the presentation and invited six friends over to try it out.
And it went really well!
We ordered some pizzas, had a few beers, went through the presentation, skipped one or two subtopics because of time, and then just stayed around discussing it. Asking questions, sharing what they thought, talking about what could happen next.
It was a really chill evening.
And they all told me how much they'd learned, and how little they actually knew about what was happening in AI. And these are engineers, people working in sales, finance, etc. Not exactly people disconnected from technology.
What surprised them most was how predicting the next word, at scale, is enough for a model to generate coherent text, solve complex logical tasks, and even call tools and take actions. Playing first with a base model, which is mostly unhelpful, and then with an assistant model that can genuinely change your daily life, made that jump very concrete for them.
How you can use it
So I decided I wanted to keep giving this talk, and I already have plans for it. But more importantly, I wanted other people to be able to give it too.
That's why I've put the whole presentation online, along with its open-source GitHub repository.
The presentation is interactive: for instance, you can play with an embedding model, a base model, an assistant model and a tokenizer, and there are animations to help explain how things work. People attending can also connect with their phones to access a glossary and play with the models themselves.
I've also written a script explaining the storytelling behind the presentation, with additional notes in presenter mode, so you don't have to figure out how to present everything yourself.
And of course, I used AI to build it. It helped me a lot with the content and the animations, but the storytelling is my own, and I spent hours testing it, checking the sources and reworking it.
And I'd really love for people to use it.
You can give the full two-hour talk, take only the parts you find useful, or fork the repository and make your own version. Maybe for younger people, older people, or with a completely different angle.
If you decide to give the talk, please let me know on Signal! I'd be happy to help you prepare, and I'd love to hear what worked, what didn't, and how people reacted. It would also be great to keep track of how many people we manage to reach.
And if you fork the project, let me know too. If you want to make your version available to others but can't host it yourself, I'm happy to help with that as well.
Six months ago, despite having worked in machine learning for almost a decade, I had no idea about most of these risks.
All it took was a friend taking the time to explain them to me.
I'd love for this presentation to help other people have that same conversation.
Open questions
I'll leave you with a few questions that I'd genuinely love to hear your thoughts on: Do you think AI safety should be everyone's concern, or is it too technical and something that should be left to technical experts and people working in governance? Could there actually be a danger in having everyone involved in the conversation, or do you think it's the only way to pressure AI companies into slowing down?
Presentation: https://ai-exhibit.samverbo.com/
GitHub: https://github.com/sverbo/ai-exhibit
Signal: sverbo.39