Caoimhe Murphy - Italian Tech Week 2025

See you later! Bye!

So I guess we're out of luck for getting Niko's autograph, right?

Later.

Caoimhe, thank you so much for making the time. We're so excited to hear about Anam. I actually didn't hear this Sims story in the background, so I'd love for you to give the audience a little bit of background about Anam, what it is that the business does, and also more importantly, how you, how you got into it.

Yeah, yeah, of course. So played a ton of Sims growing up. I was obsessed with building. I mean, Stardoll was huge. Like, we dial-up internet, and I was like trying to figure out like how to play these games. It's like terrible graphics. And that whole concept of like designing something that fits your idea and how it interacts, and like, sounds a bit controlling, but like seeing it engage with other beings, like, really interesting. And that was just something that like throughout my life, I loved Stardoll. I loved like those games that you can go on and like build worlds. I was like, that's really interesting. Um, and then I had an opportunity to work at a company called Synthesia, which is a text-to-video avatar company, incredible unicorn based out of London. And while I was there, I met my co-founder Ben, and we were like, this is interesting, but what if you could engage with these videos? What if you could have a 2-way conversation that really felt like speaking to a human? Like, that would be amazing. And I was like, yeah, I've definitely thought about this for a long time. Imagine you could build this persona that you could control what it looks like, how it sounded, its language it spoke, and interact. That would be far better than reading a big long PDF or waiting for someone to speak to you in 2 weeks' time. Massively pushing my own agenda because I don't have much patience and just always wanted to talk to someone, but nobody was there. And now we're using AI to build that new interface.

That's incredible. And You know, you've raised money from some of the who's who of Silicon Valley, but also from some world-class entrepreneurs, including Matti from ElevenLabs. Just to contextualize this for the audience, could you give an anecdote of how somebody would interact? How would somebody see Anam? Would they see— how would they know that they're interacting with one of your personas?

Yeah, cool. Well, hopefully you won't, and you'll think it's a human. No, but really, What we're building is— I mean, Matti Kaimaan talked about the voice as an interface. We're the next level to that, or the next medium towards that, which is imagine a Zoom call. You're having a 2-way conversation. The thing that you're having a conversation with is AI. You don't care. You just care that it's better than reading a ton of websites, PDFs, watching a YouTube channel. And so that's the whole concept that we're building towards, is this interaction layer that just mimics how humans have communicated for millennia.

And I guess, um, would education be a use case, for example?

Yeah, so there's actually, um, a load of different use cases, fortunately enough. Uh, we see in education— so we work with Preply, which is a language tutoring company, to create AI tutors and simulation training. We work with Arizona State University to create simulation environments for their medical students to practice, like, diagnosing and how to actually improve their bedside manner. So there's Within learning and education, there's a huge amount of scenarios where people actually want to educate on the go. They don't want to wait for their teacher to be available, and our avatars have created a space for that. But we also work with companies like Vivint, who's a sales agent company. And I don't know how many people have bought SaaS software, but typically you need a solutions engineer, and we power the interface behind their AI agent. And then we also work in customer support. So you can imagine you go on Intercom and you chat to a chatbot. We brought that to now essentially feel like a Zoom call speaking to a human.

And I mean, one of the things that I've noticed specifically when playing around with Anam is historically when I would interact with AI personas, it'd be very flat. But I've noticed that yours responds to human cues. Could you just talk a little bit about how you've— you know, you've thought about innovating on this technology and where we are in terms of, you know, I guess previously being text or audio only to being a full lookalike for a human, and how should we think about that?

Yeah, I mean, it's an incredibly difficult challenge, and we joke a lot being like, we've got such an incredible opportunity, but God, we picked like the hardest challenge. What I mean by that is like any normal person understands what great communication feels like, right? You know what it feels like if we get on well, if we click, We're trying to reverse engineer that human connection of when you really click with someone and you're like, wow, they were so engaging. To do that takes not only incredible AI models to be able to build the structure of the avatar, but also real insight to user sentiment and human understanding of connection. And so throughout conversations, what we do is we work with our researchers to understand, like, what are the things that make humans click? Typically, it's like, I'll respond to you in less than a second if we get on really well. If we don't, I'll probably be a bit more reserved. I'll be slower in my conversation. If we connect, I'll be much more open in my eyes. And all of that information is how we use that to train our models to then create these interfaces.

And, um, so you have your own models or you work with some third-party models? How?

Yeah, so we build our own video foundational models in-house. Um, that's— my co-founder Ben is an incredible technical engineer. He's built them from scratch, which power the avatar structure that you see when you interact with on the video stream. We've also built the, like, engine that brings it all together, which helps to deliver this. We leverage text-to-speech models like ElevenLabs and large language models. Essentially, they're the, like, brain and the voice, and then we become the face.

It's quite impressive and exceptional to see how Europe is taking such a strong lead on building some of these world-class foundational models. One of the things that you touched on that I'm curious about is as you foray further and further down this journey of making your personas closer to human-like, presumably there are some ethical considerations that come up. As you know, we've heard about it's AI chaos. No, I forget the term.

The uncanny valley?

Yeah, it's when you basically interact with AI for so long that you Yeah, you forget really who you're interacting with.

Yeah, yeah. So there's different terms there and they're all just like jargon within the industry. But yeah, I think that like there's tons of ethical problems with this and like how we think about it is really important for us to actually be able to offer something that has value. One of the things that I consistently say to the customers that we work with, be it Henkel, Siemens, Oracle, is that like people don't care that GPT is AI, they care that it's helpful. Like you should be upfront and say like, this is your AI avatar. It's just a new interface and medium for you to engage with. In saying that, like, people can use our tool, take a picture of themselves, and create their own avatar, we bring in a lot of verification within that flow to ensure that the person is who they say they are so that the digital replica of them is truly them. But there's tons of considerations that have to go in, and that's on the proactive side. We also do a retrospective view of moderating it to ensure that, like, what's being said and that communication doesn't hit typical flags that would look like abuse or misuse. Um, so we're really thoughtful in terms of how to actually tackle this. It's never been built before, so we're learning as we go. We work with our customers to understand, you know, what might their reservations be and how can we tackle it together. Because, you know, as we're building this, like, we're learning how it's being used, and I think that just does happen quicker than regulations. We're very thoughtful on how to bring that in with companies.

And at the moment, um, I'm correct in— or am I correct in understanding that Today, Anam is predominantly B2B. It's, it's, you're selling to other businesses and brands for a variety of use cases.

Yeah, exactly. We don't build our own like consumer app. We power apps essentially. We power a ton of companies, either, they're all businesses though. And within that, it's the face to AI agents, the face to customer support. It's basically the face to their product is the best, I guess, basic way to think about it. But we haven't built like a consumer app. We want to power those rather than being the person that builds that.

And within these organisations, who typically are you targeting? Who is your counterpart? Who's the person that you're selling to? Whose problem is it for this to solve? Engagement of their customer? Could you talk a little bit about that?

Yeah, it's a really good question. And so typically right now, because as you can imagine, this is still kind of sci-fi, so some people— we heard on this step earlier on, Jeff was talking about, you see sci-fi and then maybe we'll build towards that. People do try and understand this, but it's hard to then bring it into an organization if you're not at exec level. And so typically it's the people who are responsible for the product roadmap and revenue. So we always end up working with the CEOs and CTOs. They're typically our buyers but also decision makers. Um, but when we get into those conversations, what we're really trying to prove is value, right? Because you could build this and like, you're like not really clear how it really drives any true business value. And so when we work and become the face to a product, we've typically seen we can drive retention up to 44% of the users who engage with an Anon persona. We've seen conversion rates from free to paid between 10% and 25%. So it's really driving revenue and retention for these businesses who create a new interface to engage their users. And that's happened across education to sales agents. It's not one market that it fits. It actually can It's pretty horizontal.

What's, I mean, anecdotally, so you shared education. What are some other anecdotes you can share of industry applications? And also more importantly, what are some areas where you think that this is not relevant or not there yet today?

Yeah, so there's definitely like the traction what we see today, which is within language, within learning and development. And that is just because that when people, there's such a need for this. And so the need is so, so strong that the quality of the avatars, while they do feel real, if they're not completely natural, right, if the eye movement for whatever second conversation moves left, like, you don't really mind as much because it truly solves the problem. Sales agents is another use case that we work with Henkel, who, um, their like company Schwarzkopf initially that started as like AI hair coaches. Now we're moving into a conversion use case to work in e-commerce and retail with them to help users know, understand what they're purchasing and how to use it. We also see customer support and I guess like customer enablement within that. The tail of use cases that we still haven't unlocked yet, or we've got like early signs of signal, but I don't think that the model is good enough just yet, which is like AI therapy, AI companionship. We do have customers using us for like healthcare navigation, which is really interesting in terms of like trying to really understand If you've got hypertension, you know, you're not going to see your consultant every week, but it's important that you stay on consistent medication and exercising to the schedule that they've given you. And so we've created an AI persona that acts as the concierge through that experience for helping that patient through that navigation of their illness. And so that's been super interesting too. It's not just, oh, we exist here in L&D and it works well for AI tutors. It's like, Actually, most products could do with a better interface, be that voice or face or text. And we're just seeing these companies now really open up their mind to that. And so it's been incredible to see, but we definitely have that tailwind of where we're not quite there yet because the trust level is just a little bit higher.

And we've seen so much advancement in text-to-speech, text-to-video. OpenAI launched Sora 2 recently. Where are we in this journey? You know, and if you were to articulate what you anticipate, I mean, nobody loves predicting the future, so I don't think anybody here is going to hold you to it.

Yeah.

But over the course of the next year, 2 years, where do you anticipate this industry going?

Yeah. And so it's like, I mean, we talk about it all the time. So the way that I would see it is like text-to-speech is quite adjacent to us, text-to-image adjacent, and text-to-video. You're still building foundational models, but they're not— the constraints are different. The outcome is different that you're looking for. What we have to build is something that is a very specific model that captures like expressivity, that feels like that natural conversation. And that model needs to run at like 170 milliseconds. If it's any larger, the whole experience breaks down because it's going to take much longer for the persona to respond to you. And so I'd say we sit in our own kind of market there. So I couldn't really speak to like text-to-image or text-to-video models for what we're building. And within that whole experience, like, I mean, it's probably going to be our forever work. Um, but I think within the next like year to 18 months, the level of improvements that I've already seen from our model that's going to be released in the next few weeks It's just so fascinating to the point of where you'll have a persona that's like, hey, really nice to meet you. I was doing my hair the other day and I saw whatever, right? They'll be way more expressive, similar to how like 2 people communicate. I gesticulate a lot. I'm waiting for our personas to get to that level of gesticulation.

And one of the big things that we've seen, especially in foundational models that, I mean, how do you think about the barriers to entry? What are you training your models on? How do you think about the competitive landscape and what, yeah, what keeps you up at night?

Yeah, I mean, all of it. No. So yeah, so when we build those models, I mean, my, again, co-founder Ben has been phenomenal at assembling a very high talent dense team. We've got 4 incredible research engineers. What they've built doesn't exist in research papers. Actually, we joke with Peter and we're like, you need to write a research paper. Like, this is phenomenal, but also we need to build the models. And so they've actually come up with a specific way to create these avatars that nobody else in the market has done. That's through like thousands of thousands of experiments. And so that alone gives us a competitive edge, which is phenomenal. Um, we also train on data. Obviously data outlives models. Um, the data that we use to train our models is like 2-way conversational data. So, uh, how we structure that is we go to these data companies, we'll give them an agenda of like what we need. They'll go capture really high-quality two-way— hopefully this content might— we might be able to use this content, um, of really high quality of 2 people speaking and engaging. And then we use that to train our models. Um, we can't use anything like straight off the internet the same way that other companies can, for want of— for good or bad, I guess, um, because the quality is so— it needs to be so high. And that's one of the things that we're realizing. But that most in and of itself, of the research models that we've built, um, is— I mean, I don't know, like, I don't know what great other companies are in labs, but as of today, that's a huge, um, part of our head start, I would say. The other part is that, like, all the technology that brings it together, we've built as well. And so I think that gives us, from a technical perspective, a large moat. But I don't know, like, another small company somewhere in Europe, I'm sure, that's building this as well. Um, and so what we've seen is like we started this 2 years ago and everyone was like, well, of course this is the future. Like there's gonna be so many companies doing this. And there's really only like ourselves and one other in SF that are really trying to push the barriers here. Um, so yeah.

And, um, I think I have 2 more questions if we have time. One is for the builders in this room, how should they think about working with Anam? How should they think about integrating or— Yeah. Yeah. Is that—

We've got an API. Off you go. It takes like 30 minutes to get up and running. I think the best way— there's 3 different ways of deployment of our product. One, turnkey solution, we'll do everything. You drop in your system prompt to tell the brain, the LLM, what to do, and then you just build your sim in the most simplistic terms. Second version is you've got your own custom LLM. Pretty much every company we speak to is investing in that right now. They're trying to understand, how do I take an LLM and use it for my needs or how does it solve this problem? Um, if you have built a custom LLM, you basically can just use the API and we'll be the interface to that. And then the third one is where you can go to a company like LiveKit and just use us as a face provider. So there's very different ways. I guess if like builders in the room, like go to anim.ai, you can play around with it, API, deploy it into your website and off you go. It's pretty simple.

Caoimhe, this has been so exceptional. Thank you so much for making the time. Please, uh, everybody put your hands together for Caoimhe Murphy.