top of page

5 Levels of AI Assistants


Hi everyone. My name is Alex from rasa and actually have been here now the third time in a row, so, I was here two years ago very first conference, which was great up in the Presidio last year as well and this is the third time in SF. We also, had a great event in New York. so, I'm a big fan of this conference and al so, because it's bit smaller and like actually allows for a lot of conversations in between. so, I'm happy that you're all here and I really want to talk a bit about like expectation setting.

In conversationally I and the title of the presentation is like between super damn and super intelligent and you're going to hear more about like what I actually mean with it, but maybe just really quick on Raza and like what we do and so, we an open source framework for building conversationally.

We started two and a half years ago and pretty much have been growing ever since very fast across the developer communities. We actually had a developer conference here in this venue. Yes. day some people still wear the rather shirts from yesterday as if you interested to learn more about like how they use rasa and the wild pretty good to approach them. And other than looking at my slides. I mean I can obviously promise you the best tool that you've ever seen but it's always good to talk to real developer users. And so, as I said, we've been going fast across different Industries and actually yesterday and this number is not updated here, but we announced that we have one and a half million downloads now, so, this number is from May so, it gives you some perspective on like how fast we are growing. Really excited about this and I also, going to talk more about like what we actually work on. But as I said, the topic is more about like expectation setting and I think that's just something that we've seen early on. In this space like pretty much ever since Facebook Messenger opened up the API. Everyone was excited and people were like Bots the new apps, right? And we probably still remember that back then Nadella from Microsoft was very vocal about it. And so, here we are now three years later. It still hasn't happened yet. And so, I think the question is al so, a bit like Why didn't it happen and from our experience al so, talking to developers but al so, Executives in big companies, and there was always been like a lot of This for these chatbots, right? It's kind of human level intelligence. Ideally you upload all your two million transcripts that you have from human to human interaction and then it magically works and just figures out everything right? so, that's kind of the expectation level in the early days. And I think still a lot of people are on those expectations at the moment and that's really what we kind of defined here as like the autonomous organization, right and so, as much as this is like a super exciting Vision to work towards to and obviously like for business Executives something really interesting as well because you can automate lot of the work force that you have at the moment, which that's totally separate side conversations. Well, which I'm sure most people here think about as well as like the ethical side of this and but then as much as this is exciting to build and work towards so, from a technical standpoint, it's just really hard right and so, I think most people got al so, very early on delusion lies with like building a chatbot and realizing that it's actually really hard and that's how we actually started rather. so, we were in my kitchen in Berlin. Go phone and I and back then we used with the day. I dialogue floor back then called API dirty. I and a bunch of other tools and really try to just build a chatbot. Right and everyone was saying chat Bots the new apps and we just like I really stupid here because we couldn't really figure it out how to build something more interesting. And so, that's I think really something where the conversations at the moment more and more mature in a way that people have more expectations that are real for that. Right? And so, what we try to do with these five levels of AI assistance is really to set a bit of the frame for the conversation right and maybe it's five levels maybe six or seven. It actually doesn't matter. We stole these five levels from autonomous driving because it's just a big people can kind of relate to and But I think it's a good framework to start thinking about so, if you think about like notifications is like one-way conversations and that happened pretty much over the last 10 years ever since the iPhone was released. That's a very good way to think about just the easiest chatbot that you can build right just send a message and then I think a lot in the market right now is very much like FAQ based. It's like one question and one answer and that's what we called level 2 and that's I think what most people would al so, call chat Bots and then what we've been back then in my kitchen and ever since try to build and try to let other developers build is what we call a contextual assistant and I'll talk more about what that actually means and then there is a whole other level to this like to make it more personalized right have an assistant that actually knows you really well and knows your habits and can make recommendations and all of that and then the last piece is really an entire company pretty much running on and on different AI assistance and we had lemonade here yesterday presenting what they've been automating and around customer support and I think that's Exciting way to think about it, right like customer support is one sales marketing Etc. And so, ideally, they all work together. But this is I think a very, very long Vision in the future to make this happen. Right? And so, it's much as like IBM Watson will tell you that this already works now from our experience from the developer side of things it doesn't get and that's pretty much what we want to build as a company we want to build like the operating system for those autonomous organizations at some point. And so, we are so, I think as much as like three years have passed now in the early days of conversationally AI and so, as I said, it's like very much like one question one answer. so, if it cues and so, one way to think about it here I needed I need to renew my renter's insurance. How much will it be? And then you just give a response and send it to somebody right? so, in this case, all you need is like a good NLP engine and then you map that to intense maybe pull out some entities or so, and that's it and the reality Those that most people don't ask questions like that. And even if they do there is more context that you have to fetch and that's really what we mean with like level 3, right and so, having this full conversation and getting this full quote for your renters insurance is I think what most users expect right so, they expect these natural conversations, but then at the same time where you can really I think convince them that this is something cool for them to use is once you do something else in a different system, right? so, once you calculate Relate the actual renter’s insurance Price. Once you let them buy it through that right once you do a checkout process once you save something into another system, and so, that's really all I Quince you. Turn the light bulb on also, right, and so, all of that stuff is where people are like wow, that's magic. And so, the most successful AI systems. I think that we've seen out there do something like that, which usually al so, requires some form of locked in experience. And so, Google I think had a good example of that one and a half years ago and where they had Google duplex on stage very famously and it's also, obviously launched now available and that's super cool. Right because you can actually get this appointment scheduled you can have like something in your calendar for this and so, it's interconnecting with all of those systems and so, as much as everyone once said that most companies struggle with building and so, one other good example is Adobe here and if you can quickly That's a fan, please. Fighting imager video can be like finding a needle in a haystack using machine learning content faster find photos of flowers with people more red a sombrero? And you had found somewhere images more authentic Urban edit Photoshop. And so, I think that's a good example of how it keeps context. Right? so, you pretty much have a search query at the beginning and then you add something to it and you take something out of it again. so, that's a very specific use case but a good example of contextual in this case actually bread the head like not so, much like voice input or like the longer like text input but it's obviously something to add as well. And so, the question is really like why is it so, hard to build all of that and I think what we found back then When we started working on Raza, I was really that it's not just a know you that you need for those systems, right? so, as much as like it's important to understand what somebody says the other big piece to this and I think most developers in the audience would agree and is really that it's all about like, what do you say after so, like the dialogue management if you will right and back then what we did we just had a lot of handcrafted rules. So, we would say things like, you know, if somebody asked for this with the search query that Brett just did with the image, right? so, if He asked for a woman and in this context, he then al so, asked for flowers and combine this and in this case search for flowers and women, right and that works for a demo, but it really usually breaks when it's in production and especially once you have multiple use cases and so, the same way as for nlu, you can't just go and make up all the rules that are in the language which obviously failed over the last couple of years and since like deep learning has been around and obviously with bird. Our people really excited about like most sophisticated natural language understanding the same is true for dialogue management because you can't really just come up with all the different rules that you have in your system. You have to pretty much learn from conversations and ideally and that's really, I think where a lot of people get excited about when it comes to AI is that you can generalize to new situations that you haven't seen before right. so, if somebody interrupts the conversation and the system is intelligent intelligently able to learn from those interruptions and then in a different context al so, help handling these interruptions and its really what like Raza and our dialogue management is all about because we use deep learning for all this dialogue management. And so, as much as deep learning is great, right? I think it's also, note here to make it doesn’t solve everything right and working with big companies. We also, realized that as much as deep learning is something that they appreciate a lot on the other hand. They al so, appreciate rules a lot, right? so, they appreciate that you have to as a customer. Tentacle yourself through like five different things and you don't want ml to predict that this is really Alex you really want make sure that it is Alex. so, I think merging that with business logic is a whole other topic. We talked about four hours and but an interesting one as well and possible in Raza. so, then that's my favorite one. My team would rather doesn't like it anymore because I've been showing it over the last four years, but I hope there some new people in the audience. so, I hope you appreciate it. But it's Steve Ballmer on stage back then at Microsoft shouting developers. And I think as much as this is true, right so, building level 3 building contextual assistance, you need developers. I also, more and more think that they are only one small part of the team as much as like in the early days of the web you had only Developers Shipping websites, right and only developers developing websites. And as we know now there are whole teams of people that work on this right you have a designer you have ux designers. You have product managers all of those different people that ship websites is something and we just had a panel on that yesterday. It was smit here is something that we see more and more across organizations that ship the best conversationally I and so, obviously from Silicon. Known famous product team of like different people, right? They actually didn't have as many designers but just as a quick example, so, you pretty much need a team. I think that's what it boils down to right and this team might have more developers and data scientists or they have less so, that I think depends what you really want to ship but in the end it's not just a one person show and it's really al so, the big Trend that we see in the market right now and for level 2, it's little different right? Because if you think about it, all you need there is actually some UI where you match some intend to a response right so, that's actually fairly easy to do and you can almost and there are lots of tools that do that allow a business user who's like the domain expert to just come up with the responses and this person can obviously al so, map it to the input. But then once you go beyond that and you want to integrate into your back-end systems and all of that it's becoming really tricky and that's really what we think conversationally. I need some more and more and so, I was talking about like teams right but those teams al so, use different tools and that's exactly what we talked about. Yesterday and Vittorio’s here. For example, I saw somebody else running around earlier. I can't find them anymore and but folks from those companies are al so, here today and it is really like in the end if you think about it as like product management the same for website, right? You probably have a design in sketch or whatever tool you use and then you use a test it after this user test. You probably talk to your engineers how you can implement it and then you implement it then you get data from this implementation and then you probably improve it. Hopefully if you're a product company and so, those iteration Cycles is something that we see more and more happening and something that I think the industry very much needs, right? The problem is it is really hard right? so, it's just really aspect to my point. It's like very early days. so, it's not like in the web world nowadays like a designer can go to an engineer and say well, can you please implement this footer or this navigation bar on top right and everyone knows? You're talking about and that's really something we talk more about this. I think big challenge in space at the moment. That's really like one way to think about it in conversations is like as I just said like you have this like navbar in the website. What is it? What is the equivalent to that in conversations? Well, it's different dialogue elements pretty much right. so, maybe somebody asked the question back and you want to give you want to call it like a reference question or something like that, right or somebody agrees to some question and it might be another element and what we're really interested in rasa is like how you can use these elements al so, for the machine learning for the dialogue right? Because Again, that's kind of about pattern matching as well. If you know that in certain like states of the conversation people always like divert then this is something really interesting to learn from and improve over time. Yeah, we need advances in both applied research and tools and I think that's also, really what it's all about. so, what's new and research and just a quick update on that a lot of new research and bird and GPT to are still all over the place. I think every week there’s something new cool happening in NLP at the moment, which is a set to my point earlier is really al so, only just one part of the whole like conversationally I stack if you will and GPT to obviously as well. So, any Around like natural language generation is super exciting and from our experience and even further away in the future just because in most companies you actually don't want to neural net to create your messaging right because it's a bit risky at times, right? I mean today from Microsoft I think is a really good example for that still and but it's happening and it's super exciting and we are doing a lot of research al so, on that side mostly on the nlu side right now and then glue is out super. Louis in so, what does it even mean so, glue has been around like an NLP research for quite some time and it's pretty much like a way how you can measure how well your model. Across some certain standard tasks and so, because it has been so, much advancements in the last couple of months and super glue is now like a new challenge kind of if you will for NLP systems to be understood, right? But NLP systems do a lot of other things they do question-answering text extraction and all of that. And so, that's all part of that. And as you can see like just the benchmarks are getting much better, better, better. And then al so, what's really interesting I think is like future Generations generalization across dialogue test and that's kind of what I was alluding to earlier and how can you learn from a dialogue Behavior? If somebody interrupts your assistant, how can you not pre-program that again into the other next Interruption but this person can say handed right and humans are pretty good at that. so, if you teach somebody how to handle an angry customer who's calling into your call center, and this person probably really good at handling this person next time I didn't Think better over time and it's really what we mean with that. so, what does it mean from a developer view M Allen my co-founder wrote this blog post couple of years ago that we don't know how to build conversational software and it's kind of how we started with rasa al so, and what we've done over time? Is trying to solve this problem of like kind of do it yourself NLP for but developers so, really like not using with ODI anymore but like packing it into your own system like doing your own animal you for level 2, right? And then as I said earlier, we launched there was kind of the first rather t-shirt here by the way still very much in love, and then we launched this dialogue management, which pretty much solves this problem that I just talked about for conversationally. Building contextual assistance and there was like two years ago. We're still very much working on this. It's like in the 1.0 release now so, pretty mature and we're really excited about this so, much stuff that still has to be solved and but it's really a guru. This won’t solve all of that. And so, that goes back to this like workflow and Tool question that I brought up earlier and I think we don't have enough time to go into this but you know, this is kind of the way how training data looks oops, but it's super hard to manage. so, what we've done is like build a UI on top of that. And that pretty much allows you to learn from real conversations and like back to the point really like throwing something out there as fast as possible and then you can come to me afterwards if you want see that demo. And so, important to understand like what it's not and what it is because a lot of tools in the market do your eyes and it is definitely a tool to learn from real conversations, but it's not just a click and drag but tool that's not the target audience is we're working with and I think it's really hard to build level 3 with just a click and drag tool and at some point in time we might have that but this is not what rather axis our new product. Transfer learning, I talked about this already really important for like what else is needed in level 3. We're really excited about end-to-end training as well. so, pretty much dialogues or not having nlu and data management separate and then really also, the personalization aspect as I said cool. I'm here the rest of the day come and find me. We are also, hiring here in San Francisco and Berlin as well pretty much across the world. And if you're interested in that topic come talk to me as a company. Our goal is to build the standard infrastructure layer for conversationally. Pretty much the same way as use HTML JavaScript CSS to build web applications. We think something like that is needed to build ai lessons across all channels. And if that’s something that's interesting to you come find me. There are some people from the rasa team as well here and tomorrow tie from our team al so, gives a workshop. so, if you more curious about that. Come join tie. thank you, Alex one quick question. one question Two three, we got five minutes. so, we have to be really quick with it. All right, no questions one question. There you go. Thank you. Alex question for you to reach the level 5 autonomous organization does deep learning have a role in that just to you know, Echo some of Peter Voss has concerns. so, is your question of deep learning will bring us a level 5. Yeah, it's a really good question. I have to say I don't know so, and level 5 fast is something that is just a way how we think about the topic and I think the world is still far away from level. Five overall and we'll see if deep learning will bring us are, I think it's a probably 10 to 15-year project for the future and overall level 5 and we don't know yet. I think there’s some good chance, but we'll see. I think it's more about like the types of companies that will get there. And I think that’s something maybe just one sentence around that and what we do see is that there are more and more companies who see themselves as like ml first companies or at least powered by him out companies and that's definitely something that I think is really required, right? I really quick we have like one minute. When it comes to training the model, how does how does the process work for you? Like is it more manual or is it you can look at Aransas of real conversations and then Dragon drop to these intents with or do you have to like use command lines? How does that work for you? And so, what most of our users do is that they kind of come up initially with some form of like beta version of their assistant and they in the beginning like the team kind of so, the product team that Mentioned M comes up with like different utterances and like design right using something like pot society for example and plugging it into rasa and then there's usually a phase of like internal testing where they collect more training data and like improve the flow and then they would we recommend and still most companies. I think I'll bit like cautious about this is to put it out there in the wild as fast as possible. And I think the reason why a lot of companies are conscious is because they see then conversations fail and that's totally fine. Right because nobody ever compared in a Distant to your FAQ page, right? so, you don't know if you have the equivalent of an FAQ page on your website. Nobody talks about bounce rates on FAQ pages. Right people just assume mean they work in that sense, but do they really solve your customers problem?

I would say most teams not right. That's why they didn't call in on your phone. And so, I think a lot of companies are still afraid and I wish not many and I think it's more and more changing al so, through the success stories that we see in the market and but it's still something where people are cautious about Okay. See you at mention a few different ways in which level 3 is hard. How do you see those being easy as in what are the similarities amongst the different companies who have actually achieved it already that you see consistently and so, you mean what makes it easier to do it? Yeah. Yeah. so, there's a few different examples of companies that have achieved level 3. What do you see as the similarities amongst those companies that consistently allow them to achieve that level while other companies have been able to do that yet? Yeah, so, I think the Main similarity is that they have more than one person and it's usually more like 5 to 15 people and they have different roles. so, there's a product designer. There's a ux designer that or conversational designer and you have a product manager you have developers you have data scientists who then like improve the models over time, right? so, that's definitely consistency that we see and then those teams tend to usually work with something open source, and it was you have a bias on what they should be using. Razza open source framework, but and it is I think that's kind of bull sound. I mean the same way as like the most successful mobile applications are al so, not shipped by one developer right there. I mean a B&B al so, is a great example of like tons of different product teams working on that. Yeah. Let's get a round of applause for Alex. Thank you.



bottom of page