So um next in on in on in the line um is a company which is not lost in translation. It's Steven Masc from DL. He's the chief scientist at DL mathematician and AI scientist who actually fell in love with deep learning according to uh his LinkedIn and he will speak on building an AI powered organization. Lean my friend please with a warm welcome say hello to Thank you. Thanks for the kind invitation. Thanks for having me. Um, within the next 24 minutes, I'd like to take you on a little bit of a journey. The title says, okay, building an AI power organization. That's exactly what Debbel has done for the last almost decade here in Europe.
And now like 20 hours into this conference with a great speaker dinner yesterday uh and all the discussions I've had today I heard a bunch of questions there was there are questions about like AI infrastructure there are questions about like access to data there are questions about like generating value how to approach customers how to deal with security concerns and at the end of the day the question really is okay how do you successfully build an AI company within Europe. So we have a couple of proof points that this is possible. There are the alifas, the mistroud, the black forest labs of this world. There's also DeepL and while I don't think that I can give you a general answer that you can just copy paste to your own organization.
We've been thinking about this question for a while and I think we have a couple of answers and more importantly guiding principles to navigate this landscape. Before I dive into that, let's quickly Sorry, I don't think that this one is working. Okay, let's quickly introduce DL. Who is DL? There are many angles you can take on this, right? Like DL is an AI native company founded 2017 in Cologne, Germany that has been competing with the hyperscalers of this world in the field of language AI. Um it's also by now a multinational company approaching a lot of business com customers specifically. It has expanded its portfolio quite a lot. It is vertically vertically integrated. So it owns its own um infrastructure.
It has its own data access. It builds its own products. And that's really the reason I joined Dwell to begin with. So when I got interested in AI and specifically in NLP, I was looking for a company that does it all but does it all with impact. And what I mean by this is I wanted to find a place where I can apply my knowledge uh in AI in a way that is meaningful from a technological point of view. So keep pushing the boundaries of what is possible to do with these models at any given point in time. But I also didn't want to end up in a research lab that is basically remote from actual impact in this world.
I am a strong believer in like the positive impact of technology and as such I want this impact to show up for me personally but also for the users that I'm ultimately uh building these technologies for. So on my hunt for a suitable company, DL came up and it basically ticked all these check marks, right? DL back then nowadays does cutting edge research in uh deep learning and is building its entire product suite on top of this research. That's challenging in in and of itself, but it doesn't stop there, right? Like deep is not a generic model provider for language AI. Instead, it takes this journey a step further and figures out how to leverage the techn technology to actually deliver value to customers, which is the entire philosophy that Debbell has been built on.
If we look at what how this looks like in the past, this is a screenshot taken in 2017. Fortunately, our web UI doesn't look that way anymore, but I think it's nice to see where we started. In 2017, we launched a translator and that's the product that we are most well known for. It does what it says, right? It translates text to text and takes into consideration all the nuances of natural language and for the first time ever like globally reached the level of quality where you can't tell apart native text from translated text anymore. I invite everyone in this room to like install the browser plugin that we have for uh DL. Go to a random website whose language you don't understand and just read it and see whether you can tell the difference.
I can't anymore. Many of our users can't. And that in itself is a great achievement. We obviously didn't stop there. Once we had established baseline quality and had pushed it um to a level where this technology is not just possible but actually useful, the question is of course how can we do more? How can we do more for our customers? How can we solve more problems that are just thorny and difficult and that keep uh adoption of this technology limited? The glossery is an important answer to that. The idea basically is we want to understand not only language but the actual context that our customers are using our tech technology in their business backgrounds right there are specifics in terms of how they want to uh deal with language and over the last 8 years we have been building incrementally more and more technology to basically grant access of users who are not AI experts into our systems so that they can tell us what we should do.
The glossery is an important milestone in that regard and I'll touch on a couple of more developments in a bit. Next in line obviously like skipping a bunch of steps in 2023 Debbel Mercury went online. DBL Mercury is the first 8100 super pot in Europe and what this is at the end of the day is an AI factory right. uh we have had our own data centers before 2023 but this is the first time that we leaped to the very bleeding edge of what is possible in terms of compute and the reason we we did this um is that LLMs obviously were the way to go in order to improve our products. This I will cycle back on but keep it in mind for now.
Finally 2024 and beyond we have launched deeper right. So we are not just a translation provider, we also help with writing assistance. Uh since then deeper voice has become generally available as well. So you see that we expand both the breadth and depth of our technologies in uh an attempt to help individuals but more and more importantly also enterprises address their communication challenges. There's a tiny diagram on the right where you can see this basically summed up the high level view of this is again it's vertically integrated. We try to take cutting edge research and figure out how to exactly apply it in a way that delivers impact to our customers with all the complications that are required to actually make this impactful and scalable at the um yeah at the scope that we are trying to achieve.
This is very much the exception. This is not only the exception in Europe but worldwide. There are very few companies especially in AI but in the grander technology scape that try to vertically integrate themselves. And the reason is obvious right this is very hard to do. You're dealing with all sorts of challenges. There are infrastructure challenges. There are data challenges. There's the cutting edge research that you're trying to do. There's product development, there's engaging with customers, there's sales, there are many, many moving parts. So why are we doing this to ourselves? The reason is that we think this is a strategic advantage that we have over basically everyone else except for a few hyperscalers. So here's how we operate.
I think the arguably most important thing in technology and especially in AI is that you need to know why you are in this game. For DeepL, the mission is very simple, breaking down language barriers. So everything we do, we try to advance our mission and the mission really means we want people and companies to communicate more efficiently and help them succeed in their domains. That means that we are not doing a bunch of stuff that we could do, right? Like we are not interested in image generation. We are not trying to uh become the generic model provider for everyone else to hook into. There are a bunch of nos for every single yes that we come up with.
And I think that is super important because otherwise you just keep overwhelming yourself and don't gain traction in any of these dimensions. Second, and I think this is a part that many companies get wrong, especially startups. Um, we try to deliver solutions, not technology. Technology is the fundamental driver behind our solutions. But when we think about product development, it's with a customer problem in mind. Right? Here's a challenge that current technology is not sufficient to solve for our customers. What can we do in terms of research, in terms of product development to alleviate that problem? And this is a cycle that we rinse and repeat all the time. Which means that we remain very agnostic to the shape our solution has at the end of the day.
Sometimes it means that you need to build sovereign LLMs, right? Other times it means that you just need like this little tweak and everyone everything becomes a lot better all of the sudden. This if you if you pay attention to what DL is doing, this shows on the outside as well. This is not just internally informed. For example, you will notice that deeper doesn't publish scientific papers and the reason are basically twofold. One is we want people to be in deep doing research there for the sake of product development and product development has different agendas than um academic research. And the other one is we want people to care about the impact they are having for customers not some number on a benchmark they might be able to publish.
These two things are very closely related right like everyone knows that uh quality which essentially boils down to both fluency and adequacy in translation is super important and there are many leader bots out there that try to capture this notion. What we have found over the last eight years is that leaderboards or and synthetic benchmarks, no matter how well intentioned they are, usually don't capture the actual underlying problems well. They are easily measurable. It's nice to compete against. They make for great headlines, but that doesn't mean that they actually have lasting impact. So um as much as we can we try to focus on the letter not the former which means that at the end of the day what matters to us is applied in the intended setting what does this product do for our customers?
How does it help them succeed? If that means that we look great on benchmarks that's great that's a bonus and it certainly is part of the consideration in model development but it's not the point. Um and that I think is one of the reasons we have been able to compete against the tech giants of this world as well. So how do we do this? For the entire time I have been in AI and I think this is only increasingly true. AI development has been very very noisy. There are trends popping up all the time. There is a sheer endless number of things that you could potentially do. There are advantages, advancements, twists, evolutions, all sorts of things that basically can keep you occupied for your entire career.
And figuring out which ones of these developments will actually matter in the long run is an important task that we have debel. That means that we will not jump on every trend when it appears, right, and then dies off again, but instead consider the larger picture. Ultimately, what is the limit of this technology, right? like if you had infinite capability, what would you do? And what can we do today to build incremental steps toward this goal? This um allows us to gather more of the signal of what everyone is else doing of all the internal ideas that we have while trying to again the best of our abilities. We're not perfect in this by any means, not be distracted by short-term advances.
Um and ultimately what this allows us is to do more product development with way way fewer resources. Similarly, there is this notion that failing is a bad thing and people talk about this all the time. I think the stance has improved quite a lot. To me I don't enjoy doing poor in an area but I think it's the best teacher that we essentially have. So whenever we are releasing a new product, a new model iteration, we pay very close attention to first of all, have we improving on all the dimensions that we set out to improve and this is usually true because quality assurance is super important for us, but also despite all the advances that we have made now, what are the remaining challenges, right?
Where do we fall short beyond behind expectations and what are the next challenges that we can anticipate that our users might not even consider right now that we think we can unlock with current technology. Clarify depicted here is an interesting example of that. So for the entire history of uh AI translation people have sort of thought about this as an input output relation right text comes in in some language text comes out in a different language hopefully it's adequate it's fluent and everyone is happy this only gets you so far at the ultimate limit is not to provide the best AI translation system but that but to provide the best translator period. And that means a bunch of things.
Uh, one of the things is that first of all, context is super important, right? You can't just take the text in isolation. You need to know more about your users. Features like style adaptation, formalities, the glossery are all attempts to capture more of the relevant metadata in order to do translation well. Even if you do all of this perfectly, you will still fail. And the reason you are failing is because sometimes the knowledge required to do the job well is not captured by the metadata that you considered. So even if you put the world's best uh language specialists all in a room, gave them infinite resources and time, they wouldn't be able to figure out how to satisfy your requirements without them having to talk to you.
Okay, that's what everyone is doing already when they are dealing with human translators. But no one is doing this with AIS. Instead, they basically say, okay, like AI translation gets us this far and then we follow the usual process. So we have a human looking at the actual output and they manually annotate and edit that text. That's great. That solves a bunch of problems. But I think we can do more. And what we specifically uh tried to do with this technology and are now actually doing in production is provide a better interconnect between language specialists that we know are our core users in companies anyway and the AI systems that try to help them as best as they can.
So clarify essentially looks at this this corpus that you want to translate and tries to figure out okay like are there any areas in this text where I was forced to come up with a translation but I'm not super sure that I got it right essentially is the information missing that if a user could tell me I would be able to do a better job and that's exactly what clarify does. So it looks at your text, it looks at all the metadata and it continually asks itself, okay, could I have done better? What would I have been granted more information? And the answer you you answer very often is um yes. So instead of basically trying our best and guessing our way to an acceptable answer, we want to capture more user input and help users to provide the missing bits and pieces of information to then come up with a great answer.
That's what clarify does. And I think it's an important example in the general mindset that we have that we view progress as these like layered curves of sigmoid functions. If you allow me to show my math background a little bit. So essentially every technology that we adopt usually doesn't work in the beginning and then all of the sudden it gains traction and it accelerates for a while. Right? This is what exponential growth looks like. And then all of a sudden people are surprised that progress starts. And the reason progress starts is just well it's an S shape an S shape at the end of the day. So it only carries you so far. Which to us means while we want to capture current technology and accelerate this as much as we can.
We always keep up keep an eye on the horizon and anticipate okay what's next? what's the next S curve that we need to hop onto to move beyond the capabilities that we can foresee and understand today. Let's look at this in practice. So this is a it's it's a rough diagram but I think it's important to see this in in this context because this has been taken from a company event we had in 2022. 2022, three years ago, almost to the day, is a very different time in AI development. ChatGpt didn't exist, which is kind of unimaginable today. And we were asking ourselves the question, how can we do more? That's always what we keep asking, right?
We keep asking asking, how can we do more for our customers? And the realization was basically, okay, like we're doing pretty good on quality. We are solving the problems that we can uh capture with the technology that we have today increasingly well and we will continue to do this. How can we move beyond that and there are two obvious directions. One is we want to have better human AI collaborations which to us means building an assistant and the other one is of course beyond just transformation of text we want to help people generate content and those are the two angles that we have been thinking about and tackling to various degrees. Specifically as a realization, as an outcome of this very very simple diagram, we decided that 2022, early 2022 was the right point in time for us to move away from the custom model architecture that we had employed so far toward LLMs.
At the time, this was not entirely it was not entirely clear that this is the right bet. It was very clear that it would require a lot of infrastructure to get done and that we were missing a bunch of skills to do this well. Nevertheless, we decided, okay, whatever. March 2022, we think that we can do more if we move beyond transformative AI into generative AI. So, we called up Nvidia and basically told them, okay, we need help, right? Here is what we currently are able to do. Here's where we want to move. Let's build deep mercury at the end of the day. Let's build a super pot to unlock capabilities that we currently don't dare to dream of.
And the actual outcome of this is captured in this slide. So 2022 the decision was made we will reshape the entire organization from its technical setup toward LLMs. And this was done in a couple of steps. But at the beginning of 2023, we had ordered our supercluster. We had transformed 80% of our research initiatives into this uncharted territory of developing LLMs and we were awaiting the installment of Deep Mercury. In 2023, Deeper Mercury came online. That's this is our first 8100 Super Pot. It resides in Sweden in a nice echofriendly data center and has since then powered all of our product development essentially. So again what this has done is open up the possibility for DL to train frontier models which in turn allow us to build much more assistive and generative features into our products.
You will see this pattern over and over again, right? We start with a problem and then figure out what tech to build, what skills to acquire, how to challenge ourselves to get this done. April 2024, the first actual user outcome of this initiative. We published Debbite Pro and this would have not been possible with our previous technology. It kind of worked and we had a bunch of successes, but it was very very difficult to scale up because language at the end of the day is very varied. And if you want to if you need to build in every single feature basically manually, this is a game you will never win. There's just too much to do and your tools are not powerful enough to grant you leverage.
LLMs do and hence we were able to release Debite Pro. uh soon after that June 2024 we released our nextg models. So initially there were some questions and concerns about okay these models are more powerful they are way higher quality but we have spent a bunch of time improving the previous model generation and it's just a lot more efficient which means like lower costs and in terms of actual inference but importantly also lower latency and some of our customers a lot of our customers actually really care about like a snappy user experience. Um therefore we decided okay great like let our users tell us what they actually need and let's release them side by side.
So we continue to host our previous model generation um and offered the nextG models in addition. The very clear signal we got um as a result of that was twofold. People love the gains in quality and capability and they hate the downsides of slower inference which then informed our next research agenda right let's figure out how to do this LLM thing but in a way that is actually quick and that's what we have been doing um starting June 2024 again in close collaboration with Nvidia and not only DL but the entire ecosystem has made tremendous progress figuring out how to offer LLM MS in a way that is efficient and quick. The state of uh the art in terms of the software infrastructure 2024 to now doesn't even compare like the the number of advances are uh mind-blowing to be frank.
Then 2024 I've already talked about this uh clarify was released. this feature. We had known that we wanted to build this when I started at deep in 2020, right? This was not really a debate. Obviously, we want to enable human collaboration. But similar to deep, right, the problem is we need a tool powerful enough to do this in a way that addresses enough of the problems that we see out there that is actually worthwhile. And the previous generation did work for the few instances that we addressed, but failed to generalize. Now that we had models that generalize much better, we were able to launch um deep clarify which is first of its kind in the entire language AI ecosystem.
Today the story of course continues. We are building a new super pot now based on the Blackwell architecture. It's a DGX B200 super pot I believe. No, it's there's an NVL72 missing in there. What it essentially means it's this is our next gen AI factory. So while the entire LLM program has gotten us a lot of advances in uh our capabilities already, we are not nearly done in terms of the ambitions that we have to address language AI challenges for our customers. And to us this means being prepared for the next model generation. Bigger models, more capable models. But also very importantly we want to figure out how to extend our reach into the context of customers more so that they can help us help them right.
Clarify of course is a very explicit way to do this. You explicitly tell me what you would like to do. Can we do this less manual and that's what the challenge that we have ahead of us. If you want to see um how this actually turns out, I'm afraid you will have to invite me again. So, thank you. Thank you very much for really interesting um keynote. What I wanted to ask you is that from a point of a mathematician and someone who studied AI, what have you learned in DeepL about languages which you didn't expect or where did you did appreciate in languages? I mean a lot of things of course. Yeah. Um the one of the very first projects that I started working on contributing to at deep was creating new translation networks for both Japanese and Chinese.
Wow. I don't speak a single word Chinese or Japanese. So that's it's an interesting experience. The first thing I did is sat down with language experts in both Japanese and Chinese and basically asked them okay like I'm a mathematician. I can train AI models but I don't know anything about your language. Tell me everything that I need to know. And uh one of the outcomes that I remember very vividly is I realized that in Chinese specifically the way you address amounts of items changes depending on the kind of item it is. Wow. So it's not like three sheeps three cups. It's twice cheap and three cups. Yeah. Oh wow. Okay. Yeah. That is really something.
what what also is um in in in the point and I wanted to ask you that take for example um like the words digital sovereignity. So we think when we talk about digital sovereignity not in a way in Germany um we say digital sovereignity and you talk to a Frenchman and sovereignity nume but what we realize in the political talks about digital sovereignity that the French defined sovereignity completely different than the German with the word sovereignity. When do you think we will be at the point where also how can I say cultural or political implication of words could be detected or translated uh by deep in a lot of ways I think we have made substantial progress on this right um metaphors are a great example like the expectation to translate a metaphor doesn't have to be that you basically take your metaphor and then try to figure out how to capture the same structure and words into a different language.
In the very ancient past, we've basically done this, right? Like um the very early days of AI translation. Nowadays, these models do understand your intent already much better. So, they can come up with drastically different ways to capture the same sentiment. And there is an interesting balance how much leeway you want to give these models in in this like creative endeavor. Yeah. Um, but the technology to do this exists today. Now it's more a question of how can we expose this technology to users in a way that is useful, not overwhelm them with like an endless amount of choice, but also anticipate their needs really well and make this seamless. That was a really fascinating keynote.
Jeff Meas, everyone, thank you so much for being here. So, ladies and gentlemen, I don't need need my AI glasses for the next thing.