Every November, founders, investors, global leaders and reporters flock to frigid Helsinki for Slush. Since we first partnered with Supercell in 2013, we’ve been joining this chilly, student-run event and learned it’s well worth braving the snow and ice to connect over a sauna, dinner or a fireside chat. Having opened up our London office this August, the full team made the relatively short trip up North to connect with many of our favorite people, gathered all in one place.
DeepL’s CEO and founder Jarek Kutylowski joined IVP’s Eric Liaw on the Founder Stage to share how DeepL powers its intuitively accurate AI translator, which competes with both Google and old school, manual translation. Here are some things worth remembering and knowing from their talk (keynote video here):
Why language is so hard to crack
Different content requires different levels of accuracy and fluency. For instance, a technical document depends on a more straightforward, precise translation whereas marketing text needs to flow as if you were having a casual conversation with a friend - getting the gist, comprendo - not stiff and robotic.
DeepL spent years getting the training data and neural architectures right, for different situations. Other LLMs include an internet’s worth of data, while DeepL’s neural network is specially trained for translation and language.
Great ideas can come from anywhere and anyone, and many companies are best conceived and built in Europe
Language beguiled Jarek since his family moved from Poland to Germany when he was a teen, knowing fast kein Deutsch. He realized how tricky - and important - languages are.
He applied his computer science and mathematics education to reframe the challenge of translation, leveraging breakthrough neural networks and deep learning to create AI-powered translation. DeepL helps solve this huge problem for people around the globe.
Don’t build your own LLM unless it is essential to your business
There were no transformers in 2016, so DeepL had to build its own LLM from scratch. This decision established DeepL’s technological foundation, which is the best language model for translation. On the other hand, if your product doesn’t depend on a differentiated LLM, better to adopt models off the shelf.
There’s too much to capture from our time on the ground so take a look behind the scenes with our team.
Kiitos ja nähdään ensi vuonna!
