DeepL, long a leader in translation tech, finally embraces LLMs

Hello and welcome to Eye on AI.

DeepL, the Germany-based startup that’s been leading the way on AI translation technology, yesterday released its next-generation model. The company was valued at $2 billion in a recent funding round and has made quite a splash in the translation subset of the AI boom, stating that its technology is already used by over 100,000 businesses, governments, and other organizations. But believe it or not, this is DeepL’s first model to be based on the large language model technology that’s underlined the AI hype and just about every AI product out there.

“We’ve been working with similar neural networks since DeepL was conceived in 2017, but there’s been obviously an evolution in how this architectures works, and for the first time, we’ve moved that toward LLM technology,” CEO Jarek Kutylowski told Eye on AI, adding that “LLMs are going to be taking over translation.”

The model is not a general LLM, such as GPT-4 or Llama 3. It was built entirely by DeepL in-house, on its own infrastructure, and is tailored specifically for translation. The unnamed model —”We’re not doing names,” Kutylowski said—is meant to do everything the company’s previous model was doing, only better, plus hopefully even more down the road. The company hopes the new LLM architecture will pave the way for new capabilities such as multimodal translation and the ability to translate text into multiple languages simultaneously in real-time as the user types.

Anyone can use DeepL’s translation model via its website and various browser extensions and software integrations, but the company is specifically targeting global companies that want to translate everything from Slack messages to sensitive legal documents. Many translation tools only offer the ability to write in your own language and then prompt the model to translate your text in its entirety, but DeepL’s model also includes interactive writing features, suggesting different options for how you might want to phrase something so you can ensure it will be linguistically correct in the other language while still making it your own.

“You want to make sure that this is your text and that even if translated, this is something that you’ve written,” Kutylowski said. “Kind of fine-tuning it, making it yours, is an important thing that our users value.”

As with all AI software, data is key. And it is clear that DeepL views its access to top-class translation data, as well as the humans it has in the loop of its training process, as key differentiators. DeepL uses a combination of publicly available data (which expert linguists assess for quality so only the “right” data is actually fed into the model), as well as licensed datasets and synthetic data. Human experts provide feedback on the model outputs in order to improve its performance, a process referred to as reinforcement learning through human feedback (RLHF). According to Kutylowski, the jump in quality from the human feedback is “quite significant” and DeepL wouldn’t be where it is without it.

Of course, the company has significant competition from Big Tech, with all the regular players including Microsoft, Google, Apple, Amazon, and Meta investing in similar technologies. As the Semafor Technology newsletter has written, “companies like Google can snap their fingers and get the ability to translate more than 100 languages and offer that technology for free.” In total, DeepL supports 32 languages with its previous generation software (only English, Japanese, German, and Simplified Chinese so far with the next generation), while Google’s recent additions bring Google Translate to a whopping 243 languages.

Leading AI companies such as OpenAI are getting in on the race, too. Earlier this week, Cohere also announced a “significant investment” and strategic partnership with Japanese tech giant Fujitsu to build LLMs with Japanese language capabilities for global enterprises.

Still, DeepL believes its models provide the best quality translations. The company said that in blind tests, language experts preferred its next-generation model 1.3 times more often than Google Translate, 1.7 times more often than ChatGPT-4, and 2.3 times more often than Microsoft. DeepL, however, has not published this research and declined to offer specifics on the methodology such as how many language experts were involved in the tests. In 2021, an independent evaluation found DeepL to be the highest-performing machine translation model across 13 language pairs, outperforming Google and 16 others.

While many use cases for AI are still being debated, the benefits of using AI for translation are easy to see. Businesses are becoming more global by the day. It’s also an area where there’s a lot to gain with little to lose. Few professionals had access to human translators to begin with, but now everyone—from growing companies to vacation travelers—has a translator in their pocket that can open up entirely new opportunities.

“From my perspective, it’s really incredibly exciting. And it was not clear for quite some time if LLMs were going to be the way forward technologically when it comes to translation,” Kutylowski said. “I think right now in this step, it really shows that they are capable of outperforming other models, but also, what I’m really excited about is the additional capabilities we will be able to infuse into those models in addition to the normal translation that you know nowadays, just because [LLMs] are not only capable of translation but can do more.”

And with that, here’s more AI news.

Sage Lazzaro

sage.lazzaro@consultant.fortune.com

AI IN THE NEWS

The FTC is looking into Amazon’s deal with AI startup Adept. That’s according to Reuters. The inquiry revolves around last month’s announcement that Adept CEO David Luan along with other top talent from the company are joining Amazon, which would also license some of the startup’s technology. While the inquiry is currently unofficial and won’t necessarily result in an official investigation or enforcement action, it reflects the FTC’s growing interest in Big Tech’s dominance in AI. In June, the FTC opened an investigation into Microsoft for its similar absorption of Inflection AI, and earlier this week, U.K. antitrust officials followed suit with their own investigation of Microsoft’s actions with Inflection AI.

Meta said it will withhold multimodal AI models in the EU, citing a lack of clarification from regulators. That’s according to Axios. While Meta plans to incorporate such models into various products, the company said it will only offer text-based models in the EU and will withhold its upcoming multimodal model as well as future ones. The move would limit the products and services available to EU users and would also mean that European companies would not be able to use Meta’s multimodal models released under open licenses. Meta’s issue stems from GDPR protections around how it can use customer data to train models, not from the recently passed EU AI Act. Apple has similarly stated it won't release its new Apple Intelligence features in the EU over regulatory concerns. Together, these moves suggest tensions are growing between European regulators and the U.S. companies leading the AI boom.

An AI model from Cambridge outperforms clinical tests in predicting the progression of Alzheimer’s disease. Using MRI results and data from cognitive tests, the model was able to accurately predict if people with early signs of dementia would develop Alzheimer’s and how fast the disease would progress in four out of five cases, which is three times more accurate than the current state of care. The trials included a three-year longitudinal study, which was validated using independent data that included almost 900 individuals who attended memory clinics in the UK and Singapore. “This has the potential to significantly improve patient wellbeing, showing us which people need closest care, while removing the anxiety for those patients we predict will remain stable. At a time of intense pressure on healthcare resources, this will also help remove the need for unnecessary invasive and costly diagnostic tests,” Cambridge Psychology professor Zoe Kourtzi, senior author of the research, said in a Cambridge news article.

Anthropic rolls out new $100 million venture fund with Menlo Ventures. OpenAI has long helped seed the market for its own AI models through a venture fund. Now rival Anthropic is doing the same thing. It has partnered with Menlo Ventures to launch the Anthology Fund with $100 million to back AI startups, including those building software development tools and middleware layers around Anthropic’s Claude LLMs. You can read more on Menlo Venture’s blog here.

OpenAI unveils new GPT-4o Mini model that is much cheaper to use. The company has released a new, smaller version of its GPT-4o model that it says is more capable than its GPT-3.5 Turbo model but cheaper to use than any of its larger GPT-4 and GPT-4o models, according to a story in The Verge. The new model should appeal to cost-conscious AI developers, who have sometimes balked at how expensive it can be to use applications that are built on top of OpenAI's larger models. GPT-4o Mini achieved an 82 percent score on the Measuring Massive Multitask Language Understanding (MMLU), a benchmark exam consisting of about 16,000 multiple-choice questions across 57 academic subjects, the tech publication said, noting that this score was higher than that achieved by other miniaturized AI models such as Anthropic's Claude 3 Haiku and Google's Gemini 1.5 Flash.

FORTUNE ON AI

Are AI’s lofty valuations justified? Some VCs think it is —by Luisa Beltran

Why business leaders view AI as an opportunity to take ‘toil out of our work’ —by Leo Schwartz

Runway CEO says Hollywood should embrace AI, not fear it: ‘These are exceptional tools for great artists’ —by Jeremy Kahn

Google chief scientist Jeff Dean: AI needs ‘algorithmic breakthroughs,’ and AI is not to blame for brunt of data center emissions increase —by Sharon Goldman

We used satellite imagery and AI to see who’s keeping their climate pledges. Here’s the shocking reality we found —by Antoine Rostand (Commentary)

AI CALENDAR

July 21-27: International Conference on Machine Learning (ICML), Vienna, Austria

July 23: Google earnings

July 30-31: Fortune Brainstorm AI Singapore (register here)

July 31: Meta earnings

Aug. 12-14:

Ai4 2024 in Las Vegas

Dec. 8-12: Neural Information Processing Systems (Neurips) 2024 in Vancouver, British Columbia.

Dec. 9-10: Fortune Brainstorm AI San Francisco (register here)

EYE ON AI NUMBERS

2.5%

That’s how much of the global AI market Africa currently represents, according to a new report from the GSMA examining AI development in Africa. The report argues AI has immense potential for Africa’s economy and ability to support its Sustainable Development Goals (SDGs), citing recent estimates from AI4D Africa stating that AI could boost the continent’s economy by $2.9 trillion by 2030 (the equivalent of increasing annual GDP by 3%). Seeking to address the current development gap, the report identifies over 90 AI use case applications across Kenya, Nigeria, and South Africa, which are some of the leading tech hubs on the continent.

“The agritech sector is seeing most of the AI innovation, especially in Kenya and Nigeria where agriculture continues to play a significant role in the economy,” reads the report.

Agriculture makes up 49% of the current percentage of use cases by sector, while climate action makes up 26% and energy makes up 24%. To reap the benefits, the report also outlines how Africa will have to focus on governance, fostering partnerships between stakeholders, unlocking the finances at scale to support AI, and supporting R&D.