Hello and welcome to Eye on AI.
DeepL, a Germany-based startup pioneering AI translation technology, unveiled its next-generation model yesterday. The company was valued at $2 billion in a recent funding round and is making waves in the AI booming translation field. The company says its technology is already being used by more than 100,000 companies, governments, and other organizations. But believe it or not, this is DeepL’s first model based on the large-scale language model technology that is the foundation of the AI boom and almost all AI products.
“We’ve been working on similar neural networks since DeepL was conceived in 2017, but the way this architecture works has clearly evolved and for the first time we’ve moved it over to LLM technology,” CEO Jarek Kutylowski told Eye on AI, adding that “LLM will now take over translation.”
The model isn’t a typical LLM like GPT-4 or Llama 3; it was built entirely in-house by DeepL on its own infrastructure and is specifically for translation. The unnamed model (“we’re not going to name it,” Kutylowski says) aims to do everything the company’s previous models did, plus better, and potentially even more in the future. The company hopes the new LLM architecture will pave the way for new capabilities like multimodal translation, or the ability to translate text into multiple languages in real time as the user types.
While DeepL’s translation models are available to anyone through its website and various browser extensions and software integrations, the company specifically targets global enterprises looking to translate everything from Slack messages to sensitive legal documents. While many translation tools only offer the ability to write a sentence in your language and ask the model to translate the entire text, DeepL’s models also include an interactive writing feature, suggesting different phrasing options so you can be sure that your translation will be linguistically correct in other languages while still being in your own language.
“This is your writing, and you want to make sure that even if it’s translated, it’s still yours,” Kutzyrowski said. “Tweaking it and making it your own is a big part of what users value.”
As with all AI software, data is key, and it’s clear that DeepL sees access to top-class translation data, as well as humans involved in the training process, as key differentiators. DeepL uses a combination of publicly available data (quality assessed by expert linguists, and only the “right” data actually gets fed into the model), licensed datasets, and synthetic data. Human experts provide feedback on the model’s output to improve performance, a process called reinforcement learning with human feedback (RLHF). Kutylowski says the improvement in quality with human feedback is “pretty significant,” and that without it DeepL wouldn’t be where it is today.
Of course, the company faces stiff competition from big tech companies, with regulars like Microsoft, Google, Apple, Amazon, and Meta also investing in similar technology. As Semafor Technology’s newsletter puts it, “Companies like Google can translate over 100 languages with a snap of their fingers and give their technology away for free.” DeepL supports a total of 32 languages in its previous generation software (the next generation only supports English, Japanese, German, and Simplified Chinese for now), while Google’s recent additions bring Google Translate to a whopping 243 languages.
Other big AI companies, such as OpenAI, are also joining the race. Earlier this week, Cohere also announced a “significant investment” and strategic partnership with Japanese tech giant Fujitsu to build a Japanese-language LLM for global companies.
DeepL still believes its models provide the highest quality translations. In blind tests, language experts preferred its next-generation model 1.3 times more than Google Translate, 1.7 times more than ChatGPT-4, and 2.3 times more than Microsoft, according to the company. However, DeepL has not published the findings and did not provide details about its methodology, such as the number of language experts involved in the tests. In 2021, an independent evaluation found that DeepL was the best-performing machine translation model in 13 language pairs, outperforming Google and 16 others.
While many use cases for AI are still under debate, the benefits of using AI for translation are clear. Business is becoming more global every day. It’s also a sector where there is little to lose and much to gain. Whereas originally few professionals had access to human translators, now everyone, from growing businesses to vacationing travelers, has a translator in their pocket, opening up entirely new opportunities.
“From my perspective, this is really very exciting. It hasn’t been clear for quite some time that LLMs would be the way forward technologically when it comes to translation,” Kutyrowski said. “I think at this point we’ve really shown that LLMs have advantages over other models, but what really gets me excited are the additional capabilities that can be infused into these models on top of the regular translation we know today. [LLMs] It can do more than just translate.”
Now for some more AI news.
Sage Lazarus
sage.lazzaro@consultant.fortune.com
sagelazzaro.com
AI in the News
The FTC is investigating Amazon’s contract with AI startup Adept. it is ReutersThe investigation comes in response to last month’s announcement that Adept CEO David Ruan and some of the company’s top talent would join Amazon, which would also license some of the company’s technology. The investigation is informal at this point and won’t necessarily lead to a formal investigation or enforcement action, but it reflects the FTC’s growing interest in the dominance of big tech companies in the AI space. The FTC launched an investigation into Microsoft’s similar absorption of Inflation AI in June, and earlier this week, UK antitrust regulators also launched their own investigation into Microsoft’s conduct with Inflation AI.
Mehta said he would refrain from introducing multimodal AI models in the EU due to insufficient clarification from regulators. it is AxiosMeta plans to incorporate these models into various products, but the company said it will only offer text-based models in the EU, not the upcoming multimodal models or future models. The move limits the products and services available to EU users and also means that European companies will not be able to use Meta’s multimodal models released with an open license. Meta’s problems stem from GDPR’s protections over how customer data can be used for model training, not the recently passed EU AI law. Apple similarly said it will not release new Apple Intelligence features in the EU due to regulatory concerns. These moves signal growing tensions between the US companies leading the AI boom and European regulators.
University of Cambridge’s AI model outperforms clinical trials in predicting Alzheimer’s disease progression. Using MRI results and cognitive testing data, the model was able to accurately predict whether people with early signs of dementia would develop Alzheimer’s, and how quickly the disease would progress in four out of five cases – three times more accurate than current care. The trial involved a three-year longitudinal study and was validated using independent data including nearly 900 people who attended memory clinics in the UK and Singapore. “This has the potential to significantly improve patient wellbeing, revealing who needs the most careful care and removing anxiety for patients who are predicted to be stable. At a time of intense pressure on healthcare resources, this could also help remove the need for unnecessary invasive and expensive diagnostic tests,” said study lead author Zoe Curzi, professor of psychology at Cambridge. Cambridge News Article.
Anthropic has launched a new $100 million venture fund with Menlo Ventures. OpenAI has been trying to bring its AI models to market through venture funding for years, and now rival Anthropic is doing the same: It partnered with Menlo Ventures to create the $100 million Anthology Fund, led by Anthropic’s Claude LLM, to back AI startups building software development tools and middleware layers. You can read more about it on the Menlo Ventures blog. here.
OpenAI has announced a new GPT-4o Mini model that comes with significantly reduced costs of use. According to a report from The Verge, the company has released a new, smaller version of its GPT-4o model. The company claims that the model outperforms the GPT-3.5 Turbo model while being cheaper than the larger GPT-4 and GPT-4o models. The new model should appeal to cost-conscious AI developers who have been put off by the high cost of using applications built on OpenAI’s larger models. According to the tech magazine, GPT-4o Mini achieved a score of 82% on Measuring Massive Multitask Language Understanding (MMLU), a benchmark test consisting of about 16,000 multiple-choice questions across 57 academic subjects, a higher score than other smaller AI models, including Anthropic’s Claude 3 Haiku and Google’s Gemini 1.5 Flash.
The fate of AI
Are AI’s high valuations justified? Some VCs think they are — Luisa Beltran
Why business leaders see AI as an opportunity to “make work easier” —Leo Schwartz
Runway’s CEO said Hollywood should embrace AI, not fear it: “These are great tools for great artists.” —Jeremy Kahn
Google chief scientist Jeff Dean: AI needs ‘algorithmic breakthroughs’, not main driver of rising data center emissions —Sharon Goldman
We used satellite imagery and AI to find out who is keeping their climate pledges. The findings are shocking. —Antoine Rostand (commentary)
AI Calendar
July 21-27: International Conference on Machine Learning (ICML), Vienna, Austria
July 23: Google Revenue
July 30-31: Fortune Brainstorm AI Singapore (registered here)
July 31: Meta Revenue
August 12-14: Ai4 in Las Vegas 2024
December 8-12: Neural Information Processing Systems (Neurips) 2024 will take place in Vancouver, British Columbia.
December 9-10: Fortune Brainstorm AI San Francisco (Register here)
Focus on the AI numbers
2.5%
This shows how much of the global AI market Africa currently accounts for. New Report from GSMA Examining AI development in Africa. The report argues that AI holds immense potential for African economies and has the capacity to support the Sustainable Development Goals (SDGs). Recent estimates by AI4D Africa suggest that AI could boost the continent’s economy by $2.9 trillion by 2030 (equivalent to a 3% increase in annual GDP). To address current development gaps, the report identifies over 90 AI use case applications in the continent’s key tech hubs: Kenya, Nigeria and South Africa.
“The agricultural technology sector is seeing the most AI innovation, particularly in Kenya and Nigeria, where agriculture continues to play a key role in their economies,” the report states.
In terms of current use cases by sector, agriculture accounts for 49%, followed by climate action at 26% and energy at 24%. The report also outlines that for Africa to reap the benefits, there needs to be a focus on governance, fostering partnerships between stakeholders, unlocking large-scale funding to support AI, and supporting research and development.
