From: redpointai
DeepL, a company supporting over 100,000 businesses in AI translation worldwide, has been engaged in cutting-edge AI research long before it gained mainstream popularity [00:00:40]. The company’s CEO and founder, Yarkovski, has a specific take on why specialized models will eventually become generalized models [00:00:12].
DeepL’s Philosophy: The Power of Specialization
DeepL’s success in the AI translation market, even against giants like Google Translate, is attributed to a strong academic-level research foundation combined with a strategic focus on a particular use case: high-value translation for businesses [00:08:25]. This focus has enabled them to deliver superior results [00:08:34].
Advantages of Specialized Models
- Higher Quality and Accuracy: For specific, high-value use cases like professional translation, specialized models consistently deliver better accuracy and quality compared to general models [00:06:03]. This is crucial for industries such as law firms, manufacturing, retail, media, and government [00:05:54]. For instance, translating an operating manual for a nuclear power plant demands extreme accuracy, making a specialized model with human oversight essential [00:32:03].
- Customization and Personalization: Specialized models allow for deeper personalization and steerability of the output. DeepL enables customers to embed specific terminology into their models, influencing the translation to align with their preferred language and style, which is crucial for businesses controlling their communication [00:11:38]. This is difficult to achieve with prompt engineering alone [00:10:46].
- Tighter Feedback Loop: By owning the entire vertical stack, from product to engineering and research, DeepL can effectively “trickle down the problems” and integrate feedback from the application directly into model development [00:10:30]. This tight feedback loop, as described by the speaker, allows for building a better, more tailored product [00:11:07].
- Addressing Data Variability: Different language pairs have vastly different amounts of available training data (e.g., German-English has much more than Polish-English) [00:12:53]. Specialized models can be tailored in size and architecture to optimize performance for these varying data scales, making them more efficient in terms of inference compute [00:14:04].
- Strategic Advantage: For “super big use cases” like translation, specialized models make financial sense. They offer a ready-made, vertically integrated solution that businesses can plug in and use reliably [00:37:11].
Limitations of General Models for Specific Tasks
While large general models like GPT-4 or GPT-5 are impressive, they are not always the optimal solution for highly specialized tasks like translation [00:38:05]. The speaker believes that generalized models are “overhyped” compared to “underhyped” specialized models, where “value gets created right now” [00:48:46]. A model trained specifically for translation with good labeled reinforcement learning can outperform a general model that does “a ton of different things” [00:38:20].
Moreover, general AI players might not be incentivized to invest in developing smaller, smarter model architectures because they benefit from their “monopoly on huge compute” [00:25:29]. The pursuit of more with less compute through smarter architectures is seen as a crucial next step [00:25:12].
The Role of Human Data and Continuous Innovation
Human input remains vital for the development and quality assurance of specialized models. DeepL has been running large-scale data annotation projects internally for years, utilizing human translators to train models and ensure quality [00:15:04]. This is particularly important for specialized models where customers have high expectations for consistent quality [00:15:31]. Sourcing native speakers globally for specific languages is a core part of this process [00:16:09].
Building an innovative company in the AI space requires constant innovation and a willingness to “throw away a lot of results” [00:32:03]. Every failed attempt in creating new model architectures provides a better understanding of the problem, leading to the next level of development [00:21:08]. This philosophy is key to navigating the rapid advancements in AI and ensuring competitive advantage [00:20:17].
Future of AI Translation
While AI text translation is highly advanced for well-resourced languages, especially for one-to-one communication, areas like publishing marketing websites or critical operating manuals still benefit from human oversight [00:31:50]. The next frontier for DeepL is spoken language and voice translation [00:39:11]. The goal is synchronous speech translation, allowing anyone to understand anyone else, regardless of language [00:00:16]. This would transform business operations by removing location barriers for international teams [00:40:48] and making educational resources more accessible [00:41:13]. However, challenges such as latency, ambiguity, and the unstructured nature of spoken language need to be overcome [00:46:03].
The speaker expects early products for synchronous speaking models to be out “pretty quickly,” but perfecting the technology, similar to text translation, will take several years [00:45:29]. This indicates that even with the advent of general models, specialized solutions will continue to push the boundaries of what’s possible in specific domains.