Hacker News Reader

Comment by Rinzler89 | Hacker News Reader

What's DeepL's plan to survive when competing with ML translation from big tech?

In the era pre-ChatGPT they had a USP which it made it stand out, but today it's a different landscape.

Their translation is still substantially better than what most LLMs can offer. Certainly for languages where LLMs have very little training data for.

I will disagree here. I found ChatGPT better at English <-> German translation than DeepL. Especially at translating slang and online speak where DeepL would shit the bed.

I found that specifically when it comes to proper flow and grammatical rules, chatGPT and other LLMs tend to break down. Certainly when translating from English to another language. From another language to English often goes a lot better, but that really depends on how much training data in that language the LLM has been trained on.

German would be one of the languages where LLMs likely will perform fairly well, as it is one of the languages that LLMs training data often contains quite a lot of.

With DeepL if I am using it to translate a language I am not too familiar with and can't validate the outcome as well I therefore will have more trust in its translation. Because I know that the languages they support are actually specifically implemented for translation reasons.

I can't say for en <-> de, I can read both without problem. But for Japanese DeepL is much better, often times the tone from gpt4 are little bit "rude"/direct or "weird" in en to jp, deepl is as good as an example sentence they would put into a textbook. For jp to en, most gpt4 or any llm is much superior, because you can ask follow up questions on certain slangs, or anything that's not very clear.

Doesn't DeepL use an LLM? At least based on the company name that's what I would assume

It is(at least was) pre-LLM deep learning NLP. So was Google Translate, and those two as well as LLMs all has their unique forte and failure modes.

Sometimes: https://support.deepl.com/hc/en-us/articles/14241705319580-N...

To be fair, top-of-the-line proprietary models suck at i18n. Try gemma-27b it's much better

I've used chatgpt to translate into one the local dialects here - very impressed. DeepL is OK if you know the language you're translating to - and knowing which alternative words work better in the translation.

IMO, the quality of the translations has dropped significantly. When they first launched it was noticeably better than almost anything else. Now it's comparable to the alternatives. And I don't think the alternatives got that better, it's DeepL that is not as good anymore.

I still use it for most of the translations I need (mostly English <> German). And sometimes I have to check the translation because it can mess up the actual meaning of the sentence. Sometimes I can catch it with my limited German, sometimes I run it through another translator. And it "hallucinates" often enough, unfortunately.

As an individual, I still go to DeepL while actively using LLMs. I think the main reason is that the translations are very good, with alternatives in one click, without having to type "translate xxxx into yyyy".

My previous company ran tests on translations, for their specific use case, DeepL API was overall better that OpenAI or Claude.

Are you paying for DeepL?

Not personally, my company was. Why?

Because my question was how was DeepL planning to survive monetarily? It doesn't help if people are praising it but if nobody's paying for it. Google, Microsoft, et-al can subsidize their AI offerings longer than DeepL can stay solvent.

That’s interesting. In my tests of translation of formal speeches from Japanese to English, the latest versions of ChatGPT, Claude, and Gemini were all better than DeepL. While DeepL’s output wasn’t bad, the fact that the LLMs could be prompted in detail about the purpose of the translation and had sufficient context windows to maintain pronoun reference and other forms of cohesion made a significant difference.

Exactly this. LLMs are more context aware than DeepL.

You could add context to DeepL as well as easily, with the "context" parameter of the translation.

Thanks for that clarification. I don't think that feature was available when I did my comparison tests a few months ago.

I created a web site that uses AI to translate books, and in my testing, Deepl was way better than other AIs I tried at translating.

It is still better than Google Translate like it always has been, almost as if Google did not care about their translator, especially "post-ChatGPT" it is not supposed to be this inaccurate.

I think they will just go the AI/LLM/ML route, too.

I truly don't know what the Google Translate team is doing day-to-day

If OpenAI wrappers can flourish I'm sure DeepL can too. My guess is that they will have sales team that will reach out to customers and provide them with solutions to their specific needs and use cases.

You know how ChatGPT was supposed to be able to act like a translator in voice mode, they even had an amazing demo at the keynote? I tried to demonstrate it to my father who was supposed to speak in Bulgarian and I in Turkish and see how cool the feature looks like. It didn't work remotely as advertised, it instantly screwed up and we both knew it because we both speak these languages and proving that its not to be trusted. The tech behind it is amazing but the UX still needs to be crafted to be valuable for more than tech demos, so if DeepL has the tech and know the market they will be in much more advantageous position than OpenAI(has the tech, doesn't know the market) and others who might know the market but use OpenAI(their margins will be thinner if OpenAI can't reduce costs dramatically more than DeepL).

Yes, ChatGPT's Advanced Voice Mode is great if you don't understand the target language at all and can't catch how far from ground truth it is.

I agree, LLM translations are not only more convenient but also much more capable. I often find myself giving instructions on how to translate text, such as asking the LLM to use formal language in the target language or to apply specific gender-neutral wording. Additionally, it can translate text while preserving the structure (e.g. values in a JSON object) or even adapt to a new target structure. It's just so much more convenient.

I’ve been wondering the same thing. As near as I can tell from the explanations on the website and in the videos, this DeepL Voice does not seem to be based on a multimodal large language model. Rather, it’s probably using text-to-speech and speech-to-text models linked through a translation engine. If that’s true, it means it can’t “hear” the tone of voice of the speakers. More important, it doesn’t seem to be promptable—that is, it cannot be told what the situation is, who the speakers are, or the type of translation desired. Those are all possible with OpenAI’s Advanced Voice mode, and they can make the difference between successful and unsuccessful interpretation.

That said, DeepL is focused on providing translation services, which the large LLM companies are not. Even if DeepL’s translation engines are not as powerful as the strongest commercial LLMs, they might be able to compete in other ways, such as security guarantees, on-device operation, and training and support.

They have a next-gen model that is an LLM. But it's only available for a few languages and only for paying users: https://support.deepl.com/hc/en-us/articles/14241705319580-N...

Shouldn't it use much less compute than an LLM? Or is that not a significant percentage of the total cost?

Reply on news.ycombinator.com