Hacker News Reader

Comment by colordrops | Hacker News Reader

Isn't ChatGPT better than Google translate or deepl? Why doesn't google use a modern LLM to translate?

Because Google nor any other company is rich enough to use LLMs at that scale. Consider the quality vs. cost per query & latency trade off.

ChatGPT is a transformer model. Google pioneered transformer models specifically for translation.

If ChatGPT outperforms google translate, I'd guess it's simply because chatGPT is a bigger model with more training data.

Must be something else, don't you think? If there's a company in the world who can just "adjust their model to make it bigger" that's Google. But they haven't done so. Perhaps Google doesn't care.

Google Translate is a relatively small product with a relatively small budget. It doesn't bring in a huge amount of users or revenue.

The engineers working on it seem to be focussing on supporting small languages (ie. languages that only a few million people speak - africa and india have lots of those, both big growth markets). They're also working on better apps - for example being to translate text inside images, text from a video feed, live translation of captions, etc.

It makes sense that they probably wouldn't be deploying a huge chunk of compute there to get a few marginal quality gains.

Ironically, the amount of training data in non-English languages for ChatGPT is tiny. In the order of 10% for all of them (IIRC).

Reply on news.ycombinator.com