https://paulgraham-translated.vercel.app/

languages: English, French, Spanish, Portuguese (BR), German, Japanese, Hindi, and Chinese (Simplified)

models/methods: Google Neural Machine Translation, gpt-4o-mini, gemini-1.5-flash, claude-3-haiku

I was reading pg’s “The Need To Read” and noticed there are links to translations in the footnotes. I thought that browser translation tools would have solved that problem but it turns out Google Translate isn't that good. I ran a few tests with french speakers, and turns out that gpt-4o-mini captured some elements of his style more effectively than Google Translate.

How many other languages/translations does this apply to?* I made this website to find out. You can switch between translations, compare them, and rate them, or flag an issue.

Curious to see if this is something that people would use over browser translation.

* This is only possible due to the incredible value of gpt-4o-mini-like models. With the gpt tokeniser, there are 727,362 tokens across 223 essays. Translating them all assuming 1:1 input/output (approximation) amounts to $0.55, very reasonable for an experiment like this.

The same job for 1 language with gpt-4o would be $14.5 using the same approximation.

If this project gets traction, I’ll explore adding more languages and experimenting with better models: GPT-4o, Claude-3.5-Sonnet, Gemini-1.5-Pro, LLama-3.1, etc.