As part of the AI Coach and Ambassador Program, I had an interesting homework assignment: to experiment with different language models. I decided to use my recent article, “Unicorn Platform: Easy Website Building for Everyone, Whether You’re a Beginner or Pro!” for this purpose.
The task was to translate the article using four different AI language models. With my native language being Czech, I decided to translate from English to Czech (🇨🇿 konečně něco pro české čtenáře). Below, you’ll find a comparison table highlighting the differences between translations generated by four different AI language models. I’ve also included the full results of these translations for your review.
Comparison of different language models
| Section | GPT-4o | Claude-3-Sonnet | Gemini-1.5-Flash | Mixtral-8x7B-Chat |
|---|---|---|---|---|
| Translation Quality | Incorrect grammar – minor issues but requires changes (“Snadné tvorba”, “okamžitě zaujalo design”). ❌ This expression does not seem correct: “mezi sebou konfliktovat.” ❗ Uses a slightly more formal style, which may seem professional but slightly less accessible.❗ | Consistently accurate with straightforward grammar. ✅ No significant errors, preferred translation of “Versatility” as “všestrannost” is personal preference.❗ | Clear and correct grammar. Slightly more dynamic and engaging.✅ | Date format is incorrect (should be 10.7 instead of 7.10).❌ No major errors in grammar.✅ Tone is more informal than others but maintains a good balance between professional and conversational.✅ |
| Specific Terminology | Specific terminologies translated correctly. ✅ The name of theme was kept as it was.✅ | Specific terminologies translated correctly.✅ | Specific terminologies translated correctly.✅ | Translated “NoCode” humorously as “Bezkód” and “Nikód”. ❌ The term “reaktivní” for “responsive” could be better as “responzivní” to align with common industry terminology.❌ |
| Size and Efficiency | Translated the whole article quickly, no errors.✅ | Translated the whole article quickly, no errors. | Translated the whole article quickly, no errors. | Had to split the article into two parts as the model did not finish the whole translation in one go.❌ |
| Formating | Formated (bullet points, bold text).✅ | Plain text.❗ | Formated (bullet points, bold text).✅ | Plain text.❗ |
| Ethical | Translated as “he”.❗ | Translated as “he”.❗ | Translated as “he”.❗ | Translated as “he”.❗ |
[wptabs id=”2136″]
Conclusion
I was pleasantly surprised by all translations; all models did a good job. In my opinion, Claude and Gemini provided similar results. GPT-4o had the most grammatical errors, though it sounded the most professional—perhaps almost too professional—due to some complex phrases. Claude performed the best in my opinion; I could use this translation as it is. Gemini provided very similar results to Claude and was a little more dynamic, so I would rank it second. Mixtral made me laugh with its translation of some industry-specific words like “NoCode” and “Responsiveness”. Additionally, it did not get the date correct. However, I liked the tone it used.
Interestingly, GPT-4o and Gemini formatted the text, which I really liked. All models translated as “he.” Since I did not specify this in the prompt, it’s acceptable, but we can see a bias here, right?
I encourage you to try different models yourself, as there are many available, and each excels in different areas. Try Poe ! A platform that allows you to access a wide range of AI chatbots.
