09/04/2026 1:49 PM

⬅️ [08/04/2026 1:13 PM](<./08_04_2026 1_13 PM.md>) | ⬆️ [2026 - April](<./README.md>) | [14/04/2026 10:30 AM](<./14_04_2026 10_30 AM.md>) ➡️

---

09/04/2026 1:49 PM

AI generated translations make native speakers go "Huh that's not how I would say that"
Might bring that in to the argument.

Haha the benchmarks were from generated data which meant that using auto-translated data for training increased performance on benchmarks, but reduced real world performance. But training on real data decreased performance for benchmarks.

Quality and diversity more important than quantity.

---

⬅️ [08/04/2026 1:13 PM](<./08_04_2026 1_13 PM.md>) | ⬆️ [2026 - April](<./README.md>) | [14/04/2026 10:30 AM](<./14_04_2026 10_30 AM.md>) ➡️