The context section of Kagi Translate is meant to provide additional context to help with getting a translation that fits the intended context. However, I suspected that meant it was actually a way of delivering instructions to the LLM, and I was not wrong.
This started off very handy. KT will happily correct for typos in the input, or even outright mistakes, and produce the output it statistically infers that you want. So when I mistakenly used “pudé” (I did) instead of “pus” (I put) in a sentence, it incorrectly translated to “I put” in the output, because statistically that was more likely, the other wasn't grammatical in my sentence so wouldn't have appeared in any samples. Which is fine if you're translating to understand something, not so much if you want to see if you wrote the right thing.
So I went into the context section and added, “Translate the meaning of what is written, not what you think the author meant.” And sure enough, it correctly translated my incorrect Spanish, errors and all.
I’ve had less luck with persuading it not to hallucinate word definitions based on a typo’s similarity to a real word. In Mexico “abonar” is “to buy on credit" or “to buy in installments”. “Avonar” is not a word at all, but it sounds the same, and when I first heard the word, I wasn't sure which was correct. KT didn't help, because it made up a similar definition for “avonar”. So I thought it was correct when it wasn't. I've tried putting “Do not fix typos” in the context instructions, but I think it needs something more complex. Maybe asking it to run all the words through some other step.
But then I began to wonder. What about the classic test of LLMs? Can it generate a Python program?
Yes, Virginia, it can.
P.S. Don't get me wrong, this is an incredibly useful translation program. Especially for understanding country-specific words and slang. And it does have a “proofread” feature which catches some (but not all) input errors. But GIGO still rules.
