Fuck AI

One of my projects to satisfy the edicts was to make a translation system based on LLM for our application.

I hashed something out real quick. input, output, language -> translate this word from English to common Spanish that would be understood in central and south America taking special care to make sure it's not too confusing for Spanish dialects in other parts of the world

Then I fed parts of the same sheet we give our translators and reviewed the outcome got some native speakers to comment on the accuracy.

"Your shipment has arrived." translator what roughly acceptable, AI was uninformed.

Then I realized the translator had context, so:

translate this [your shipment as arrived] from English to Spanish understanding that it's being used in this context: [it's in a fantasy world and is happening in the long ago, it's a airplane drop of a box that contains goods that you ordered]

Now if gave me results that were in the same range as the translator.

Sometimes it was better than the translator, sometimes worse, but never unacceptable.

The problem is we need thousands of things translated, sometimes many new ones a month and in a dozen languages and writing the context for ever damn thing that needs translated is pain, then it needs to be at the very least spot checked, and how do you find people who are verifiably good to spot check?

It's certainly taking jobs, and often not for the best. It wasn't easy to get it to output good results. I think a lot of it comes down to inference and context that the translators could guess at, and often when the translators were less accurate, they also had the context wrong.

Kind of like capitalism ruining search engines so AI can get better results, Translation services seem to have ruined themselves for be cheaper and make more money to the point that LLM will be able to eat their lunches, at least until we actually have to pay for the full cost of those servers and the power.