Please tell me I'm getting this wrong somehow:
1. We invent the transformer model to build the best translation engine possible
2. We instead use it to create an everything machine
3. We proceed to prompt the everything machine to badly imitate a translation engine
That sounds … stupidly inefficient
My understanding of transformers is basic, but by training one on the task of predicting a translation output directly, it should be possible to obtain a model that is actually good at translating, and doesn't suffer from prompt injection or alignment problems. It would probably also be much smaller than current LLMs for the same performance (though still quite big admittedly)
Jul 13, 2026, 1:03 PM@piegames i want to disagree with the wording "good at translating", but thinking it through? it depends on your definition of translating. transporting words from a to b? yeah, quite good. transporting the same emotions in the translation? probably doable okay-ish. the more translation requires human creativity and understanding the more i see it fallible.
but i didn't sadly study computerlingustics. so my opinion is uninformed.
@piegames I got so confused because I read "transformer" as the electrical component and "translation" as physical movement 🙃