As somebody who got to work adjacent to some of these things for a long time, I've been wondering about this. Are LLMs and transformers actually better than these "old" models or is it more of an 80/20 thing where for a lot less work (on developers' behalf) LLMs can get 80% of the efficacy of these old models?
I ask because I worked for a company that had a related content engine back in 2008. It was a simple vector database with some bells and whistles. It didn't need a ton of compute, and GPUs certainly weren't what they are today, but it was pretty fast and worked pretty well too. Now it seems like you can get the same thing with a simple query but it takes a lot more coal to make it go. Is it better?
As somebody who got to work adjacent to some of these things for a long time, I've been wondering about this. Are LLMs and transformers actually better than these "old" models or is it more of an 80/20 thing where for a lot less work (on developers' behalf) LLMs can get 80% of the efficacy of these old models?
I ask because I worked for a company that had a related content engine back in 2008. It was a simple vector database with some bells and whistles. It didn't need a ton of compute, and GPUs certainly weren't what they are today, but it was pretty fast and worked pretty well too. Now it seems like you can get the same thing with a simple query but it takes a lot more coal to make it go. Is it better?