Well, designed is maybe too strong a term. It’s more like stumbling on something that works and expand from there. It’s all still build on the fundaments of the nonsense generator that was chatGPT 2.
Given how dramatically LLMs have improved over the past couple of years I think it’s pretty clear at this point that AI trainers do know something of what they’re doing and aren’t just randomly stumbling around.
A lot of the improvement came from finding ways to make it bigger and more efficient. That is running into the inherent limits, so the real work with other models just started.
Well, designed is maybe too strong a term. It’s more like stumbling on something that works and expand from there. It’s all still build on the fundaments of the nonsense generator that was chatGPT 2.
Given how dramatically LLMs have improved over the past couple of years I think it’s pretty clear at this point that AI trainers do know something of what they’re doing and aren’t just randomly stumbling around.
A lot of the improvement came from finding ways to make it bigger and more efficient. That is running into the inherent limits, so the real work with other models just started.
And from reinforcement learning (specifically, making it repeat tasks where the answer can be computer checked)