alt Hacker NewsDo we really? Is that the human ‘moat’? I don’t see why AI couldn’t begin to handle that as well. What is limiting AI fundamentally from doing it?
Do we really? Is that the human ‘moat’? I don’t see why AI couldn’t begin to handle that as well. What is limiting AI fundamentally from doing it?
My believe is that current LLMs performs very well in the constraint environment with lots of data available, e.g. we have lots of code on github, public discussion with feedbacks, compiler and tests failures as feedback, so LLMs can train on trillions data records.
That's not true for many unconstrained real world scenarios, and LLMs currently perform poorly there. Another breakthrough is needed to overcome this.