Next generatoin of open source is the skills. Open source is basically collection of opinion and idea. When the summerize it very well in the form of skills, it will replace current source code distribution.
Stage 0 - AI can't do coding. It's just better completion engine.
Stage 1 - It can generate code, but not usable.
Stage 2 - Better engineer can steer AI better.
I can not belive me anymore.
@RegularJoe_Ceo Correct, it's covered on the linked article.
And basically it's a bug, not a feature. It costs some, so they didn't fixed. And actually it's not important(yet).
The behavior of model is deterministic, but it's not predictable. With single more space, whole result can be changed.
I find that they said LLM is non-deterministic by design. But it's not. This mis-conception cames from mostly by the temperature. As a chatting engine, veriation was required. So they produced temperature value. When it's non zero. It does probabilitc sampling. So answer was(1/4)
And I think I never heard what agent did wrong on code in detail. I believe it came from lack of expression. If he know what should be done in clear language to agent, agent will did it as asked. But user did not ask because he do not know how to express it.
Maybe now is right time to formalize architecture pattern. So many people complain about wrong architecture of agent generated code. But they do not tell what's wrong. That means they don't know how to explain. To make agent works, they need to explain with architecture patterns.