Whenever someone says how LLMs have improved, all I can think is that the changes during all this time have felt very minor to me.
-
Whenever someone says how LLMs have improved, all I can think is that the changes during all this time have felt very minor to me. The one thing I would consider a real breakthrough is a new architecture that doesn't make hallucinating the core mechanism for how it works. Otherwise I just don't see them ever become reliable.
-
Whenever someone says how LLMs have improved, all I can think is that the changes during all this time have felt very minor to me. The one thing I would consider a real breakthrough is a new architecture that doesn't make hallucinating the core mechanism for how it works. Otherwise I just don't see them ever become reliable.
@volpeon
According to most researchers, hallucinations will never go away. -
Whenever someone says how LLMs have improved, all I can think is that the changes during all this time have felt very minor to me. The one thing I would consider a real breakthrough is a new architecture that doesn't make hallucinating the core mechanism for how it works. Otherwise I just don't see them ever become reliable.
@volpeon some of the MOE models that don’t need to run all the parameters in the model seem interesting in that they will run on more modest hardware.
I had one of the quantised quen 3 30b models running on a 24core first gen epyc and getting okish performance.
Getting good (relatively anyway) models that don’t need globs of hardware would be good progression although I suspect it will still burn a shitton of resources training them.
-
Whenever someone says how LLMs have improved, all I can think is that the changes during all this time have felt very minor to me. The one thing I would consider a real breakthrough is a new architecture that doesn't make hallucinating the core mechanism for how it works. Otherwise I just don't see them ever become reliable.
@volpeon "Look, NetBeans doesn't needs 5s for autocompletion, now it only takes 2s!" -
Whenever someone says how LLMs have improved, all I can think is that the changes during all this time have felt very minor to me. The one thing I would consider a real breakthrough is a new architecture that doesn't make hallucinating the core mechanism for how it works. Otherwise I just don't see them ever become reliable.
@volpeon not hallucinating is fundamentally impossible for what an LLM is, which is a probabilistic model that takes in a string of tokens and predicts what should come next. we're a long long way away from something which actually reasons in a meaningful way -
@volpeon some of the MOE models that don’t need to run all the parameters in the model seem interesting in that they will run on more modest hardware.
I had one of the quantised quen 3 30b models running on a 24core first gen epyc and getting okish performance.
Getting good (relatively anyway) models that don’t need globs of hardware would be good progression although I suspect it will still burn a shitton of resources training them.
@Dragon The quality of models you can run on personal hardware certainly has improved a lot, but even that feels minor to me. I just don't find myself using LLMs beyond experimenting because I always double check what it says, because I know how they work. As long as these doubts exist, they won't be all that useful to me
-
@volpeon not hallucinating is fundamentally impossible for what an LLM is, which is a probabilistic model that takes in a string of tokens and predicts what should come next. we're a long long way away from something which actually reasons in a meaningful way
@chjara Exactly!
