What kind of new architecture do you think will be common now that auto-regressive transformer based LLMs has been prove that will not reach any form of usable IA.

>>106005184 (OP)
after the 145 kruger point comes the realization that llms are artificial resoners
i agree that we will probably have and benefit greatly from arch changes but there is nothing in current llms that makes them incapable of becoming "agi" after 10 years of similar improvements that we had up until now

>>106005241
>naive extrapolater naively extrapolates

>>106005260
>midwit begs the question by assuming the extrapolation is naive because it just is
there are years of datapoints by this point and no signs of stopping, neither practical nor theoretical

>>106005184 (OP)
Leftist meme

>>106006703
>3 sentences = wall of text

ADHD

>>106007191
It takes up 75% of the picture. He's right. Learn 2 meme, you retard.

>>106005184 (OP)
Autoregressive is fine, the problem is that the training is not autoregressive.

>>106007191
It's a meme, not a dissertation. Understand the medium you're trying to use.

>>106005241
>that llms are artificial resoners
they arent, they still require to use classical algorithms to guide them using prompts and are uncapable to escape the trial and error loop because can't reason to create novel answers. No matter what temperature you use, is a lie that they are general reasoners , cant even beat simpler neural networks like lc0.