Hacker News Reader

Transformers are highly parallelized, and generative AI appends.

Even if you ignore that graph problems typically have more time complexity than LLMs offer during inference, they are circuit like and not turing machine like.

Without getting lucky with pre-training matches, it would be surprising if they were good at it at all.

They can’t solve a Sudoku, why do you expect it to be deterministically good at actual reasoning/algorithmical tasks?

same, it’s funny how bad chatgpt is at css… in my experience it can’t even do simple css refactors

There are a lot of bad examples on the internet, "try this, it worked for me" but only because of some other rule up the cascade that the person answering isn't aware of.

there's plenty of this for JS too and somehow it's much much better at that... I wonder what the difference is, JS is more logical?