News for hackers. Stuff that matters.
Re:GPT-5: Overdue, overhyped and underwhelming. And that's n... by screye on Saturday August 09, 2025 @09:41PM (#44852089)
To be fair, Gary Marcus pioneered the "LLMs will never make it" genre of complaining. Everyone else is derivative [1]. Let the man have his victory lap. He's been losing arguments for 5 years straight at this point.

[1] Due credit to Yann for his 'LLMs will stop scaling, energy based methods are the way forward' obsession.

Re:GPT-5: Overdue, overhyped and underwhelming. And that's n... by lorenzo_medici on Saturday August 09, 2025 @10:07PM (#44852226)
I was a grad student at NYU, its been much longer than 5 years of this, not just with LLMs but with just ML in the past.
Re:GPT-5: Overdue, overhyped and underwhelming. And that's n... by disgruntledphd2 on Monday August 11, 2025 @09:54AM (#44864088)
I mean, his conceptual arguments are pretty good, ML/LLMs/statistical learning methods have real problems with out of distribution inputs and samples, and there's really no way to work around this (that I know of).

I get what you mean though, even if he's right it must've been pretty annoying hearing the same thing all the time.