Promt: I've also tried synchronous implementation and saw the same behavior
Claude and Grok: The issue appears to be a classic async timing problem
We are close to AGI guys!
Follow
@matrix Read an article last week, and some guys were claiming we're maybe 2 years away from AGI.
On one hand I doubt we're THAT close, on another it shouldn't take a genius to realize that the pattern seeking approach that LLMs use is what human brains do too, so maybe we're closer than people realize. I wouldn't be surprised if it took years to realize we have an AGI on our hands after the AGI was created.
@alyx @matrix the G for General makes the bar pretty high, LLMs are good at some tasks but abysmal at others. I found GPT-4 Turbo pretty valuable for language learning but when asked to count syllables in each verse of a poem, it got about a third of them wrong and couldn't even correct most of the mistakes unless I pointed out specific lines or words
this seems to fall under the broader claim that "LLMs can't count" and that seems to be more of a structural problem that you can't just throw more performance at
this seems to fall under the broader claim that "LLMs can't count" and that seems to be more of a structural problem that you can't just throw more performance at
@alyx It's possible that LLMs are soon going to be refined so much that they 100% short queries, but all of them start shitting the bed with growing context and I doubt that's going to change anytime soon.