I would say it's good journalism to take the claim of "general purpose" intelligence by AI companies at face value and see that it doesn't hold up. Making outlandish claims allows for silly comparisons
I wouldn't say a normal CPU is inefficient at graphics or cryptography, rather that a specialized GPU is particularly efficient at those tasks.
We only consider a CPU slow at these tasks because of how much faster a GPU is with them, but we never see how much worse a GPU is at general conputation tasks, because of how stupendously bad it is.
As soon as operations need to share info, the GPU speed advantage is gone. Branching paths bog a GPU down with redundant execution. Latency is quite poor too. And exceptions & interrupts are basically impossible at the system level. Trying to run normal programs on the GPU would be a disaster.
“Their learning curve was actually negative,” Lindell said. “The more time we gave them to complete basic land-based tests, the more pitiful their efforts became, with many of them opting to bask in the sun rather than perform a simple task.”
Never mind the fact that keeping a squid in a tank they’ll kill themselves within the day as well. We do not know how to keep adult squid in captivity.
27 replies
How journalists sound when they compare an LLM to specialized software that plays chess
Journalists are not the source of the claim that LLMs are the path to AGI.
They say en passant was invented to justify an illegal move from LLM
Holy hell!
I would say it's good journalism to take the claim of "general purpose" intelligence by AI companies at face value and see that it doesn't hold up. Making outlandish claims allows for silly comparisons
I think this depends on how you define "general purpose".
The CPU is a "general purpose" processor but it's not very efficient at graphics or cryptography tasks.
I wouldn't say a normal CPU is inefficient at graphics or cryptography, rather that a specialized GPU is particularly efficient at those tasks.
We only consider a CPU slow at these tasks because of how much faster a GPU is with them, but we never see how much worse a GPU is at general conputation tasks, because of how stupendously bad it is.
As soon as operations need to share info, the GPU speed advantage is gone. Branching paths bog a GPU down with redundant execution. Latency is quite poor too. And exceptions & interrupts are basically impossible at the system level. Trying to run normal programs on the GPU would be a disaster.
Yeah, that's an unfair comparison because LLMs aren't smart enough to play chess. They're better at Candyland
For a fair comparison, I need to see the chess program give me my grandma's family recipe for napalm.
https://theonion.com/study-dolphins-not-so-intelligent-on-land-1819568299/
That squid is cosplaying as an octopus.
Never mind the fact that keeping a squid in a tank they’ll kill themselves within the day as well. We do not know how to keep adult squid in captivity.
Source?
Look it up literally all the first results on the web say this.
I dunno what you looked up, but my search returns multiple sites with instructions for keeping squid at home?
This is a variation of an idiom that goes something like "if you judge a fish by its ability to climb a tree, it will think it's stupid".
But as someone who used to raise killifish, I actually found out that there is at least one species of fish that can climb trees.
There is an aquarium supply store near me that has an open-top mudskipper display with a tree in it. Those guys definitely can climb trees.
Relevant Flying Circus
Slrpnk.net images are not loading for me, even in the browser.
This loaded. Thanks!
The solution to the problem is often logging out, clearing SLRPNK.net cookies, and then logging back in.
I'm from another instance and using Boost on Android. But, thanks 👍🏻
@[email protected]
Love Wulff & Morgenthaler