3 ms·

I'm not sure what it means but a lot of people on this forum have a strong tendency to denigrate human beings whenever people bring up that LLMs kind of suck. I

by __loam 2y ago

I'm not sure what it means but a lot of people on this forum have a strong tendency to denigrate human beings whenever people bring up that LLMs kind of suck. I think this undersells what humans are capable of and distracts from the discussion about what LLMs can do. I'm really tired of this discussion immediately getting sidetracked into some philosophical quagmire about the nature of thinking and the mind when someone is trying to bring up a point about LLMs themselves. When this Apple Intelligence feature comes out, I won't have a tiny human in my phone telling me things, I'll have a ChatGPT api call. LLMs are being integrated into these systems, not human beings. When someone talks about other systems, like search engines, nobody talks about how they compare to humans. It's weird that we're dead set on making the comparison with these systems in particular. Maybe that's down to poorly chosen names for things like artificial intelligence and neural networks inviting the comparison.

seadan83 2y ago
> It's weird that we're dead set on making the comparison with these systems in particular. Maybe that's down to poorly chosen names for things like artificial intelligence and neural networks inviting the comparison. It could be that those names are not poorly chosen. People want grant money, companies want venture capital money. What better way to do that than to use specifically chosen names (also known as propaganda!) Next, there is certainly the "main character syndrome", and the hope of AGI and that these things are on that path, and we are the ones to unlock this. Someone with that point of view, would have a lot of incentive to make those comparisons.
ein0p 2y ago
No, humans legit just suck at associative memory for textual data. This happens to be one of the few things LLMs excel at. LLMs suck at sample efficiency, they do not model most types of cognition, they suck at multimodality (so far), in-context learning, etc, etc. But they absolutely blow humans out of the water on associative memory and taking advantage insanely large amounts of it instantaneously. Humans will not be outdone on other things in my or my children’s lifetime.
- seadan83 2y ago
  What use is associative memory for textual data without the ability to also validate an answer? Honest question. The answer I come to is "fancy autocomplete." If you know there is a significant chance (non zero) of a wrong answer, you have to validate every answer. To me, that describes auto-complete. Useful, but not something you can just use blindly (which greatly limits the utility, ie: human still required).

by __loam 2y ago

seadan83 2y ago

> It's weird that we're dead set on making the comparison with these systems in particular. Maybe that's down to poorly chosen names for things like artificial intelligence and neural networks inviting the comparison. It could be that those names are not poorly chosen. People want grant money, companies want venture capital money. What better way to do that than to use specifically chosen names (also known as propaganda!) Next, there is certainly the "main character syndrome", and the hope of AGI and that these things are on that path, and we are the ones to unlock this. Someone with that point of view, would have a lot of incentive to make those comparisons.

ein0p 2y ago

No, humans legit just suck at associative memory for textual data. This happens to be one of the few things LLMs excel at. LLMs suck at sample efficiency, they do not model most types of cognition, they suck at multimodality (so far), in-context learning, etc, etc. But they absolutely blow humans out of the water on associative memory and taking advantage insanely large amounts of it instantaneously. Humans will not be outdone on other things in my or my children’s lifetime.

seadan83 2y ago
What use is associative memory for textual data without the ability to also validate an answer? Honest question. The answer I come to is "fancy autocomplete." If you know there is a significant chance (non zero) of a wrong answer, you have to validate every answer. To me, that describes auto-complete. Useful, but not something you can just use blindly (which greatly limits the utility, ie: human still required).