Very good article. I think about how language and the ability to convey ones self has nothing entirely to do with intelligence quite often. I frequently have days where I sound like I have forced speech issues and I sound "stupid". My father has deep brain stimulators for essential tremor and when he has those on his high setting that controls his tremors in both hands very well, he has forced speech. Sounding "stupid" has nothing to do with what is going on in the brain. We need to stop judging others and ourselves on how fluidly we can speak or type or write.
The problem most people have including the author of this article is not being able to distinguish between different types of cognition or related characteristics of animals such as humans and accept that some may exist without the others.
Sentience, subjective experience, the ability to reason or understand abstract meaning in a functional way, emotions, self-awareness, autonomy, being alive, fast adaptation to environment, spatial understanding, etc., are all different things.
GPT-4 has proven that you can have useful reasoning ability without a lot of the other things. It also shows that some spatial/visual understanding does not require eyes or a body.
GPT-4 probably doesn't have the same richness or level of understanding of some concepts as even a dog does since it's world model is mostly text and some static images, but functionally in many contexts it can still understand and be an effective problem solver.
It's true that LLMs are not close to simulating digital humans and are missing all kinds of basic capabilities of animals. Still, the leading edge LLMs do have a useful type of intelligence.
Actually, people are eventually going to find out the hard way that it was a _good thing_ that all of those animal characteristics were missing from large multimodal models. Because not only are researchers going to keep finding ways to make them more animal-like (most if not all of the missing characteristics are those shared with other animals) but also they will keep making them faster.
Eventually they will be dozens or more times faster at thinking than humans and quite lifelike. Again, GPT-4 shows the lifelike characteristics are not necessary to get useful intelligence. And we have enough movies and TV shows demonstrating why it's a very stupid path.
So many jumps to conclusions I had to quit early. Assuming we have a clear definition of “understanding”, “consciousness”, intelligence, ai.. etc.
“It’s just auto-complete”.. more of the same takes dabbling with the same unknowns but with high confidence. My subjective experience is that I’m basically just a spicy multi-modal auto-complete. Rarely do I ever feel like I truly “understand” something. I mostly operate on varying degrees of “confidence”. Maybe I’m just dumb like LLMs though. Maybe I have no soul?
Though I agree in spirit I with what I think the author is getting at. There are diverse intelligences. Language is a medium for communicating our intelligence.
5 comments
[ 3.2 ms ] story [ 19.1 ms ] threadSentience, subjective experience, the ability to reason or understand abstract meaning in a functional way, emotions, self-awareness, autonomy, being alive, fast adaptation to environment, spatial understanding, etc., are all different things.
GPT-4 has proven that you can have useful reasoning ability without a lot of the other things. It also shows that some spatial/visual understanding does not require eyes or a body.
GPT-4 probably doesn't have the same richness or level of understanding of some concepts as even a dog does since it's world model is mostly text and some static images, but functionally in many contexts it can still understand and be an effective problem solver.
It's true that LLMs are not close to simulating digital humans and are missing all kinds of basic capabilities of animals. Still, the leading edge LLMs do have a useful type of intelligence.
Actually, people are eventually going to find out the hard way that it was a _good thing_ that all of those animal characteristics were missing from large multimodal models. Because not only are researchers going to keep finding ways to make them more animal-like (most if not all of the missing characteristics are those shared with other animals) but also they will keep making them faster.
Eventually they will be dozens or more times faster at thinking than humans and quite lifelike. Again, GPT-4 shows the lifelike characteristics are not necessary to get useful intelligence. And we have enough movies and TV shows demonstrating why it's a very stupid path.
“It’s just auto-complete”.. more of the same takes dabbling with the same unknowns but with high confidence. My subjective experience is that I’m basically just a spicy multi-modal auto-complete. Rarely do I ever feel like I truly “understand” something. I mostly operate on varying degrees of “confidence”. Maybe I’m just dumb like LLMs though. Maybe I have no soul?
Though I agree in spirit I with what I think the author is getting at. There are diverse intelligences. Language is a medium for communicating our intelligence.