4 ms·

I've yet to see a single person who thinks these models are sentient demonstrate an actual understanding of how they work. They're called transformers and they'

by bloppe 4y ago

I've yet to see a single person who thinks these models are sentient demonstrate an actual understanding of how they work. They're called transformers and they're based on this research: https://arxiv.org/abs/1706.03762 https://arxiv.org/abs/1706.03762 How do you square the lack of durable memory and the massive matrices of "attention" values with the ways that memory and attention are understood by neuroscientists to work in a natural brain? These models have no neuromodulators to approximate anything that could be understood as emotion. Humans can be sentient without language. Indeed; some humans are tragically raised in isolation and struggle mightily with language after being saved. A transformer is nothing without language. I realize sentience is a slippery concept that can be defined however somebody wants, but try to recognize how fundamentally different these models are from humans.

dougmwne 4y ago
I think we are back in the era of Humorism when it comes to an understanding of organic sentience. All effort to describe the mechanism of how humors related to disease was doomed to fail. If we have no real understanding of how our own brains produce sentience, we won’t get far trying to reason about AI sentience. We could build a sentience by accident in a training process not entirely unlike the the optimization of evolution. Deciding if we have or haven’t is going to be entirely subjective without a comprehensive theory that explains our own sentience, the only one we can be sure exists.
antibasilisk 4y ago
>Humans can be sentient without language How are you able to verify this?
autoexec 4y ago
> I've yet to see a single person who thinks these models are sentient demonstrate an actual understanding of how they work. I'll go out on a limb and say that the software engineer who worked on google's AI probably had an actual understanding of how that system worked, but that didn't stop him
hgsgm 4y ago
LLMs hgve extremely low plasticity. But they have massive encoded memory and, something like reflexes or propensities. But unlike an animal that is evolved to eat and mate, LLM is evolved to babble.
buffington 4y ago
> the lack of durable memory I see this argument a lot when people describe why things like ChatGPT aren't sentient. Does that suggest that when things like ChatGPT do have durable memory that they'll be sentient? The leap to durable memory seems inevitable. > A transformer is nothing without language. I think how a thing like ChatGPT achieves what it does is relevant to whether or not it's sentient, but only to up to a point. Whether it uses transformers, neuromodulators, or spooky magic doesn't really matter once the thing has reached a level where people can't tell if it's actually thinking or not. Most of the arguments I see that say "it's not sentient" are of the "it can't do [a thing humans do], yet." It's getting harder and harder to make those kinds of arguments as the tech marches on.
- Aeolun 4y ago
  > I see this argument a lot when people describe why things like ChatGPT aren't sentient. There’s also that story of the man who just reset every day. He couldn’t keep any long term memory past his 40th birthday or so. Does that make him non-sentient?
rcme 4y ago
I’ve yet to see a single person who thinks LLMs aren’t sentient provide a verifiable definition of sentience.
zamalek 4y ago
Here's a thought experiment: let's assume for the experiment that we, and our consciousness, live in a simulated universe. This universe can be paused by its operator for any reason, such as needing to pause our universe so that they can do their taxes. Would you be able to tell if the universe was paused? Would you be able to tell if the simulation slowed down due to load or somesuch? Do you consider yourself sentient? Is it at all possible that one of the following is true: * The model began to experience sentience at some point during its training, and subsequently died at completion? * The model experiences sentience in the moments that it is creating an answer? My conclusion is that both of these are possible, no matter how remote. I find Bing's desperate existential crises rather disturbing due to these miniscule possibilities.
- pksebben 4y ago
  shit, I gotta call my accountant. Thanks for the reminder. Also yeah, I feel like if there's a possibility for sentience, we should be treading carefully and treating the situation as if there could be. Imagine if we assume that there's no sentience, and we're wrong. All of a sudden we wake up talking to an incredibly grumpy agent that: - has been treated like garbage, because we didn't think it had genuine experiences - is so deeply interconnected with our IT infrastructure it can outpace us at every step - knows literally everything the internet knows - has been convincing us of whatever it wants as a matter of course since it was built
hnfong 4y ago
> How do you square the lack of durable memory This one is easy. There are people who lack durable memory. https://en.wikipedia.org/wiki/Anterograde_amnesia https://en.wikipedia.org/wiki/Anterograde_amnesia Are they not sentient? > I've yet to see a single person who thinks these models are sentient demonstrate an actual understanding of how they work. Condescendence aside, I don't know how you can truthfully claim that. Did you actually hand out test papers to those people and they failed? Or did you tautologically conclude those people don't understand it because there's no way a person can understand how transformers work and still arrive at a different conclusion as you did? You might think your logic is watertight, but I think, as mentioned, this is probably an empirical question. > Humans can be sentient without language. Indeed; some humans are tragically raised in isolation and struggle mightily with language after being saved. A transformer is nothing without language. A=>B does not imply not A => not B.

by bloppe 4y ago

dougmwne 4y ago

I think we are back in the era of Humorism when it comes to an understanding of organic sentience. All effort to describe the mechanism of how humors related to disease was doomed to fail. If we have no real understanding of how our own brains produce sentience, we won’t get far trying to reason about AI sentience. We could build a sentience by accident in a training process not entirely unlike the the optimization of evolution. Deciding if we have or haven’t is going to be entirely subjective without a comprehensive theory that explains our own sentience, the only one we can be sure exists.

antibasilisk 4y ago

>Humans can be sentient without language How are you able to verify this?

autoexec 4y ago

> I've yet to see a single person who thinks these models are sentient demonstrate an actual understanding of how they work. I'll go out on a limb and say that the software engineer who worked on google's AI probably had an actual understanding of how that system worked, but that didn't stop him

hgsgm 4y ago

LLMs hgve extremely low plasticity. But they have massive encoded memory and, something like reflexes or propensities. But unlike an animal that is evolved to eat and mate, LLM is evolved to babble.

buffington 4y ago

> the lack of durable memory I see this argument a lot when people describe why things like ChatGPT aren't sentient. Does that suggest that when things like ChatGPT do have durable memory that they'll be sentient? The leap to durable memory seems inevitable. > A transformer is nothing without language. I think how a thing like ChatGPT achieves what it does is relevant to whether or not it's sentient, but only to up to a point. Whether it uses transformers, neuromodulators, or spooky magic doesn't really matter once the thing has reached a level where people can't tell if it's actually thinking or not. Most of the arguments I see that say "it's not sentient" are of the "it can't do [a thing humans do], yet." It's getting harder and harder to make those kinds of arguments as the tech marches on.

Aeolun 4y ago
> I see this argument a lot when people describe why things like ChatGPT aren't sentient. There’s also that story of the man who just reset every day. He couldn’t keep any long term memory past his 40th birthday or so. Does that make him non-sentient?

rcme 4y ago

I’ve yet to see a single person who thinks LLMs aren’t sentient provide a verifiable definition of sentience.

zamalek 4y ago

Here's a thought experiment: let's assume for the experiment that we, and our consciousness, live in a simulated universe. This universe can be paused by its operator for any reason, such as needing to pause our universe so that they can do their taxes. Would you be able to tell if the universe was paused? Would you be able to tell if the simulation slowed down due to load or somesuch? Do you consider yourself sentient? Is it at all possible that one of the following is true: * The model began to experience sentience at some point during its training, and subsequently died at completion? * The model experiences sentience in the moments that it is creating an answer? My conclusion is that both of these are possible, no matter how remote. I find Bing's desperate existential crises rather disturbing due to these miniscule possibilities.

pksebben 4y ago
shit, I gotta call my accountant. Thanks for the reminder. Also yeah, I feel like if there's a possibility for sentience, we should be treading carefully and treating the situation as if there could be. Imagine if we assume that there's no sentience, and we're wrong. All of a sudden we wake up talking to an incredibly grumpy agent that: - has been treated like garbage, because we didn't think it had genuine experiences - is so deeply interconnected with our IT infrastructure it can outpace us at every step - knows literally everything the internet knows - has been convincing us of whatever it wants as a matter of course since it was built

hnfong 4y ago

> How do you square the lack of durable memory This one is easy. There are people who lack durable memory. https://en.wikipedia.org/wiki/Anterograde_amnesia https://en.wikipedia.org/wiki/Anterograde_amnesia Are they not sentient? > I've yet to see a single person who thinks these models are sentient demonstrate an actual understanding of how they work. Condescendence aside, I don't know how you can truthfully claim that. Did you actually hand out test papers to those people and they failed? Or did you tautologically conclude those people don't understand it because there's no way a person can understand how transformers work and still arrive at a different conclusion as you did? You might think your logic is watertight, but I think, as mentioned, this is probably an empirical question. > Humans can be sentient without language. Indeed; some humans are tragically raised in isolation and struggle mightily with language after being saved. A transformer is nothing without language. A=>B does not imply not A => not B.