Only the best and brightest are working for Anthropic.
-
@Mabande thus, Anthropic's J-space paper seems sound: neuroscientists have a working/operational definition of (a form of?) consciousness that yields testable predictions. (Else, not science?) They find that Claude exhibits some of those predicted properties.
They don't take a stance on (the hard problem of) consciousness, merely whether Claude exhibits behaviour predicted by one definition of consciousness.
To me, it seems that, to disagree, we need to show a flaw in their derivations.
3/3
@rowat_c Sorry, but it's useless to try to have a conversation if one of us redefines or throws out well-established definitions on the fly.
It simply makes it so we can't agree on anything, even the basics of what the conversation's about. -
@rowat_c Sorry, but it's useless to try to have a conversation if one of us redefines or throws out well-established definitions on the fly.
It simply makes it so we can't agree on anything, even the basics of what the conversation's about.@Mabande I think we agree: if we want to make non-tautological statements about C- or Si-based intelligence or consciousness, we need to operationally define those terms: "X has property Y if it exhibits behaviour Z".
-
@Mabande I think we agree: if we want to make non-tautological statements about C- or Si-based intelligence or consciousness, we need to operationally define those terms: "X has property Y if it exhibits behaviour Z".
@rowat_c Problem is when there's a vast discrepancy re. what "exhibiting behaviour" means.
Successful mimicry of a property (eg the baseline for passing the Turing test) is not the same as having the property. -
@rowat_c Problem is when there's a vast discrepancy re. what "exhibiting behaviour" means.
Successful mimicry of a property (eg the baseline for passing the Turing test) is not the same as having the property.@Mabande sure. All I'm saying is that, if someone wants to declare "X has/doesn't have property Y", then they need to have a definition of Y that has testable implications that they're willing to stand behind, rather than moving whenever they get a result they don't like.
Why do we think I'm intelligent, or conscious, rather than just passing all the tests through mimickry + appropriate regularisation so I can generalise beyond my training data?
-
@Mabande sure. All I'm saying is that, if someone wants to declare "X has/doesn't have property Y", then they need to have a definition of Y that has testable implications that they're willing to stand behind, rather than moving whenever they get a result they don't like.
Why do we think I'm intelligent, or conscious, rather than just passing all the tests through mimickry + appropriate regularisation so I can generalise beyond my training data?
@rowat_c Usually it's the one making the claim who needs to prove it, and extraordinary claims need some more solid proof than passing a test passed by chatbots since the sixties (or a paper by a company reliant on hype to bring profits).
So if you claim LLM:s are intelligent it's up to you to provide the evidence for it (other's works are fine), and then anyone can examine the evidence to verify/falsify the claim.
This far, there's nothing supporting claims that true AI ever been created. -
@rowat_c Usually it's the one making the claim who needs to prove it, and extraordinary claims need some more solid proof than passing a test passed by chatbots since the sixties (or a paper by a company reliant on hype to bring profits).
So if you claim LLM:s are intelligent it's up to you to provide the evidence for it (other's works are fine), and then anyone can examine the evidence to verify/falsify the claim.
This far, there's nothing supporting claims that true AI ever been created.@Mabande I don't know what "true" AI is, but worry that words like "true"/"truth" (whether or not social)/"pravda" risk being used in the same way that "democratic" does in a country's name.
On consciousness, q.v. my comments above on Anthropic J-space paper.
On intelligence, my working definition is ≈ some of the thinking I need to do can now be done by an LLM.
Do you see the statement that "LLMs are not intelligent/conscious" as a claim? If so, how do people making it pass your test?
-
@Mabande I don't know what "true" AI is, but worry that words like "true"/"truth" (whether or not social)/"pravda" risk being used in the same way that "democratic" does in a country's name.
On consciousness, q.v. my comments above on Anthropic J-space paper.
On intelligence, my working definition is ≈ some of the thinking I need to do can now be done by an LLM.
Do you see the statement that "LLMs are not intelligent/conscious" as a claim? If so, how do people making it pass your test?
@rowat_c Now this is also part of the problem of discussing things with someone who doesn't believe in definitions: there's always an aspect of bad faith arguing.
Going off on a tangent about pravda? Miss me with that nonsense.
True AI as in AI proven to be an actual manufactured intelligence, passing the relevant tests.Your working definition of intelligence means a Casio fx-180Pv's intelligent.
But okay, explain the J-space paper to me. In which way does it prove consciousness in Clause?
-
@rowat_c Now this is also part of the problem of discussing things with someone who doesn't believe in definitions: there's always an aspect of bad faith arguing.
Going off on a tangent about pravda? Miss me with that nonsense.
True AI as in AI proven to be an actual manufactured intelligence, passing the relevant tests.Your working definition of intelligence means a Casio fx-180Pv's intelligent.
But okay, explain the J-space paper to me. In which way does it prove consciousness in Clause?
@rowat_c
* "Do you see the statement that "LLMs are not intelligent/conscious" as a claim?"
- It's a refutal of the claim that LLM:s are intelligent/conscious. -
@rowat_c Now this is also part of the problem of discussing things with someone who doesn't believe in definitions: there's always an aspect of bad faith arguing.
Going off on a tangent about pravda? Miss me with that nonsense.
True AI as in AI proven to be an actual manufactured intelligence, passing the relevant tests.Your working definition of intelligence means a Casio fx-180Pv's intelligent.
But okay, explain the J-space paper to me. In which way does it prove consciousness in Clause?
@Mabande **Consciousness**. I won't explain the Anthropic paper as well as it does, so encourage you to read it: https://www.anthropic.com/research/global-workspace
n.b. It (and I, above) do *not* say it claims that Claude is conscious.
**Intelligence**. Yes, good point: by my working definition, I would see a calculator as slightly intelligent. That seems fair: it helps me think in some particular ways.
If you're ever comfortable sharing your definition of "true" intelligence, I'd be keen to hear.
Thanks for chat!
-
@Mabande **Consciousness**. I won't explain the Anthropic paper as well as it does, so encourage you to read it: https://www.anthropic.com/research/global-workspace
n.b. It (and I, above) do *not* say it claims that Claude is conscious.
**Intelligence**. Yes, good point: by my working definition, I would see a calculator as slightly intelligent. That seems fair: it helps me think in some particular ways.
If you're ever comfortable sharing your definition of "true" intelligence, I'd be keen to hear.
Thanks for chat!
@rowat_c So, you can't say what in the Anthropic paper that "find that Claude exhibits some [consciousness] properties."
The primary reason I want you to explain it is I want to understand if you understand it and not just having being awed, the secondary is that I haven't got the time to read their PR.
So, a calculator is intelligent, thus an abacus is intelligent, as are scratches in soft clay, beans dropped in a recess etc.
Don't confuse the tool of intelligence with intelligence itself. -
G gambolputte@expressional.social shared this topic
