Matt:* By that definition, no objective test exists.*
Todor: Yes, IMO there is no objective test.
Маtt: *"It amazes me that people have not yet accepted that AI has already
surpassed human intelligence."*
Todor: Indeed it has surpassed it in many ways, it depends on the definitions.
Even GPT2 Medium generates better texts than most humans. Consciousness
(subjective experience) and intelligence are different dimensions, as explained
here and "everywhere". However the localization of the agents and "what they
are" is problematic.
Matt: *"Humans are social animals that evolved to suffer when they see others
suffer." *
Todor: This is far from universally true, there are much more severe cases than
the Stanford experiment (it might have been fraud as well). There is no one
monolithic thing called "humans".
*Human systems *and humans *as tokens and cogs in these machines *evolved also
to **want to cause other humans to suffer** if that would give them, both the
system or them personally, an advantage. Humans closest relatives - the chimps
- also sometimes exhibit this behavior.
Signs of "consciousness", even ones matching some official definitions, do not
save other people from being slaughtered by other "empathetic and conscious"
apes, and sometimes it rather **encourages* *the sadistic agents to apply more
violence, because they *want their victim to suffer. *If the oppressors believe
that the victim "is not sentient", they would be less aggressive - that is like
beating a dead body or a puppet.
The human "empathy" is addressed in AGI safety context by the early 2000s (or
by early 1960s by others, or by early 1920s by Karel Capek, and by philosophers
like Kant, Schopenhauer or Marx by 18-th and 19-th century).
Matt: *"removing the amygdala"*
You remove *manifestations*, behaviors that you *associate with*.
External behavior or observable phenomena are not equal to the subjective
experience. If they were, then the objective=subjective, therefore they match.
Humans often cannot even get that.
If a person suffers a stroke, she may have her facial nerves paralyzed, other
subcortical structures damaged and neural pathways broken, motor areas damaged,
cerebellum damaged, neurotransmitter damaged. They cannot exhibit the behavior
as complete agents, "human beings", that they had before the trauma. You do
not know if "they", whatever is locked "inside", does not keep suffering and
experiencing pain, however being unable to express it in a way that you, or any
evaluator-observer, considers the "legitimate" way to express having
consciousness, sentience, feeling pay etc.
*EXPRESSING* pain in a "codified" or "conventional way" is what the "vanilla"
homo sapiens expect, e.g.. they search for matches to their own known behaviors.
In addition, referring to the theory of the Integral of infinitesimal selves
(see "Analysis of the meaning of a sentence ...", 2004 which addresses this
concept and the multi-level, multi-scale, multi-agent cumulative reward
hypothesis for modeling of complex human-like intelligent agents; another paper
from 2013? on "Artificial Mind" which also addresses the lack of an Integral
self, but claims as the original one that the integrity is constructed by the
boundaries of the body and it represents the predictive model of the
evaluator-observer who tries to explain he behavior by his compressed and
incomplete models.
Matt: *"**If it passes the Turing test then it is indistinguishable from having
subjective consciousness"*
Todor: *IMO there is no such logical entailment.* The concepts should first be
defined, as well. What is "subjective consciousness"? How precisely the Turing
test is defined? What exactly counts as "passing" - human judges ("random" or
any kind) may cheat, their opinions are not a serious measure. There must be
explicit measure, and if there is one, in a *text-only* test (code-only) it can
be passed.
What this proofs is that the system "passes it", nothing about the
consciousness. You can believe any object is "consciousness", a simpler dialog
program "passes the Turing test" (depending on how many questions, which
questions, how much and what requirements you have etc. - for someone a very
simple conversation is "human-like", for another one - it should be like if he
talks to Kant; however if the machine speaks like Kant, the first human may
say: "a human cannot talk like that, this is fake".
The evaluators (=Eva) may watch for different marks, signs, details for their
judgment, for example:
1. Is the text of the Interlocutor (Max) internally coherent,
2. Does the text in each message of Max properly reflects what Eva said?
2.1. In the previous message. 2.2. In earlier message. 2.3. How deeply in the
past? 2.4. How complex? 2.5. What is the memory of Eva? 2.6. Is Eva allowed to
review old messages and to verify them with other tools, more than just the
"teletype" keyboard?
Matt: *"consciousness, which requires two agents, one to predict human behavior
and the other to carry out those predictions in real time.** "
*
This is not consciousness as sentience, awareness, phenomenal experience etc.
and "self-awareness" in spiritual sense. "Real time" is elastic, complexity is
elastic (what aspects of "human behavior" - how many bits to describe it; why?
what precision? text is a very few bits).
When one predicts and can cause other "entity" (sub-universe, data, whatever)
to change in the desired ways, this is "control", connection, mapping,
reflection, influence - not consciousness.
"Humans" themselves are "elastic" as a concept and exist in many variations and
partially.
Matt: *"The third option is just to convince ourselves that machines don't
suffer.** "
*What is a machine? Where exactly its boundaries lie? When does it suffer and
when it doesn't? When a human suffers - if this is subjective, you cannot know,
if may be silent about it. One may be very noisy and suffer less than someone
who is mute. An agonizing man might be in a hell of pain, but already unable to
move, and as he cannot "return from the outer world" and tell - an observer may
believe that the agonizing person was already saved from the torture of life.
There was a case of a man, sentenced to death with an injection, where the
procedure went wrong and they tortured the poor fellow for half an hour or an
hour with terrible way, burning him from the inside, semi-awaken etc.
Indeed, the "justified executions" by the judicial systems are an example of
humans who want to see another human suffering. Some of the official bullshit
is that the procedures are designed "to cause less suffering" - the guillotine
also had a "copywriter" who advertised it like that.
If the sadist, wearing suits, and the apes, wearing any kinds of clothes,
really wanted the sentenced to death not to suffer, the procedures would be
with methods which make the victim calmly fall asleep, be tired etc. and die in
his sleep. Yet they still *don't know* what their victim feels, as it cannot
return from the deaths to tell.
The poor guy who was tortured woke up between the steps of the sadistic
procedure and he did express the "path" if the brains capability to record the
sensation and feelings, whatever they are, wasn't shut off yet, but appeared
off for some time.
*References:
*"death sentence injection gone wrong man painful" -- see in another message.
------------------------------------------
Artificial General Intelligence List: AGI
Permalink:
https://agi.topicbox.com/groups/agi/T1b6c21f92326684e-M88eb12e780c8823b0a995dbf
Delivery options: https://agi.topicbox.com/groups/agi/subscription
