Last year, I wrote an essay "The hard problem of qualia in the age of AI", https://zenodo.org/records/20549564 (11 pages PDF). I want to create a linkpost to that essay while taking a fresh look at the subject.
Namely, he describes a stratification into Camp 1,
people for whom neither the notion of qualia nor the claims that the Hard Problem of
Consciousness is exceptionally difficult make much sense, and Camp 2,
people for whom the notion of qualia is well defined and the extraordinary
difficulty of the Hard Problem is obvious.
This means that there is room for camp-specific texts and for intercamp communications (his essay is an example of a very useful intercamp communication). In this sense, the present text and the linked essay are Camp 2 texts, they rely on various notions which seem to make sense only for Camp 2 people. The linked essay contains a longer Section 2, "Methodological remarks", discussing various related details.
What are we taking about:
We say that an entity or a process has phenomenological consciousness if the question "what is it like to be that entity or that process?" has an answer different from "nothing".
To solve the "hard problem of phenomenological consciousness" means:
2.1) to be able to correctly describe which entities and processes have phenomenological consciousness;
2.2) to be able to describe (or, better yet, to experience with reasonable fidelity) what it is like to be that entity or that process for a fairly diverse set of entities/processes.
The main point of the essay is that in order to make progress in solving the "hard problem of phenomenological consciousness" we need to find ways to approach this via well-known standard scientific methods and to overcome the key obstacle that there is no straightforward way to "look at a potentially conscious entity or process from the inside" in order to directly see what (if anything) it is like to be that entity or that process.
In this sense, the essay proposes a two-pronged approach.
The introductory section 1.1, "Theory", starts as follows:
The essence of theoretical scientific methods is being able to make novel, nontrivial,
empirically verifiable predictions.
If someone tells me, “there is this new theory A, and it says that if you
do this weird thing B, then you’ll have this strange perceptual effect C, and,
by the way, no one had even heard about C three months ago,” and if I then
do that weird thing B and actually experience a very strange perceptual effect,
and it sounds to me like C is a good description of my experience, this would
be a strong argument in favor of theory A correctly capturing some important
aspects of reality I am in.
...
Quoting from the introductory section 1.3, "Experimental modulation of conscious experience":
We should experimentally investigate varieties of subjective experience and validate
our theories of consciousness via various techniques that allow us to modify
our subjective experience.
1.3.1 Synchronizations
...
1.3.2 Tight coupling and merges
Tight coupling between electronic circuits and biological entities is particularly
promising as a way for us to try to “look at the dynamic of electronic circuits
from the inside” and as a means to establish better synchronization between
ourselves and other entities (for example, via coupling to the same electronic
circuit serving as an intermediary).
If we require that our theories have novel non-trivial predictive power, this would allow us to differentiate between all those theories of consciousness (there are already dozens if not hundreds of them, and the new ones appear at rapid pace as the interest in the subject is being further catalyzed by the highly non-trivial issues of the possibility and the nature of "AI consciousness").
If we develop an experimental platform in the spirit of "tight coupling", at the very least this would provide us with a novel and rich source of observable phenomenology (and, in particular, might make it much easier to generate novel, non-trivial predictions).
In some sense, this might be a good place to say, "if this looks interesting, one might want to read the essay at this point".
But I'll try to briefly indicate a few more things.
The first two sections of the essay propose to follow a qualia-centric approach ("qualia realism") with minimal possible assumptions ("polysolipsism").
Section 3 gives one example of how a predictive theory might look (just one possibility out of many possible ways, just to make it all less abstract).
Section 4, "Experimental modulation of subjective
experience and tight coupling", is a relatively long section advocating for pursuit of non-invasive BCI (brain-computer interface) projects and their use to induce novel subjective phenomena and investigate relevant issues.
In this sense, I am happy to see the relatively new Altman-backed Merge Labs and its efforts, https://merge.io/blog and https://openai.com/index/investing-in-merge-labs/, but it would be good to have a less proprietary and more open community effort with more affordable tech and more democratic access to the platform.
Pursuing that kind of a research program is, by no means, risk-free. I include rather extensive Section 5, "Risk management and risk-benefit balance", and I end up thinking that we should proceed, rather than refrain from doing so, but I can easily be wrong about that.
The essay concludes with these two paragraphs:
This essay advocates proceeding cautiously but rapidly despite formidable
risks, mostly because it would likely be much better for humans and for AIs
if we actually understand whether the increasingly powerful AI systems we are
creating are sentient, and what it is like to be such an AI system if it happens to
be sentient (what kind of subjective reality it might have, what kinds of qualia
are involved, does it feel well or does it suffer, and so on).
In particular, it would probably be safer to understand this before we start
transitioning to superintelligent AI systems, and before AI systems begin, on
their side, to initiate similar efforts aimed at understanding what it is like to be
a human.
On one hand, solving the "hard problem" is a long-standing dream of many people.
And we do see a lot of research into the issues related to "AI consciousness" predicated on the premise "we can't solve the 'hard problem' fast enough, but still need to address AI consciousness somehow". It would really be nice to solve the "hard problem" fast enough instead (although we should not bet on this and should continue all these other research efforts).
And our ability to formulate what might constitute a "good future" needs to explicitly address both how subjectively conscious entities and processes feel and the properties of the overall reality as a whole, so it would be really nice to understand how subjective experience and the properties of reality actually relate to each other.
But is it existentially risky to actually succeed at solving the "hard problem"? Probably, yes, it's easy to imagine various ways how it can be so (depending on what we actually discover and on what will be done with those discoveries). However, our lack of understanding of what (if anything) it is like to be this or that unfolding process of a computational nature seems to create considerable existential risks as well.
Last year, I wrote an essay "The hard problem of qualia in the age of AI", https://zenodo.org/records/20549564 (11 pages PDF). I want to create a linkpost to that essay while taking a fresh look at the subject.
Epistemic note: In 2023, Rafael Harth published a seminal post describing what seems to be a fundamental stratification of people into two camps with respect to the problem of phenomenological consciousness: https://www.lesswrong.com/posts/NyiFLzSrkfkDW4S7o/why-it-s-so-hard-to-talk-about-consciousness.
Namely, he describes a stratification into Camp 1, people for whom neither the notion of qualia nor the claims that the Hard Problem of Consciousness is exceptionally difficult make much sense, and Camp 2, people for whom the notion of qualia is well defined and the extraordinary difficulty of the Hard Problem is obvious.
This means that there is room for camp-specific texts and for intercamp communications (his essay is an example of a very useful intercamp communication). In this sense, the present text and the linked essay are Camp 2 texts, they rely on various notions which seem to make sense only for Camp 2 people. The linked essay contains a longer Section 2, "Methodological remarks", discussing various related details.
What are we taking about:
We say that an entity or a process has phenomenological consciousness if the question "what is it like to be that entity or that process?" has an answer different from "nothing".
To solve the "hard problem of phenomenological consciousness" means:
2.1) to be able to correctly describe which entities and processes have phenomenological consciousness;
2.2) to be able to describe (or, better yet, to experience with reasonable fidelity) what it is like to be that entity or that process for a fairly diverse set of entities/processes.
The main point of the essay is that in order to make progress in solving the "hard problem of phenomenological consciousness" we need to find ways to approach this via well-known standard scientific methods and to overcome the key obstacle that there is no straightforward way to "look at a potentially conscious entity or process from the inside" in order to directly see what (if anything) it is like to be that entity or that process.
In this sense, the essay proposes a two-pronged approach.
The introductory section 1.1, "Theory", starts as follows:
Quoting from the introductory section 1.3, "Experimental modulation of conscious experience":
If we require that our theories have novel non-trivial predictive power, this would allow us to differentiate between all those theories of consciousness (there are already dozens if not hundreds of them, and the new ones appear at rapid pace as the interest in the subject is being further catalyzed by the highly non-trivial issues of the possibility and the nature of "AI consciousness").
If we develop an experimental platform in the spirit of "tight coupling", at the very least this would provide us with a novel and rich source of observable phenomenology (and, in particular, might make it much easier to generate novel, non-trivial predictions).
In some sense, this might be a good place to say, "if this looks interesting, one might want to read the essay at this point".
But I'll try to briefly indicate a few more things.
The first two sections of the essay propose to follow a qualia-centric approach ("qualia realism") with minimal possible assumptions ("polysolipsism").
Section 3 gives one example of how a predictive theory might look (just one possibility out of many possible ways, just to make it all less abstract).
Section 4, "Experimental modulation of subjective experience and tight coupling", is a relatively long section advocating for pursuit of non-invasive BCI (brain-computer interface) projects and their use to induce novel subjective phenomena and investigate relevant issues.
In this sense, I am happy to see the relatively new Altman-backed Merge Labs and its efforts, https://merge.io/blog and https://openai.com/index/investing-in-merge-labs/, but it would be good to have a less proprietary and more open community effort with more affordable tech and more democratic access to the platform.
Pursuing that kind of a research program is, by no means, risk-free. I include rather extensive Section 5, "Risk management and risk-benefit balance", and I end up thinking that we should proceed, rather than refrain from doing so, but I can easily be wrong about that.
The essay concludes with these two paragraphs:
On one hand, solving the "hard problem" is a long-standing dream of many people.
And we do see a lot of research into the issues related to "AI consciousness" predicated on the premise "we can't solve the 'hard problem' fast enough, but still need to address AI consciousness somehow". It would really be nice to solve the "hard problem" fast enough instead (although we should not bet on this and should continue all these other research efforts).
And our ability to formulate what might constitute a "good future" needs to explicitly address both how subjectively conscious entities and processes feel and the properties of the overall reality as a whole, so it would be really nice to understand how subjective experience and the properties of reality actually relate to each other.
But is it existentially risky to actually succeed at solving the "hard problem"? Probably, yes, it's easy to imagine various ways how it can be so (depending on what we actually discover and on what will be done with those discoveries). However, our lack of understanding of what (if anything) it is like to be this or that unfolding process of a computational nature seems to create considerable existential risks as well.