Ask A Genius 1049: The Paul Cooijmans Session

Author(s): Rick Rosner and Scott Douglas Jacobsen

Publication (Outlet/Website): Ask A Genius

Publication Date (yyyy/mm/dd): 2024/07/30

Rick Rosner, American Comedy Writer, www.rickrosner.org

Scott Douglas Jacobsen, Independent Journalist, www.in-sightpublishing.com

Scott Douglas Jacobsen: This is from Paul Cooijmans. “If possible at all, how can we verify whether a system or being is conscious?”

Rick Rosner: AI will claim and act as if it is conscious long before it’s conscious. Because AI takes, in these large language models, the thoughts of millions of people, tens of millions, expressed verbally and comes up with a probability model, what will the following words be? What will the following sentence, paragraph, and entire essay be as prompted by the user typing in a query? Since all these words that it’s working from are the product of conscious humans, AI will deliver output that claims to be mindful or appears to think that it’s conscious, even though it doesn’t think.

This doesn’t answer your question, but now, I’m going to poke from a different direction: we have tacit consciousness. Our brains work as if they’re conscious to the point where we are conscious. I use the analogy when you go to a museum and look at a painting. You stand six feet away or ten feet away from a giant Rembrandt. At ten feet, your eye can only look with high definition at an area of the painting that might be eight inches across. So your eye scans the whole thing.

You keep enough of these scans as your eye covers the painting. You eventually don’t need to assemble a complete image of the painting in your consciousness. You send enough information to your brain. That it behaves as if it has a completeversion of the painting. The associations triggered in your brain are the things that pop up, the memories, the thoughts. It managed to get an entire picture of the painting into your consciousness, but you didn’t even do that. So, our consciousnesses are as if things. I should do more work to develop a more coherent, less hand-wavy way of talking about it, but we need a complete picture of the world that we’re in the way we think we do. We’re constantly patching together enough of the world for little snippets of time that the associations that come out of our brains—because our brains are association engines—are pulled out of our stored information.

The experience of the world is complete and continuous and not glitchy, which is to say that our consciousnesses are undeniably conscious, but our consciousnesses are not as conscious as we think we are in the way we have all the emotions that go into consciousness. It’s just that the informational structure that’s being built moment to moment in our minds is frickin threadbare.

This presents issues when trying to determine if other entities are conscious. Another issue is that we don’t need the mathematics of consciousness, which we will eventually do. I’ve been lazy my whole fucking life and haven’t been able to substantiate this. Still, the locally three-dimensional and temporally one-dimensional universe is the most efficient way to present the information within, in the case of the universe, a vast consciousness.

You can leave consciousness out, but the most efficient way to map or present the information within a vast, self-consistent information processing system looks like the universe. If you could map the information in a human mind, it would have many of the same physics. The three spatial, one temporal, the same particles as the universe. Frickin words, but I think that when we have—maybe I’m wrong, maybe it’s some other structure, but I doubt it—an understanding of that structure, we will be able to model consciousness mathematically. 

People like Max Tegmark. He has a half-assed idea of how to model consciousness based on bandwidth, more or less. Though, he’s not far off. It needs to have multimodal bandwidth. It would help if you had a geometry that goes with it. But anyway, if we ever get to do that, we’ll at least be able to model consciousness in the abstract. We should be able to, working with that, assign an index of the amount of information in any given consciousness, at least abstract consciousness, and apply that to the engineered consciousnesses.

You can do that in the future, but we’re already running into the problem of them needing to be more black-boxy, like Google Translate. It depends on how many servers are connected. I think people at Google speculate that it’s got this. The base language is no human language, but it is just like these word nodes in some abstract word space that if you’re trying to translate “I love you with the intensity of a million suns” from English to Ukrainian.

It doesn’t go directly from English to Ukrainian. It goes into this word space or phrase space or some freaking information space. That is where all the different word concepts live. That’s informationally more efficient, and consciousness is an efficient way to model the world. Google Translate is not conscious of this. However, the issue I’m bringing up is that we need to be tuned to know what is happening inside Google Translate. It’s black boxy, so if you’ve got artificial intelligence, it’s an engineered thing that lives across 1,200 servers on a barge floating off the San Francisco coast. How do you tell what’s going on inside of there? 

The Go AI trained itself to play Go unless you develop a technology that lets the engineered system interrogate itself. Now, it’s better than any human player, but it makes moves unintelligible to humans. Ditto for chess. Those things aren’t set up to be interrogated. Also, I’m not conscious at all, just probability nets. Also the Turing test, in the last two or three years, we’ve seen that the idea of the Turing test has been blown out of the water because there’s no one Turing test. Growing up in the seventies, we learned about the Turing Test. 

It was: You’re just typing, and something on the other side of the wall is typing. If you can’t tell that it’s not a person typing, then it’s conscious. It’s thinking. It’s some freaking thing. Now, with all the different versions of AIs coming out and how convincingly they can write, how well they can write, not originally, they can’t or do art. When you first see the product of the art generated by AI, you can’t tell that a human illustrator didn’t make it. So it’s passed the Turing test, just for half a second, or 10 minutes. Because if you look at the images for 10 minutes, for 20 minutes, it could be a better image.

Then, you get an idea of the standard output of an AI graphics engine. I look at a lot of AI porn because A, I like naked ladies and B, I like seeing the rapidly changing product generated by the combined horniness of tens of thousands of people typing in prompts. Things started like a year ago. All the naked ladies looked like Japanese anime. All the lonely guys in their basements were anime guys just trying, and they had an anime product. Then, it went through all these iterations very rapidly.

Now, you see the guys into old ladies typing in prompts. You see that there are the guys who are into tentacles. You can see the different emotions in the anime era; all the women were happy to see you and be naked for you. You see many emotions as people add their kinks to the prompts. This means that whatever version of the Turing test you try to apply is a moving target. Eight years ago, I started looking at an app called This Person Does Not Exist. You’d just hit the button, giving you the face of somebody who didn’t exist, assembled from a probability net based on pictures and photographs of hundreds of thousands of people.

At first glance, that’s a real person who passes the Turing test for being real, but if you look at a few of them, you start seeing what it couldn’t do. Initially, it couldn’t do glasses. It didn’t understand that glasses have to have physical integrity, the same way porn couldn’t do underwear. It didn’t understand that you can’t just have a piece of cloth floating on somebody’s midriff. 

It has to have a strap around the waist, or it’ll fall off. So it didn’t understand glasses, didn’t understand underpants, and famously didn’t understand hands and how many fingers people have. So, this person is fine with glasses or earrings because there’s a great variety of earrings and teeth. Backgrounds are important because there’s a multiplicity of backgrounds presenting their understanding of issues. Over the years, it’s improved. I have yet to look at it now since the coming of stronger AI, but the last time I looked at it. It had no problem with glasses, still a little problem with earrings, and still not great on backgrounds.

However, about the Turing test, this is not a person circa 2017 who could pass the Turing Test for five minutes or for looking at five fake photos. But if you looked at 20 of them, you’d start to notice what was wrong with them, and then it would flop. So, we haven’t even touched on the Turing.  We’ve discussed the practical obstacles to deciding whether something is conscious. There’s also the

Jacobsen: May I interject?

Rosner: Yes, please, because I’m about to confuse myself. 

Jacobsen: There is a separation between conscious and non-conscious processing. How would you differentiate that non-conscious processing? 

Rosner: We’ve been focusing more on the practical. The separation between conscious and subconscious, or non-conscious processing, is, often, what your brain decides needs attention, so that you can let some processing tasks proceed without your awareness. Usually walking, almost always breathing, you’ve got to breathe all the time. You generally don’t run into obstacles to breathing. Walking: You can often walk without paying attention because we have built a world of flat surfaces to walk on. You have to pay more attention if you’re hiking or in a neighbourhood with crappy sidewalks or traffic. 

But consciousness is your brain. There’s a hot thing in brain science that seems obvious, but it wasn’t previously obvious: your brain’s job, conscious and unconscious, is to prepare you for the next moment. It’s always trying to model reality and put you in the best position to respond to what it predicts you’ll face next so you don’t die. 

So often, the division between conscious and subconscious, or unconscious, processes is an allocation of resources that your brain–that you’ve evolved–to have a brain that wants to devote its resources in such a way that you don’t mess up, that you minimize your errors, maximize actions that maximize your well-being, and some of this stuff is hardwired, e.g., breathing is super not conscious because you always need to be doing it. 

It usually doesn’t present issues. You’re usually in a place where you can get air. There’s a thing in sports where a process that’s become less conscious.Because you’ve repeated it, you’ve done the action so many times. In baseball, it’s called the yips, where somebody’s been a pitcher for 15 years, starting in junior high. Now, they’re in the minor leagues. They have these motions that are pretty hardwired, but then, all of a sudden, they start thinking about them.

The motion falls apart, and they suddenly become very bad at it. Simone Biles got a case of this in the 2021 Tokyo Olympics. In gymnastics, it’s called the twisties because they flip through the air thousands of times. They have an intuitive, not entirely conscious, understanding of how their body travels through space. But for some reason, she started becoming self-conscious of her actions. They started falling apart slightly. But it’s all resource management. Your brain does the best it can do, given that it’s got limited resources.

So, walking is right on the boundary between conscious and unconscious. Many aspects of vision are pre-processed unconsciously before they get to our consciousness. Your brain picks out straight lines and shading and imposes perspective, and a lot of that is not formed consciously. You don’t often run into a situation that defies your unconscious perspective engine. Some rooms trick your perspective engine—the one where it looks like a cubicle room but is like a big parallelogram. So, you put two guys at opposite corners, and a six-year-old can look bigger than an adult.

Because the room has been skewed with a false perspective, there’s the whole deal where you see somebody lurking in a doorway because your brain has taken limited information and bet you better let the person think somebody might be there. These border issues are things that, under normal circumstances, you don’t have to pay attention to. What else? We could talk about how sex defies your brain’s normal behaviour because it’s tricking you into doing stuff that’s not in your self-interest, but I don’t want to. That can be discussed and can be found elsewhere. 

Jacobsen: Here’s a question. Would the verification process of this system or being as conscious differ if you were to deal with an evolved system instead of an engineered one?

Rosner: Chris Cole has an idea. Because we’re evolved creatures, there are many different feedback systems in the human body at all levels. Organ to organ, all reasonably followed shortcuts or helpful connections have been made in the fullness of time. He is trying to develop an atlas of all these feedback loops. So we’ve got many of these based on a couple of billion years of evolution. An engineered system is likely to have fewer feedback loops.

Because A, we’re engineering it, and B, it has yet to have a billion years or generations of organisms to work its way through, it has more data. An engineered intelligence has different data. We have sensory information. A large language model has letters, words, and sentences.

So, because we’ve already established this, we’ve proved that we’re conscious because we have the experience of consciousness. Every person knows they’re conscious because they feel conscious, which is based on the multimodal real-time intake of sensory information experiences. We know we’re conscious even though we think we’re more conscious than we are. That doesn’t prove it to the satisfaction of people asking for proof, but we know. Eventually, we’ll have some mathematics that verifies that.

If we have the mathematics and a model we’ve discussed, you and I have discussed this, where many of the things we do within consciousness are similar to what happens in AIs. It’s like we have a probability net based on accumulated experience. When we’re saying a sentence, we don’t continue and carefully consider every word. We have these probability nets for grammar. I didn’t have to think.

For example, when I said “probability nets for grammar,” I didn’t have to think about what came after “nets.” I knew the right word was “for” rather than “or,” “as,” or some other connector word like that. Half the words in our sentences or more just come from processes in our brains, like AI. But we have the multimodal thing in consciousness: you have all these differences.

Nodes that each take a different angle and focus on it. They focus on a different aspect of the world, the model of the world we’re building, and they share the information. Every mode shares, and all these nodes share their analysis withAI, which is nothing like that now. However, some companies try incorporating verbal and graphic input when you start doing that. So it’s a two-node thing.

And again, that doesn’t give you consciousness because you’ve got just two fricking nodes. The AI that’s working from that system still doesn’t know anything. It’s just doing probability. But if you did enough of it with enough real-time, actual sensory input, we now know the elements of consciousness.

And some of them aren’t necessary, but we know based on that we’ve got real-time sensory input, multimodality, and agency. We can move around and move things in the world; we have value judgments and emotions. If you canengineer a system that has all that, and it’s not pathetically small in terms of the amount of sensory input, if it’s got as close to as much information flowing into it and through its shared systems, then you can argue by analogy that that thing is probably conscious. Then, if you’ve got a mathematics that provides a context for all this, you can prove it. 

Jacobsen: As I noted in annoyance about the common conflation in professional and common discourse, what about the distinction between sensory modality, different nodes of word production, and videos on YouTube? 

Rosner: So, what do you mean by a node?  Like, is part of your brain that–as you walk from room to room, or you go outside, and you look at trees and buildings and cars and people talking on their phones–part of your brain that takes a bunch of flat pictures, hits your eyes, and makes them three-dimensional in your imagination? Is that just a node? Is that a bunch of nodes? What is a fricking node? Does it matter whether there are nodes? Is the straight-line detection part of your processing a node in the same way that the full-on perspective processor is a node? Does anything become a node regardless of the number of neurons contributing? 

Jacobsen: Is there a distinction between modalities, such as YouTube videos mixing with language like chatbots, and sensory inputs like physical senses coming in from your leg and entering your central nervous system or visual input from your eye to your occipital lobe?

Rosner: Yes and no. We’ve developed our integrated modality by moving around the world and seeing the associations. We use all our senses on stuff. Once we develop sophisticated enough systems, and they don’t have to be that sophisticated, we’ve learned to treat images of reality as if they’re reality. Because we have a deep understanding and deep modelling of the world from our accumulated experience and analysis, we have no problem distinguishing between YouTube and reality.

Looking at stuff on TikTok, let’s be modern and understand them as the products of reality; even when they’re not, we can look at animation and analyze it like we analyze actual reality.  It is based, at least in part, on our deep experience of the world. For example, when we go watch a movie, we can forget ourselves and immerse ourselves in the characters in a movie theatre or at home in a safe environment as if we are the characters. So, that’s just because those modes of analysis are accessible by the filmed experience, just as they are by ourselves in the world.

So, if stuff on TikTok or in a movie dredges up associations, it’ll happen. Because we’ve got a versatile consciousness, it will happen. There are going to be a ton of issues for AIs. Will that one be particularly tricky? I don’t think so. But we can talk more about it.

Jacobsen: Thank you, take care. 

License & Copyright

In-Sight Publishing by Scott Douglas Jacobsen is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License. ©Scott Douglas Jacobsen and In-Sight Publishing 2012-Present. Unauthorized use or duplication of material without express permission from Scott Douglas Jacobsen strictly prohibited, excerpts and links must use full credit to Scott Douglas Jacobsen and In-Sight Publishing with direction to the original content.