2021.09.06

“ 5am? Shit.”

Su rubbed her eyes and roller back to the other side of the bed. She knows trying to get back to sleep probably is a waste of time and soon she will be frustrated enough to get up. This is the third sleepless night this week.

Su is a neuroscientist. Graduated from medical school, she now specializes in nightmare treatment for veterans. Her schedule is always chaotic. She spends many nights with sleeping patients and monitoring their brain waves. Across the observation room glass, Su watches the subject either sleeping sound and motionless, or tossing and turning, or snoring, or sweating, or murmuring, or crying …. The sleeping statues of a person are revealing. Every 90 mins or so, the subject cycles back to the rapid eye movement stage in their sleep, where they are surfacing from the unconscious sleep and reaching to awakeness. They dream. Their brain waves become rapid, low-voltage desynchronized. Their body abruptly loses muscle control, legs kicking around.

“This is very intimate.” Su thought, as she lay on the subject’s bed and knowing someone else was about to watch her falling asleep in 30 mins.

2021.08.02

Su was contacted by Prof. Sebald, the director of the Sebald laboratory, about two months ago. The Sebald lab was founded about fifteen years ago, when the AGI systems had evolved to replace most manual labor in repetitive, efficiency-oriented industries, such as service, agriculture and manufacturing. It was also the guiding light on resource distributions and energy control as issues like climate change were at their bottleneck in the first part of the 21 centuries. Only thanks to the massive computational forecast, humans were able to prevent environmental catastrophes. Everyone was excited and put their hopes in artificial intelligence’s promising ability to model massive-scale systems, and fix them. But not long after adapting to automation and this belief, the public and enterprises became not satisfied that AI systems only solve their problems, but need to understand its internal logic. A sense of distrust emerges. More vigilant groups proclaim the “ultimate rebel” one day. The Sebald laboratory was founded in this context. Rather than developing effective intelligence, its focus is to understand it. They occupy two massive warehouse data centers right off campus. The twin brick buildings have no windows but only air venting fences wrapping around, through which the humming sounds of servers inside emits. Su has never set foot inside there, and only nodded to Dr.Sebald at staff meetings. Thus, his call surprised her.

“Dr. Su? Sebald here. How are you?”, an unfamiliar voice over the school intercom. “I want to bring up a potential collaboration and hear your thoughts.”

“Sure. What is it?”

“Well, as you know, my team has been chasing the thoughts of AGI for the past decade. We managed to record a great deal of it and had many attempts to interpret the data. Much like trying to understand another person, we identify actions that show certain patterns, expecting these details to accumulate and fill in a loose puzzle that eventually reveal the underlying image of the person. ”

“You can do that with the machines?” Though her daily life was filled with autonomous coffee shops and chatty apps in her phone, Su was still not used to the way people talk about the AGI as if they are people.

“To a degree, yes, we have succeeded. We were able to observe the periodical emergence of semblable behaviors in the data, or we can say: its ‘thoughts’. ” He paused for a second, making sure Su can follow, “These thoughts are born out of self-supervised algorithms, networks whose operation hinges on a kind of iterative, suffusing process. ”

“Like neurons in the brain?”

“Exactly! I am so glad that this makes sense to you.” Prof. Sebald sounded suddenly excited, “ What we tried to do, somewhat learnt from neuroscience. We speculated about the AGI system behaving like human brains, that there is a spatial relationship related to its contextual information. So we encoded spatial relationships between semantically related data, and these ‘thoughts’ quickly took form into a massive, condensed high-dimensional swamp of information. We were able to observe clusters of thoughts actively emerge and disappear, and we call those : ‘islands’”.

“You are saying that the system self-organized its data like a biological brain. Each island is an independent ‘cortex’ that manages particular jobs, then shares and exchanges information among them. ”

“Yes, that is what we thought and we have tried many ways to navigate this swamp of information and map out its topology. However, this is where we are stuck. The high-dimensional digital swamp is far more complicated than our three-dimensional brains. And this is why I am asking for some help. ”

“I am not sure how I am useful. I am not an expert in neuron function or mapping. In fact, I work in the cognitive side more… ”

“No,” Prof. Sebald interrupted Su, “this is, unfortunately, not a neuron mapping question anymore. It is hard to explain over the phone. Would you mind paying a visit to our lab sometime? I can show you some work we have done here.”

“Ok. I can swing by next week.”

Hesitated, but also curious, Su agreed.

2021.07.05

Sitting in Prof. Sebald’s office, Sun can’t help but feel a little funny. There is a big gap between this and the computer scientist office she had previously imagined. There are no flashing LEDs or carbon fiber covered servers. The room is sun-lighted with a window on the ceiling. Next to the coffee table, there are a few plants and a big white bookshelf. Prof. Sebald offered her some green tea and poured himself a cup too.

“Before understanding, we have to be able to observe. ” Prof. Sebald started talking straight in, “ High-dimensional space is not sensible to us. We discuss it only in its concept, not in our experience. So we decided to reduce the dimensionality of the digital swamp ‘brian’, by projecting it to three-dimensional space. There are many techniques for projection, and in all of them, we found regionalized data that create micro islands and these islands reflect certain semantic and perceptual correlations.”

“Then your speculation is accurate. The system has cortexes. That’s incredible.”

“That is what we thought too. ” Prof. Sebald smiled politely, “However, we quickly notice that there are linkages between these islands that share similar priority. To explain this simply, these linkages themselves seem to and should regroup and become islands. So we reprojected the space, adding these new “linkage islands” into our map. But again, we would find new linkage correlations in the new maps. In fact, with only a few dozens of iterations, it is clear that we would either be recreating another high-dimensional swamp through these add-ons, or only seeing the shadow of a rotating, morphing, structure, of which the islands in the projection are nothing but the result of the projection methods itself. ” Prof. Sebald paused, looking back at Su.