on
Gordon Pask's Experiments
Gordon Pask is one of these unknown figures from AI history that you never really hear much about despite the influence he had on cybernetics. The main reason why is that he had a philosophy of doing things and building things rather than emphasizing writing about what experiments he did. The fact that he has very little writing to his name leads to him being passed over in history, unlike figures such as Beer who was a prolific writer.
Beer on the other hand has writing that is mostly unintelligible new age babble in my honest opinion. I’ve tried to read several of his texts but most of them just seemed to be futurist writings devoid of real arguments or content. Gordon Pask seems very similar but he is through and through an experimentalist. Most of his life was devoted to trying to put his theory into practice through out-there experiments in building various devices.
Gordon Pask first encountered cybernetics via reading Norbert Wiener and interacting with Wiener on campus at Cambridge where he went to school. He was well known at the time for his erratic and strange experiments as well as his difficult to understand theories of philosophy and cybernetics. As a scientist, he characterized the devices he made as “maverick machines” which he defines as machines that employ some form of dynamic feedback system within them.
One example of these that was turned into a product was a keyboard that was able to teach the user how to type out specific punch cards. The keyboard did not simply give rote examples to test the student’s knowledge but instead attempted to learn particular traits about the student and help them progress.
The machine would create exercises that were tailored to the information that was learned specifically about a given student. At least that is the claim made about the machine, in reality, I’m not sure how successful such a machine actually was at doing this. Personally, I’m highly skeptical about many of the claims about such machines that Pask puts forward.
Similarly, Pask invented a music machine that took in sound input from playing musicians and outputted a specific light show. The light was calibrated similarly to learn from the musician and play off of the musical style. If the musician played too repetitively the machine would stop engaging, this forced the players to keep changing up songs and refreshing the machine as they went.
Again, I would like to see a video of this machine to see what exactly it looks like or what it is outputting. I believe that it does something beyond simply being an EQ visualization but the claims made about the extent to which it is really ‘learning’ from what the musicians are playing are kind of lofty.
The most famous examples of experiments in this area were electrochemical examples created by Pask which attempt to harness electricity to showcase feedback systems. Less notably, he also developed two theories of the mind Conversation Theory and Interaction of Actors theory. I won’t claim to understand what either of these theories is or what they are about, but they don’t seem to have had much influence on later thinkers on consciousness.
The main experiment people talk about in AI circles is one in which Pask and Stafford Beer attempted to develop a self-organizing protocol of metallic threads within a dish of solution. The threads have an electrical current that flows through them around the dish by the use of high-voltage electrical sources applied to the metal which is situated in a highly resistive solution.
Over time as the threads form and branch out competing with one another, they form into semi-stable structures and self-allocate branches in more optimal ways. By placing different barrier structures within the dishes they can cause the behavior of the system to evolve in different ways.
They take this further and allow the system to have access to sound waves by introducing electrodes tied to microphones as stimulants within the system. If you are confused by what I’m saying there are many videos of people reproducing this experiment online that you can look through. It should be noted that this experiment is prohibitively unsafe due to the chemicals and electricity involved, and thus not many people have attempted to recreate this.
The goal of what the two researchers were trying to do was to create sets of components or materials that can be combined to create a wide variety of what Pask calls “entities”. These are systems through which the interaction of the entity within some self-reinforcing system allows for change and reward loops based on some behavior.
Pask has a unique philosophy toward such experiments. He argues that instead of acting as passive observers that attempt to perform some methodological abstraction in their experiments researchers should take on an active role within the systems and artifacts that they create.
Experiments for Pask should be performed via constructing problems and constructing systems with which we can interact and experiment actively. Scientific problems should be created as part of the scientific process through the design of custom systems.
As part of this philosophy, Pask takes on a position that systems or rulesets that minimally specify the design or artifact are optimal. From his perspective, we should instead attempt to design cybernetic systems that generate complexity as a result of their simple functionality, rather than attempt to explicitly specify the more complex functionality ourselves manually. He argues for a type of “law of downhill synthesis and uphill analysis”, where we build simple systems that in turn generate complexity that is difficult to understand.
His arguments are kind of hard for me to understand. Particularly, he makes a claim that the threads exhibit an ‘elementary form of learning’. It seems to me largely to be a deterministic process based on where the electricity is situated in relation to the metal components.
He says that when you change the current points the network redistributes and regrows afterward to align with the new set of current points. This turns into a claim that the network is doing something intelligent or exhibiting intelligence.
Frankly, this doesn’t seem like intelligence to me, it just seems like a deterministic electrochemical reaction to the electricity moving to a different area. If you see intelligence as a product of a deterministic material process, this makes sense but I don’t think he is trying to make that claim at all. He doesn’t see the experiment in that light and doesn’t talk about it in that way at all.
The learning claim is very confusing to me as there is no evidence of anything being retained here on repeat experiments. He says that if the current is then set back to the original position the network “tends to regrow its initial structure”.
What does this mean, it ‘tends to regrow’? Does it sometimes not regrow the initial structure? Does it regrow it in exactly the same way it was before or does it merely regrow some structure vaguely similar to the original? If it is regrowing in much the same way as it was initially this seems more like evidence of a deterministic process playing out under very repeated material circumstances.
I wanted to hear much more about the mechanics of this experiment and what the claims were. However to Pask, creating replicable experimental data hardly seems to be the point of his work, so maybe I’m missing the point by asking that we try to create some explanation for the dynamic that forms.
The reward loops described are somewhat primitive. While it is clear that we need some kind of conditioning or vector that allows us to enhance our performance, the encodings of this within Pask’s set of maverick devices seem very primitive.
The feedback loop is trying to optimize toward some vector that we calibrate to be the reward mechanism. This is not clearly more complex than operant conditioning within animals and it is not clear that the complex behavior rises above that to be something more intelligent than a system where the learning has no conscious thought behind it.
It feels like the systems often reach some feedback equilibrium much like a thermostat example. Maybe we can attribute this kind of thermostat model (which I would argue is most categorically similar to these maverick machines) to have some few bits of intelligence if we really want to but I’m not convinced that a thermostat exhibits meaningful levels of intelligence.
Pask and Beer are clearly prototyping the ideas that would later become evolutionary computation, dynamical systems, and other forms of reflective self-configuration. Their work laid some of the foundations for studies wherein researchers attempt to create complex systems by creating a small set of movements within a delineated space that evolves into complex dynamical processes.
They were predating things like cellular automata or rule 110. They attempt to do similar things not through the specification of digital rulesets that give rise to complex action but through creating interesting configurations of spaces that happen to give rise to complex phenomena.
It is unclear how much Pask’s concept of the experimenter or researcher being embedded in the system that is created has ended up being cultivated by later researchers. It is also unclear what the point necessarily is or how we could justify it within a modern academic setting. We, generally speaking, have a want for automata or learning processes to exist on their own in isolation or to self-modify with only limited outside help through their various processes.
The same goes for evolutionary computation, we want internalized reward structures and internalized self-modification, and we don’t want the researcher forming part of the cybernetic relation with the intelligent processes that are formed under experiment. Though arguably the researcher is part of any given experiment that they perform whether they like it or not.
This concept of the interplay with the researcher or other objects within a larger cybernetic system seems to have been totally lost to time and doesn’t play out in research designs. There are obvious exceptions such as automation processes within factories, governance, or logistics operations. We don’t generally conceive of smaller units of man and dynamic machine combinations as cybernetic systems despite the fact that we use them constantly.
It seems clear to me that someone like Pask, if alive today, might see the smartphone or the way we use Google search, or Youtube algorithms as a cybernetic feedback system through which we and the computer algorithm combine together to create some dynamic. Pask would see the world around us as massive cybernetic feedback loops filled with intelligent processes simultaneously creating feedback into each other.
The work of Pask is a radical departure from this kind of sterile observational-based or thought-based experiment. Instead, he tries to actually construct dynamic looping systems himself from scratch.
He tried to actualize the theories everyone talks about through some experimental means. Other authors, attempt to characterize the systems that exist in the world or that we seem to experience ourselves. These are descriptive examples by philosophers who never attempt to create ways to put theory into practice.
This reminds me of soap bubble computer experiments where it was claimed that soap bubble solutions situated in a specific plastic peg configuration reach an optimal solution to an NP problem in polynomial time. This claim as it turned out was not actually correct as was proven experimentally by Scott Aaronson [0], but it is an example of this kind of dynamic self-organizing system using natural elements that has somewhat surprising and perplexing behavior.
The reason this example is interesting to me is the claim that it had a greater intelligence than our most advanced algorithms or human thought processes through some naturalistic self-organizing physical system similar to Pask’s metallic threads.
In fact, it does work as described under contrived circumstances. These kinds of experiments are few and far between. There are also many great examples of experiments being done to design automata between intelligent mushrooms that fall into this category [1].
[0] NP-complete Problems and Physical Reality - https://www.scottaaronson.com/papers/npcomplete.pdf
[1] Fungal Automata - https://arxiv.org/abs/2003.08168