This is the final part of a three-part series. Check out the first and second parts here:
What Makes a Mind? Part 1: Exploring Intelligence through Juxtaposition of Man and Machine.
What Makes a Mind? Part 2: The Framing Problem and Parasitic Parrots.
Imagine yourself in Ancient Greece, surrounded by the sights and sounds of a bustling city. The air is filled with the scent of olive oil and incense, and the distant hum of the marketplace blends with the chatter of people. You find yourself amongst a growing crowd who listen intently to the words of a rhapsode perched upon a makeshift stage.
“Gather round, citizens of Athens! Listen to the tale of Talos, the bronze giant, forged by the divine hands of Hephaestus! In his fiery forge, deep within Mount Olympus, he crafted this marvel to protect the island of Crete.”
As the rhapsode speaks, you close your eyes and imagine Hephaestus, the god of blacksmiths, hammering away. Sparks fly, and the rhythmic clang of his hammer against metal resonates in your mind. You see the imposing figure of Talos, his bronze skin gleaming as he takes shape under the god's skilled hands.
“Talos, the mighty guardian, was no ordinary creation. Hephaestus imbued him with life, a single vein running from his neck to his ankle, filled with ichor—the lifeblood of the gods.”
Murmuring begins, and the crowd stirs with excitement at the grand mental image painted by the skilled performer. So what is this image? Of course, in many ways this tale is simply one of many myths told in Ancient Greece, amongst many more told by collective humanity. However, in some ways, this story isn’t a myth at all. Echos of Hephaestus’ lifelike creations are beginning to take form as we speak.
The Challenge
For thousands of years, we’ve pondered what it might look like for our tools to come to life. The mythical Greek automatons were but one example of many. However, thus far, we’ve seen but a glimmer, a hint of life. The landscape has begun to change. The quest for artificial life is full-steam-ahead, with hundreds of billions of dollars being poured into the development of new advanced Artificial Intelligence. But on such a quest, we must ask ourselves, what is the destination?
Relevance realization is the feature of human cognition that allows us to infer the relevance of various stimuli. This is a primary feature of our intelligence generally. We use it to focus our attention, categorize, represent things, and communicate. Replication of this particular human function is known as the frame problem in Artificial Intelligence research.
The State of the Art
In the first and second part of this series, we explored the current state of artificial intelligence. The epitome of current efforts seem to be language models, and other transformers. These complex neural networks use statistical representations to determine the similarity between terms, and these representations can be used to approximate language, or other modalities.
These statistical representations, however, are derived from an external source. Humans. The current state of the art, therefore, acts to replicate the built-in intelligence of human generated data, it doesn’t produce its own. This copycat behavior is suitable for many tasks, but the lack of underlying intelligence is a concern. The frame problem has not been solved.
Not only has it not been solved, but this particular direction of research might be a dead end by this metric. Here’s an extreme opinion for you: neural networks will never be intelligent. Of course, this involves plenty of semantics. What is intelligence? And all that. But it remains clear, a neural network is nothing more than a system capable of approximating a function.
For transformers, a more complex variation of a neural network, this function is usually some form of sequential data processing. Language generation, for example. This approximation of a function, is pretty cool. It lets us take a rather arbitrary set of input/output examples, and train a machine to make accurate predictions on them. However, the utility of this method has its challenges.
We don’t have to worry about the frame problem much with neural networks, because they are only ever concerned with relevant information. Training data is relevant by default, and inputs are carefully processed to be interpretable by the model. Thus, the frame problem is bypassed, however the scope of the intelligence is limited to a specific domain. Language models, therefore, give the illusion of a more general intelligence, only because they deal with a more general domain.
(If you disagree with my assessment, or have something to add, I’d love to have a discussion! Leave a comment down below, message me directly, or find me on Substack Notes.)
The Path Forward
The state of the art may not have completed our challenge, but we are still on the quest. In order to reach general intelligence, artificial intelligence will need to realize relevance all by themselves. How do we even get to something like that? Well, frankly, it’s complicated. We’re still trying to understand the underlying mechanisms of our own brains.
Relevance itself is not a measurable phenomenon. In the same way that biological fitness cannot be quantified, and therefore a general theory of biological fitness cannot be developed. However, a theory of the mechanisms which give rise to biological fitness can and has been identified. In the same way, relevance can be theorized through its underlying mechanisms (Vervaeke, Lillicrap, & Richards, 2009).
Those underlying mechanisms, if properly synthesized, could give rise to real intelligent lifeforms of artificial nature. But what do these mechanisms look like?
Self Organization
At the heart of an intelligent being is the ability to self-organize. We humans, along with every other organism on the planet, start from an initial conception, and organize ourselves through to maturity, and beyond. This phenomenon, known as autopoesis, is fundamental to our ability to act as agents in an arena.
This autopoesis is what allows organisms to adapt to their environment, and is a central driver of evolution as a whole. A language model, is not autopoetic. However, machines could theoretically be designed to self-organize in similar ways to organisms. There are various theoretical approaches to such a machine.
However, we may find that such a machine isn’t possible at all. Perhaps the nature of a machine, as in a mechanism, is that it is not and cannot be autopoetic. Perhaps the ingredients of self-organization can only be found in biological organisms, and so true machine intelligence requires becoming biological.
Cognitive Economy
Cost and benefit. These two concepts seem to self-similarly repeat themselves at various scales of analysis, like a fractal. There is a cognitive economy to be maintained. Internal processes expend energy, and this expenditure must be balanced with benefit gained. This balance is critically maintained by us humans for the sake of… well, I suppose, not starving to death.
This managed cognitive economy, also favors efficiency, and allocates cognitive resources to various tasks. Adaptability arises, as cognition favors efficient systems and builds them autopoetically. Resilience is built as the efficient system also must stand the test of time. The mind favors that which repeatedly works, and builds systems through this learned experience.
Opponent Processing
The flexibility of the human mind’s comprehension is unparalleled. We are capable of wrapping our heads around a vast amount of information, and putting it to practical use shortly after. However, this flexibility requires a careful and dynamic balance between cognitive processes, in favor of adaptation to various situations.
For example, when the mind encounters some information, it must decide at what scope to approach it. Imagine a continuum between compression and particularization, where a balance must be struck between compressing information for general use, or particularizing details for specific tasks. When parsing new information, the mind must find the balance on this continuum to suit the particular situation.
Other continuums exist too. Narrow vs. broad focus to capture finer or broader details. Short-Term vs. Long-Term Planning. Risk-Taking vs. Caution. Urgency vs. Importance. These balances are struck automatically by the mind, but are done through interaction with the environment. An agent is in part defined by its relationship to an environment.
The Horizon
Our machines have not yet become intelligent, and perhaps will not for a long time. However, while the transformer architecture and neural networks likely won’t give rise to general intelligence, there are efforts being made in other areas. Perhaps in the coming years, the billion-dollar AI labs will fail to yield significant results, but other areas of research will continue to advance.
Xenobots, for example, are small cellular automatons designed by computers to perform specific functions. These little things are built from the skin and heart cells of an African Clawed Frog. These neat little creations could someday help clean up the oceans, or perform other operations on a large scale without generating pollution.
Granulobot, developed by researchers at the University of Chicago and Illinois Institute of Technology, is a system inspired by… well sand. The system was designed to address fundamental problems of adaptability and self-organization. This idea targeted the question of how an assembly of simple building blocks, such as grains of sand, could form emergent intelligence when given actuation abilities.
While I can’t cover everything, Software and Synapses, as well as its sister series Byte Sized, will continue looking into various new areas of artificial intelligence research as things progress.
Why does this matter?
So, we’ve spent three weeks breaking down the human mind, the nature of artificial intelligence, and various aspects of philosophy. The question might occur to you dear readers:
Why does this matter?
The truth is, the question of whether machines can think matters in both very concrete and abstract ways. We live in a time where many are unsure about their job security in the coming decades, where conversations of universal basic income are being had to address concerns of human replacement. The future is clouded by a combination of grifting, hype, and unanswered questions.
So, this series exists as my attempt to bring a more clear footing to the wider conversation. What will machine intelligence really look like? If machines aren’t intelligent, and likely won’t be for a while, then we humans might be safe in our jobs for a bit. We will likely see automation occur, and human’s jobs might become easier, but we will remain necessary.
Humanity is resilient, so while we can’t be sure about the future, I think we’ll end up being fine. However, this publication will continue its attempt to shine a light on the landscape.
Author’s Note
This series has been really fun but also challenging to write. There are so many avenues to explore, rabbit holes to go down. I’d definitely love, however, for this to be a broader discussion than just what I can talk about here. If you have thoughts, feedback, comments, concerns, don’t hesitate to message me! I’d love to hear what you all have to say. I’m also present on Substack Notes, and I’d love to have some open conversations there!
P.S. I’m going to be starting a small Discord server for this community soon. If you’re interested in joining, let me know :)
As always, thank you for reading, and I’ll see you next time! Goodbye.
Further Reading
Despite my efforts, a three-part series is not nearly enough to fully tackle whether machines can think. While Software and Synapses will continue to explore this question in the future, I’ll also be providing some resources for those of you interested in doing a deeper dive.
Credits
Thumbnail:
Bilal Azhar - https://substack.com/@intelligenceimaginarium
Background Music:
Track - Marshmallow by Lukrembo, Source - https://freetouse.com/music, Free Music No Copyright (Safe)
References
Vervaeke, J., Lillicrap, T. P., & Richards, B. A. (2009). Relevance Realization and the Emerging Framework in Cognitive Science. Journal of Logic and Computation. Retrieved from http://www.ipsi.utoronto.ca/sdis/Relevance-Published.pdf.