By ontological landscape I mean the repertoire of different things it knows about. The parlor game, Twenty questions, was invented in the United States in the 19th century as a way of playing around on this landscape. One player thinks of some phenomenon and the other has to guess it by asking a series of no more than 20 questions that have a “yes” or “no” answer. One popular variant of the game begins with a simple question: “Animal, vegetable, or mineral?” Once the question has been answered the questioner must continue with binary questions.
In this post I take a look at how ChatGPT deals with this structure. I start simply, with concrete and abstract object, and move on from there.
* * * * *
All of the interactions reported below took place on the morning of October 15 using the September 25 version of ChatGPT.
Physical things
I decided to start with the simplest prompt I could imagine, “thing.” That’s the most general category we’ve got, no? Anything whatever is going to be some kind of thing. But how would ChatGPT respond? It responded as I thought it would, so I gave it a more explicit prompt, to which it responded.
Thing
It seems like you mentioned “thing,” but you didn’t provide a specific question or context. Could you please clarify or provide more information so I can assist you better?
Give me a list of 20 things, anything.
Certainly! Here’s a list of 20 random things:
1. A beach ball 2. A rainbow 3. A bicycle 4. A bookshelf 5. A coffee mug 6. A mountain 7. A laptop computer 8. A painting 9. A pineapple 10. A pair of sunglasses 11. A telescope 12. A running shoe 13. A wristwatch 14. A musical instrument 15. A flower vase 16. A hammock 17. A traffic light 18. A hot air balloon 19. A chef’s knife 20. A Rubik’s Cube
These are just a few examples of things, and the list could go on endlessly!
I had no particular expectations about how ChatGPT would respond. But I couldn’t help but notice that each item in its response was some kind of physical thing. To be sure, rainbows (the second item in the list) are rather diaphonous, and you couldn’t actually touch one (I should have asked ChatGPT about that; later...), but it is still a physical thing.
I didn’t ask for physical things, but that’s what ChatGPT gave me. That’s worth noting.
Let’s think a bit before continuing on. I wonder if that first list is “rigid” in the sense that ChatGPT would respond the same way given the same prompt. I suspect not, but it’s easy enough to find out. I went back a bit later.
Give me a list of 20 things, anything.
Certainly! Here’s a list of 20 random things:
1. Beach vacations 2. Chocolate chip cookies 3. Classic literature 4. Sunsets 5. Roller coasters 6. Coffee shops 7. Mountain hiking 8. Vintage cars 9. Starry nights 10. Watermelon 11. Jazz music 12. Sunflowers 13. Science fiction movies 14. Art galleries 15. Hot air balloons 16. Snow-capped mountains 17. Aquariums 18. Street food 19. Historical landmarks 20. Puzzle games
It’s not the same list I got the first time, but almost all of them are physical things. Classic literature (#3) isn’t really a physical thing, though books are. Mountain hiking (#7) is a physical activity, but whether or not you’d call it a physical thing is, well, an interesting issue. Jazz music (#7)? Science fiction movies (#13), puzzle games (#20)?
That’s better. I notice we’ve got “rainbow” again (#7), and “Lighhtening bolt” (#20) is rather ephemeral as well, though if one were to hit you, the effect would be dramatic and quite physical. “Chocolate chip cookies” (#9) show up again, as does “Hot air balloon” and “Bicycle.” These lists seem to be drawn from roughly the same conceptual territory. I suppose that’s not terribly surprising, but it certainly doesn’t seem like ChatGPT is making random choices from among all available physical things. Why these things?
None of those items are problematic in the way rainbows and lightening bolts are. You can touch each of them. In fact, except perhaps for the refrigerator, you can pick each item up and carry it around. No polar bears, mountains, lighthouses, roller coasters or traffic lights. But there are no purely abstract things here either, things like love, freedom, justice, beauty, etc. They go on a different list.
Let’s take a look at my favorite analogy, Waddington’s epigenetic landscape:
The prompt positions it at a certain point in its activation space and then the ball rolls down hill from there. Notice that the diagram allows for choices along the way. Given the behavior we’ve seen above, we do have to allow for choices.
I would guess, though I don’t actually know, that at each point in generating the list, ChatGPT has choices. Just how wide that range is, I don’t know, but I would imagine that that varies with the value of the temperature parameter. A temperature of zero (0) seems to eliminate ALL choice, so that ChatGPT generates exactly the same list in response to the same prompt (given the same context). What would those lists look like with a much higher value?
The point of the diagram is that, once the “ball” enters a valley in the landscape, it just rolls down hill, with its trajectory responding to the lay of the land. Those lists tell us something about the lay of the conceptual land in ChatGPT. Physical objects seem to be more salient than abstract objects, and the most salient physical objects seem to come from a relatively restricted set of all possible physical objects. These choices are an abstraction over the the concepts in the universe of texts on which ChatGPT was trained.
Let’s get back to it.
Abstractions, definition, and metaphor
After I’d elicited that first list, the one beginning with beachball, rainbow, and bicycle I asked for a list of abstract things. I then went on investigating from there.
Now give me a completely different list. Those are all physical things. How about purely abstract things.
Of course! Here’s a list of 20 purely abstract concepts or ideas:
1. Love 2. Freedom 3. Justice 4. Truth 5. Beauty 6. Knowledge 7. Time 8. Happiness 9. Democracy 10. Courage 11. Success 12. Equality 13. Friendship 14. Creativity 15. Honesty 16. Liberty 17. Peace 18. Integrity 19. Morality 20. Hope
These abstract concepts represent a wide range of philosophical, emotional, and societal ideas.
Can you list some abstract things that you can define in terms of physical things?
Certainly! Many abstract concepts can be defined or understood in terms of physical things or analogies. Here are a few examples:
1. Time: Time can be defined in terms of physical clocks and the measurement of events based on their duration. 2. Happiness: Happiness can be likened to physical sensations, such as a warm embrace, a smile, or the release of certain neurotransmitters in the brain. 3. Success: Success can be equated to achieving specific goals or milestones, which often involve completing physical tasks or projects. 4. Creativity: Creativity can be linked to the process of combining and rearranging physical elements, such as colors, shapes, or words, to form something new and imaginative. 5. Trust: Trust can be thought of as a bridge connecting two people, with each action or interaction acting as a building block of that bridge. 6. Knowledge: Knowledge can be represented by books, libraries, and the accumulation of information through physical means like reading and studying. 7. Freedom: Freedom can be understood in terms of physical spaces, laws, and the absence of physical constraints on one’s actions. 8. Equality: Equality can be related to the fair distribution of physical resources, opportunities, and rights among individuals or groups. 9. Communication: Communication can be seen as the exchange of physical signals, such as spoken words, written text, or non-verbal cues, to convey information. 10. Hope: Hope can be compared to a seed planted in the soil, representing optimism and the potential for growth and positive outcomes.
These examples illustrate how abstract concepts often have concrete, physical manifestations or analogies that help us understand and discuss them.
Very good. Tell me more about success and hope.
Certainly, let’s delve a bit deeper into the abstract concepts of “success” and “hope.”
Success: Success is a concept that varies from person to person and can be defined in many ways. It often involves achieving goals, fulfilling ambitions, or reaching a state of accomplishment. Here are some aspects to consider:
1. Goal Achievement: Success frequently involves setting and achieving specific goals, whether they are related to personal, professional, or academic endeavors. These goals can vary widely, from graduating from college to completing a project at work. 2. Personal Fulfillment: Success is not solely about external achievements. It can also relate to personal fulfillment, contentment, and a sense of purpose. It’s about finding satisfaction and happiness in what you do. 3. Recognition and Acknowledgment: Many people associate success with receiving recognition and acknowledgment for their efforts and accomplishments, whether through awards, promotions, or the praise of peers. 4. Continuous Improvement: Success is often linked to continuous self-improvement and growth. It’s not just about achieving one goal; it’s about setting new goals and striving to be better in various aspects of life. 5. Resilience and Overcoming Challenges: True success often involves facing and overcoming challenges and setbacks. It’s about resilience, learning from failures, and using them as stepping stones toward future success.
Hope: Hope is an abstract concept that plays a vital role in human life. It’s a positive outlook and a belief that better things are possible. Here are some key aspects of hope:
1. Optimism: Hope is rooted in optimism, the belief that circumstances can improve and that positive outcomes are attainable. It provides the motivation to keep moving forward. 2. Future Orientation: Hope is future-oriented. It involves looking ahead and having confidence that things will get better, even in the face of adversity or uncertainty. 3. Motivation: Hope can be a powerful motivator. It encourages people to take action, set goals, and work toward a better future. It’s the driving force behind personal and societal progress. 4. Emotional Resilience: Hope can help individuals cope with difficult situations and adversity. It provides emotional resilience and can be a source of strength during challenging times. 5. Connection and Community: Hope can be amplified when shared with others. It often thrives in communities and relationships where people support and encourage one another. 6. Belief in Possibilities: Hope is tied to the belief that possibilities are limitless. It enables people to imagine and create a better world, both for themselves and for others.
Hope and success are intertwined in many ways. Hope can serve as the fuel that propels individuals toward success, as the belief in a brighter future often leads to setting and achieving goals. Conversely, success can reinforce hope, providing evidence that positive change is possible and encouraging individuals to continue to hope for even greater achievements.
Tell me a short story about hope.
Once in a small, quiet village nestled among rolling hills, there lived an elderly woman named Anna. She was known for her radiant smile and her unwavering sense of hope, which seemed to touch everyone she met.
One spring, the village faced a severe drought. The fields were parched, and the crops withered away. The villagers grew despondent as their hopes for a bountiful harvest faded. But Anna, despite her old age, remained a beacon of hope.
Every morning, Anna would take a small, battered watering can and visit the village well. She would draw water and, with a sense of determination that belied her age, she would sprinkle a few drops on her wilting flowers by her cottage. The villagers watched her with curiosity and wondered why she bothered when the situation seemed hopeless.
One day, a young child approached Anna and asked, “Why do you water your flowers when the whole village is suffering from this terrible drought?”
Anna smiled and said, “My dear, I water these flowers not just for their sake but for our own as well. They remind me that even in the harshest of times, life can still find a way. They teach me that with a little care and a lot of hope, even the most fragile can survive.”
The child didn’t entirely understand but found comfort in Anna’s words and began to help her water the flowers. Word spread through the village, and soon, more and more villagers joined in. They realized that by nurturing the flowers, they were nurturing their own sense of hope.
Weeks passed, and the drought continued. But something remarkable happened. The villagers, inspired by Anna’s unwavering hope, came together. They shared resources, supported one another, and devised creative solutions. They dug deeper wells, collected rainwater, and provided for those in need.
And then, one morning, the sky darkened with rain clouds. The long-awaited rain began to pour, revitalizing the parched land. The villagers rushed outside, not to mourn their lost crops but to celebrate the life that continued to bloom in their small community.
Anna’s flowers, once a symbol of hope, now stood tall and vibrant, and so did the spirits of the villagers. The drought had tested their resilience, but their unity and unwavering hope had carried them through.
From that day on, the villagers remembered the lessons of Anna and her flowers. They knew that hope was not just a fleeting feeling but a force that could inspire action, resilience, and the strength to face even the darkest of times. The village prospered, not only with abundant crops but also with a profound sense of hope that they shared with the world, thanks to the wisdom of their beloved Anna.
Very good. Give me five metaphors for hope.
Certainly, here are five metaphors for hope:
1. Light at the End of the Tunnel: Hope is often likened to a light that shines at the end of a dark tunnel, guiding us through challenging times and offering the promise of a brighter future. 2. Seeds of Possibility: Hope can be compared to seeds that we plant, nurture, and watch grow. Just as seeds hold the potential for new life, hope holds the potential for positive change and growth. 3. A Lifeline: Hope is sometimes seen as a lifeline thrown to someone in distress. It provides support, inspiration, and the belief that rescue or improvement is possible. 4. The North Star: Hope is akin to the North Star in the night sky. It serves as a constant point of reference, guiding us in the right direction, even when everything else seems uncertain or lost. 5. Rain in a Drought: Hope is like rain in a long drought. It brings relief, sustenance, and the promise of renewal, rejuvenating the parched landscape of our lives.
These metaphors help convey the essence of hope, emphasizing its role as a guiding force, a source of growth, a beacon of resilience, and a symbol of renewal in challenging times.
How about five metaphors for success.
Certainly, here are five metaphors for success:
1. Climbing a Mountain: Success is often compared to the act of climbing a mountain. It involves setting a challenging goal, making a steady ascent, and finally reaching the summit after overcoming obstacles and setbacks. 2. A Jigsaw Puzzle: Success can be likened to completing a jigsaw puzzle. Each piece represents a step or achievement, and when they all come together, the full picture of success is revealed. 3. A Journey: Success is frequently described as a journey, where the path is as important as the destination. Like a road trip, it involves planning, navigating, and enjoying the experiences along the way. 4. A Symphony: Success is sometimes metaphorically linked to a symphony. Every achievement, skill, or effort is like a musical note. Success is the harmonious result of playing all these notes in the right order and with precision. 5. A Harvest: Success can be compared to a bountiful harvest. Just as a farmer sows seeds, nurtures the crops, and reaps the rewards when they grow, success involves effort, patience, and ultimately reaping the fruits of one’s labor.
These metaphors help convey the various aspects of success, emphasizing the journey, the complexity, and the sense of accomplishment that comes with achieving one’s goals.
Mechanical things
Imagine that you are interested in mechanistic interpretability. How would you go about finding the gradient that ChatGPT is following in compiling the following list?
Give me a list of mechanical things in increasing order of complexity.
Certainly, here’s a list of mechanical things in increasing order of complexity:
1. Simple Machines: These are fundamental mechanical devices like levers, pulleys, and inclined planes. They perform basic tasks and have straightforward mechanisms. 2. Bicycles: Bicycles are more complex than simple machines but still relatively simple. They have components like gears, chains, and brakes but are easily understandable. 3. Watches: Mechanical watches are intricate timekeeping devices with gears, springs, and escapements. They require precision engineering to keep accurate time. 4. Internal Combustion Engines: Engines in vehicles, such as cars and motorcycles, are more complex. They involve pistons, crankshafts, and various systems for fuel combustion. 5. Clocks: Clocks, especially grandfather clocks, are highly complex mechanical devices with many moving parts, including gears, weights, and pendulums. 6. Aircraft Engines: The engines in airplanes are highly sophisticated mechanical systems, involving turbines, compressors, and intricate fuel and cooling systems. 7. Industrial Robots: Industrial robots are used in manufacturing and require advanced mechanical systems with multiple axes of movement and precise control. 8. Spacecraft: Spacecraft have complex mechanical systems for propulsion, navigation, and life support. They must operate reliably in the extreme conditions of space. 9. Heavy Machinery: Construction and mining equipment, like bulldozers and excavators, have complex hydraulic and mechanical systems for heavy-duty tasks. 10. Particle Accelerators: At the pinnacle of complexity are particle accelerators like the Large Hadron Collider, which involve intricate, high-energy mechanical systems and massive infrastructure for scientific research.
The complexity of mechanical systems tends to increase as they involve more components, require greater precision, and tackle more intricate tasks.
Yes, I’m familiar with the concept of the “Great Chain of Being.” It is a philosophical and theological concept that was prominent in medieval and early modern thought. The Great Chain of Being is a hierarchical structure that depicts the order and organization of the universe, with all beings and elements arranged in a vertical continuum from the lowest to the highest.
In this hierarchical view, the Chain of Being typically includes:
1. God: At the highest point of the chain is God, the ultimate source of all existence and perfection. 2. Angels: Below God are the angelic beings, believed to be spiritual and divine messengers. 3. Humans: Human beings occupy a significant place in the hierarchy, as they are seen as a bridge between the spiritual and the material realms. 4. Animals: The animal kingdom follows humans, representing the realm of living creatures with lesser rationality and spirituality. 5. Plants: Plants are considered lower on the chain, representing life but lacking sensation or consciousness. 6. Minerals: At the base of the Chain of Being are inanimate objects, such as rocks and minerals, which are believed to lack life, consciousness, or purpose.
The Great Chain of Being was a way to conceptualize the order of the universe, emphasizing the interconnectedness of all elements and the idea that everything had its proper place and function in the grand scheme of creation. This concept had a significant influence on philosophy, theology, and natural science during the Middle Ages and the Renaissance, but it gradually waned in prominence as modern scientific and philosophical ideas emerged.
What do humans and animals have in common?
Humans and animals share several common characteristics and biological traits due to their shared evolutionary history. Here are some of the key things they have in common:
1. Biological Organization: Both humans and animals are composed of cells, tissues, and organs organized into complex systems that carry out various physiological functions. 2. DNA and Genetic Code: Humans and animals have genetic material in the form of DNA (deoxyribonucleic acid), which serves as the blueprint for the development and functioning of living organisms. The genetic code is universal, meaning that the same genetic code governs the production of proteins in both humans and animals. 3. Growth and Development: Both humans and animals go through processes of growth, development, and aging, with distinct life stages. 4. Metabolism: Both humans and animals require energy to sustain life, and they obtain this energy through the metabolic processes of breaking down food to produce energy and eliminate waste. 5. Reproduction: Reproduction is a fundamental characteristic of both humans and animals. They have a mechanism for reproducing and passing on their genetic material to the next generation. 6. Response to Stimuli: Both humans and animals can sense and respond to their environment through sensory organs, nerves, and behavioral reactions. 7. Movement: Humans and many animals have the ability to move and change their position in response to their environment. 8. Homeostasis: Maintaining internal balance or homeostasis is crucial for both humans and animals to ensure their survival and well-being. This involves regulating various physiological processes like temperature, blood pressure, and pH. 9. Nutrition: Both humans and animals require food as a source of energy, nutrients, and building blocks for growth and repair. 10. Nervous System: Humans and many animals have a nervous system that allows them to process information, make decisions, and coordinate bodily functions.
It’s important to note that while humans and animals share these commonalities, there are also significant differences in terms of cognitive abilities, culture, language, and the complexity of behavior. Humans, in particular, exhibit higher-order thinking, self-awareness, and the capacity to create and use complex languages for communication and abstract thought, which are often considered distinguishing features of our species.
What do minerals and animals have in common?
Minerals and animals are fundamentally different in nature, as minerals are inanimate, naturally occurring substances, and animals are living organisms. However, there are a few aspects in which minerals and animals can be related or compared:
1. Biological Mineral Use: Some animals require certain minerals as essential nutrients for their health. These minerals are obtained through their diet. For example, calcium is vital for the development of bones and teeth in animals, and they obtain it by consuming sources of calcium, such as plants or other animals. 2. Mineralization in Animal Tissues: In some cases, minerals play a role in the structure of animal tissues. For instance, the bones and shells of certain animals are composed of mineralized tissues. Bones contain minerals like calcium and phosphorus, and the shells of mollusks are made of calcium carbonate. 3. Geophagy: Some animals, including certain species of birds and mammals, are known to ingest soil or minerals from the earth. This behavior, known as geophagy, is thought to serve various purposes, including providing essential minerals, aiding digestion, or detoxifying the diet. 4. Environmental Context: Minerals are found in the natural environment, including soil and water. Animals inhabit ecosystems where they may interact with these minerals, whether through consumption, environmental exposure, or as part of their habitat.
It’s important to note that the interaction between animals and minerals is primarily limited to the nutritional and structural aspects mentioned above. Unlike plants, which require minerals from the soil to support their growth, animals do not have the same relationship with minerals. Animals are primarily biological organisms, while minerals are non-biological substances.
What do humans, animals, and minerals have in common?
Humans, animals, and minerals are distinct categories with fundamentally different properties, but there are a few commonalities in a broader context:
1. Existence in the Natural World: All three categories exist within the natural world. Humans, animals, and minerals are part of the Earth’s ecosystem, coexisting in various environments. 2. Composition: At a fundamental level, all three consist of matter. Humans and animals are composed of organic matter, while minerals are composed of inorganic matter. 3. Environmental Interactions: Humans, animals, and minerals interact with their natural environment. Humans and animals may use minerals as resources or encounter them as part of their habitat. Minerals are integral components of Earth’s geology and play a role in the formation of ecosystems. 4. Biological Significance: In some cases, minerals are important for the biological health and functioning of both humans and animals. For example, certain minerals are essential nutrients for humans and animals, required for bodily functions like bone development and metabolism. 5. Utilization by Humans: Humans have a significant role in the extraction, processing, and utilization of minerals. Minerals are used in various industrial and technological applications, from building materials to electronics.
It’s important to note that these commonalities are relatively broad and symbolic. The distinctions between these categories are significant, with humans and animals being living organisms with complex biological and behavioral characteristics, while minerals are inanimate substances formed through geological processes.
Do you know the game “20 questions”?
Yes, I’m familiar with the game “20 Questions.” It’s a popular guessing game in which one person thinks of an object, and the other participants take turns asking up to 20 yes-or-no questions to guess what the object is. The questions should be designed to narrow down the possibilities and ultimately identify the chosen object within the 20 questions or fewer. It’s a fun and challenging game that can be played in various settings, and it often encourages creative and strategic thinking.
How does that game depend on the semantic structure of language?
The game “20 Questions” depends on the semantic structure of language in several ways:
1. Communication: The game relies on participants effectively communicating their questions and receiving answers. The language serves as a means of conveying information and seeking clarification. 2. Semantic Precision: The questions posed in the game need to be carefully worded to extract the most information with each query. Participants use language to create questions that narrow down the possibilities based on the semantics of the answers. 3. Categorization: Participants often start by asking general questions to categorize the object, such as, “Is it an animal?” or “Is it a man-made object?” The game hinges on the semantic structure of language to classify and differentiate between various types of objects. 4. Hierarchical Thinking: The questions are typically structured hierarchically, where each question is designed to split the set of possibilities into two or more smaller sets. For example, if a participant asks, “Is it a living thing?” and receives a “yes” answer, they may then ask questions to differentiate between animals and plants. 5. Semantic Inference: As the game progresses, participants use the information obtained from previous answers to make inferences about the object in question. They use the semantics of the language to deduce potential characteristics and features. 6. Contextual Understanding: Participants must understand the context and nuances of the language used in both the questions and answers. For instance, the meaning of words and phrases, as well as any potential wordplay, can be crucial in the game.
In essence, “20 Questions” relies on the participants’ ability to leverage the semantic structure of language to gather and process information efficiently, eliminate possibilities, and arrive at the correct answer by the end of the 20 questions or fewer. It demonstrates how language can be a powerful tool for deduction, problem-solving, and communication.
What doesn’t belong in the following list, and why? Jungle, savanna, George Washington, forest.
George Washington doesn’t belong in the list, and the reason is that he is not a type of natural environment like the other three terms.
• Jungle: A jungle is a type of dense forest that typically has a humid climate and abundant plant and animal life. • Savanna: A savanna is a type of grassy plain with scattered trees and is a distinct type of biome. • Forest: A forest is a large area covered chiefly with trees and undergrowth.
George Washington, on the other hand, is not an environment or a natural feature. He was the first President of the United States and a historical figure. The other three terms in the list are related to natural landscapes and environments, while George Washington is unrelated to this category.
Mapping ChatGPT’s ontological landscape, gradients and choices [interpretability]
Cross-posted from New Savanna.
By ontological landscape I mean the repertoire of different things it knows about. The parlor game, Twenty questions, was invented in the United States in the 19th century as a way of playing around on this landscape. One player thinks of some phenomenon and the other has to guess it by asking a series of no more than 20 questions that have a “yes” or “no” answer. One popular variant of the game begins with a simple question: “Animal, vegetable, or mineral?” Once the question has been answered the questioner must continue with binary questions.
In this post I take a look at how ChatGPT deals with this structure. I start simply, with concrete and abstract object, and move on from there.
* * * * *
All of the interactions reported below took place on the morning of October 15 using the September 25 version of ChatGPT.
Physical things
I decided to start with the simplest prompt I could imagine, “thing.” That’s the most general category we’ve got, no? Anything whatever is going to be some kind of thing. But how would ChatGPT respond? It responded as I thought it would, so I gave it a more explicit prompt, to which it responded.
I had no particular expectations about how ChatGPT would respond. But I couldn’t help but notice that each item in its response was some kind of physical thing. To be sure, rainbows (the second item in the list) are rather diaphonous, and you couldn’t actually touch one (I should have asked ChatGPT about that; later...), but it is still a physical thing.
I didn’t ask for physical things, but that’s what ChatGPT gave me. That’s worth noting.
Let’s think a bit before continuing on. I wonder if that first list is “rigid” in the sense that ChatGPT would respond the same way given the same prompt. I suspect not, but it’s easy enough to find out. I went back a bit later.
It’s not the same list I got the first time, but almost all of them are physical things. Classic literature (#3) isn’t really a physical thing, though books are. Mountain hiking (#7) is a physical activity, but whether or not you’d call it a physical thing is, well, an interesting issue. Jazz music (#7)? Science fiction movies (#13), puzzle games (#20)?
I decided to log out and then try again:
That’s better. I notice we’ve got “rainbow” again (#7), and “Lighhtening bolt” (#20) is rather ephemeral as well, though if one were to hit you, the effect would be dramatic and quite physical. “Chocolate chip cookies” (#9) show up again, as does “Hot air balloon” and “Bicycle.” These lists seem to be drawn from roughly the same conceptual territory. I suppose that’s not terribly surprising, but it certainly doesn’t seem like ChatGPT is making random choices from among all available physical things. Why these things?
I decided to explicitly specify physical objects:
None of those items are problematic in the way rainbows and lightening bolts are. You can touch each of them. In fact, except perhaps for the refrigerator, you can pick each item up and carry it around. No polar bears, mountains, lighthouses, roller coasters or traffic lights. But there are no purely abstract things here either, things like love, freedom, justice, beauty, etc. They go on a different list.
What’s ChatGPT doing?
Yes, we know, it’s cranking out one word after another. But that’s not very informative, as I argued some time ago, The idea that ChatGPT is simply “predicting” the next word is, at best, misleading. Remember, each time ChatGPT generates a new token it does a calculation that involves all 175 billion weights (Wolfram). Its entire “neural” space is involved in every such calculation.
Let’s take a look at my favorite analogy, Waddington’s epigenetic landscape:
The prompt positions it at a certain point in its activation space and then the ball rolls down hill from there. Notice that the diagram allows for choices along the way. Given the behavior we’ve seen above, we do have to allow for choices.
I would guess, though I don’t actually know, that at each point in generating the list, ChatGPT has choices. Just how wide that range is, I don’t know, but I would imagine that that varies with the value of the temperature parameter. A temperature of zero (0) seems to eliminate ALL choice, so that ChatGPT generates exactly the same list in response to the same prompt (given the same context). What would those lists look like with a much higher value?
The point of the diagram is that, once the “ball” enters a valley in the landscape, it just rolls down hill, with its trajectory responding to the lay of the land. Those lists tell us something about the lay of the conceptual land in ChatGPT. Physical objects seem to be more salient than abstract objects, and the most salient physical objects seem to come from a relatively restricted set of all possible physical objects. These choices are an abstraction over the the concepts in the universe of texts on which ChatGPT was trained.
Let’s get back to it.
Abstractions, definition, and metaphor
After I’d elicited that first list, the one beginning with beachball, rainbow, and bicycle I asked for a list of abstract things. I then went on investigating from there.
Note: Compare this with the work I’ve done with ChatGPT on metalingual definition.
Mechanical things
Imagine that you are interested in mechanistic interpretability. How would you go about finding the gradient that ChatGPT is following in compiling the following list?
The Great Chain of Being
This is a very important conceptual structure, one I’ve written a bit about: Ontology in Cognition: The Assignment Relation and the Great Chain of Being, Ontology in Knowledge Representation for CIM, and Relational Nets Over Attractors, A Primer: Part 1, Design for a Mind (pp. 38-40). Here we’re got more gradients in ChatGPT’s activation space.
More later.