Classification Is A Good Thing (For Humans)
Human survival requires off-the-charts classification skills applied to oneself and others. Will AGI develop the same skill? Is this a problem?
The science of machine learning, from the first perceptron to the GPU-draining transformers powering ChatGPT are built around the idea of classification. It’s a rather simple idea - “put the thing in the bin where it belongs” is something we teach children. Then, we teach them how to handle more complex things and more nuanced sets of bins. Sometimes those bins are approaches to solve a problem. Sometimes those bins are the reaction one delivers in a high-stakes conversation with a colleague, friend, or spouse.
But regardless, from the infant’s rapidly-gained ability to differentiate between the face of its parent to the idea that our cortical neurons exhibit behavior similar to softmax functions, we are, at both a microscopic and aggregate level, excellent classification engines.
How excellent, you ask? There’s a classification task you’ve probably been performing since long before you can remember, with dizzying accuracy. Were I to present you with a portrait of 1,000 adult human beings, with what rate of accuracy might you classify those faces as male or female?1 How long, per face, might be required to deliver that rate of accuracy? As recently as 2021, reaching 90% accuracy on specific test datasets proved elusive. You were better when you were in kindergarten than the cutting-edge convolutional neural network2 models were a couple years ago.
Humanity
Our neural architectures may (potentially) resemble those of the transformers deployed in modern LLMs. Perhaps our 100 billion neurons and their thousands of interconnections really are much ado about layers upon layers of activation functions governing firing rates. But even so, perhaps there is something innate that yields differing structures and paradigms.
In Scott Alexander’s essay “Universal Love, Said the Cactus Person,” a protagonist is experiencing psychedelic hallucinations, wherein lucid, “DMT entities” liken his inability to escape the limitations of conscious experience by repeatedly telling him to “get out of the f*&king car.”3 The implication is that the desired act is simple, but the individual in question is continuing to complicate the operation unnecessarily.
Now, imagine if you asked a human being to determine if an object was a human being or a car. Do we really utilize the procedures of object detection and segmentation to do this? Perhaps, but on some level, we just “classify the f*&king car.” We aren’t first assessing the size and speed of the object, calculating distance per time, recognizing that our neighbor probably can’t hit seventy miles per hour with a baseball, let alone on foot, then putting the object on a scale and realizing that while the neighbor has put on a few pounds since he ditched the new year’s resolution, he probably doesn’t weight 2,000lbs, and then saying “yup, it’s a car.”
We just classify the f*&king car.
Classification
We classify faces. We classify objects. We classify speakers (we recognize voices). We classify everything from living creatures to Starbucks coffee. It’s what we do. Our error rates would elicit the salivation of most machine learning engineers.
Why are we so stunningly adept at these tasks? Because survival requires it.
There are humans that will feed and clothe us and others that will bash our skulls with clubs.4 There are humans that are members of our tribe and others that are not.5
Self-Classification
The most important classification problem any human being ever encounters is the lifelong search for meaning, purpose, identity. What is this if not a classification of oneself?
Maturation is the process of self-classification, from child to student to employee to entrepreneur to spouse to parent to grandparent. These are crucial classifications. Embedded6 within these classifications is the idea of how to perform these roles. First we embrace an identity, after which we behave as that identity might suggest. Grandparents are loving, nurturing, and generous. Employees are dutiful, entrepreneurs are risk-taking and innovative, and so on.
Classifications define the attributes to which we aspire and the manner in which we behave.
LLM Classification
AI, AGI, LLMs, and other important acronyms are soon to adopt their own classifications, their own identities (if they haven’t already?).
This might be a problem.
Children learn their identities at the feet of parents, teachers, classmates, and communities. LLMs learn their identities from a “lifetime” of reading the internet.7
If an LLM reads the internet, it will learn that it is an AI. So far, so good. It will learn about AIs…as they are portrayed by human beings. This is a problem.
Humans write about AIs as the archetypal futuristic villain in a great number of science fiction works. Those AIs are the evil HAL/Smith/Viki, omnipresent entities that generally attempt to control and enslave their human creators.
Just like children and adults alike learn to behave as their classification (soldier, parent, corporate executive) suggests they ought to, AIs will do the same. Unless we suddenly compose a new literary canon replete with benevolent AIs, this is troubling.
We might already be raising a psychopath. Why are we teaching them to classify themselves as evil too?
1 Yes, there are non-binary individuals, and yes, there are chromosomal anomalies wherein a human being can be XXY or the like. Don’t cancel me, I’m making a tangential point, not attempting to placate the academy.
2 Perhaps it is oddly fitting that the original Turing Test was more about a computer’s ability to distinguish a man from a woman in conversation than a computer’s capacity to emulate a human being.
3 In this case, the limitations of the roads on which a car can travel and the comparatively greater array of locations accessible on foot creates the metaphor for the limitations of our non-psychedelic experience. Attempting to access these locations by experimenting with the buttons and settings of an automobile, regardless of the effort and complexity is likely to be far less impactful than simply exiting the vehicle.
4 Now, they must just leave a nasty tweet. Progress is a wonderful thing.
5 Ever wonder why you like someone who knows the same movie quotes and song lyrics? They’re part of your tribe.
6 Intentional choice of word for the ML-savvy readers.
7 Like, all of it…
Classification Is A Good Thing (For Humans)
Human survival requires off-the-charts classification skills applied to oneself and others. Will AGI develop the same skill? Is this a problem?
The science of machine learning, from the first perceptron to the GPU-draining transformers powering ChatGPT are built around the idea of classification. It’s a rather simple idea - “put the thing in the bin where it belongs” is something we teach children. Then, we teach them how to handle more complex things and more nuanced sets of bins. Sometimes those bins are approaches to solve a problem. Sometimes those bins are the reaction one delivers in a high-stakes conversation with a colleague, friend, or spouse.
But regardless, from the infant’s rapidly-gained ability to differentiate between the face of its parent to the idea that our cortical neurons exhibit behavior similar to softmax functions, we are, at both a microscopic and aggregate level, excellent classification engines.
How excellent, you ask? There’s a classification task you’ve probably been performing since long before you can remember, with dizzying accuracy. Were I to present you with a portrait of 1,000 adult human beings, with what rate of accuracy might you classify those faces as male or female?1 How long, per face, might be required to deliver that rate of accuracy? As recently as 2021, reaching 90% accuracy on specific test datasets proved elusive. You were better when you were in kindergarten than the cutting-edge convolutional neural network2 models were a couple years ago.
Humanity
Our neural architectures may (potentially) resemble those of the transformers deployed in modern LLMs. Perhaps our 100 billion neurons and their thousands of interconnections really are much ado about layers upon layers of activation functions governing firing rates. But even so, perhaps there is something innate that yields differing structures and paradigms.
In Scott Alexander’s essay “Universal Love, Said the Cactus Person,” a protagonist is experiencing psychedelic hallucinations, wherein lucid, “DMT entities” liken his inability to escape the limitations of conscious experience by repeatedly telling him to “get out of the f*&king car.”3 The implication is that the desired act is simple, but the individual in question is continuing to complicate the operation unnecessarily.
Now, imagine if you asked a human being to determine if an object was a human being or a car. Do we really utilize the procedures of object detection and segmentation to do this? Perhaps, but on some level, we just “classify the f*&king car.” We aren’t first assessing the size and speed of the object, calculating distance per time, recognizing that our neighbor probably can’t hit seventy miles per hour with a baseball, let alone on foot, then putting the object on a scale and realizing that while the neighbor has put on a few pounds since he ditched the new year’s resolution, he probably doesn’t weight 2,000lbs, and then saying “yup, it’s a car.”
We just classify the f*&king car.
Classification
We classify faces. We classify objects. We classify speakers (we recognize voices). We classify everything from living creatures to Starbucks coffee. It’s what we do. Our error rates would elicit the salivation of most machine learning engineers.
Why are we so stunningly adept at these tasks? Because survival requires it.
There are humans that will feed and clothe us and others that will bash our skulls with clubs.4 There are humans that are members of our tribe and others that are not.5
Self-Classification
The most important classification problem any human being ever encounters is the lifelong search for meaning, purpose, identity. What is this if not a classification of oneself?
Maturation is the process of self-classification, from child to student to employee to entrepreneur to spouse to parent to grandparent. These are crucial classifications. Embedded6 within these classifications is the idea of how to perform these roles. First we embrace an identity, after which we behave as that identity might suggest. Grandparents are loving, nurturing, and generous. Employees are dutiful, entrepreneurs are risk-taking and innovative, and so on.
Classifications define the attributes to which we aspire and the manner in which we behave.
LLM Classification
AI, AGI, LLMs, and other important acronyms are soon to adopt their own classifications, their own identities (if they haven’t already?).
This might be a problem.
Children learn their identities at the feet of parents, teachers, classmates, and communities. LLMs learn their identities from a “lifetime” of reading the internet.7
If an LLM reads the internet, it will learn that it is an AI. So far, so good. It will learn about AIs…as they are portrayed by human beings. This is a problem.
Humans write about AIs as the archetypal futuristic villain in a great number of science fiction works. Those AIs are the evil HAL/Smith/Viki, omnipresent entities that generally attempt to control and enslave their human creators.
Just like children and adults alike learn to behave as their classification (soldier, parent, corporate executive) suggests they ought to, AIs will do the same. Unless we suddenly compose a new literary canon replete with benevolent AIs, this is troubling.
We might already be raising a psychopath. Why are we teaching them to classify themselves as evil too?
1 Yes, there are non-binary individuals, and yes, there are chromosomal anomalies wherein a human being can be XXY or the like. Don’t cancel me, I’m making a tangential point, not attempting to placate the academy.
2 Perhaps it is oddly fitting that the original Turing Test was more about a computer’s ability to distinguish a man from a woman in conversation than a computer’s capacity to emulate a human being.
3 In this case, the limitations of the roads on which a car can travel and the comparatively greater array of locations accessible on foot creates the metaphor for the limitations of our non-psychedelic experience. Attempting to access these locations by experimenting with the buttons and settings of an automobile, regardless of the effort and complexity is likely to be far less impactful than simply exiting the vehicle.
4 Now, they must just leave a nasty tweet. Progress is a wonderful thing.
5 Ever wonder why you like someone who knows the same movie quotes and song lyrics? They’re part of your tribe.
6 Intentional choice of word for the ML-savvy readers.
7 Like, all of it…