A few months ago, I spent 28 days asking ChatGPT to generate code diagrams and musical compositions. I was impressed. You can check out my experiments on Quality Containers.
What kind of model is AI — particularly Large Language Models like Chat GPT?
Influences on my thinking today include:
Paul Ricoeur — Freud and Philosophy (Ricoeur coined the term hermeneutics of suspicion)
Eve Sedgwick — Touching Feeling (This book has an excellent section on the hermeneutics of suspicion)
Colin Drumm and his take on triangles (the Oedipus myth is a triangle) and Sedgwick’s late-in-life interest in mathematics and geometry (triangles?).
Ray Brassier — Dialectics between Suspicion and Trust
Michael Weisberg— Simulation and Similarity: Using Models to Understand the World
Roman Jakobson — “The Metaphoric and Metonymic Poles”
Thomas Nagel — What is it like to be a bat?
The Neural Net
Weisberger distinguishes between three types of models. I interpret these as concrete, computational, and simulation models.
Concrete models are scale replicas.
Computational models are like equations from physics — force = mass * acceleration
Simulation models run scenarios.
What sort of model is a Large Language Model, or more broadly, a neural network?
Is a Neural Network a simulation model? No
A neural network does not simulate a process or run a scenario. A neural network tries to train itself to produce “the right answer.”
Neural Networks are outcome-focused, not process-focused.
My friend Kai says that a neural network is like dog training. All you do is punish or reward the model for answers and hope the model learns.
But Neural Networks are based on the brain? No
Although neural networks use the word “neuron,” its relationship to the brain is a metonym, not a metaphor. A neural network is not like the neurons in the brain. Using a threshold value is not enough to hold this metaphor.*
Instead, a neural network stands in for computational reasoning we cannot access. Similar to how we do not know whether a dog will roll over or not — upon command.
Are Neural Networks Algorithms? I would say no.
Algorithms are a list of steps, or instructions, that specify how something happens. While I can have an algorithm to create a neural network, neural networks are not necessarily algorithmic. We cannot access how the steps the neural net follows generate the outcome.
Hermeneutics of Suspicion
Ricoeur refers to Marx, Nietzsche, and Freud as the “masters of suspicion.” These thinkers analyzed aspects of society and humans by looking at and interpreting symptoms.
Suspicion opposes skepticism — doubt and the ability to disprove something.
Scientists like Newton observed the world and proved or falsified theories and hypotheses. Theories and hypotheses can be coded up as algorithms in a computer.
Interpreting a system is a different activity.
Why did the Large Language Model provide a particular answer to a prompt? We cannot go to the algorithm to find out; we must interpret the symptom.
Note on Metynomy and Metaphor and Neural Networks
There is a relationship between the concept of a threshold value in neurons — which causes the neuron to fire, and a threshold in neural networks, which causes the neurons in the network to fire, but threshold values are everywhere in computation. It is not a unique property of neurons in the brain.
—
Written by Meredith
Witness
Leave a Reply