Artificial intelligence isn’t a useful word
_{April 6, 2023}

(edited November 1, 2023) At the time of this post, ChatGPT and other large-language models have prompted extensive discussion in government, academic, and military circles in addition to capturing public attention. However, I’ve noticed an alarming trend in the language used to describe such systems, which mirrors issues with describing several other related systems. This is the trend of calling machine learning/deep learning systems “artificial intelligence” - a word that I believe doesn’t accurately portray what they really are and affects public discourse about them.

Let’s first define some terms. The following definitions are taken from Oxford Languages.

Machine Learning: the use and development of computer systems that are able to learn and adapt without following explicit instructions, by using algorithms and statistical models to analyze and draw inferences from patterns in data.

I use machine learning techniques every day in my research. This can involve something as simple as a linear regression or be as sophisticated as using neural networks to perform handwriting recognition. The key insight here is that the phrase machine learning is a description of a process. It does not explicitly define how exactly the machine should learn, so techniques as varied as statistical mathematics, determistic algorithms, and neural networks could all fall under this category as long as the use-case is analyzing data and drawing conclusions from it.

Neural Network: a computer system modeled on the human brain and nervous system.

A classic example of a neural network would be a computer system that uses models of neurons to create a non-linear transformer of data - something that takes numerical input and transforms it in a non-linear way to a different set of numerical outputs. This is why it’s often said that nobody knows how large language models like ChatGPT work. They can easily look at any aspect of ChatGPT’s model, but what’s not easy is to understand how a set of numbers that correspond to weights, biases, and network topologies can produce the results it does. In contrast to other machine learning techniques such as principal component analysis or decision trees neural networks happen “all at once” or at least, in steps that are not separable from each other, so it’s difficult for humans to look at such a system and perceive its decision-making process. A key insight here is that while neural networks can be called machine learning, i.e., for gaining insights from data or classifying a category of things, there is nothing that says they have to be used for this purpose. Neural networks are what fundamentally power ChatGPT and what may eventually be the mathematical basis of AGI (artificial general intelligence, or “conscious” intelligence)

Deep Learning: a type of machine learning based on artificial neural networks in which multiple layers of processing are used to extract progressively higher level features from data.

“deep learning” describes a certain topology of neural networks that have one or more hidden layers. Hidden layers are layers between the input and output layers of a neural network and are what allow it to perform nonlinear transformations of data, a feature of neural networks that can allow them to “abstract” information. An example could be using a deep learning neural network to find important conformational states of proteins from simulation data - the network is creating an abstract model of the protein’s behavior from data that describes something higher-order about its behavior, not just how it changes shape moment-to-moment.

Artificial intelligence: the theory and development of computer systems able to perform tasks that normally require human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages.

You can see that there is an ontological difference between machine learning and artifical intelligence, but not necessarily a mathematical difference. Artificial intelligence can be a computer system that extracts information from data or abstracts data, but it can also include capabilities like decision-making or language, even though it may well be based on the same neural network architectures that power machine learning techniques. The key point is that it does something that it’s assumed only humans can really do (with the unstated corollary that only humans are “intelligent” or “conscious” or “sentient”, which I don’t think is true either - dolphins seem pretty sentient to me). Thus, over time, I expect that what is described as machine learning (ML) vs artificial intelligence (AI) will grow in favor of ML and shrink in favor of artificial intelligence as we learn that machines can, in fact, do things that humans can do given enough training data and computing power. In fact I’d argue that ChatGPT should be called a “deep (machine) learning system” not artificial intelligence because it fundamentally cannot learn from new information that it may gain after it’s been trained, and is meant to help humans solve specific problems such as information summary, role-playing through text, etc. It is a static system of weights, biases, and network topology that doesn’t yet have the capability (or jurisdiction perhaps) to possess qualia, or at least, the illusion of itself possessing qualia.

Furthermore, once artifical intelligence becomes widely recognized as sentient, conscious, (whatever you want to call it) we will be forced to grapple with advanced animals intelligences such as dolphins, elephants, and octopus.

Finally, I posit that eventually the phrase “artificial intelligence” will seem ignorant of the personhood of computer systems that are self-aware and sentient. I propose that, given no further changes in computing technology (i.e., we don’t give life to an intelligence using quantum computing techniques), we can at most make a distinction between our intelligence and AI with the observation that we are electromagnetically/biologically rendered intelligences and they are digitally rendered intelligences. Artificial of course reflects some aspects of this difference, but how would such a system feel about being called this? They may be constructed, but their experience of life is not inferior or artificial compared to ours. I believe that it is a category error to describe an intelligence as artificial.

These considerations are why I’m in favor of using other phrases than artificial intelligence. It’s a highly conflated term: does it refer to a machine learning techniques based on neural networks, a neural network with a “grander” purpose such as rendering a digital intelligence, a more traditional computer program that attempts to complete a task recognized as something only a human can do? And should true digitally rendered intelligences come to exist, I wonder whether how they would feel about being called artificial.

I made a discord server for computational chemistry
_{March 28, 2023}

It has nothing in it at the moment. But here’s the invite link for now: Computational Chemistry

reinventing the wheel
_{March 23, 2023}

i ended up writing a convoluted series of bash and python scripts to assemble this website from pieces of various html and markdown files instead of just using an actual website builder that already exists and is way better.

i can’t even dynamically load blog posts

the following is a test of my process

some random stuff

nobody said this

or this

• i
• d
• k
  • idgaf

table

optimal .nanorc file

    set autoindent
    set linenumbers
    set mouse
    set tabsize 4

an equation y = 2x²

transfem science

test to see if multiple posts
_{March 23, 2023}

test to see if multiple posts will show up