Flavors of AI

Artificial intelligence can mean different things and

has evolved over time. It includes a variety

of techniques, ranging from simple

classifying strategies to complex generative models. The biggest distinction

across applications of artificial intelligence

is whether it's being narrowly

or broadly construed. Narrow artificial

intelligence is designed to accomplish specific tasks

within a limited domain. For example, an email filter

designed to determine if a particular email is spam

is an example of narrow AI. It only does that specific task. In contrast, broad

AI involves systems capable of performing

multiple tasks across various domains, such as virtual

assistants that can handle diverse

queries and tasks or generative algorithms

that produce novel text in response

to human input. Over time, artificial

intelligence applications have become increasingly

sophisticated. While early applications

of what we now consider AI were virtually

indistinguishable from other statistical models, today, artificial intelligence

uses a toolbox that looks somewhat different from many other statistical

applications. Some of the earlier approaches that evolved into

what we today think of as artificial intelligence

are classifying systems. That is systems that sort input into predefined

categories. For example, a credit

risk assessment can use relatively basic

statistical procedures to predict how likely

someone is to default on a loan based on aspects of that individual's

financial risk. As these models have

become more sophisticated, credit risk assessments now use AI tools to analyze patterns

in the data and learn from real-life loan

defaults to estimate financial risk both more accurately and in a

more granular basis. Artificial intelligence

models learn using statistical

learning techniques. They start with relatively

simple basic models and use them to identify

patterns and make predictions. To give an example of

how this might look, imagine that you were in the business of

selling ice cream. It doesn't take long to notice that the hotter

the temperature, the more ice cream

you'd be likely to sell on a particular day. A statistical regression is

a tool that can help you predict how much ice cream to make based on

the temperature. But many problems are

more complex than a direct relationship between temperature and ice cream sales. For instance,

predicting the weather requires numerous

pieces of information, and the presence of

the same factor may yield different predictions

in different circumstances. As the relations

between inputs and outputs get

increasingly complex, two things begin to happen. First, more and more

data is needed to figure out how the parts of a model relate to one another. Second, it becomes less and less clear how those

predictions are made. This process results in a

model that can generate excellent predictions but is

functionally a black box, where we may not know why the model predicted what it did. To figure out the predictors and outcomes of complex models, computer scientists have

derived a series of approaches that are collectively referred to as deep learning. As an example, one

common strategy is called generative

adversarial networks. These GANs train

artificial intelligence by involving two models: a generator and a discriminator. The discriminator tries

to classify things while the generator tries to create new content to trick

the discriminator. For example, a discriminator

tries to decide if an email should be classified as spam or if it should

get into your inbox. Each time the

discriminator improves at determining which email

should be sorted out, the generator model

produces new content with the goal of fooling

the discriminator model. As the discriminator improves at predicting

certain categories, the generator

simultaneously improves at figuring out what might

trick the discriminator. Over time, this competition

improves both models until the discriminator can

classify as well or better than humans without

regular human input. As this process moves forward, the models get better and

better at their purposes. But they also get

increasingly complex, and the multitude of factors involved make them

relatively opaque. Applications like ChatGPT

or its Google, Meta, and Anthropic

counterparts fall into a group of AI products referred

to as foundation models. These are general

purpose algorithms that work with a broad series

of natural language, visual and/or auditory inputs, and outputs on a large variety

of different problems. Foundation models can then be honed to accomplish

a specific goal. One of the major

developments in AI was the realization that

foundation models that were later tailored to a

particular purpose tended to perform better than models that were trained only on data from other

similar problems. For this reason, many

contemporary applications of AI involve applications of foundation models to

specific problems rather than the development

of new AI models. In sum, contemporary artificial

intelligence is built on models that attempt to identify statistical

relations between concepts. The models themselves

sometimes are designed and built to

accomplish a narrow goal and sometimes involve

multi-purpose algorithms that can either be used as general assistance or may be honed toward specific use cases. The way that many

of these models are constructed can be

quite complex to the point that it is

often difficult to understand why they produce

the outputs they do.