What AI Is

What We Mean By Artificial Intelligence

Artificial intelligence, or AI, is the study and building of computer systems that do tasks which normally need human judgment. Recognizing a face in a photo, translating a sentence, and suggesting the next word in a message are all examples. The word covers a wide family of methods, not one single technology.

It helps to separate three ideas. A program is a list of instructions a computer follows. An algorithm is a clear procedure for solving a problem, such as sorting a list of names. A model is a stored set of numbers that a program uses to make predictions. Many modern AI systems are built from a model plus a small program that runs it.

People often use AI to mean the chat assistants they see in the news. Those assistants are one part of the field. Spam filters, map routing, voice dictation, and photo search are older and quieter examples. In this course we start with the whole field and then narrow to large language models, which power chat assistants.

A Short History, 1950 To Today

In 1950 Alan Turing published a paper called Computing Machinery and Intelligence. It asked whether machines can think and proposed a test based on conversation, now called the Turing test. In 1956 a summer workshop at Dartmouth College gave the field its name, artificial intelligence.

The early decades focused on symbols and rules. Researchers wrote programs that reasoned with logic and wrote down expert knowledge by hand. These expert systems worked in narrow areas, but they were expensive to build and broke when they met a case nobody had written a rule for.

Funding and excitement rose and fell in cycles. Periods of disappointment are often called AI winters. Progress continued in the background, especially in statistics and in a method called machine learning, where computers find patterns in data instead of following hand-written rules.

Two events marked a turn. In 1997 the chess computer Deep Blue beat the world champion Garry Kasparov. In 2012 a neural network called AlexNet won an image recognition contest by a large margin, which drew attention to deep learning. In 2017 researchers at Google published the paper Attention Is All You Need, which introduced the transformer design behind today's language models. We study these in weeks 3 to 5.

Milestones In AI

Year	Event	Why it matters
1950	Turing publishes Computing Machinery and Intelligence	Poses the question of machine thinking
1956	Dartmouth workshop	The term artificial intelligence is adopted
1966	ELIZA	First famous chatbot
1997	Deep Blue beats Kasparov	Machine wins at chess
2012	AlexNet	Deep learning takes over vision
2017	Transformer paper	Basis of modern language models

Dates are the commonly cited ones. Source: Standard AI histories

Rules Versus Learning

There are two basic ways to make a computer do a task. In the first, a person writes the rules. In the second, the computer finds the rules by studying examples. Both are useful, and good engineers know when to pick each one.

Consider a program that decides whether an email is spam. A rules-based version might say: if the subject contains the phrase free money, mark it as spam. This is easy to read and easy to fix. It also fails quickly, because spammers change their wording and there are too many patterns to list.

A learning-based version receives thousands of emails that people already labeled as spam or not spam. It adjusts itself until its guesses match the labels. It then applies what it found to new emails. Nobody wrote the final rules, and the system can notice patterns a person would miss.

Learning needs data and gives answers that are sometimes wrong. Rules need human effort and give answers that are fully predictable. Many real products combine the two: a learned model proposes an answer and a few firm rules check it.

• Use rules when the task is clear, the logic is short, and mistakes are costly.
• Use learning when the task has many fuzzy cases and you have plenty of examples.
• Combine both when you want flexible predictions with firm safety checks.

From Rules To Learning

1. Person writes rules
2. Program follows rules
3. Fails on unseen cases
4. Show examples instead
5. Model learns the pattern

The shift that defines modern AI.

Rules Versus Learning

	Rules	Learning
Who writes the logic	A person	The data
Strength	Predictable, easy to explain	Handles messy, fuzzy tasks
Weakness	Breaks on cases nobody listed	Needs data, can inherit its flaws
Example	If subject contains 'prize', mark spam	Learn spam from labeled emails

Narrow AI And General AI

Almost every AI system in use today is narrow. It does one kind of task well, such as playing a board game or translating text, and cannot easily move to a different kind of task. Large language models are broader than older systems because they handle many text tasks with one model, but they still have clear limits.

Artificial general intelligence, or AGI, means a system that can learn and perform nearly any intellectual task a person can. No one has built such a system, and experts disagree about when or whether it will exist. In this course we focus on what current systems actually do, which you can test yourself.

The Landscape Today

Today's AI tools can be sorted by what they take in and what they produce. Knowing the categories helps you pick the right tool and ask the right questions about it.

Most of this course is about the first category, language models, because their ideas carry over to the others.

• Language models read and write text. Chat assistants are built on them.
• Image generators turn a text description into a picture.
• Speech systems convert audio to text, or text to audio.
• Vision systems label or find things in photos and video.
• Recommendation systems rank videos, products, and posts for each user.
• Multimodal models accept more than one type of input, such as text and images together.

Who Builds AI And How You Use It

A small number of companies train the largest models, because training costs a great deal of computing power. Many other groups publish open-weight models, which means the trained numbers are available to download and run yourself. Universities and independent researchers also contribute methods and benchmarks.

You can reach these models in three ways. You can use a chat application in a browser. You can call a model from your own program through an API, which is a defined way for programs to talk to a service. Or you can download a smaller open model and run it on your own computer. Week 10 covers building with an API.

Whatever the route, treat the output as a draft from a fast but fallible assistant. It can write fluent text that is wrong. Your job is to check what matters.

Habits For Learning AI Well

Learn by testing. When someone makes a claim about what AI can do, try it on a small example. Keep a notebook of prompts that worked and prompts that failed, and write one line about why.

Keep your own judgment active. Ask what data a system learned from, who checks its answers, and what happens when it is wrong. These questions matter more as you build real tools, and we return to them in weeks 8 and 11.

Finally, expect the details to change quickly while the core ideas stay stable. Names of products and models will change many times during your life. The concepts in this course, such as data, training, evaluation, and context, will remain useful.

Throughout this course, keep one simple test in mind for any AI claim: what is the input, what is the output, and how would we know the output is good? Asking these three questions turns a vague promise into something you can try on a laptop. It also protects you from both excess worry and excess excitement, because you judge each system by what it demonstrably does.

Key Terms

Artificial intelligence

The field of building computer systems that perform tasks which normally need human judgment.

Algorithm

A clear step-by-step procedure for solving a problem.

Model

A stored set of numbers that a program uses to make predictions.

Machine learning

A method where a computer finds patterns in data instead of following hand-written rules.

Expert system

An early kind of AI program that encoded human expert knowledge as hand-written rules.

Narrow AI

A system that performs one kind of task well and does not transfer easily to others.

AGI

A hypothetical system able to learn and do nearly any intellectual task a person can.

Open-weight model

A trained model whose numbers are published so anyone can download and run it.

API

A defined way for one program to send requests to another service and receive answers.

Exercise: Rules Versus A Chat Assistant

1. Pick a small sorting task, such as labeling 10 short customer comments as positive, negative, or neutral. Write the 10 comments yourself.
2. Write down 5 simple rules a program could use to label them, for example looking for the word broken.
3. Apply your rules by hand to all 10 comments and record how many you label correctly.
4. Open a free chat assistant and paste the 10 comments with the instruction: label each as positive, negative, or neutral.
5. Compare both sets of labels against your own judgment and count the matches for each approach.
6. Write 5 sentences on where the rules failed, where the assistant failed, and which approach you would trust for a real product.

Further Reading

• Computing Machinery and Intelligence, Alan Turing, Mind, 1950
• Artificial Intelligence: A Modern Approach, Stuart Russell and Peter Norvig, Pearson (chapter 1)
• Elements of AI, Introduction to AI, University of Helsinki and MinnaLearn, free online course
• Attention Is All You Need, Vaswani et al., Google, 2017 (read the abstract only this week)