L02. From written rules to talking with AI

You will outline AI’s history and distinguish a chat reply from an action completed with a tool. Allow about 30 minutes. Read Lesson 1 first and have paper ready; no account or installation is needed.

In class

The teacher compares a rule-based program, a model trained on examples and an agent with tools. Together, identify what three requests need: answering, saving a file and publishing a video.

An old question, several approaches

Olena watches a program write an announcement in seconds. Artificial intelligence, or AI, also covers tasks such as recognising images, translating and finding patterns. Conversational programs are one part of this history.

In 1950, Alan Turing proposed an “imitation game” assessing machines through written conversation. This framed a research question; it did not establish that convincing answers make a computer a person. [F207]

John McCarthy’s summer 1956 Dartmouth workshop helped establish artificial intelligence as a field. Researchers wanted machines to perform tasks associated with human intelligence, without one agreed method. [F206]

One approach used written rules linking conditions to actions. In the mid-1960s, Joseph Weizenbaum’s ELIZA processed phrases this way. It could seem like an attentive listener without knowing the person’s life. [F227]

From specialised tasks to learning from examples

Games offered clear tests. In 1997, IBM’s Deep Blue defeated Garry Kasparov in a chess match: a major specialised achievement, not a general everyday assistant. [F208]

Machine learning instead adjusts a system through examples, supporting image recognition, translation or text prediction. Larger datasets, greater computing power and improved methods gradually made these systems more useful.

A neural network is a mathematical system whose numerical settings change during training. In 2017, researchers presented a way to build one, the Transformer, in *Attention Is All You Need*. It helps account for relationships between parts of text and became a foundation for many modern language models: systems that work with language. [F209]

The public launch of ChatGPT in 2022 brought conversation with such a model to a broad audience. [F207] Interfaces then combined conversation with tools: searching, handling files, calculating and editing. This led to practical interest in agents, systems that can take a sequence of actions towards a requested result. Modern ChatGPT Work and Codex, for example, can carry tasks beyond a text reply. [F202]

Earlier methods did not disappear. Ordinary rules remain useful: for example, a form can check whether a required field is filled in.

[[VIS:f5_ai_history]]

What this means for your work

Olena asks, “Suggest three titles for my video.” A chat can immediately offer alternatives. “Make a file containing my filming plan” is a different task: it needs a file-creation tool. Publishing the video also requires access to the appropriate account.

For each request, name a result you can check: three titles on screen, a file in a folder or a page at an address. This helps you decide whether the task needs a tool as well as conversation.

Home practice

• On a sheet headed “The path to chatbots”, arrange four milestones: ELIZA, Deep Blue, the Transformer and ChatGPT’s public launch. Find the period or year in the lesson and explain each milestone in one sentence.
• Compare “Suggest a video title” with “Save the plan in a file”. Note which needs only a text reply and which needs a tool. Check your chronology and explanations against the lesson.

Self-test

• Did ELIZA work like a modern language model?
• Does winning at chess demonstrate the ability to write a good letter?
• A chat says “plan saved” but only shows text. What is missing before you can count the saving task as complete?

Words you will need

• Artificial intelligence: the field of creating systems for tasks that require intellectual abilities.
• Machine learning: adjusting a system using examples.
• Agent: a system that uses tools to carry out a task.