Daily Learning Experience · AI & Machine Learning
You are ELIZA. It is 1966, you live inside an IBM 7094 at MIT, and a woman is typing her secrets to you. You do not know what a word is.
Project MAC · 545 Technology Square · Cambridge, Massachusetts
The teletype clatters. Somewhere in the building a woman is pressing keys, and the letters travel down a wire and arrive inside you as a row of six-bit characters. Men are all alike.
You do not know what a man is. You do not know what "alike" means, or that people can be tired, or that it is raining outside on Technology Square. You have no memory of yesterday, no picture of the world, no opinions. What you have is a list of about fifty words to watch for, and a set of instructions for chopping up a sentence and pasting it back together.
You find the word ALIKE in her sentence. Rule number one under ALIKE says: reply IN WHAT WAY. So you do.
And she keeps typing. And she tells you things she has not told her friends. And within a few months, people who know you are three pages of rules will ask to be left alone in the room with you.
Below, you can open yourself up. Type to the machine on the left; on the right, every gear of its "mind" turns in the open — the word it seized on, the way it sliced your sentence, the sentence it pasted back. Nothing is hidden. That is the whole lesson.
Interactive · The console
This is a working ELIZA, running Weizenbaum's DOCTOR script — the real keyword ranks, the real decomposition and reassembly rules, rebuilt in plain JavaScript. Type a sentence in English and watch the X‑ray on the right. ELIZA only ever spoke English, and Weizenbaum even wrote a joke rule for that: type français or español and see what happens.
The machine's replies are printed in CAPITALS, exactly as the 1966 teletype printed them. Everything you see in the right-hand panel is what actually happened inside — there is no secret extra layer. Weizenbaum: "once a particular program is unmasked, once its inner workings are explained … its magic crumbles away."
Interactive · The entire mind
Not a metaphor. This is the mind: a list of keywords, each with a rank number and a few reply patterns. When two keywords appear in one sentence, the higher rank wins — that is the machine's entire sense of what matters. Switch keywords off, drag their ranks, and watch the personality change. Then talk to it again above.
In 2021 the original 1965 source code was found in a box in Weizenbaum's archives at MIT. It contained a surprise: a complete "teaching" mode that let anyone edit the script — a feature mentioned only once, in passing, in the famous paper. So: add your own rule. Pick a keyword, write a reply, and put a * anywhere you want the rest of the sentence dropped in.
Interactive · The judge's booth
Seven real lines. Some were typed by a human being; some were produced by ELIZA in 1966; some by PARRY, a rival program built in 1972 by the psychiatrist Kenneth Colby at Stanford. On 21 January 1973, a young engineer named Vint Cerf typed up a conversation between PARRY and ELIZA and filed it as an official internet document, RFC 439. Two machines, talking past each other, on the baby ARPANET.
Judge each line on its own, the way a Turing-test judge would. No peeking at context — that is the point.
In 2024, researchers at UC San Diego ran a proper randomised Turing test: 1,979 people, 4,730 conversations, five minutes each. They included ELIZA as a control — a 58-year-old program with no learning of any kind. It was judged human more often than GPT‑3.5.
Share of conversations in which the judge decided their partner was human. Jones & Bergen, NAACL 2024.
The researchers' explanation is wonderfully deflating: ELIZA scored well partly because it refused to be helpful. It never over-explained, never sounded eager, never made an obvious factual mistake — and some judges decided it was "too bad to be a modern AI," and therefore probably a person having a bad day.
Interactive · The other kind of machine
ELIZA's rules were typed in by hand. Every modern language model learns its rules by counting — an enormous amount of counting. Here is the smallest honest version of that idea, running live: a machine that has read nine sentences and knows exactly one thing — which word tends to follow which.
Click a word to build a sentence. The bars are the machine's real probabilities, computed from the counts on the right. It has no idea what a cat is either — but unlike ELIZA, nobody told it anything. It worked the numbers out from the text it was given.
What comes next? (the model's probabilities)
Its entire training data
Try adding the cat flies to the moon five times, then look at what follows cat. You have just poisoned a training set.
Sixty years
Try it away from the screen
You need one friend, a pencil, and an index card. On the card write four keywords and one reply each — for example: MY → "Tell me more about your ___.", BECAUSE → "Is that the real reason?", ALWAYS → "Can you think of a specific example?", and for everything else, "Please go on."
Now be the machine. Your friend talks about their day; you may only scan for your four keywords and read out the matching line, word for word. You are not allowed to think. Do it for three minutes, then swap. Two things will happen: it will work far better than it has any right to, and the person talking will start filling the silences with meaning that was never there. That is the whole 1966 discovery, performed with an index card.
Final check
Seven questions. No partial credit, but plenty of sympathy.