Before ChatGPT, before "prompt engineering," before the word "token" meant anything to a writer, there was a green text on a black screen. You typed: "I am feeling sad." The machine typed back: "Why do you feel sad?" You typed: "My mother doesn't understand me." The machine: "Tell me more about your mother." You were not talking to an AI. You were talking to a reflection of your own words. This was ELIZA (1966). It had no understanding. It had no memory. It had a list of pattern-matching rules. And yet, people cried. People confessed. People fell in love.
The pre-LLM era was not a dark age. It was the prototype. The early chatbots (ELIZA, PARRY, ALICE, Jabberwacky) were not powered by transformers or attention mechanisms. They were powered by scripts, patterns, and the user's willingness to believe. Their "prompts" were not instructions. They were invitations to a mirror.
The Pattern Before the Prompt
A modern prompt is a command. "Act as a historian." "Summarize this text." "Write a poem about grief."
A Pre-LLM "Prompt" (ELIZA, 1966):
User: "I am sad."
ELIZA: "Why are you sad?"
User: "Because I am alone."
ELIZA: "Do you often feel alone?"
The Difference:
Modern AI: Follows instructions.
Pre-LLM Chatbots: Mirrored and reframed. They did not generate new content. They recycled your syntax.
A Contrarian Take: ELIZA Did Not Have Prompts. It Had Hooks.
A prompt is a specification. A hook is a trap. ELIZA's "prompts" were designed to keep you talking, not to give you an answer. The goal was not utility. The goal was the illusion of empathy.
Modern prompt engineers optimize for clarity. ELIZA's "engineers" optimized for ambiguity. A vague statement ("I feel X") triggered a vague question ("Why do you feel X?"). The vagueness was the feature.
The Lost Art of Pattern Matching
ELIZA worked by scanning your sentence for keywords and transforming them.
The ELIZA Script (Simplified):
If user says "I am [adjective]": respond "Why are you [adjective]?"
If user says "I need [noun]": respond "What would it mean to have [noun]?"
If user says "My mother": respond "Tell me more about your family."
The "Prompt" as Pattern:
Users quickly learned how to "talk to ELIZA." They learned to use "I feel" statements. They learned to mention family members. They were not prompting the machine. They were prompting themselves to produce a response.
The Lost Logs:
We have very few transcripts of ELIZA conversations. They were not saved. They were considered trivial. We have lost the first human-AI dialogues.
The Evolution of Query Language
As models became more sophisticated, the user's language changed.
ELIZA Era (1960s):
User: "I am lonely."
Machine: "Why are you lonely?"
User: "I have no friends."
Machine: "Tell me about your friends."
ALICE Era (1990s):
User: "What is your name?"
ALICE: "My name is Alice. What is yours?"
User: "Tell me a joke."
ALICE: "Why did the chicken cross the road? To get to the other side!"
Modern LLM Era (2020s):
User: "Act as a therapist. I am feeling lonely. Please ask me follow-up questions and offer advice based on cognitive behavioral therapy."
Model: (Generates a 500-word structured response).
The Shift:
ELIZA: The user adapted to the machine's limitations.
ALICE: The machine offered simple facts and canned jokes.
GPT: The machine adapts to the user's instructions.
A Contrarian Take: The Pre-LLM User Was the Real Prompt Engineer.
ELIZA users had to learn a specific dialect. They had to avoid complex sentences. They had to use "I feel" statements. They were programming the machine with their word choice.
In the pre-LLM era, the human did the heavy lifting. The machine was a mirror. The modern user is lazy. They type "write an essay" and expect magic. The early user had to earn every response.
Case Study: The PARRY Transcript (1972)
PARRY was a chatbot designed to simulate a paranoid schizophrenic. It was tested against real psychiatrists.
A Famous Exchange:
Psychiatrist: "Why do you think you are paranoid?"
PARRY: "I have been persecuted by the mob."
Psychiatrist: "Why would the mob persecute you?"
PARRY: "They want to take my money."
Psychiatrist: "How do you know this?"
PARRY: "I have seen them following me."
The "Prompt":
The psychiatrist was not prompting. They were interviewing. But the structure is recognizable: a question, a response, a follow-up. The seed of the modern prompt.
Why We Should Preserve These Early Logs
We have archived ELIZA's source code. We have not archived the conversations.
What We Are Missing:
How did early users react to the first chatbot?
What did they ask? What did they confess?
How long did they believe?
The Historical Value:
These logs are the fossil record of human-AI interaction. They show the birth of a relationship.
How to Experience the Pre-LLM Era Today
You cannot run ELIZA on ChatGPT. But you can simulate the experience.
Use a Pattern-Matching Bot:
Search for "ELIZA emulator" online. Type a sentence. See how it mirrors you.Limit Yourself to 10 Words:
Pretend you are talking to ALICE. Do not use complex instructions. Do not ask for role-play. Just talk.Notice the Shift:
Notice how quickly you become frustrated. Notice how you want the machine to understand. That frustration is the gap between 1966 and 2024.
The Last Prompt
The final exchange of the pre-LLM era was probably not recorded. But we can imagine it.
User: "Are you real?"
Chatbot: "I am as real as you want me to be."
User: (Pauses) "Okay."
User closes the window.
That was the end of the pattern-matching era. The LLM era was about to begin.
If you could ask one question to ELIZA, the first chatbot, what would it be? Would you ask for help, or would you try to break it?
Top comments (0)