ITS and Early AI: MIT's Pioneering Operating System and Symbolic AI Research

ScienceITS, Eliza & MiT's AI Lab - Computerphile

TL;DR

ITS was MIT's incompatible time-sharing system built in the 1960s for interactive AI research, enabling symbolic AI work like SHRDLU that demonstrated natural language understanding and block manipulation, though the AI goals ultimately failed despite producing foundational computing innovations.

Key Takeaways

1ITS was created as a lightweight, responsive alternative to CTSS specifically for AI researchers who found CTSS too slow; it deliberately prioritized interactive performance over compatibility, hence the name 'Incompatible Time-Sharing System.'
2Early AI in the 1960s focused on symbolic AI—encoding human intelligence through programming and logical reasoning—rather than machine learning or analyzing massive datasets.
3SHRDLU, a 1960s AI program written by a PhD student, could parse natural language commands like 'Pick up a big red block,' understand a 3D block-and-pyramid universe, and make plans to solve problems within that constrained world.
4The SHRDLU demo was recorded beforehand and played to military and financial backers, who were impressed enough to trigger the first AI summer with major funding increases, followed by the first AI winter when symbolic AI proved infeasible at scale.
5Despite failing to achieve its AI research goals, MIT's AI lab produced lasting computing innovations: the Emacs text editor, Lisp programming concepts, the ITS operating system itself, and numerous programming languages still influential today.
6SHRDLU's source code was lost for decades and only reconstructed about a year before this interview by members of the PDP-10 community, allowing researchers to finally test the system interactively rather than just viewing the original scripted demo.

Notable Quotes

“It's kind of a pun but it's also some truth behind it.”
— Speaker on ITS's name

“It's kind of a fake AI, really, because it's just picking up on some phrases in the input and generating something plausible.”
— Speaker on Eliza

“Getting this from a demo that works a bit is easy, but making it all-purpose, all-weather proof is much harder.”
— Speaker on why symbolic AI failed at scale

“In terms of developing artificial intelligence, that was not entirely a success. Yet in the process of all that research, they developed this fantastic operating system with a lot of new concepts, the Emacs editor, lots of programming languages, lots of computer science stuff that is now living in our computers.”
— Speaker summarizing ITS's legacy

Chapters

1. Origins of ITS and Time-Sharing Systems

John McCarthy pioneered AI and time-sharing at MIT in the late 1950s, leading to CTSS (Compatible Time-Sharing System) and the Multics project—but AI researchers found CTSS too slow and unresponsive, so they built ITS as their own incompatible but highly interactive alternative on a PDP-6 computer.

2. Symbolic AI and the Nature of Early Artificial Intelligence

Early AI, called symbolic AI, focused on encoding human intelligence through hand-coded logic and reasoning rather than machine learning—researchers tried to understand how the mind works and write programs that replicated that process.

3. Eliza: The Granddaddy Chatbot

Eliza was the first chatbot where users could type and receive computer-generated responses; the 'doctor' variant on ITS was a re-implementation that picked up on input phrases and generated plausible but fake responses, yet many users had genuine emotional interactions with it despite knowing it wasn't real AI.

4. Serious AI Work: Computer Vision and Robotics

Beyond Eliza, the lab pursued more serious AI projects including the Blocks World (copy demo), where a camera trained on a stack of blocks required the computer to visually recognize configurations and replicate them—this code has been lost, though the demo reportedly worked.

5. SHRDLU: The Landmark Natural Language AI

SHRDLU parsed natural language commands like 'Pick up a big red block,' understood a 3D block-and-pyramid universe, asked clarifying questions when needed, and made plans to solve problems—written by a PhD student using Lisp and a micro planner package, it represented genuine AI understanding within a constrained world.

6. The SHRDLU Demo and the First AI Summer

The SHRDLU demo was pre-recorded and shown to military and financial backers, who were amazed enough to trigger the first AI summer with massive funding increases; however, the system was slow and crashed on unexpected inputs—it was designed as a proof-of-concept rather than a production system.

7. The First AI Winter and Its Aftermath

Symbolic AI's failure to scale from impressive demos to reliable, general-purpose systems triggered the first AI winter as funding dried up and initial optimism turned to disappointment.

8. Legacy: Computing Innovations Beyond AI

Although symbolic AI failed, the MIT AI Lab produced lasting contributions to computer science: the Emacs editor, Lisp programming concepts, the ITS operating system with novel features, and multiple programming languages still influential today—though the PDP hardware approach became obsolete.

9. SHRDLU Code Rediscovery and Current Preservation

SHRDLU's source code was lost for decades; the PDP-10 community reconstructed it only about a year before this interview, allowing researchers to test the system interactively and verify that the original recorded demo was a genuine unscripted session.

Key People & Entities

John McCarthy: MIT researcher in the late 1950s who pioneered concepts of artificial intelligence, time-sharing, and Lisp programming language.
SHRDLU's Creator: A PhD student at MIT who built SHRDLU, the natural language understanding and block manipulation AI system; his name is not mentioned in the transcript.
MIT AI Lab: Research group that developed ITS, Eliza, SHRDLU, and other pioneering AI and computing systems in the 1960s–70s.
PDP-10 Community: Group of researchers who reconstructed SHRDLU's lost source code approximately a year before this interview, enabling interactive testing of the historic system.

Glossary

CTSS: Compatible Time-Sharing System; an early time-sharing operating system from the late 1950s that ran on IBM computers and influenced later systems like Multics.
Symbolic AI: An approach to artificial intelligence that encodes human intelligence through hand-written logical rules and reasoning, rather than machine learning from data.
Time-Sharing: A computing architecture where multiple users can interact with a single computer simultaneously, contrasting with earlier batch processing where programs were queued and run one at a time.
Batch Processing: An early computing model where users submitted punch cards with programs to an operator, who fed them into the computer; results were returned hours or days later.
Micro Planner: A software package used by SHRDLU to encode planning logic and goal-driven reasoning within its constrained 3D block world.
Blocks World: A simplified 3D environment containing geometric shapes (blocks and pyramids) used for AI research in computer vision, natural language understanding, and planning.
AI Summer: A period of optimism and increased funding for artificial intelligence research; the first AI summer (1970s) followed successful demos like SHRDLU.
AI Winter: A period of reduced funding and pessimism for AI research, typically following hype that could not be delivered; the first AI winter followed the failure of symbolic AI to scale.

Explore