Incompatible Time-Sharing System (ITS): Early AI, Eliza, and SHRDLU

The Incompatible Time-Sharing System (ITS) emerged from the early days of computing and artificial intelligence (AI) research, offering a unique perspective on the development of interactive computing and symbolic AI.

Origins: From Batch Processing to Time-Sharing

In the late 1950s, John McCarthy at MIT pioneered concepts like artificial intelligence, time-sharing, and the Lisp programming language. While some researchers focused on AI, others picked up on time-sharing.

Prior to time-sharing, computing was a cumbersome process. Users would bring a batch of punch cards to an operator, who would feed them into a computer. Results, often in the form of a printout, would only be available the next day. This "batch processing" was slow and inefficient.

The Compatible Time-Sharing System (CTSS) was an early experiment in time-sharing, running on IBM computers and compatible with their existing batch software. Its success led to Project Multics, an ambitious endeavor to create a time-sharing utility envisioned as the future of computing across the country. Multics was a serious academic undertaking, though it is now primarily remembered for its role in the history of Unix.

The Birth of ITS: An "Anti-CTSS" Approach

While Multics was being developed, another group of researchers at MIT, focused on AI, found CTSS to be too slow and unresponsive for their needs. They acquired their own computer, a PDP-6, which initially operated as a single-user system. After a couple of years, they realized the need for a time-sharing system but desired something highly interactive and responsive, unlike CTSS.

This desire for a system that was the antithesis of CTSS led to the creation of the Incompatible Time-Sharing System (ITS). The name itself was a pun, but it also reflected their goal of building a lightweight, responsive, and interactive system tailored for AI research.

Early AI: Symbolic AI and Eliza

The AI research conducted on ITS was vastly different from the deep learning and neural networks prevalent today. This early form was known as "symbolic AI," which focused on encoding human intelligence and knowledge by hand, rather than analyzing massive datasets through machine learning. The goal was to understand how the mind works and then write programs that emulated those processes.

One of the most famous early AI programs, and often considered the "granddaddy of all chatbots," was Eliza. Eliza was a groundbreaking program that allowed users to type in questions and receive seemingly intelligent responses. However, Eliza was a "fake AI" in the sense that it didn't truly understand language. Instead, it picked up on keywords and phrases in the user's input and generated plausible, often canned, responses. Despite its simplicity, many users found themselves having surprisingly heartfelt interactions with Eliza, even when they knew it wasn't a "real" AI.

Beyond Eliza: Computer Vision, Robotics, and SHRDLU

Beyond chatbots, ITS also facilitated more serious AI research in areas like computer vision and robotics. One notable project was the "copy demo," where a camera observed a stack of blocks, and the computer had to interpret the visual input to understand the block arrangement. It then had to replicate that arrangement using other blocks. While the code for this specific demo has been lost, it represented a significant step towards real AI.

Another remarkable achievement was SHRDLU, a program that demonstrated natural language parsing and interactive problem-solving within a simulated "blocks world." SHRDLU allowed users to interact with a virtual 3D environment of colored blocks and pyramids using natural language commands. For example, a user could command, "Pick up a big red block," and SHRDLU would respond and execute the action within its virtual world.

SHRDLU was written in Lisp and utilized a pre-existing software package called Micro Planner. It was a highly complex system for its time, capable of understanding the user's intent, asking clarifying questions, and planning actions within its limited universe. The original description of the SHRDLU demo was preserved, but the source code was only reconstructed about a year ago by members of the PDP-10 group, allowing modern users to interact with it directly.

The SHRDLU demo was incredibly impressive, particularly for a PhD student's work. The goal was to create a functional demo that showcased the potential of AI, even if the underlying system was prone to crashes in real-world scenarios. A famous anecdote recounts how military officials, upon seeing the SHRDLU demo, were "blown away," leading to a surge in AI funding, often referred to as the "first AI summer."

The First AI Winter and ITS's Legacy

However, the initial excitement surrounding AI quickly led to disappointment, ushering in the "first AI winter." The leap from a compelling demo to a robust, all-purpose AI system proved far more challenging than anticipated.

Despite the eventual "failure" of symbolic AI to achieve its ambitious goals at the time, the research conducted on ITS had a profound impact. While the original funding goal of creating a fully realized AI did not materialize, the process of developing ITS led to the creation of a fantastic operating system with many new concepts, the Emacs editor, numerous programming languages, and other foundational computer science advancements that are still relevant today.

The hardware on which ITS ran, the PDP-10, eventually became obsolete. ITS itself was eventually shut down at MIT, though its files remained available on FTP sites for original users. While ITS may not have delivered on its promise of a fully intelligent AI, its legacy lies in the pioneering work it enabled in interactive computing and the foundational contributions it made to computer science.