The machine wasn’t understanding anything. It was spotting patterns and rearranging them. At the time it felt like a clever trick with language. Only later did it become clear that the same basic idea—recognizing patterns instead of understanding meaning—would begin turning up in many other places.
But the illusion could still be convincing. If someone typed “I hate my history teacher,” the program might respond, “Tell me more about your teacher.” People sometimes reacted as if the machine were actually listening.
At the time it felt like a toy.
My attention kept drifting somewhere else: the operating system.
One Sunday afternoon at Dial-Data I had the SDS-940 all to myself and started doing what young hackers do when nobody is watching—poking around inside the system to see how it worked.
One thing caught my eye.
The system spent an enormous amount of time shuffling information between memory and disk. Watching it run, it felt wasteful.
So I tried a small experiment.
I spent the evening tweaking code in the operating system so it saved only memory that had actually changed. Around three o’clock Monday morning, the test came together, and the system suddenly felt different. The frantic shuffling had almost disappeared. Programs ran noticeably faster.
The whole machine felt lighter on its feet.
Satisfied with my experiment but totally exhausted, I went home for some sleep. What I didn’t do was put the operating system back the way I had found it.
The next morning at 8 a.m., the incessant ringing of the phone woke me.
“Southworth, were you here yesterday?”
I said yes.
“Get in here. Now.”
I assumed I had broken something and was about to be fired. On the drive in, I rehearsed my mea culpa and planned how quickly I could restore the original code.
When I walked in, my first words were ready.
“I’ll put the old code back,” I said. “Don’t worry.”
The response came instantly.
“No. No! Just tell us what you did.”
Not long after that I received an offer to work at Harvard’s Aiken Computation Laboratory. The pay was modest, but the job came with a wonderful benefit: I could audit Harvard courses for free.
A few months later another call came, this time from MIT—with a real salary and MIT classes thrown in.
They were building a new operating system called Multics, part of Project MAC, a collaboration involving MIT, Honeywell, Bell Labs, and the Defense Department’s Advanced Research Projects Agency.
Project MAC lived inside 545 Technology Square in Cambridge, then a fairly new building.
On the ninth floor was the Artificial Intelligence Lab where researchers like Marvin Minsky and Seymour Papert explored early ideas about machine intelligence.