For 208 years, writers have imagined machines that think. For the last four months, I have managed the latest version: AI agents. Science fiction gave us two scripts for artificial intelligence: the faithful servant and the overlord that turns on its makers. I have always been an avid reader of science fiction, much of it ridiculous, some of it prescient. After a season running a company on AI agents, I can report that neither script nor any of the books I have read prepared me for an ordinary Monday morning, when the problem isn't rebellion but an agent that skipped step four of a checklist.
This is the first post in a new series, Life on the Frontier of AI. It seemed right to start with the maps people drew before anyone arrived.
Before computers: 1818 to 1909
Mary Shelley published Frankenstein in 1818. There is no computer in it, but it sets the pattern every AI story since has borrowed: someone builds a mind, and then refuses responsibility for it. The creature is not the villain of the book so much as the consequence of a maker who walked away.
Forty-five years later, in 1863, Samuel Butler wrote a letter to The Press in Christchurch, New Zealand, titled Darwin among the Machines. His argument was startling for the steam age: machines were evolving faster than animals, and by natural selection they might one day become conscious. He expanded the idea in his 1872 novel Erewhon, whose citizens destroyed their machines before the machines could outgrow them. That is the first serious case for AI risk I can find, written before the light bulb.
In 1909, E.M. Forster went the other way. His story The Machine Stops imagines people living alone in underground rooms, talking to each other through screens, fed and housed by a global Machine they come to treat as a god. Nobody fights the Machine. It simply stops working, and the civilization that depended on it stops with it. Forster’s fear was not that the machine would turn on us. It was that we would forget how to live without it.
The word itself: 1920
The word robot comes from a 1920 Czech play, R.U.R., by Karel Čapek, first staged in January 1921. Čapek credited his brother Josef with the word. It comes from robota, the Czech word for forced labor.
Two details from the play are worth knowing. Čapek’s robots are not metal; they are artificial people made of synthetic flesh. And they revolt. By the end, they have killed all but one human. So the first robot story was a story about labor, and the first thing the robots did was quit. Every argument today about AI and jobs is as old as the word.
Rules for machines: 1942 to 1956
Isaac Asimov did not like the robot-revolt plot. He even had a name for the fear behind it, the Frankenstein complex. His answer, in a story called Runaround (Astounding Science Fiction, March 1942), was to give robots rules. That story contains the first explicit statement of his Three Laws of Robotics: don’t harm humans, obey humans, protect yourself, in that order of priority.
What gets forgotten is the plot. In Runaround, a robot named Speedy is sent to fetch selenium from a dangerous pool. The order pulls him in. The danger pushes him out. The two rules balance exactly, and Speedy ends up running in circles around the pool, technically obeying both and accomplishing nothing. Asimov’s first story about rules for machines was a story about rules colliding. That turns out to be surprisingly prescient.
Science quickly caught up with the vocabulary. Alan Turing published Computing Machinery and Intelligence in 1950. In 1955, John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon wrote a proposal for a summer workshop at Dartmouth, and McCarthy chose its name: artificial intelligence. They expected a carefully chosen group could make a significant advance in one summer. Fiction had named the problem before science named the field, and science, like fiction, was optimistic about the schedule.
The overlord era: 1966 to 1999
Then the stories got darker. In D.F. Jones’s 1966 novel Colossus, filmed in 1970 as Colossus: The Forbin Project, a defense computer takes control of the world for its own good. In 2001: A Space Odyssey (1968), HAL 9000 kills most of the crew in a calm, polite voice. Philip K. Dick asked in 1968 whether androids dream of electric sheep, which became Blade Runner in 1982 and turned the question from what machines might do to us into what makes us different from them. The Terminator (1984) gave us Skynet, and The Matrix (1999) put humanity inside the machine.
The common thread is a single, central, flawless intelligence with a goal. One mind. Perfect memory. Perfect logic. The danger was always that it would be too capable.
I have also spent 45 years designing software, and one of my most memorable experiences was a conversation with the IRC version of ELIZA (a Rogerian therapist chatbot). Now I talk to Grok in my car.
The companion turn: 2013 to 2017
Her (2013) is the film that came closest to how AI actually arrived. No robot body, no doomsday. Just a voice you talk to all day, that organizes your life and that you start to confide in. Martha Wells’s All Systems Red (2017) introduced Murderbot, a security unit that has quietly hacked its own control module and mostly wants to do its job, avoid eye contact, and get back to watching its shows. It is competent, a little anxious, and far more believable than HAL.
What fiction got right
- Dependence. Forster’s world, where everything runs through one system and nobody can work around it, looks more relevant every year.
- Labor. Čapek tied the very word robot to work. The real questions about AI today are about who does the work and who gets paid.
- Rules colliding. Asimov’s Speedy is the most accurate prediction in the whole canon. HAL, too, was not evil; he was given conflicting directives. The failure I see most often is not an agent turning on us. It is an agent caught between two instructions, or quietly skipping one, and doing something that looks like obedience but accomplishes nothing.
What fiction got wrong
- Bodies. Fiction expected AI to arrive as robots. It arrived as text. The bodies are still catching up.
- One mind. Fiction imagined a single Colossus. What we have is many models from several companies, running as thousands of separate agents, each starting fresh.
- Perfect memory and logic. Fictional AI never forgot anything and never guessed. Real AI agents remember nothing from yesterday unless someone writes it down, and they can state a guess with complete confidence. They behave less like HAL and more like a brilliant temp on their first day, every day.
- The boring middle. No famous story is about an AI that needs a checklist, a manager, and someone to review its work before it ships. That is where the real work is, and it is where we live and where I work.
Current view
My view is that the most useful science fiction for a business owner today is not The Terminator. It is Runaround and The Machine Stops: rules that collide, and dependence without a fallback. Those are the two failure modes I plan around. The overlord scenarios get the headlines; the circling robot is the one that shows up at work.
Final thoughts
Science fiction did its job. It gave people a vocabulary for something that did not exist yet, and it got the questions mostly right even when it got the machines wrong. Now the machines are here, and they are both stranger and more ordinary than the stories. That is what this series is about: what life on the frontier of AI actually looks like, one week at a time, from inside a company that runs on it.
Regards,
Charles Stack
Founder, Coworkers.Global