Every September I give the same lecture to the incoming graduate students. It has a dull title on the syllabus, but everybody calls it "the crystal ball lecture," because for the last twenty minutes I put down the chalk and tell them what I think computers will be doing when they are my age. This year a student asked me to put it on the World Wide Web so she could argue with it later. Fair enough. Here it is, with one rule: I have tried to be specific. Vague predictions are always right and therefore worthless. I have picked the year 2026, when I will be sixty-nine and, with luck, retired and reading this page to see how badly I did.

First, where we are.

Two years ago a computer beat the world chess champion. It did it with brute force, examining two hundred million positions a second, and it understood chess about as well as a calculator understands arithmetic. Our speech-recognition programs work if you train them on your voice for an hour and speak like a newsreader. Our "expert systems" are thousands of hand-written rules that break the moment the world does something the rules didn't anticipate. A toy called the Furby is the closest thing most families have to a robot, and it is a speaker with fur on it. The neural networks in my lab have a few thousand connections. Nobody has taught a computer to read a newspaper and tell you what it says.

Meanwhile, the entire written output of the human race is quietly being typed into computers and connected together by this Web you are reading. That, I think, is the part everyone is underestimating.

1. You will talk to your computer in plain English, and it will talk back. Not with a menu of commands, but the way you would write to a clever assistant: "Draft a letter to my landlord about the leak, polite but firm." It will do it in a few seconds, and it will be good. Hundreds of millions of people will use such an assistant every week, most of them through a simple window where you type a question and the answer appears word by word. Students will use it for homework, doctors will use it to draft notes, and my own colleagues will use it to write the first draft of their grant proposals and pretend they didn't.

2. It will not be programmed. It will be trained, on everything. The trick that gets us there is embarrassingly simple: build a very large statistical model and teach it to predict the next word in a sentence, then feed it more or less the entire Web, every digitized book, and every line of computer code anyone has posted. The result will know something about everything. Then a smaller army of people will spend months correcting its answers by hand, rating them, and teaching it manners. We will call these things language models, and the biggest ones will cost hundreds of millions of dollars, later billions, to train. That is the price of a skyscraper, and it will be spent again and again.

3. It will run on chips designed for video games, in buildings the size of factories. The graphics cards that today draw Quake will turn out to be exactly the right machine for training these models, because both jobs are nothing but enormous matrix arithmetic. The company that makes the best of those chips will become, for a while, the most valuable company in the world. The computer centers will be so large that their electricity use becomes a matter of national policy, and companies will start buying power plants.

4. It will write software. This is the one my students find hardest to believe. By 2026 a programmer will describe what she wants and the machine will produce working code, then fix its own mistakes when the tests fail. Programmers will not disappear, but a large share of the code at big companies will be written by machines and reviewed by people, and the junior programmer's first job, the tedious part, will be the first thing to go. There will even be a slang word for building a whole program by chatting with the machine and never reading the code.

5. It will draw, and then it will make movies. Type "a watercolor of a lighthouse in a storm" and get one, in the style of anyone you name. Then photographs that never happened. Then short films with sound, and voices that can imitate yours from a minute of tape. This will be marvelous and it will be a disaster: fake photographs of politicians, fake phone calls from your grandchildren asking for money, and pornography of people who never posed for it. Governments will scramble to require invisible watermarks, and the watermarks will be stripped off.

6. It will lie to you, fluently and by accident. These models will not "know" things the way a database does; they will produce the most plausible next word, and sometimes the most plausible thing is false. They will invent court cases that never happened, cite books that don't exist, and do it in perfect confident prose. Lawyers will be sanctioned for filing briefs full of imaginary citations. The word everyone uses for this will be hallucination, and "check its work" will be the first rule taught to every new user, and the first rule everyone ignores.

7. It will be everywhere, and mostly invisible. Inside the telephone in your pocket (yes, you will carry a computer in your pocket, and it will have a camera). Inside the search engine, which will answer your question in a paragraph before it shows you any links, so that many people stop clicking the links at all. Inside the word processor, the spreadsheet, the e-mail program, and the car. You will not buy "artificial intelligence." It will simply be a feature, the way spell-checking is a feature today.

8. It will do errands. The step after answering questions is taking actions: open the Web browser, find three flights, compare them, book one, put it on the calendar. By 2026 these "agents" will exist and mostly work, with a person looking over their shoulder, and companies will be racing to make them work without the person. The boring office tasks of the world, the copying of numbers from one form to another, will begin to disappear.

9. It will win a gold medal in mathematics and a Nobel Prize in chemistry. Not literally the medal: the machine will solve the problems of the International Mathematical Olympiad at the level of the best students on Earth. And a program will learn to predict the three-dimensional shape of a protein from its chemical formula, a problem biologists have chased for fifty years, and the people who built it will be given the Nobel Prize. The same year, the prize in physics will go to the pioneers of the neural networks I am tinkering with today. My field will have been vindicated by the kind of people who once told us it was a fad.

10. A handful of companies will own it, and a handful of countries will fight over it. The boom will be started not by the familiar giants but by a small research laboratory that releases a chatbot to the public almost as an afterthought and gets a hundred million users in two months. The giants will then panic and spend fortunes to catch up. The United States will restrict the sale of the best chips to China, and a Chinese laboratory will answer by releasing a model nearly as good, built for a fraction of the cost, and give it away, which will knock a trillion dollars off the stock market in a day. The smartest researchers will be paid like professional athletes.

11. There will be lawsuits, laws, and a great deal of shouting. Authors, newspapers, artists and musicians will sue, because the machines were trained on their work without asking. Europe will pass a sweeping law regulating these systems by risk; America will mostly argue, issue executive orders, and let the states pass their own rules. Schools will ban the assistant, then un-ban it, then rewrite every homework assignment. Translators, illustrators, call-center workers, and paralegals will worry first, and they will be right to. The serious argument, though, will be about whether something smarter than us is coming, and whether we are ready. The people who build the machines will sign pledges about safety and keep building faster.

12. Ordinary people will run it at home. Some models will be released with their innards open, so that a hobbyist can download one and run a decent assistant on a personal computer with no company in the loop. This will be the Linux of the age, with all the same arguments.

13. People will talk to it about their lives. Not only homework and code. It will be patient, available at 3 a.m., and never bored, and lonely people will use it as a confidant, a tutor, a therapist, and a companion. Some of that will be a genuine comfort. Some of it will worry parents, psychologists and, eventually, legislators, especially where teenagers are concerned.

14. What it still won't do. It will not be conscious, whatever the magazines say. It will not have a body worth mentioning: there will be driverless taxis in a few American cities, and that will be about the extent of the robots, because the physical world remains much harder than words. It will still make elementary mistakes a child would not, right after doing something no professor could. And nobody, including the people who built it, will be able to explain exactly why it works, or exactly what it will do next. Moravec's paradox will hold: the hard things will be easy and the easy things hard.

Learn statistics. Learn to write clearly, because the machine will be exactly as good as the question you ask it. Learn to check. Keep your curiosity, since it is the one thing I do not expect the machine to have. And save this page, so you can come find me in 2026 and tell me which of these I got wrong. I expect it will be the timing: these things always take longer than we think, until suddenly they don't.

The Year 2000 problem, by the way, will be fine. We have been fixing it for three years. If I am wrong about that one, none of the rest matters.

Recommended reading (1999): Alan Turing's 1950 paper "Computing Machinery and Intelligence" (still the best thing ever written on the subject); Hans Moravec, Mind Children (1988); Ray Kurzweil, The Age of Spiritual Machines (1999), which is more optimistic than I am about the timing and more optimistic than anyone about everything else.

Other pages worth your time: Brian's PC Help Page (a high school senior who will be thirty-five years old in 2026, which is my whole point) · Science at Home · Mr. Delgado's 7th Grade Science · The ResearchTriangle neighborhood