This time around I decided on a guest column. The guest is a 37 years ago younger me, and this post is one that I wrote in for Manufacturing Engineering in March of 1988. It was for the last page of the trade magazine, in a regular feature titled “THE LAST WORD”. You can a download a pdf of the original from my MIT website, which has a pencil drawn picture of the 33 year old me (over half my life ago). Below is just the text of that piece, with few places that I have highlighted as being particularly relevant to today, or otherwise notable. There is also a footnote to give the basis for a couple of the weirder things that pointed out at the time.
AI: great expectations
Artificial intelligence (AI) has repeatedly inspired great expectations in people who see the possibilities of applying its techniques. Sometimes it delivers. Sometimes expectations are dashed.
The idea that machines can be rendered intelligent has always been seductive, and demonstrations of limited scope tend to raise greater expectations than hindsight analysis shows were warranted. In his 1949 book GIANT BRAINS or Machines That Think, Edmund Berkeley ponders the amazing ability of machines such as ENIAC carrying out 500 multiplications of two 10-digit numbers per second, and envisions machines that would act as automatic stenographers, translators, and psychiatrists.
This pattern is still evident. A few years ago there were high hopes that robots would revolutionize factories. In a way they have, but not in the grand manner predicted in the business plans of start-up companies six or eight years ago. Programming problems, combined with a lack of flexibility, made it impossible to overcome the systems-level problems of integrating assembly robots into the world of manufacturing. Great expectations raised by demonstrations of such robots glossed over other critical aspects of a complete operational enterprise–not the least of which is capital. Robots, indeed, have proven to be useful, but not as useful as was first predicted.
We recently entered the bust side of another set of boom expectations: expert systems. Though mildly successful in industrial applications, once again the expectations of the expert systems industry have not been borne out over time. The too simple representations of the problem domains of expert systems make them extremely brittle when the world diverges from the narrow range of applicability of their knowledge rules. Broadening those representations has been frustratingly difficult. Expert systems are here to stay–though not yet ready to solve all the world’s problems.
With every bust there is a new boom, and in the past year just what that new fashion will be has become clear—neural networks. These networks incorporate an appealing idea in that instead of having to work out all the details of a task we’ll simply let some randomly organized network of neuron models “learn” through trial and error how to do the right thing. Although neural networks have rarely accomplish anything beyond a computer simulation, business plans are being cranked out for new start-up companies to apply the technology.
But the current neural networks phenomenon is more than just another set of high expectations. This is the second time around for neural networks. It happened in the early ’60s. In 1962 a distinguished Stanford professor predicted that computer programming would be obsolete by 1963 because, by then, users would simply converse in English with the front-end neural networks. Since then, there have been a few technical improvements, and computers are much faster, broadening the scope of the applicability and likely successes of neural networks. But, again, they really can’t be expected to solve the world’s problems. The old-timers, “immunized” the first time around, seem less enamored than the new converts.
I recently worked with a group from industry, offering a detailed explanation of a technical AI method. After some time, the lead technical member of the group—who had no previous exposure to AI—exclaimed, “But that’s not intelligence! All you’re doing is writing a compete program to solve the problem.” Well folks, I’m sorry–but that’s all there is. There is no magic in AI. All we do is tackle areas and tasks that people previously were unable to write computer programs to handle. Because we have developed sets of tools and methodologies throughout the years to accomplish this, AI approaches have had a number of good successes. But there is no universal set of magic ideas.
Every so often a new AI development comes along and great excitement ensues as people stumble over themselves, convinced that the key to intelligence has been unlocked. Maybe it will happen someday, but I rather doubt it. I don’t think there is a single key to intelligence but rather that, unfortunately for both the philosophers and dreamers, intelligence is a vast, complex collection of simpler processes. To develop truly intelligent computers and robots, we’re going to have to unlock those processes one by one—causing flurries of great expectations, followed by more modest real successes. This may sound boring and unimaginative, but I find it exciting. Intelligence really is a complex interaction of many things. As we unlock its secrets in the next few years and decades, we will see a constant flow of ideas that have real and immediate practical applications. Finally when we truly understand AI, it won’t seem like just a computer program but will appear as a wondrous testament to the creative genius of evolution.
<signature: Rodney A Brooks>
Dr. Rodney A. Brooks Associate Professor Electrical Engineering and Computer Science Dept. Artificial Intelligence Laboratory Massachusetts Institute of Technology Cambridge, MA
A COmmentary 37 years on
When I recently reread this I was pleased to see how optimistic thirty three year old me was, while being completely aware of the amount of hype going on. The orange text in the last paragraph is full of optimism, and that optimism about AI was well founded as it turned out.
The three pieces of text in blue described the cyclic nature of hype, and the third one of them pointed out how each new set of hype drivers really do believe that that have cracked AI once and for all. Those traditions are alive and well, and woe to you should you challenge one of the true believers about the current two maelstroms of hype, generative AI and humanoid robots. I analyzed the sources of those whirlpools about three months ago with a previous blog post.
Finally, I highlighted two pieces of text in red. Here I can imagine people today saying the predictions that I used as examples of things that didn’t age well have in fact aged well 37 more years on. And they have. These two are good examples of one of my seven deadly sins of predicting the future of AI, which can be found on this blog or in a more tightly edited form in MIT Technology Review, both published in 2017.
The first of those seven sins was over and under estimating the power of AI, which is just an instance of a more general technology prediction sin, known as Amara’s Law.
We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.
The first of the items in red was where I said bundled Edmund Berkeley’s expectations for “machines that would act as stenographers, translators, and psychiatrists” with those that in hindsight seemed unwarranted. He made those predictions in 1949 based on the speed of multiplications in digital computers at the time. Today, in 2025, seventy five years later we have his first two predictions, but not his third. And a 10 cent microprocessor is about 100,000 times faster at multiplying than the large digital computers of the day, but it is not those machines that are stenographers or translators. It is machines that a further million times faster than today’s 10 cent microprocessor. He was directionally right, but without Moore’s Law, first elucidated sixteen years later, and certainly not dreamt of at that time, and without more than fifty years of AI research, he had no real way of knowing if or when such machines might come about.
The second was where I had said that an eminent Stanford Professor had predicted in 1962 that we would talk to neural nets and they would write programs for us. Again this person was directionally right, as that is one of the more consistently successful applications of LLMs today. But I thought then, in 1988, and do now that the prediction was at the time unwarranted. What was I talking about and what did he say? I was referring to Bernard (Bernie) Widrow who is now an emeritus Professor at Stanford. You can download the paper (original citation: Bernard Widrow. Generalization and information storage in networks of ADALINE neurons. In Marshall C. Yovits, Geogre T. Jacobi, and Gordon D. Goldstein, editors, Self-Organizing Systems, pages 435-461. Spartan, Washington, 1962) from his website here (many thanks to Leena Mathur for tracking this paper down). It is scanned from a printed copy and the scanning has pages rotated, so it takes a little patience to read it.
The ADALINE (from the paper title) was an analog implementation of a linear threshold perceptron (the ones he shows have 16 inputs, 16 weights, and a threshold, 17 weights in all) in an analog circuit made from components he calls “memistors” (not to be confused with today’s “memristors”) made from pencil leads. An ADALINE is one neuron from today’s neural networks of hundreds of billions of them.
Towards the end of the paper he reports on the speed up achieved by connecting 6 of these analog ADALINES (in a machine called MADALINE) to an IBM 1620, with a total of 102 (= 6 x 17) weights in all, compared to having the 1620 do all thoese multiplications itself. And after 8 months of operation and learning. they diagnosed that 25% of the weights were not being adapted due to solder failures, and sloppy manufacturing quality control in general. But, and this was the important point he says:
Yet the machine was able to adapt around its internal flaws and to be trained to make very complex pattern discriminations. These errors were corrected, and the capacity of the machine increased accordingly.
Then he says that they are going to expand to 49 input ADALINES, and have 1500 analog weights connected to the IBM 1620, in the next year (for a total of 1500 (= 30 x (49 +1)) weights, so I infer that means 30 of these 49 input perceptrons). Then the last paragraph of the paper is this one, which I am guessing is what I must have been referring to when I talked about the distinguished Stanford professor.
The fundamental objective in connecting adaptive neurons to a computer is to develop a new type of computer, one as different from the digital computer as the digital computer is different from the analog computer. This new type of machine might be called the Adaptive Computer. The basic “flip-flop” for this machine is the ADALINE. The adaptive computer is taught rather than programmed to solve problems. The job of the “programmer” is to establish suitable training examples. This machine will be taught by men (so that it will solve the problems of men) in the environment and with the language of men, not with machine language. The learning experience derived from human teachers will provide reasonable initial conditions, upon which the machine could subsequently improve from its own system experimentation and experience gathering.
I inferred, it seems, that the “language of men” was English. And I inferred that he was saying that the new set of 30 neurons was what would be use for conversing in English with the machine. In hindsight I plead guilty to overstating the things that I said he was overstating. However, I also think that as with Edmund Berkeley above, I was directionally right in saying that the claims were wildly overblown in that they required decades of research by thousands of people and needed more than a billion times as many weights as his proposed machine would have. And even today those LLMs are not continually adaptive as he suggested they would be.