Hi everyone! My name is Sarah Dean and I am assistant professor here at Cornell in Computer Science. I work in the fields of Machine Learning and Feedback Control on topics related to prediction and decision-making in the presence of dynamics. Two application areas that I work in are robotics and personalized recommendation systems.
Today's film is 2001: A Space Odyssey. Coming out in 1968, the film presents a futuristic vision of the turn of the millennium including a space journey orchestrated by HAL 9000 an Artificial Intelligence computer system that runs the ship and chats with the crew. Just a year after it came out, humans landed on the moon for the first time. But in the real 2001, we had neither voice assistants or autonomous vehicles - those arrived one and two decades later. Where is the technology today? Why did it take so long? And why was Stanley Kubrick able to foresee and tap into these ideas almost 60 years ago?
Part 1: History of AI
Let's start with a history of Artificial Intelligence, the field that (sometimes under different names) has given rise to these technologies. I actually want to trace this history to the 1940s and 50s, and argue that the intellectual foundations had been laid well before Kubrick made his film.
The first piece of this foundation is cybernetics, a term coined by Norbert Wiener who worked on anti-aircraft targeting systems during WWII. The key idea here is that of a feedback loop: measure the state of the world, take an action, observe the result, adjust and repeat. This simple idea is in some sense the skeleton of all autonomous systems built today.
The second piece of the foundation is information theory, a mathematical way to view language and information developed by Claude Shannon after working on secure communication systems during the war. The key idea is that for a machine, meaning need not be understood, but merely predicted. This is illustrated in a game the Claude and Mary, his wife, played together at home. Claude would pick up a book and turn to a page at random. Mary would guess letters, and when incorrect Claude would supply them to keep things going. Like a game of hangman, the goal was to require as few provided words as possible.
These prediction games are possible because there are statistical regularities in our language. Once you write down these regularities, all sorts of things become possible. Consider, for example, generating new sentences according to letter and word frequencies: Claude Shannon did just that in his famous 1948 report "A Mathematical Theory of Communication".
First-order approximation (symbols independent but with frequencies of English text).
OCRO HLI RGWR NMIELWIS EU LL NBNESEBYA TH EEI ALHENHTTPA OOBTTVA NAH BRL.
Second-order approximation (digram structure as in English).
ON IE ANTSOUTINYS ARE T INCTORE ST BE S DEAMY ACHIN D ILONASIVE TU-COOWE AT TEASONARE FUSO TIZIN ANDY TOBE SEACE CTISBE.
Third-order approximation (trigram structure as in English).
IN NO IST LAT WHEY CRATICT FROURE BIRS GROCID PONDENOME OF DEMONS-TURES OF THE REPTAGIN IS REGOACTIONA OF CRE.
First-order word approximation. Rather than continue with tetragram, . . . , $n$-gram structure it is easier and better to jump at this point to word units. Here words are chosen independently but with their appropriate frequencies.
REPRESENTING AND SPEEDILY IS AN GOOD APT OR COME CAN DIFFERENT NAT-URAL HERE HE THE A IN CAME THE TO OF TO EXPERT GRAY COME TO FURNISHES THE LINE MESSAGE HAD BE THESE.
Second-order word approximation. The word transition probabilities are correct but no further struc-ture is included.
THE HEAD AND IN FRONTAL ATTACK ON AN ENGLISH WRITER THAT THE CHAR-ACTER OF THIS POINT IS THEREFORE ANOTHER METHOD FOR THE LETTERS THAT THE TIME OF WHO EVER TOLD THE PROBLEM FOR AN UNEXPECTED.
This is the same idea that underlies the large language models powering today's AI chat systems.
The third and final piece of this foundation is the Perceptron, invented here at Cornell by Frank Rosenblatt in the late 1950s. It was the first computer system to learn to recognize patterns from examples, doing so by changing the strength of electrical connections between components. At the time it received breathless coverage from the New York Times - "will be able to walk, talk, see, write, reproduce itself and be conscious of its existence" - with rhetoric that reminds me of what we sometimes see today. The Perceptron didn't pan out, and the techniques fell out of favor for a couple of decades.
So, by 1968, the conceptual foundations had been laid: feedback, statistical prediction, and pattern learning.
Part 2: How AI Systems Work Today
So why did it take us so long to get where we are today? The short answer is scale. To make them practical, we needed to scale them up: more computing power, more memory, more data. It took several decades to get this scale - during which time we moved from IBM workstations and punch cards to PCs and laptops to smartphones.
The "AI" part of the systems that power ChatGPT and Waymo cars are basically made from two key components:
First, there is a large model, a.k.a. a deep network, also called a neural network. These models are direct descendants of the Perceptron, but with billions of parameters (a.k.a. weights or connections) rather than a few hundred. They are more complex, but they operate on the same principle: inputs go in, the model produces an output prediction, and the parameters are adjusted accordingly. This process is called "training". For language models, the input is text and the output is the next word.
Second, there is a massive amount of training data. In the case of large language models, we basically use the sum total of all written text - or at least all the text ever uploaded onto the internet. The first and largest phase of training is supervised prediction: predict the next word. Essentially, it is the Shannons' prediction game, but at an enormous scale. The second phase of training is a trial-and-error style mix of generation followed by verification. Much of the time, humans serve as the "verifier", for example, by rating the outputs of an AI chat assistant. This stage tends to have proportionally less data, and its goal is to tune the statistically plausible patterns learned in the first phase towards more useful outputs.
The same story roughly holds for autonomous driving. In this case, the inputs are camera and sensor data while the outputs are steering and braking decisions. Training data comes from fleets operating first with human "supervisors" and then autonomously (at least in SF and AZ).
Part 3: Issues That Remain
I don't want to make it seem like we've perfectly reached the future vision of 2001: A Space Odyssey, or indeed that it is a vision for a perfect future.
First, let me highlight three important issues that are irrevocably entangled with the story of technological progress that I just told you. There is the question of energy and infrastructure. Scale may have been necessary, but it is not free, and the data centers running these systems consume huge amounts of electricity. There is also the question of data and surveillance. The training data for these models comes from wherever the developers can find it - often from people who did not know about or explicitly consent to this use. And then there is the question of labor: humans who do micro-tasks like labeling, feedback, and intervention under opaque conditions and low wages. I don't think Kubrick anticipated the importance of scale or any of these issues that come along with it.
I want to end with a final question: do our AI systems actually work? Many public debates about these technologies suffer from "the fallacy of functionality". Compare AI to commercial aviation, another inheritor of cybernetics. Aviation gets five nines of reliability: 99.999% of the time there is no accident. Machine learning systems, in terms of accuracy, do well to get one eight (80%).
One reason for this gap is that AI is unlike aviation, where systems are built up from components that can be inspected and certified. Instead, the deep models (our modern Perceptrons) are black boxes. Rather than carefully designed, they are trained by optimization of various metrics - e.g. maximizing accuracy. This process works extremely well! But computational optimization is like a genie or a toddler: it finds ways to score well on metrics, sometimes in ways that completely miss the point.
Here is an example that I teach in my reinforcement learning class: an agent learning by trial and error to win this boat race video game. It turns out that one can gain more points by repeatedly doing donuts than by actually winning the race:
People often refer to this as the "alignment problem". If we must use the language of optimization to design our systems, how can we make sure the ultimate behavior is aligned to what we meant, not only what we measured? Kubrick anticipates this problem with HAL 9000, as you will all soon see. Enjoy the film!
I was asked to give this talk in place of Joe Halpern, a legend in AI (and other areas of CS), who passed away in February. I was lucky to know him during my first years at Cornell.