Thank you for this opportunity to be here with as I know you're experiencing as I am this amazing gift of so many wonderful people and messages. Timothy asked me to reflect on the uncertainty engine theme, and the uncertainty engine beguiles. Now the original idea of an engine comes from a word that meant any contrivance that was created by human ingenuity. Now a little pun on engine, but I am an engineer by training, and so I have invented a lot of contrivances which I think will allow me to expand a little bit on many different ways that we might conceive of an engine. Most of them to an engineer have this idea of an input being acted on by an engine and creating some output. So of course we all are familiar with the automobile, right, with gasoline or electricity today going in and some engine creating turning movements and motion.
But what would be an uncertainty engine? Could it be that the input is something certain and the output is made uncertain? At Templeton, I imagine sometimes you feel this way when you put resources into something and you're uncertain what's going to happen.
This is a movie. Yep, there it goes. Where the human creator of the animation injected one of these with something that was designed to make its movement look a little bit more lifelike. How many of you think the one on the left looks more lifelike? And how many of you think the one on the right looks more lifelike? Seeing more hands going up at the right, but I also see a lot of uncertainty. Appropriate. All right. The one on the right has had the movement had extra uncertainty added to it using a kind of shaped noise we call it on the movement. That hereby got the name Perlin noise after Ken Perlin who won an Academy Award for creating this kind of noise that is widely used in animations today to give them more of the illusion of life.
Now I'm wearing a smartwatch that takes input that is continuously changing physiology and movement, runs an AI on board, and when this AI detects a certain combination of patterns, it can detect a potentially life-threatening seizure called a grand mal or generalized tonic seizure and issue an alert that goes through a paired smartphone to a caregiver. This process might be seen as an engine to reduce uncertainty. It's known now over actually about the last seven years that seizures are five times more deadly these grand mal seizures if you are alone when they happen. Bringing somebody there at the moment to reposition you can restore breathing and bring other needed first aid. So this reduction of uncertainty engine, if you will, continuously monitors for this possibly terrifying event that disables you and makes you unconscious so you can't call for help on your own and brings somebody to your rescue, hopefully brings somebody to your rescue.
The first time that I saw this actually in action in the wild was when a mom had had was testing an early version of the device with her daughter who was in bed sleeping wearing the device. Mom was in the shower and her phone was on the counter next to her when the phone went off saying her daughter might be having a seizure. So she interrupts her shower, runs to her daughter's bedroom, finds her face down in bed, blue and not breathing, flips her on her side, and as she repositions her, her daughter takes a breath and another breath, and her color was restored. And mom emailed me about this. I was reading my email, probably turning white. And as we spoke, she said, you know, I was saying, 'It's uncertain. It couldn't actually get you there.' Like, 'What if the battery broke? What if?' She says, 'I know no technology is perfect, but this got me there in time to save my daughter's life.' And wanted to send pictures back then of the little girl in the arms there and later as they're growing up and celebrating happy occasions together.
Kurt Eichenwald who has epilepsy and is an award-winning journalist and author says one of the biggest problems with epilepsy is the fear living with it every day and this uncertainty reducing device is a huge step in helping people combat the fear. You know you won't disappear. You know if you fall in the street you will be found. It's a massive step forward in helping people with epilepsy get liberated.
Now, most of my work focuses on reducing uncertainty, reducing stress, reducing fear, but the means by which we do that actually involves injecting uncertainty into our processes to improve them. And I develop a lot of different AI systems and AI is routinely developed with lots of use of uncertainty. Now I am not going to go into the details in this talk because when I do I immediately jump to a lot of equations and probability. So I deliberately shrunk this. So I know I'm terribly disappointing some of you who want me to go into all the math. We can do that over lunch or at a class at MIT. But the idea is that we use a lot of probability formulations to manage uncertainty and these particular ones actually come from a citation that OpenAI that created ChatGPT back when ChatGPT was open and they wrote about how they did things. This is a particular set of math they said was helpful to them in formulating it giving better responses.
Instead of the math, I want to talk about gaming because I think everybody here will get the ideas that we use inside the AI with math by looking at things we know in regular life. Like you get a deck of cards, it's too ordered, you have to shuffle it in order to deal fairly. We want fairness in our models as well. So similarly, we have different random shuffling ways to mix up the training data to initialize parameters in the model and also to draw from our test data to understand how the model's performing. In Scrabble, we also see randomly drawn letters that people rearrange to try to find a word that is worth the most points. And some of you may know, you may have a strategy where you just look at all of them and rearrange them in your mind. Some people find that by physically rearranging them, they come up with more ideas faster. It allows them to see and search the space and come up with a more high-valued word quickly. Similarly, in AI, we do a lot of search and exploration using randomness.
Now, I'm going to propose that we might also need a certainty engine, not just an uncertainty one. When you come up with a word in a game like Scrabble, unless you have a prodigious vocabulary, somebody is likely to say, 'I don't know if that's a real word.' They challenge it and the ultimate arbiter of the result is the certainty engine of a dictionary like the Scrabble dictionary. And these two words muzzjix and bazique happen to be extremely high value words that are legitimate words according to the certainty engine of the dictionary. With randomness and uncertainty, you might continue to propose other crazy words, just desperately trying to come up with something. And so you might propose this mashup of these two known words. But it turns out that's invalid.
So Tim asked me to reflect on does uncertainty aid creativity and yes, in many ways it does. It might be that this word is to become a neologism, a new word, right? We propose neologisms all the time in science when we need a new word. And this uncertainty aids us in coming up with the things that would violate the certainty engine.
I was also asked to reflect on AlphaFold. I don't know how many of you are familiar with this. It's one of the most famous examples of an AI being quite revolutionary in science. AlphaFold takes protein. It takes sequences of amino acids as an input and produces the 3D folded structure of the protein and the probabilities of its pieces being the way they are as an output. Now, some people want to think, hey, the AI is just hallucinating these randomly shuffling, and that's what some people say the AI won a Nobel, but that's not actually what's going on. The way that it's actually working is involving lots of careful knowledge from humans that have crafted a lot of specific processes highly oversimplified by the diagram in here that are not simply hallucinating an output. It's very much trained on examples from experiments and injected with knowledge of what these amino acids can and can't do. When the Nobel Prize was awarded for this amazing work, it did not go to the AI. It went rightly to the human makers who conceived of the ideas, designed the in this case, AlphaFold 2 and went through extensive work to craft and train and test and shape and reshape and iterate and make this so that it gave useful and meaningful and accurate and now revolutionary outputs.
So what was in there was not uncertainty being so creatively generate but aiding people in their creative generation. And I want to give an example maybe that's a little bit more directly tying the uncertainty to the creative generation. And I don't know how many of you have seen Refik Anadol's unsupervised work in MoMA Museum of Modern Art. I'll show a bit of it here. Here, Refik started by digitizing all of with permission of MoMA, all of the art in MoMA and then creating ways to guide using uncertainty through this art so that you can experience a new kind of art, you might say, inspired by these original artists. He also adds uncertain information from the environment, acoustic sensors, weather information, and some other infrared information from people in the gallery. The average person at MoMA who stares at Vincent van Gogh's Starry Night stays there 27 seconds. And the average amount of time a visitor spends gazing at unsupervised is 38 minutes. I'm only going to give you one minute of it here.
Refik now has just opened in LA a whole new building devoted to this experience of art from data and it's called Dataland. And here he is going beyond what he did with unsupervised. Actually now he's starting often with data from nature and communities that wish to share their data. And each person gets a completely different experience of the artwork because he's not just using the uncertainty of the environment. He wanted to get data from the individual themself and have you and your data shape this experience.
Now I wound up inadvertently being a part of this. And in fact this exact same wearable that was built to reduce uncertainty for people facing the fear of an unknown seizure is now being used in Dataland. You can get one if you go there to add your own human physiology and affective parameters to the art experience, your own shaping of it. Here he talks very briefly about this with or shows very briefly this being attached to what he calls the emotion your emotion. It's not reading your emotions, it's just reading a little physiology from you. So that's the same device and that is now being used to shape the art along with lots of other sensory information that is quite a full body experience. So you can now shape the art not only with your physiology but also your movements and information you're going through there.
All right, I'm going to wrap up with the end of uncertainty. I've mentioned several things such as how uncertainty is increased to aid the illusion of life, to aid fairness in a game, to speed up search, a certainty engine to test if something is valid. How we use reduction of uncertainty in a wearable to help reduce people's fear and stress by detecting seizures. How it's used generatively to create new kinds of art and also information from us to shape that art.
In closing, might there be a different end to uncertainty? These two images look uncertain, but the top one is actually completely determined by the random number generator. That is a piece of deterministic math. But most people can't see the difference between that and the one below that was randomly generated with atmospheric information. The second column are things like, you know, Ben Franklin's, you know, the only thing certain is death in Texas and, maybe the weather forecasters trying to do them up one with saying, hey, it was perfectly certain that it would rain that day or here, perfectly certain that it won't rain today. We know in part, however, even things that might appear to be certain are things that we know only in part. In his letter to the Corinthians, the Apostle Paul writes, 'Now we see things imperfectly, like puzzling reflections in a mirror, but then we will see everything with perfect clarity.' Now I know in part, then I shall know fully, even as I am fully known. Someday we may find that all uncertainty has a different kind of end. This end we cannot engineer and we cannot control but we humbly begin to comprehend. Thank you.
