Artificial intelligence is not a new technology. We've been doing R&D in AI for over 60 years. And just in the last two or three years, it's exploded into public consciousness. So much so that you need augmentation or stopaction photography in order to resolve the blur of events. Manufacturing is fundamentally about manipulating atoms and molecules, and intelligence drives manufacturing. Whether you're chipping flint or bone or soldering or welding or 3D printing, it's all about improving our knowledge of how to manipulate atoms and molecules.
Here are hundred startups in AI in virtually every field of human endeavor. Robotics, cyber security, automotive, fintech, commerce, internet of things, business intelligence, bots for interacting with people. And yet most of these companies didn't exist 3 years ago. This field is moving very fast. People often ask for definitions of AI. Here are a few of them. Pattern recognition techniques to solve practical business and technical application problems.
Software agents that if given some resources and a well-defined set of tasks can solve those tasks in the minimum amount of time that they have available. a vision of superhuman intelligence that hasn't happened yet, even though movies and science fiction uh stories tend to blur that line. And computer science that's accelerating all of the other exponential technologies that you'll be hearing about over the next few days. Synthetic biology, robotics, nanotech, and AI.
I was on a plane recently and a guy asked me, "How many head do you have?" He said, "Oh, you're an AI. I'm an AI, too." And clearly he meant artificial insemination. And there is a relationship between AI and artificial insemination because you can select the genes for dairy cattle or beef cattle. Hardware and software tend to get most of the attention in AI. But what you'll see when you look behind the scenes of most AI startups and even research labs is an emerging symbiosis between human intelligence and machine intelligence.
Behind the scenes, you'll see something like this. People interacting with machine intelligence with great intensity. And already in 2017, we've seen that intensity turn into results. This is Labratus that came out of Thomas Sanholm and Norm Brown's lab at Carnegie Melon. Labatus plays no limit Texas Holde poker. And this system played 120,000 hands against four world champion poker players. and it ended up with $1.77 million worth of chips. It was a very non-random result.
And why should you care? Because this system can handle bluffing and misinformation and create strategy to win. Ken Forbis at Northwestern has been working on a system called Cogs Sketch for about 30 years. And just recently, 2017, in January, he announced that his system could solve the Raven Progressive Matrices test, a test of analogical and visual intelligence, better than the average American. Take a look at this. What should go in that bottom corner cell out of these eight alternatives?
Anyone? Which alternative here? A through H D. Right. Because it matches the symmetry in the row in the column where that empty cell is. Amazon Alexa. The agent running on the Echo speaker connected speaker from Amazon was the surprise gift in the Christmas season of 2016. Look at the growth of Amazon Alexa skills. Q1 2016, 135 skills. Q1 2017, 10,000 skills. These are things that you can speak to this artificial intelligent agent and it can do things for you.
And we're now building our AIS into our machines. In many cases, big machines like cars. Alexa is now being built into Ford's Focus and Ford's Fusion as an option. The historical trends in materials and manufacturing are absolutely clear. Increasing precision eventually leading to atomic precision. We're not quite there yet. Decreasing cost, increasing flexibility, increasing velocity. And complexity is free. Why is complexity free? Because a 3D printer or other similar robot for doing manufacturing doesn't care how complex the CAD or computerated design file is that you send it.
It just operates with the feed stock that it has. Doesn't cost more. And even a master manufacturing company, a semiconductor manufacturing company like Intel is realizing this. They recently acquired Nirvana, an AI company. And Naveen Ralph, who was CEO of Nirvana, has now been made the the head of an AI products group at Intel because they want to drive intelligence into every product and process that they have. AI contributes to manufacturing in a variety of ways.
Faster actions and decisions, innovative products, services, and materials. Not just better, faster, cheaper, but different, higher precision, lower cost, exponential scale, because you can now drive intelligence to the edges of your organization. You know about IBM's Watson beating the two world champions in Jeopardy way back in ancient history, 2011. But now IBM is driving Watson into virtually every area of business. They're still looking for new partners.
And they have a Watson Explorer expressly designed for advice in manufacturing. General Electric is using big data and AI to make smarter turbines. And by the way, they've been at this for 30 years. They had an AI group 30 years ago. I visited them. And their AI group at that time was working on predictive maintenance for locomotive engines. And here's Christina Valyaki with a editorial in manufacturing automation all about machine learning being the key to sustainable manufacturing.
Why is that? Because she was suggesting that you could use machine learning to select molecules in high throughput experiments, molecules that are less toxic for the environment and less toxic for human health. and we can still have them perform in their manufacturing roles. And I'll show you an example of high throughput experimentation. I want to give you a framework for thinking about the AI revolution to really wrap your minds around it. And the framework involves seven steps.
Capital, algorithms, hardware, data, talent, applications, and responsibility. And responsibility is not just a throwaway item. you'll see that we take it seriously. Let's look at capital. 2012, 589 million invested in AI. 2016, 658 deals and over $5 billion of investment. Algorithms are also a key. Deep learning is the champion. Machine learning algorithm and its cousins, convolutional neural networks, recurrent neural networks, and others. Let me just show you a little bit about how this works.
It's a hierarchical pattern recognition system. At the first stage of the hierarchy, you have a set of pixels like in a photograph. Here you have edge detection. So it's detecting vertical lines, horizontal lines, diagonal lines. The next stage, object parts. So the parts are ears, nose, mouth, eyes, etc. And the third stage, the objects themselves, faces. And three layers of hierarchy is not enough to differentiate all of your beautiful faces. But we can build hierarchies with many levels.
Microsoft announced 152 layer hierarchy. You don't need anywhere near that much to do a great job on facial pattern recognition. And here is a article that just came out very recently from the deep mind group at Alphabet. And this system was about something more than just recognizing objects. Recognizing the spoon and the bowl and the hamburger and the Advil and the Coca-Cola. This was about being able to pattern recognize how you would have to move a robot hand to grasp these objects.
So, a neutral wrist for picking up a Coke can, a pronated wrist for picking up an apple, a tripod kind of configuration of one's hands for picking up a toothpaste tube, and a pinch for picking up something like a pen or a knife. Now, sometimes people will tell you, "Don't try this at home. It's dangerous." I'm encouraging you all to try AI and machine learning at home. You can download the R toolkit for free right into your laptop. You don't need specialized hardware and you can test this out.
And if you don't have the chops, the programming chops to do it yourself, you can invite a high school student to help you. R has got 1,800 extensions to the standard toolkit. It's very, very powerful and free. You can also use a platform like sentient technologies. They have 130 million of investment. Some of the people from the Apple Siri project, Adam Cher and others are involved in this and they are using deep learning at massive scale and also evolutionary learning.
They grow hedge fund agents and pit them against each other in a kind of simulated evolution of hedge fund agents in order to be very successful at making bets. And they also have recommendation engines that they simulate evolution and build ever better recommendation engines that way. IBM is now labeling their Watsonfueled business cognitive businesses and they're encouraging a wide variety of interaction there. Check them out. And what I'm I believe in being honest about these technologies and there are problems.
I don't expect you to read this. I just want to show you a very recent research paper that I think went largely unheralded. This is a graph from some work that was done at Deep Mind at Alphabet where they addressed a really serious problem with machine learning. If you learn task A with deep learning, that's great. If you then use that same network to learn task B, you get what's called catastrophic forgetting. So, it forgets how to do A when it learns how to do B. and they developed a system that would preserve the knowledge of the weights in the network of task A when it learned task B.
So you can see with no penalty for losing that knowledge, you get fairly low performance on this graph. And then you get much higher performance if you preserve the knowledge of how to do the prior task. And it turns out this works so that you can do more than just two. You can keep going. So this is a very big deal. It's sort of most people didn't notice. I think it's a real breakthrough in dealing with the catastrophic forgetting problem. There's another problem with machine learning and that's explanation.
Machine learning tends to be a black box. So yes, it can solve hard problems, but if you're a doctor and the thing tells you to uh provide a certain treatment recommendation to a patient and the patient says why? And you say, well, I don't know. The machine told me to do it. That's not acceptable. So, we're going to have to have good explanations. And there's now a program at DARPA, the Defense Advanced Research Projects Agency, uh, on XAI. It's run by Dave Gunning.
And it's all about opening up these black boxes so that we can provide highquality explanations. And there may be a performance hit associated with that because you have to add additional machinery to do it. Hardware is another key aspect of the AI revolution. You may recognize this device which was in stealth mode until about spring of 2016. How many of you used Google search in the last two weeks? Raise your hands. You used this machine. It's a TPU, a tensor processing unit.
It's optimized for running deep learning and other machine learning algorithms. In fact, it's an order of magnitude more efficiency, has more efficiency in computation than the GPUs and CPUs they were using before, and it plugs right into the Google data rack. Alphabet has competition in that arena. So, the Snapdragon 835 from Qualcomm runs deep learning in its native form and it even comes with a software development kit. So you can embed that Snapdragon chip in your products.
Data is considered gold or electricity in the AI world. And this company, Data Miner, has over 130 million of investment. And it can vacuum up data from all over the web. 500 million tweets per day, medical data, geographic data, weather data, financial data. It puts it through a set of machine learning algorithms and then people who subscribe to particular kinds of data, say the latest in genomics data, can get a vector of information aimed right at them and their interests.
TensorFlow is another key algorithm uh in machine learning. This was developed by Alphabet. They invested tens of millions of dollars in it and then they released it for free. So for those of you who are looking for a very powerful machine learning uh toolkit, use this. It allows you to represent your problem like a data flow graph where the nodes of the graph are algorithms and the links between the nodes are multi-dimensional data arrays. And that allows you to distribute your computation across a broad fabric of hardware.
So this merges or has a convergence effect with those TPUs that I mentioned to you before. Talent is another key area of the AI revolution. Mark Zuckerberg hired this gentleman, Yan Lun, who's an expert in convolutional neural networks. Jan is heading up the London, Menllo Park and New York labs of Facebook and they had a recent success in achieving human level and beyond accuracy in facial pattern recognition with their deep face 2 algorithm. This gentleman Andrew Ing was chief scientist at BYU.
They hired him from Corsera and Stamford about three years ago and then just in March Andrew announced that he's leaving BYU. So it's very hard to hold on to your AI talent uh and uh they really need to be courted in order to keep them at your organization. Here is a company called Viv. They have a bold vision for the future. AI and intelligence becomes a utility. A utility like Wi-Fi or Bluetooth. A utility that you can delegate work to, not just a tool.
And here is a seminal demo from DeepMind that I've mentioned a couple times. This was from 2013. Deep Mind created an Atari game player that started with zero knowledge of these Atari 2600 games emulated in software. They used a convolutional neural network which is just a fancy word for a powerful pattern recognizer and reinforcement learning which reinforces reinforces an agent for a high score in a game. Take a look at this. This system is starting out with zero knowledge of the game.
This is Deis Sabis from Deep Mind. You'll see the system missed that ball and it's learning but it's playing very poorly. After 120 minutes of training, it'll play at a mediocre human level. And you can see the mediocre human score there. But after 240 minutes of playing, the system will develop the strategy of burrowing up the left hand side of the screen and playing from the back court where the game has no defenses. Take a look at the score. So that's a breakthrough not just in game playing, but in AI and people sometimes describe AI as a gamechanging technology.
But it's more than that. AI is altering the landscape of competition for all of us. Fundamentally disrupting the landscape. Most of you have heard about the win of Alphabet's Alph Go player over the world champion in Go, the Korean player Lee Sadole. This is really an amazing event uh that took place in March of 2016. Amazing because if you compare chess to go, chess has an average number of moves per turn of 35. Go has an average number of moves per turn that are alternative moves 200 to 250.
So it's epically complex in its fan out. If you want to think ahead, you have to think with a fan out of 200 to 250. This gentleman had amazing grace under pressure. uh after losing three games in a row on move 78 of game four, he came back with a one in 10,000 move and won that game. Uh and now Go is more popular than than ever. He credited uh Alph Go with teaching him how to do that. And South Korea after this game promised to invest $3 billion for AI R&D after Alph Go shock.
You may say we don't have the bench strength in our company to do this. You can use Kaggle and put up prize money uh and your data and your problem and you can get two or 30 hundred of the best data science teams to solve your problem for you. This actually works. data science competitions or here is uh the fact is it's been so successful that Google cloud bought Kaggle and you might think oh there goes a resource no it's still available to you so Google just wanted a deeper relationship with them you can still use Kaggle on the web and if you don't want to put your data out in public you can use expertify.com and just put up your problem and get curated data scientists to bid on the solution to your problem.
It's very successful. In the spirit of full disclosure, I'm an investor in this company. Came out of the Harvard X Innovation Lab. So, applications are driving the AI revolution. I've added some companies to this. Uh Tesla, Amazon, Nvidia, Uber, and Salesforce are new players in AI R&D. They've been doing it for a while, but they're now amping it up. I did a study of over 360 innovative applications of AI at the IAAI website. I was looking for patterns.
These are the patterns that I found. And it's not just better, faster, cheaper. It's different. Expanding the range of what's possible, doing things we never knew we could do before. And you might say, well, where do we get started? And the answer is in any of these areas, design, diagnosis, manufacturing and management, customer service, sales and configuration, or quality. Here's an example of using this was the winning paper at the innovative applications of AI conference.
I don't expect you to read this in real time, but let's zoom in on this diagram. They were running at the department of energy high throughput experiments for understanding where in this phase map liquid crystals are for photovoltaic arrays. They wanted to optimize those arrays and they were using humans and the humans couldn't keep up with the high throughput experiment. They added an AI solver and now you get the best of both kinds of intelligence.
This is a very generalpurpose architecture for doing high throughput experiments. GE has the predict system that allows you to spread a broad fabric of computation across a traditional factory that allows the factory to be more predictive and responsive and connected. Rolls-Royce is using AI to leverage their manufacturing. They're calling it industry 4.0. They are manufacturing both automotive engines this way and aircraft engines. Some of you may be worried that this costs too much, but here's the Watson virtual agent for doing customer service, for example, a bot. $265 a month.
So AI is here now, not with how from 2001, but with practical business and technical problem solving. I'm encouraging all of you to build an applicationoriented AI toolbox. There are lots of free and other tools available. Let's talk about responsibility very briefly. AI is going to impact human resources. Yes, it's going to augment teams and enable increase precision and productivity. It will transform business models. It will alter the world of work just like that soccer field I showed you.
So, I think that means that we need to consider basic income models, a variety of different models. We need to experiment with them proactively and it means we ought to give everyone access to free education because yes AI and robotics will create new jobs but they may destroy more jobs. What matters is the ratio and that will change over time and there's lots of uncertainty about this and rather than be uncertain and unprepared I think it's better to be uncertain and wellprepared and proactive.
Trust is a big issue going forward. This is the adult conversation that AI comes with tradeoffs. Faster, cheaper, better, different problem solving and job disruption, human identity change and risk amplification and some risk reduction, but it's always tradeoffs. Elon Musk uh tweeted about the dangers of AI in 2014, but to his credit, he was with us at a future of AI conference in San Juan, Puerto Rico in January of 2015. We talked about the need to do R&D and verification, validity, security, and control.
There was another conference also run by the future of life organization that took place in Ayamar of January of this year and they came up with 23 Ayamar principles for making AI safer and more beneficial. Elon Musk has been working with Sam Alman on a new organization called Open AI. Uh they have a variety of reports out. They have an open AI gym. Uh Sarah later on will show you a video that they've created on a system that you can use. And there's a new organization called Partnership on AI to benefit people and society.
Lots of organizations are joining organizations that are deeply involved. Teaming with AI strengths means best-in-class performance for some tasks, fewer biases in decision-making, super fast operational velocity. You get a partner that works seven days a week, 24 hours a day. No vacation or sick leave, no health care required, no whining, but limited empathy, language understanding, and social grace. We're encouraging you to build the future boldly and to do it responsibly.
Here, if you want to know what you can do on Monday to move this forward in every one of the areas I mentioned to you, here's some operational advice. If you have small amounts of capital, you can innovate. Large amounts of capital, you can acquire AI companies. There are lots of algorithms available both for free and at the platform level. You have a wide choice of hardware available. On data, you don't have to have big data. You need to marshall relevant data.
On talent, you don't have to have all the AI talent inside your company. You can crowdsource your talent from Kaggle or Experify. And on applications, don't just rebuild and repave the cow paths that you have. Use AI to do fundamentally new things. On the security front, consider secure on on the responsibility front, consider security, empathy, ethics, policy, and eventually liability. It's important to have some sense of perspective about this. not everything's going to be great and we'll all be nirvana or it's all gloom and doom and and you know we're finished, but rather real perspective.
And this photograph that was taken at the edge of our solar system by a NASA probe is a picture of Earth. That little dot inside the ring, that's what Carl Sean called the pale blue dot of Earth. That's where all of us live. That's where everyone that you have ever loved lives. That's where everyone you've ever read about in human history has lived. And it should be obvious by now that we are all in this together. And we damn well better act like it. Thank you very much.
This transcript was generated automatically from the
video's captions and may contain errors.
