Now the second area that I think is really important is what we call bias for action. This was probably one of the hardest things that I had to learn coming from other companies, which is, speed matters. You're always here, you have to move fast, innovation is fast. But some decisions are reversible and some are not. You really don't want to go fast all the time, you want to really think about how you're taking that calculated risk. We call this the one-way door or the two-way door. A two-way door is if you try something, let's say changing colors on our website, you can change it to red, and if your customers don't like it, you can change it to blue, there's really not a lot of impact in that. But a one-way door would be, let's say you're going to change your API for a customer, that's a big deal, because every customer has to change it. You want to be very thoughtful, slow, methodical, and doing that decision. This is true as well in machine learning. Let me give you a couple of examples. This is a startup, based out of San Francisco, 360Fashion. What they do is they use a machine learning and IoT to gather data for a whole set of products and portfolio and fashion tech. You can see Anita, she's the CEO, in the dress, and what she's got, she has machine learning in that dress that learns her gait, her turns, what signals that she's talking to someone. What it does is it enables the lights to change based on learning her patterns, and her motions, so that might calm down if she's talking to someone, or if she's walking, it might blink really fast. Now, the first thing that she did, was she tried connecting this up to the heartbeat, using an IoT sensor to the heartbeat. But what ended up happening, is as people were wearing it, when their heart beat fast if they were nervous, everybody knew. That was a two-way door, because she could easily change that decision. Even though she was looking at AI and machine learning, that was a decision that she could change, two-way door. But as you're thinking about your solutions, there will also be one-way doors. This is Redfin, company based out of Seattle that has really taken real estate online. Bridget Frey, their CTO, she actually spoke here last year. She's doing some really amazing stuff with machine learning. They had this idea, talking to their customers, working backwards of putting up an estimate of your house's value. They were hiring data scientists rushing to get this out the door, and then they paused and they said, this is really a one-way door. Because if we put out an estimate of the value of a house and we say, your house is worth a million dollars on Saturday, and the next Saturday your house is worth $200,000, that's a big difference. We're not going to earn the trust that we need from our customers. They step back, they involve the real estate agents as well as the buyers and the sellers. They found that sellers will talk to the real estate agent about this, so they didn't want machine learning to be in a black box. They were able to reveal how they came up with the value of the house. This estimator now is done using machine learning and AI. It is deemed the most reliable estimate of your house's value. In fact, some banks are considering using this instead of getting the appraisal that you have to pay for using something like this to do the appraisal of your house. Thank goodness, they stepped back and they said this is not about getting first-to-market, unlike 360Fashion, which wanted to come out with that beautiful AI robotic dress. This is something we need to be thoughtful about and really take our time. Does that make sense to you guys as you're thinking through your projects? Okay, good. Thank you. For Redfin, they really engage the human side. They made the data available so the real estate agent could explain to the customer what was going on in that machine learning algorithm. This is becoming more and more important that people don't want just the answer, they want to see how you came up with the answer as well. Advice for how you might want to do this. The first thing you want to do, is you want to think about how you might build these templates and build these models. You need to look at, can you get some pre-built templates based on what you're trying to solve, if it's a customer service problem like a ChatBot for instance. You have to label your data, remember we talked about if you have an apple, you want to label as an apple, so you know the insight, and then look at how you're building those machine learning algorithms. Next, you have to do your training piece. This piece really helps you to figure out how to optimize that training. Even though it's not a lot of the work, inference is more of the work you really need to figure out how to optimize that, and then finally how to get that deployed. Funny story about Redfin. This helps you to figure out how to auto-scale and how to make sure that you're not consuming too much compute. When Redfin first turned on their estimator, they did not have some of the things turn onto optimize it. They were just running, consuming, a lot of infrastructure. This can help you figure out how to do that best area of deployment as well. There are other ways too that you can figure out how to teach and how to learn. I don't know if you guys saw our announcement of DeepRacer. This is really using reinforcement learning instead of just labeling. This is a little miniature car, it's got sensors on it to collect data. It has two batteries, one for the machine learning and one for the compute, one to power the car. What we're doing is we're setting up leagues. The leagues are really about not just competing against each other, but also figuring out how do you learn machine learning in a fun way. We're hoping that kids, boys and girls and all of you will play with something like this, and then learn some of the mechanics so that machine learning isn't so hard to consume.
