The Warby Parker of… Are we still in Web 2.0? Or have we made the jump to 3.0? Let’s just compromise and call it Web 2.74(b). Regardless, one of the defining characteristics of our current digital moment is the slow infusion of web-first business strategies to different industries. Tesla is a web-first car company. Caspar is a web-first mattress company. Bowery is a web-first arugula company. In industry parlance, these are called Digitally Native Vertical Brands – meaning that they handle all aspect of their business from production to marketing to distribution to customer service themselves and they make effective use of digital platforms. Matt Heiman, writing for TechCrunch this week, has outlined the factors that lead to a successful DNVB. You need friction to exist in the market that cannot be resolved by the incumbent brands. You need there to be a unique advantage that results from digital, be it marketing, distribution or sales. You need either a frequent purchase cycle or a high average order value. And you need high margins for the product, either for the manufacturer or the retailer, preferably both. It helps if your product is likely to be shared through social media or has the propensity to gain a lot of earned media, either through its values or celebrity involvement. It also helps if the product is timeless, meaning that it isn’t subjected to the whims of changing styles. Naturally, it is possible to build a successful DNVB without combining all of these factors. In fact, all of the most prominent DNVB’s lack one or more of these characteristics. Why does this matter? I see a lot to like in Heiman’s list. One thing I particularly appreciate is that he does not minimize the grinding difficulty of launching a DNVB successfully. This is not a matter of simple efficiency. These companies have to rebuild every aspect of an industry to meet changing consumer habits. They cannot rely on legacy marketing, legacy distribution or legacy retailers. If they do, they have left themselves open to further disintermediation by a more nimble competitor. However, I think the Warby Parker example which everyone uses to talk about DNVB’s (including me) is actually a bit misleading. Warby Parker worked because of their manufacturing relationships. Not just their web-based vertical integration. All successful DNVB’s win because they possess a technical advantage over their competitors that is not reducible to simply being “web-first.” These are the same type of advantages that have always differentiated successful companies. Either their product is better or they can manufacture for less or they have identified an unmet need. The reason we have been hearing more about DNVB’s is that the internet allows them to go around the traditional middlemen to market and sell their (inherently better) product directly to consumers. The only thing I would add to Heiman’s list is that the company must have such obvious advantages that it could have been successful even without the internet. Digital integration just allows their success to manifest itself that much more quickly. In a nutshell: Better companies win. The Internet just helps them to win sooner. Read More Winning the AI Arms Race While I remain extremely impressed with IBM’s approach to neural networks, I am starting to suspect that their rush to productize their technology may lead them to fall behind their competitors. During the early days of machine learning, they looked to have a decisive advantage. But Google DeepMind’s focus on research (versus sales) seems to be paying off. Their recent announcement of a Go-playing algorithm capable of routinely beating their old, world champion Go-playing algorithm seems to suggest that they are playing a solo game of technological leapfrog. DeepMind continues this rapid cycle of innovation with the development of Population Based Training of neural networks (link below.) No matter what task a neural network is set to addressing, researchers need to first determine a set of hyperparameters that will guide the process. Hyperparameters include what type of network to use and what type of data or methods should be used to teach it. Previously, there were two ways to set the hyperparameters. One way, random search, required that a number of neural networks be trained in parallel, each trying different models or using different data. At the end of this process, the highest performing model was adopted. While this model had the advantage of settling on the most effective available method, it wasted a lot of resources and time getting to it. The other method was hand tuning. This involved having (human) researchers guess at the right set of hyperparameters based on goals they had in mind. While this wasted fewer resources than random search, the iterative nature of hand tuning took time and risked the adoption of a local optimum based on human biases. DeepMind was able to improve on their Go-playing algorithm when they stopped training it against human players. Training against human players made a neural network slightly better than humans. Training against other go-playing neural networks made a neural network that could play on a whole new level. (Post-human – if you’ll forgive the apocalyptic overtones.) Population Based Training of neural networks is superior to both of the established methods because it uses the random search approach of testing multiple methods, but it shares its results as it goes along. As a certain set of hyperparameters shows promise, the other neural networks pick up and iterate off these hyperparameters. This allows all neural networks to arrive at the best available set of hyperparameters faster and wastes fewer resources and less time in the process. Why does this matter? I know, I know, you think I’m in the weeds here. But neural networks are being applied to some of the most difficult challenges in business – equity-trading, distribution networks, healthcare. If your company’s neural network is using a sub-optimal set of hyperparameters, your company is vulnerable to any competitor who has done a better job training their network. Because neural networks exceed human performance in certain areas, it can be difficult for companies to recognize when they have only achieved a local optimum; that is, until a competitor comes to eat their lunch. Already in financial services, an AI arms race exists. But many companies in different industries labor under the delusion that they have “a neural network” and that is enough. Machine learning is not a one-time purchase. It is an ongoing commitment to improvement. In a nutshell: Neural networks are improving and iterating faster than the iPhone in your pocket. Read More Digital Privacy – Still an Oxymoron Dan Goodin who writes on security and privacy issues for Ars Technica published an article last week on session replay scripts. Session replay scripts are provided by analytics companies. They allow a website administrator to watch the actions that a user takes on their website during the entirety of a user session. Essentially, these are recordings of user behavior on the website. If the user clicks on pages or runs their mouse over certain content, it is captured. If the user begins to input their data into a checkout and then abandons their cart, this is captured. Goodin clearly views this as a security risk. After all, session replay scripts capture inputted credit card data and other personal information, even if the user later decides not to complete the purchase. Why does this matter? The internet is not a broadcast medium. Don’t think of time you spend online as the equivalent of watching television. Think of it as walking into a store. In a store, you might browse a bit, pick up a few items and then put them back down, maybe you make a purchase, maybe you change your mind at the last minute. But you wouldn’t be shocked to learn that a salesperson was watching what you were doing. That’s their job. You should expect the same thing online. A company that depends on online transactions is being irresponsible if it does not capture session replay scripts. Successful user experiences require constant optimization based on the actions of actual users. If people are abandoning their cart after they input their credit card information that is a crucial piece of data that requires immediate action. You have no inherent right to privacy while you browse at IKEA and you have no inherent right to privacy while you browse on ikea.com. Yes, this type of session data should be carefully protected so people’s credit card numbers aren’t stolen. But the companies aren’t violating your privacy when you go to their website and start clicking around. In a nutshell: Going to a company’s website is like going into their store. Read More