Enterprise AI: What's It Really Good For?

Editor's Notes

• #3: If you’re a subscriber to the O’Reilly online learning platform, you can take live training course about AI for business,
• #4: Explore thousands of hours of video training
• #5: read the bestselling technical books on the subject
• #6: Or look at our surveys about AI adoption in the enterprise.
• #8: With our newest feature, O’Reilly Answers, your engineers can get immediate answers to technical questions posed in plain language
• #9: With our newest feature, O’Reilly Answers, your engineers can get immediate answers to technical questions posed in plain language
• #10: And your business executives and sales people can even get quick answers to “what is?” type questions so they can understand what the hell your engineers are talking about!
• #11: But in this talk, I’m going to focus on the big picture, and some general advice about how to think about applying AI to any business.
• #12: Paul R. Cohen, a former DARPA programming manager who became Dean of a new school of Information Sciences at the University of Pittsburgh, put it beautifully at a meeting of the National Academies, where we were both speaking about the future of AI. He said, “The opportunity for AI is to help humans model and manage complex interacting systems.” These vast algorithmic tools let us do things that were previously impossible. Google gives searchable access to trillions of documents – it’s not quite “access to all the world’s information,” but it’s the closest thing we’ve seen. Facebook connects billions of people. Uber and Lyft have put millions of people to work providing on-demand transportation.
• #13: Google search is a great example of this. Billions of people are creating content, billions of people are looking for it, and Google has to make the connections. It’s developed many ways to do this over the years, even before the age of AI, weighting hundreds of factors and using thousands of search engineers to balance those factors to produce consistent results. Google is constantly integrating and updating information from many complex interacting systems. As you can see, it works pretty well. I just learned that Paul Cohen is no longer the dean of the School of Computing, but stepped down this year and is now just a professor.
• #14: On the O’Reilly platform, we manage a smaller search space than Google, but it is similarly dynamic. We have tens of thousands of books, thousands of hours of video, hundreds of upcoming live trainings, katacoda or jupyter based interactive scenarios, playlists and learning paths. New ones are introduced every day by hundreds of information providers, and our users are providing signals in the form of ratings and usage for what they find most valuable. Our search team has to find the right balance of factors to produce the best results for every query. And I can tell you that we know we aren’t able to do as good a job of it as we like. Here for example, you see the first two search results for that same query I showed you in Answers. They are pretty good, but even with a lot of tuning, our most successful book on deep learning, which explains gradient descent, isn’t at the top of the search results. And so we give the user lots of ways to modify the query – what format are they looking for (book, video, etc.)? Are they looking for the most recent? Are they looking for a specific publisher? And so on.
• #15: Yet when we put the same question to our new machine-learning based Answers search engine, it not only brings up what we believe to be the best product based on many, many factors, it takes us right to the exact page where the subject is explained. This search engine is based on AskMiso, a machine-learning system created by one of your other speakers today, Lucky Gunasekara and his team.
• #16: Answers relies on a language model called BERT – it’s in the same class of system as OpenAI’s GPT3, which you’ve been hearing about in the news. Lucky and his team trained it on all the O’Reilly content, as well as questions from Stackoverflow and other sources, and the generated model is able to “understand” the corpus of O’Reilly content and match it to user intent far better than we can do with manual tuning of a traditional search engine.
• #17: You can also see the enormous power for algorithmic systems to do good in the new field that Cornell professor Carla Gomes calls Computational Sustainability. Her team has worked with the Brazilian national grid to build data models that determine which Amazon tributary to dam, solving simultaneously for the need for power generation, the fewest number of people that need to be displaced, and the impact on endangered species. In California, they are helping the water management districts time the release of water into California rice fields to coordinate with the migrations of waterfowl. Both farmers and waterfowl benefit. The possibilities are enormous. We must use these tools to confront the challenges of the 21st century!
• #18: Amazing work by Kirk Bansak,1,2* Jeremy Ferwerda,2,3* Jens Hainmueller,1,2,4*† Andrea Dillon,2 Dominik Hangartner,2,5,6 Duncan Lawrence,2 Jeremy Weinstein1,2 https://immigrationlab.org/project/harnessing-big-data-to-improve-refugee-resettlement/
• #19: This is the master design pattern for applying technology: Do more. Do things that were previously unimaginable. Think through what is possible with new technology. Yes, technology can eliminate labor and make things cheaper, but at its best, we use it to do things that were previously unimaginable! It is human decisions about what to do with technology that put people out of work.
• #20: Even in our consumer society, you can see what happens when you put people and machines together working to do what was previously impossible, rather than simply using them to fatten corporate profits by putting people out of work. Here’s what actually happened when Amazon added 45,000 robots to their warehouses, they added more than 250,000 human workers. The human workers are part of a complex ballet of human and machine, programmers and warehouse workers and delivery drivers, websites and robots, all coordinated by algorithms to work with uncanny speed and precision, delivering many products within a few hours in the luckiest zip codes. Why was this? Amazon didn’t just use the robots to do the same thing more cheaply. They packed more products into the warehouses, and used the partnership of humans and machines to get them out more quickly, so that in some zipcodes, you can get products the same day. Source: https://qz.com/904285/the-optimists-guide-to-the-robot-apocalypse/
• #21: Amazon is a complex human-machine hybrid. From it’s web or mobile front end, where software robots help you find what you want and place your order from a catalog of more than three BILLION SKUs from a network of hundreds of thousands of vendors, through its automated warehouses, where robots and humans work together in a complex dance, through its Amazon Flex on-demand delivery service (now about the size of lyft, if not bigger), it is one giant, algorithmically managed network. https://www.youtube.com/watch?v=I-n6fHfUHzA&t=60
• #22: Jeff Bezos calls this the flywheel. Lower costs lead to lower prices, which lead to more customers, which draws more sellers, offering a greater selection, which leads to better customer experience and more economic activity in a virtuous cycle. This has been true as long as market economies have been around. But you have to work at speeding up the flywheel, like Amazon does. All the parts of Amazon work together to create its value. And it keeps searching out ways to increase the speed of the flywheel.
• #23: AI requires us to change our workflows and processes. We may start out grafting it onto existing processes, but ultimately, it will challenge and change them, as summed up in this quote attributed to Marshall McLuhan (but apparently actually from one of his friends and colleagues, Fr. John Culkin): “First we shape our tools, then they shape us.” But also consider this advice from Buckminster Fuller: “If you want to teach people a new way of thinking, don't bother trying to teach them. Instead, give them a tool, the use of which will lead to new ways of thinking.” You just have to jump in and get started.s
• #24: In this regard, I like to point people to a talk that Google’s Peter Norvig, who is also the co-author of the leading textbook on AI, gave at our first AI conference in 2017. He talked about changes that AI brings to the software engineering workflow. There are a lot of people who understand this now, but there many folks in traditional IT organizations that may struggle not so much to learn the new tools, but to learn the new mindset.
• #25: In his talk, Peter summarized major elements of the change. I’m not going to go through them in detail, but I highly recommend you check out the talk, which can be found on the O’Reilly platform, if you think members of your team need help making the transition.
• #26: I would be remiss if I didn’t also call out the importance of deep engagement with AI ethics. My one big piece of advice here is not to get caught up in the idea that AI is potentially an out of control golem that is just waiting to run amok. Instead, I urge you to thank of AI as a mirror, not a master. Because AI models are trained on data we provide, when they are biased, it is because *we* are biased. If a machine learning model for hiring or pricing or sentencing is biased, we not only have to retrain the model, we have to ask ourselves about the data we trained it on. If, for example, a model is trained on our own corporate data and practices, if it is biased, what does that say about us. The Fairness, Accountability and Transparency in ML conference is a great group to engage with. We also have great resources on the O’Reilly platform.
• #27: Thank you very much.