The promise of AI (A16z mini moocs)
Frank Chen, a partner from Andreesen Horowitz, gave this lecture surveying the primary applications of Artificial Intelligence. The talk starts with the historical lesson of how relationship database technology became widely utilized in the past 40–50 years. RDBMS is the core backend for most companies today. The driving factor of such wide spread was that it made it cheap to perform CRUD (create, read, update, delete) procedures. Frank organized the applications of AI technologies under the similar premise. AI will make so many things we do today so much cheaper and therefore will become the core backend of future technologies in a way similar to how RDBMS became dominant.
A disclaimer: this blog post is merely summarizing what Frank speaks about in his talk and some personal reflection based on his points. It is rather ironic that the activity of summarizing will be quickly replaced with AI as you will see below.
As a brief interlude, this idea of reducing cost as a main vehicle of gaining mainstream adoption is nothing new. In fact, the disruption theory from Clayton Christensen summarizes this phenomenon. If you are familiar with his framework, please keep his theory in mind when thinking about these applications.
Which AI application will achieve low-end market dominance and gradually but steadily evolve upstream?
AI makes it cheap for vehicles to drive themselves.
- Beer drives itself from warehouse to store and then to your house. (Dispatcher). Consumer delivery
- Drones protect soldiers and innocent civilians. A new way to save lives. Shield.AI can go into buildings and generate map of the building.
- Drones can also deliver life saving blood. (E.g. ZIP line drone. 90 mile range.)
AI makes it cheap to understand the world cheaply
- Computer vision to help recognize the world better using generative adversarial networks. There are two neural networks in GANs. One classifier networks that tries to classify the input as best it can while the other network tries to work against the classifier network by generating fake inputs that are very similar to the real input to the classifier networks. Then there’s a competition dynamic between the two networks where the first optimizes the accuracy of classification while the second optimizes for trickery. The outcome from GANs is extremely accurate. For example, Pinterest allows you to highlight a region in an image to find images similar to that region.
- Visual search becomes ubiquitous. Pinterest recently added the visual search to identify whatever is live captured by the camera.
- Tensorflow powers cucumber sorting. In Japan, an engineer built a cucumber sorter based on raspberry pi. This is extraordinary because of the level of enhancement the AI technologies can bring to the individual. Individuals become more and more empowered. Thousands of other applications can be built like this by the individuals.
- Precision agriculture. Tractors fertilize individual plants, not whole fields. This has consequences to a lot of savings.
- Amazon Go: shopping with no checkout lines
- Robots everywhere. Robot help shoppers navigate the store. Robot security guard. Robots can check inventory to make sure that inventories are constantly on the shelf. Robots can verify the correctness of planogram.
- Algorithm can tell stories from pictures. Researchers are working on algorithms to identify a sequence of pictures and recognize the story behind the sequence of images. #Reconstructive-narrative
- Google queries that will work in a few years and will be able to answer questions that weren’t answerable today. (1. Which of these eye images shows symptoms of diabetic retinopathy? 2. Describe this video in Spanish 3. Please fetch me a cup of tea from the kitchen 3. Find me documents related to reinforcement learning for robotics and summarize them in German.)
AI makes it cheap to create contents
- Basic contents are generated from AI coverage of sporting events. (HuffPo and Toutiao from China reported the stories)
- Cookbooks. Algorithms are able to watch the videos of cooking and can create discrete steps for the recipes.
- Photo-realistic pictures from text. Input is a sentence and can output the images. (Application of GANs)
- Photo-realistic pictures from sketches. Based on the sketch and baseline input an output of realistic.
- AI can generate music that are indistinguishable from human generated music.
- Movie trailers can be generated from computer.
- DeepCoder from Microsoft to generate code based on code samples from Github
AI makes it cheap to predict the future
- Implicit, continuous authentication and the end of passwords. The way people walk is a distinguishing factor.
- The future of customer support. Automated routing, seamless escalations.
- Detecting cancer early using cell-free genomes (Freenome). Instead of tissue biopsy, Freenome can make prediction based on blood input.
- Detecting abnormal heart conditions with your apple watch (Cardiogram). Using sensor data to make such predictions and help you understand what’s possible.
AI makes it cheap to automatically optimize complex systems
- Waze suggesting the best routes. Optimizing driving behaviors
- Automatically figure out best defensive positions. Defensive analysis on the field.
- Make software run faster. Compiler takes highlevel language and turn them into assembly code. A better compiler can help run faster.
- Find more alpha. Sigopt takes the model and tune the model in such a way that the model is enhanced with machine learning.
AI helps reduce energy consumption
- Google DeepMind took 120 variables and reduced 25% energy in their data centers. Machine learning is good at calculating how much things cost.
- Shop and deliver efficiently. Instacart is able to help save 8% of total time.
AI makes it cheap to understand people better
- Speaking faster than typing. Accuracy is so good that people talk
- Smart reply (Inbox team from Google). The feature suggests contents for emails. In Feb 1st, 2016, Inbox replies are utilized a lot.
- Automated summaries from multiple documents. Oracle had a demo a decade ago that helps generate summaries from multiple documents and a New York Startup
- Better hiring through AI-tuned job descriptions. This helps create accurate job descriptions.
- Natural language powering e-discovery. Everlaw helps categorize documents based on actual contents of the documents.
AI brings new types of interaction based on emotional experiences. Real-time translation between languages. Waverly labs is the babel fish.
Based on EHRs data can help predict suicidal tendencies. AI will get into every piece of software in the same way relationship database application.
The take-away of building an AI-powered future
- Get to know the tools, which are improving week over week: http://aiplaybook.a16z.com/
- Train your people (People need to understand the technology and make sure people in the organization are trained on what’s possible)
- Give time and space to explore and create things based on the technology -> figure out where exactly how AI can be plugged in to move the needle.
The original video is here: https://vimeo.com/215926017