Open Source Artificial Intelligence

Transcript

• The open-source movement
• The artificial intelligence revolution

The open-source movement is the trend toward making the source code for software available to developers and users so that anyone can modify it and, theoretically, improve it. Examples of open-source software include the Linux operating system, the Mozilla Firefox browser, and the Android mobile platform. The concept has spread to encompass open innovation, in which companies invite everyone outside the company to help in developing new products or in solving design challenges.

The artificial intelligence revolution refers to the growth in the science of using algorithms to develop computer software that is capable of reasoning, problem solving, and decision making. Examples of AI include:

• Loan-processing systems that evaluate the creditworthiness of mortgage applicants.Image-recognition systems that can identify what is happening in a photograph and provide a text description.
• The software that allows driverless cars to recognize obstacles in their path and navigate around them.

Now, tech giants like Google, Microsoft, and Amazon are bringing the open-source concept to artificial intelligence by opening their AI code to the outside world.

In November 2015, Google open-sourced its deep learning engine software called TensorFlow. Deep learning uses neural networks to learn to do something by absorbing huge amounts of data.

TensorFlow is the software that the Google Photos app uses to "learn" to recognize photos of an object after it is fed millions of photos of that object. So if you want to find a photo of a birthday cake in your photo library, you can search for the term "birthday cake" and the app will show you every image you've stored of a dessert topped with candles.

Google uses the same software to constantly increase the accuracy of its search results; to improve speech recognition apps; and to translate written words from one language into another language.

At the same time, the company was criticized because the version of TensorFlow that it released could only run on a single computer. To achieve the scale needed to create new enterprise applications, developers must be able to use the software to analyze data quickly across a multitude of computers.

In 2015, Facebook had shared the software it uses to operate the open-source AI software called Torch on multiple computers. And in early 2016, both Microsoft and Yahoo released their own AI software that can be used on multiple servers.

Finally, in April 2016, Google released an update of TensorFlow that included the ability to use it on hundreds or even thousands of machines at the same time.

In the same month, Elon Musk of Tesla and Sam Altman of the technology firm incubator Y Combinator announced that they were launching a billion-dollar artificial intelligence company, called OpenAI, that would give away most of the AI software that it develops. OpenAI is currently focusing on a form of artificial intelligence called reinforcement learning, in which computers continually improve their performance at tasks by doing them repeatedly and adjusting their approach based on feedback.

Ultimately, however, the project will devote most of its resources to yet another type of AI called unsupervised learning, in which computers can acquire knowledge without human intervention. In the deep learning example discussed earlier, a human would have to label each of the photos of birthday cakes that a neural network would learn to recognize. In unsupervised learning, by contrast, neural nets learn without human help, simply by scanning all the information on the Internet.

In May 2016, Amazon open-sourced its DSSTNE AI software, which fuels its product recommendation system. Unlike TensorFlow, which relies on millions of inputs to promote deep learning, DSSTNE uses less data.

Why would these companies give away their proprietary AI software? There are at least three compelling reasons:

1. Deep learning was built on shared knowledge. The pioneers of the technology were professors who freely distributed what they had discovered to their colleagues. Now, many of those pioneers have been hired by companies like Google, and they've brought the open-source AI mindset to their employers.
2. Unlike previous Google AI tools like MapReduce, TensorFlow was designed to be shared. MapReduce and other tools couldn't function without Google's massive infrastructure, so even if the company released them, very few developers would have the resources to do anything with them. Today, when Google creates new software, it ensures that it is flexible enough to be open-sourced.
3. Opening the software will allow the industry to accelerate the evolution of AI-and possibly return even more value to the company that releases it. There are a limited number of AI experts in the world; most are employed by tech firms, and some refuse to leave their posts at universities. By letting people who aren't on their payroll work with the code, companies can benefit from innovations their own experts might not have considered.

According to an article in Wired, "Through open source, outsiders can help improve on Google's technology and, yes, return these improvements back to Google." The article quotes Google engineer Jeff Dean: "What we're hoping is that the community adopts this as a good way of expressing machine learning algorithms of lots of different types, and also contributes to building and improving [TensorFlow] in lots of different and interesting ways."

Based on this important trend, we offer the following forecasts:

First, the result of the open-sourcing of all this AI software will be an explosion of innovation.

Entrepreneurs can now create new businesses that take advantage of technology that was only available to companies with research budgets in the billions. Existing companies can improve their offerings and increase their revenues with AI tools like DSSTNE to predict which additional products a customer is likely to buy. Insurance firms can use TensorFlow to determine which customers pose the greatest risks so they can adjust their rates accordingly.

Second, the companies that open-source their AI software will also benefit.

As both the basic knowledge of AI and its applications expand, the tech companies with armies of engineers and big budgets for marketing (such as Google, Amazon, and Facebook) will be in the best position to profit. If someone can improve Amazon's recommendation engine, Amazon will benefit because it has millions of customers and millions of products. If someone can develop a better open-source AI system for driverless cars, Elon Musk will benefit because his company, Tesla, is currently paying an Israeli company called Mobileye for its driver-assist technology. At Amazon, meanwhile, turning the proprietary AI software DSSTNE into an open-source offering allows the company to benefit if an innovator beyond the company figures out how to use it to provide even more value to customers. Google would benefit if independent developers adapt TensorFlow to other programming languages; currently developers can only code with C++ or Python, but if it were extended to languages such as Java and Javascript, it would be easier to create new apps.

Third, dispersing the knowledge and the tools to create AI applications will prevent AI from advancing beyond humans' control.

The downside of artificial intelligence is the risk that it will become smarter than humans and decide that we are expendable. This fear has been expressed by Bill Gates, Elon Musk, and Stephen Hawking, among others. If all of the power to create artificial intelligence is monopolized by one or two multinationals, the risk would be even greater. But if hundreds of thousands or even millions of people have access to AI tools, there is a greater chance of avoiding the complacency that could lead to disaster.

Fourth, open-source AI will serve as an indispensable catalyst for a $50 to $100 trillion surge in global wealth through 2025.

McKinsey & Company recently released the report: "Disruptive Technologies: Advances that Will Transform Life, Business, and the Global Economy." Consider the economic impact over the period of 2015 to 20215 just AI and robotics. Adding up the GDP impact in 2025, we get a total range of $10-$20 trillion a year. To get the total economic impact for the 10 years from 2015 to 2025, the simplest model is linear, starting from $0 and ramping up to the 2025 level. The linear approximation is just 10 times the 2025 impact divided by 2. This gives us a range of $50- $100 trillion in cumulative GDP impact. That's a mighty big reason to get excited!

Comments