APPENDIX A
AI Experts

AI Experts

In the course of writing this book, we engaged with a number of experts in the field. We spoke with not only technical experts, such as machine learning engineers and data scientists, but also business leaders who have already taken the AI adoption journey. We asked each of our experts five similar questions, resulting in some interesting similarities and disparities among their perspectives. You are invited to glean whatever insights you may from the comparison of their responses.

Chris Ackerson

Chris Ackerson leads Product for Search and Artificial Intelligence at AlphaSense, where his team applies the latest innovations in machine learning and natural language processing (NLP) to the information discovery challenges of investment professionals and other knowledge workers. Before AlphaSense, Chris held roles in both product and engineering at IBM Watson where he led successful early commercialization efforts. He is based in New York City.

1. Given you have expertise with AI and how people interact with it, are there any insights or tips that you'd like to share?

AI development is a new and different form of software development requiring new tools, processes, and job roles. Building an effective AI development organization requires investment in each of these areas. My experience is that teams often focus too much on algorithm design and architecture at the expense of building an effective AI development organization. The result is they back themselves into a corner as technology rapidly improves and they are unable to keep up. Just check out the change in the leaderboards for state-of-the-art performance in any AI task to see what I mean. It goes without saying that algorithm design is central to AI, but teams should expect to rip and replace algorithms on a regular basis. While every project is different, a good general rule is to prioritize open source, be leery of black-box APIs promising superior performance, and invest most of your resources in a system that allows you to train and deploy completely new model architectures in regular intervals with low overhead.

2. What has been your biggest challenge while adopting AI?

Data collection continues to be the biggest barrier to broad adoption of AI. State-of-the-art deep learning models require massive quantities of clean data in order to make accurate predictions. For example: Google's influential BERT model for natural language processing, which was trained on a corpus containing billions of words; self-driving AI trains on millions of real and simulated hours of driving; and the sentiment analysis algorithms we developed at AlphaSense, which learn from hundreds of thousands of corporate earnings calls. Whether or not these types of datasets are even accessible in your AI project is the first question you should answer, but assuming you can acquire the data, the bigger challenge is developing the tools, processes, and labor to clean and label it. The research community has made enormous progress in the last decade, improving the raw predictive power of AI, but commercial applications have lagged behind because of the practical challenges of building robust datasets to take advantage of algorithmic horsepower.

3. What advances in AI do you envision over the next five years?

In the next five years, I believe we will solve the problem of generalizing AI models within the scope of a narrow domain. That sounds like an oxymoron, but the vast majority of successful AI projects to date have been developing models for just a single prediction objective. This is the narrowest definition of “narrow AI” possible. Let's say you want to build an AI to monitor social media for stock trading signals; today you would train independent models for sentiment analysis, entity recognition, topic extraction, and any other classification task relevant to your project. When IBM developed its oncology technology, the company spent enormous resources developing independent models for each individual cancer, with few economies of scale between each new target. The practical implication is that the cost of solving real-world problems that generally involve many separate objectives is prohibitive. Intuitively we understand that all of those social media–monitoring tasks are highly related, as a human interpreting the sentiment of a tweet is critical to extracting the key topics, and vice versa. Transfer learning—that is, sharing the learning from one task to a related task—is an area of intense investment in the AI community. Google's BERT is an important step in that the model is pre-trained in an unsupervised way by consuming huge amounts of unlabeled text, and is then fine-tuned with supervision for specific tasks like sentiment analysis. Having a single framework for solving multi-objective problems like social media monitoring or identification of anomalies in medical images will lead to a huge increase in the number of real-world AI applications.

4. What job functions do you see as a prime target for AI assistance over the next three years?

The next three years will be dominated by advances in natural language processing. In the same way that structured data analytics and visualization have become ubiquitous across the enterprise, all knowledge workers will be assisted by AI that draws insights from the vast number of unstructured text documents related to their job function. Consider analysts in corporate strategy; every day there are tens of thousands of new research reports, news articles, corporate filings, and regulatory actions that contain information absolutely critical to making the right investment decisions for their company. In the same vein, a sales executive going into an important meeting, an HR rep deciding on a corporate policy, a product manager prioritizing features in an application will all benefit from AI that surfaces timely insights from this sea of text. As a knowledge worker, imagine the productivity increase if you were given a team of analysts to prepare you each day. Or if you have a team, imagine everyone on your team was given a team. The result is a boost to the collective IQ that will lead to better decisions across all industries.

5. Any other thoughts about AI adoption you'd like to share?

One of the most common reasons AI projects fail is that teams underestimate the importance of great product design. The mystique of AI leads to a belief that the algorithm is all that matters. This is especially common in internal enterprise applications. The reality is that great product design is just as important in AI-powered products as in traditional software applications. Consider Google Search, where the placement and construction of every link, feature, and answer box is meticulously crafted to increase the speed of access to information. Over many years Google slowly and thoughtfully evolved how AI could enhance the search results pages for billions of queries without getting in the way of user experience. If AI is to reach its potential and become ubiquitous across software applications, it behooves companies to hire and develop product managers, engineers, and designers who understand the capabilities and limitations of AI and invest in ways that prioritize great user experience over technology choices.

Jeff Bradford

Jeff Bradford is the founder and CEO of Bradford Technologies, a developer of real estate appraisal software. He has over 31 years of experience providing appraisers with innovative software solutions and is a nationally recognized expert in computer technology and analytics. Jeff has been recognized as a Valuation Visionary by the Collateral Risk Network and as a Tech All Star by the Mortgage Bankers Association. Prior to founding Bradford Technologies, Jeff notably worked at Apple Computer, Structural Dynamics Research, and FMC Central Engineering Labs. He holds three master degrees in engineering mechanics, computer science, and business administration.

1. What has been your experience pulling together data for AI purposes and what are your thoughts on the need for building an initial prototype?

If you are a first-time, wanna-be user of AI/deep learning technology, you definitely want to start with a prototype case. Something simple. This is also not the time for trial and error and learning as you go. You want to know if this is going to work as quickly as possible for the least amount of money, so select a consulting firm that has considerable experience in AI. Have them evaluate your training data. Is it good enough? Does it have to be massaged? What AI model are you going to use? Which is best for your application? Have them run some tests. What are the results? If the results are promising, then you can make some strategic decisions about incorporating AI into your products and services.

2. What has been your biggest challenge while adopting AI?

The biggest challenge has been the training data—selecting and assembling the data into structures that can be used to train a model. How well the model works is directly related to how good the data is, so the quantity and quality of the data is critical to the success of the AI model.

3. What advances in AI do you envision over the next five years?

I think that in the next five years, AI will have advanced to the point where models can be pieced together to form very comprehensive systems that can augment many facets of work environments as well as everyone's lifestyle. Everyone may have their own virtual companion or work assistant.

4. What job functions do you see as a prime target for AI assistance over the next three years?

Any function that involves communication, data, or the providing of recommendations can be automated or augmented by a virtual assistant. These are the areas that will be targeted by AI systems.

5. Any other thoughts about AI adoption you'd like to share?

Everyone should learn about AI: its potential and its shortcomings. It is definitely going to be part of everyone's life. It is not going away. It could become Big Brother, as it is already becoming in China, or it could be used to enhance one's life and work experiences. Learn how to use it to enhance humanity.

Nathan S. Robinson

Nathan S. Robinson is an Ohio-born, China-raised, Oklahoma-schooled, Austin, Texas, transplant. Nathan began his career in product management and artificial intelligence at IBM Watson. He currently works as a product manager at Babylon Health, an organization leading the charge in making a health service accessible and affordable to all using technologies like AI. Prior to joining Babylon Health, he worked in product management at Keller Williams, where he owned Kelle, a virtual assistant and mobile application hybrid for realtors.

1. What are the key benefits of AI based on your experiences using it within your organizations?

There are many benefits of utilizing artificial intelligence in an organization. AI can help automate simple and repetitive tasks, freeing up time for team members to focus on higher value, human-centric tasks.

Artificial intelligence can be leveraged to process and understand large values of data, and unlock insights that would not have normally been found.

2. What has been your biggest challenge while adopting AI?

Aligning both the organization's and the end user's expectations to reality. Artificial intelligence is powerful and can provide immense value, but setting expectations for the end users, as well as the internal organization, is key in having a great experience.

Everyone has data, but not everyone has good data that is ready to use. Finding, cleaning, normalizing, and preparing data is often the bulk of many AI proofs of concept in an organization.

3. What advances in AI do you envision over the next five years?

There will undoubtedly be incremental advancements in existing technologies as a result of better datasets, tweaking of models, improvements to training methods, and more.

There will also be advances that come as a result of increased processing speed from technologies like quantum computing.

It is also likely to see new innovations within the field to “leap” us forward. Things such as combining technologies to simultaneously use multiple forms of input as the training data in order to gain a “1 + 1 = 3” type of value from the output. Think: using both audio and video inputs to train a model. Or finding alternative ways (not just improvements to existing methods) to train models in the first place.

4. What job functions do you see as a prime target for AI assistance over the next three years?

I don't have any new insights here. Things that are repetitive or require little creativity and unique problem solving are at risk of automation. However, more than replacement, AI will more commonly be used to augment humans doing a job. Rather than AI replacing people directly, it's more likely that it will augment fewer people to be more productive.

5. Any other thoughts you think are relevant to AI adoption?

Culture and human experience is crucial in a company's successful adoption and implementation of AI. The organization's culture can both dictate whether or not they successfully utilize AI, as well as how quickly they can achieve value from it. Setting the expectations of the end users of the AI, as well on the organization's side, building/investing in it, is crucial in an organization's long-term AI strategy.

Evelyn Duesterwald

Dr. Evelyn Duesterwald is a principal research staff member and manager of the AI Lifecycle Acceleration team at IBM Research AI. She holds a PhD in computer science and is passionate about all aspects of AI engineering and the AI lifecycle, with a current focus on securing AI models and AI operations. Her interdisciplinary research focuses on the intersections of artificial intelligence, security, and software engineering.

1. Given your expertise with AI security, are there any insights or tips that you've found to be effective?

AI security is usually not about hacking and breaking into systems. AI security threats can result from seemingly harmless interactions with exposed AI interfaces such as APIs. An adversary may craft inputs to force a particular model response (these are called adversarial inputs) or to poison the data that the underlying models are trying to learn from. For example, in speech recognition systems you just have to add a little bit of imperceptible adversarial noise to a speech recording (e.g., “Hello, how are you?”) to cause a dramatic and targeted misclassification by the speech model (e.g., the model hears “Order a pizza.”)

We don't want to be surprised by what might happen as a result of malicious users of AI, so we have to build security into our models from the ground up so as to train robust models that are resilient to these kinds of adversarial exploitations. Training robust models is a very active research area, and new approaches are being developed all the time by the larger AI research community, as well as by our IBM Research AI teams.

2. What has been your biggest challenge while adopting AI?

The significant lack of automation and process standardization that you typically see across the AI lifecycle. Everybody is training their models in their own way and without an accountable or repeatable process. There is an urgent need to bring the rigor of modern software engineering to AI. Even very basic engineering concepts like systematic version management are often lacking.

3. What advances in AI do you envision over the next five years?

We need to move AI from an art into an engineering discipline. Some of the largest recent gains in AI capabilities have been possible through deep learning, especially in speech and image recognition. But deep learning is also one of the areas where AI most resembles an art, which severely hinders adoption. We cannot trust an AI that we cannot fundamentally understand. So I expect the most impactful future advances in AI to be in the areas of interpretability and “explainability” of AI.

4. What job functions do you see as a prime target for AI assistance over the next three years?

We are starting to see some traction in making AI itself as a target for AI assistance—in other words, AI-assisted AI. For example, teams are working on new AI-powered tools that aid the data scientists in building AI by assisting in processing their data, in extracting relevant features, and in training and testing their models. Just imagine talking to your AI assistant: “Hey, I have data streams of customer comments; build me a model that can tell me what my customers are talking about.”

5. Any other thoughts about AI adoption you'd like to share?

We are seeing great advances of AI in assistance roles—personal assistants in your home as well as professional assistants in the office. Adoption of AI in more autonomous decision-making roles is further behind. Autonomous vehicles are a great example, but here we are slower to move past experimentation. To make autonomous AI a practical reality, we first need to address the various adoption obstacles that I mentioned earlier. We need more productivity in our AI engineering practice, and we need robustness assurances and safety guarantees through more explainable and trustworthy AI. AI is an incredibly fast-moving field. We will undoubtedly get there, and it's a very exciting journey be on!

Jill Nephew

Jill Nephew is CEO and founder of Inqwire, PBC, an interactive platform whose goal is to help people make sense of their lives. From scientific modeling to software language and tools construction, to domain-specific AI and IA systems, Jill has always sought out novel ways to use technology to help people think better.

1. What has been your experience pulling together data for AI purposes, and what are your thoughts on the need for building an initial prototype?

Back in the 1990s, I worked on supply chain planning and scheduling solvers. We tried out a method that has similarities to today's artificial neural nets in that it iterated, did hill climbing, started with a random landscape, and was a black-box algorithm designed to converge on a global optimum. We had widespread adoption because we could show that we produced an objectively better solution. However, what we didn't anticipate was the need for “explainability” of the results.

The schedulers and planners using the software most often weren't the ones who made the purchasing decisions and also weren't given the opportunity to weigh in on the black-box nature of the algorithm. So when they attempted to use it, they expected that if it gave an answer that they thought didn't make sense, they should report that as a bug and we should fix it. Since this was a black-box algorithm, most of the time there was no way to fix it, and this was unacceptable. The actual workers whose decisions we were supplying decision support to were held accountable by their decisions, and this meant we were dealing with a kind of unanticipated revolt from the ground up against the adoption of our solver.

If the system had perfect answers or answers that were visually, obviously correct in the absence of explanation, this pressure might have been relieved by getting global organizational buy-in that the decisions made by the algorithm could be accepted without question. However, for our system we couldn't converge on this. I am skeptical that any system can, because when a human inspects the thinking of a machine, the strong unstated assumption is that if you are saying the system is intelligent, then it has common sense and its decisions will make sense. We didn't find any organizational support for pushing back against this expectation.

Things got worse and better at the same time when we were tasked with fixing the algorithm to include common sense and explainable results. We did this by building a constraint-based solver. The core solver would not yield to this shift in approaches, so a layer was built on top that could. The global optimization solution would serve as starting conditions for the constraint-based solver that would be a kind of clean-up phase. This went a long way to “straighten out” where the core algorithm went wrong, but we still found it incredibly challenging to undo far-upstream decisions by the system. What happened next was the really big insight.

When we deployed this new clean-up phase, the users loved it and then came back with a request that we give them a way to skip the global optimization as the starting solution and just have the constraint-based system generate one. So basically, they asked for a steepest descent, explainable solver. We gave them that option and it was widely preferred over the other one, despite the fact that we could show the final solution was less optimal.

The takeaway was that, when you are thinking in objective functions and global optimizations, it seems absurd to build anything that would settle for a less optimal solution. But when humans have to take responsibility for the decisions of the system, a more explainable, suboptimal solution might just win out.

I think we are living in interesting times for algorithms. The market is just getting educated enough to start to ask hard questions about accountability and “explainability,” and anyone implementing black-box solutions who can't do either should probably think through a backup plan. Further, making sure that the actual users of the software (not necessarily the ones who are making the purchasing agreement) understand the implications of having a black-box algorithm making decisions is essential to avoid being caught in a support nightmare.

2. What has been your biggest challenge while adopting AI?

Black-box algorithms. No matter how much we explained the nature of the algorithm and that we couldn't just “reprogram” it to fix it, the users still expected it to work like all the other software systems they have used. There is an obvious commonsense problem with the solution, and therefore the system needs to be fixed. We never found any way to push back on this expectation, and it drove our team to ultimately decide to abandon a black-box approach.

3. What advances in AI do you envision over the next five years?

What I envision is very shaped by past experiences and may not at all reflect where the world actually goes. But if I had to do the thought experiment, I would imagine that as the market is educated on how global optimization currently comes at the price of “explainability,” we will abandon any AI that uses black-box algorithms and switch to systems that don't.

I think that we will also face a real crisis around what a pure data solution can actually deliver. There is a lot of faith right now that we are accumulating enough data to overcome first-principle limitations such as most of the world generating sparse, dynamic, evolving, and non-normative data. There is also a complete lack of understanding of the value of incorporating process models in AI.

I think there will still be an opportunity to keep the dream alive of using cognitive-based technologies in new ways to attack wicked problems, provided the field can pivot toward systems that can start to incorporate process models and simulation instead of optimization. If we can pivot, I would predict that AI experts who know how to work intimately with domain experts and do modeling will have budgets to build a lot of new and imaginative systems and solutions as they re-channel this market interest in cognitive technologies.

4. What job functions do you see as a prime target for AI assistance over the next three years?

By “prime target” I am speaking to where I think there is a need, not another business model. The main need I can see AI addressing is looking for patterns in domains where patterns aren't expected—particularly in the human life support system. Monitoring the natural environment isn't cut and dried. Scientists need to be alerted when something has changed. We don't currently have a way to do this. We find things, like the ozone hole, by accident.

The software in place to monitor the ozone didn't expect a hole. The scientists who programmed the algorithms didn't expect a hole was possible based on their understanding of the stratosphere, and they programmed monitoring algorithms to discard outlier data as a systematic error instead of reporting it thoughtfully. This oversight took years to track down, and we are fortunate that it eventually was. The mechanisms behind the ozone hole formation were so complex that it took multiple scientists years to figure out that it was actually coming from propellants in spray cans. The takeaway is that monitoring, along with the thoughtful reporting of anomalies, is critical for the ongoing protection of our human life support system. We need many more of these systems, and I hope someday that information technologists and scientists together can find ways to convince decision makers of this point.

5. Any other thoughts about AI adoption you'd like to share?

When I talk with newer AI enthusiasts, I hear a common misconception that if we had infinite data, we would have perfect knowledge. It has a commonsense ring to it and often goes as an unstated assumption. Anyone who has studied the nature of the physical world, feedback mechanisms, chaos, quantum mechanics, numerical analysis, the nature of counter-factual information, and how mechanisms can't be restored from pure data, or simply thought through the pragmatics of what it would take to collect perfect knowledge, knows this is not true. Confronting this head-on as a starting assumption with any discussion of adopting an AI project may cost some sales, but ultimately it may mean long-term success in terms of setting realistic expectations of project outcomes.

Rahul Akolkar

Rahul Akolkar leads Worldwide Technical Sales for Data Science and Artificial Intelligence at IBM. He has worked on some of the largest implementations of AI in past few years and established AI delivery teams on four continents. Rahul previously held a number of roles in IBM Research and Corporate Technology. Rahul has contributed to open source projects and W3C standards, and he is an author of numerous research papers and inventor of dozens of granted patents. He holds graduate degrees in computer science and mechanical engineering from the University of Minnesota.

1. Given you have expertise with AI and how people interact with it, are there any insights or tips that you've found?

There are a number of detailed recommendations that exist on every aspect of creation and evolution of AI, but the singular insight I can offer is one must focus maniacally on the elimination of friction between the interaction of AI and other actors, whether people or processes. Some of the ubiquitous embodiments of AI that we find today in terms of assistants on our phones or devices in our homes can also be the source of frustration. AI that either creates an aura that it can answer anything or is not engaging in the right user experience is bound to become shelfware over time. We need AI that people want to interact with, that is available in the right context, that is using the right modality, that provides information and value consistently, and that fades in and out of the experience as needed.

An important part of eliminating the friction in the experience revolves around the general notions of trust, transparency, auditability, and ethical behavior of AI. People who interact with AI increasingly, and rightfully, demand that AI can no longer be magic, not to them, or to regulators, auditors, overseers, when AI is providing input to influence increasingly important decisions in fields such as medicine, insurance, crime prevention, investment, and security, both IT and physical.

2. What has been your biggest challenge while adopting AI?

It's very important to lay out an organization's journey to AI while providing the right organizational and technical governance models and to begin delivering incremental value in short order. There are certainly elements such as customer engagement and the broad category of recommender systems that can go to production in short order, whereas other areas such as deep domain expertise and esoteric vocabularies of interaction often require a phased implementation. The two common fallacies I find in terms of AI adoption are (a) moonshot or nothing, where one tries to win it all by going for a large prize, without incremental gains in process, and (b) one and done, with the plethora of data science and AI platforms and tools out there, it is fairly easy to deliver quick value and celebrate success prematurely, but it's important to keep track of the principles that will deliver lasting value with AI; otherwise, a successful launch can quickly devolve into a less than optimal experience.

Those who get the best business value with AI are those who have taken the due measures to foster a fitting culture for AI adoption within their organization. Where AI differs from a lot of traditional software and systems is that it requires near constant care and feeding, as it is trying to provide value in a changing environment and adapting as the real world or its environment changes (think of retraining models or adding a new component to an ensemble recipe). One might then wonder, why bother with AI at all? Essential as tabulating and procedural computing devices are, the world has moved on in terms of expectations from technology. A vast number of afore-unthinkable experiences can now be powered by AI, and those who harness those effectively will conquer the hearts and the minds of their audiences and customers.

3. What advances in AI do you envision over the next five years?

We're already seeing advances in the transparency of AI, not just around classification and structured predictions, but also around neural nets and deep learning techniques. I see this work around “explainability” of AI, and detecting bias in AI, as one that is ripe for advances, both in operational runtimes as well as tools. We will get much better at interacting with data through virtual data pipelines and distributed querying of disparate sources to power AI. Another advancement we're already seeing the tip of the iceberg on, and I see accelerating, is the use of AI techniques in the design of AI—for example, trainless predictions in hyperparameter search spaces and neural network architecture selection, a second-order function illustrating a field at the cusp of new breakthroughs. We will also see advances in creating compelling AI interfaces in terms of the sophistication of conversational agents, visual and speech services such that they can be more effective and scale (a welcome addition to anyone who has had to repeat an utterance numerous times to their phone or home device). Finally, of immediate significance to the creators of AI, we are going to see a growing emphasis on open platforms and end-to-end governance via AI fabrics and a maturing field of AIOps, with practices similar to the analogous ones in DevOps and DevSecOps.

4. What job functions do you see as a prime target for AI assistance over the next three years?

Job functions that are based on collection and retrieval of data are prime targets for AI assistance. The second area is around providing expert assistance in context and covering vast areas of information. Predictive models that help make recommendations to human subject matter experts for any key decision is another area. There is a longer list of broad categories, but in essence any job function that requires customer engagement, or entrenched learning and incremental expertise development, is a good target. This has led to some level of uneasiness when we talk about AI and jobs, but I don't see it largely differing from, say the Industrial Revolution, where a number of job functions were lost to machines replacing human workers in industrial settings on the shop floor or in the farmlands. To summarize, I think AI will enable humans to be more productive, to make better informed decisions and focus on the less mundane tasks that AI can easily master.

5. Any other thoughts about AI adoption you'd like to share?

AI is here to stay. It is creating and will continue to create disruptive innovation at a faster pace. Organizations owe it to their future selves to explore how they can adopt AI, and there is a wealth of knowledge available for anyone who seeks to learn more.

Steven Flores

Steven Flores is an AI engineer at Comp Three Inc. in San Jose, California. He leverages state-of-the-art methods in AI and machine learning to deliver novel business solutions to clients. In 2012, Steven earned his PhD in applied mathematics from the University of Michigan, and in 2017, he completed a postdoc in mathematical physics at the University of Helsinki and Aalto University.

1. Given your expertise with AI, are there any insights or tips that you've found to be effective?

Much of the academic and popular literature paints AI in an idealistic light, which can lead to misconceptions about its practice. Spectacular and even miraculous results hide underlying limitations and a long list of failed ideas that preceded them. Indeed, when one actually uses AI, breakthroughs take a long time to achieve. Models are constantly redesigned, tuned, babysat during training, tested, and discarded, and the decisions that go into this process may have vague, unsatisfying motivations. Good results are hard-fought for over months and even years of work. They rarely work well out of the box.

Because of this reality, AI practitioners must participate in the larger AI community to be successful. They must read the research literature, attend talks, discuss their problem with experts, and listen to experts talk about theirs. They must apply insights from expert knowledge and experience to their problems, because the solutions usually do not live in textbooks. Doing so can speed up progress on work in the nebulous frontier of AI.

2. What has been your biggest challenge while adopting AI?

One of my biggest challenges is proper formulation of the problem to solve with AI and how to measure success. This is difficult because the problems that we encounter in real life are vague and qualitative, but the problems that we can solve with AI are precise and quantitative. Translating between these worlds is challenging.

Another challenging aspect of AI is that it seems to be more a bag of mathematical techniques and models than a unified theory. Practitioners of AI always want to find and use the best methods, but they may not understand why one is best. I think that making informed strategy and design decisions is difficult in this environment.

3. What advances in AI do you envision over the next five years?

Reinforcement learning is one of many areas in AI research that has seen recent breakthroughs. One of the most famous examples is attaining super-human performance on the classic Chinese game of Go. However, as impressive as these gains are, the solutions behind them often do not generalize beyond the original problem. Progress on this front is expected to grow in the coming years, with huge implications for industry.

4. What job functions do you see as a prime target for AI assistance over the next three years?

AI will have a growing impact on many sectors of our workforce. To anticipate the future, let's look at some examples of what is happening today.

In media, AI assists reporters with research, data crunching, and in some cases, content creation. This frees up reporters to work on less mundane and more nuanced parts of their work that AI cannot handle.

In the service sector, chatbots handle end-to-end routine questions made to call centers, and they route calls regarding more complicated matters to the proper human staff member.

In medicine, AI is used for diagnostics, like tumor detection, data mining in medical records, and more.

In genetics research, AI is used to significantly enhance our understanding of the human genome. Future results may lead to the eradication of genetic disease and insights into healthier life choices for certain genetic fingerprints.

Throughout industry, computer vision has seen widespread adoption. Today, computers read handwriting, detect nanoscale defects in silicon wafers, and provide safety assistance to car drivers.

In all of these cases, AI is used to enhance human ability by finding hidden patterns in complex data and making better decisions with it, and to free up human talent for more creative or subjective work by handling more routine and objective tasks.

5. Any other thoughts about AI adoption you'd like to share?

Right now, we are living in the first generation of AI integration into our lives. AI in its present state has potential applications well beyond its current use, so there is a lot of room for innovation that builds on its current status. As this happens, future breakthroughs will push the horizon of possibility further, from simple decision making based on complex models made from big data, to perhaps something smarter and more human: complex decision making based on simpler models made from less data.