Apr 1, 2017

Thoughts on Superintelligence: Paths, Dangers, Strategies

Last week I finished up Nick Bostrom's Superintelligence: Paths, Dangers, Strategies. Coming into the book, I'd often heard it as the pessimist's response to Kurzweil's The Singularity is Near. I didn't really find that to be the case though: Bostrom certainly paints some scary pictures of potential futures as artificial intelligence develops, but he's not in denial about all of the positive potential. It's more that Bostrom felt there hadn't been a sufficient treatment of the downsides (and strategies to mitigate them) in the existing literature, so he sought to create that balance.

Getting through the read was a slog. It's probably one of the longest periods of time (a couple of months) I've spent on a book and still finished it. It's a topic I'm really passionate about too, more's the pity. The reality is that it's hard to relate to the topics that Bostrom digs into to a sufficient degree to justify spending time on the detail that he goes into. He spends 100 pages digging into the different structures of AGI systems and their relative merits and downsides vis-a-vis their capabilities and their potential to destroy humanity. I would be fascinated by the blog post. 100 pages is tough.

That being said, I don't think Superintelligence is a bad book. In fact, I think it serves as a great handbook to form a baseline for practitioners' future efforts to address the riskiness of developing AGI. After an initial read-through, the book may have lasting value as a reference guide when developers and researchers are diving into the actual development of these mitigation systems.

In all, I think Superintelligence is a must-read for any serious AI advocate. Most of the topics covered and the arguments presented won't be novel for someone who has spent time in the space, but it does provide a common language and frame of reference to drive future discussion. Relevant topics: superintelligence take-off scenarios, mediums (silicon, biological, swarm, etc.), types of systems, organizations that might pursue/achieve AGI.
Mar 3, 2017

The Power of Connectedness

I’ve written before about my desire to tackle the problems that lead to the achievement gap in our society. When you boil the problems down to their core, you’re left with one consistent theme: those from the lower economic classes who want to get ahead in life are limited by a lack of opportunity and exposure. My previous article describes how an AI agent can help address those lacks through universal education. Here, I want to lay out an approach that I think may be even more powerful.

Much of the world’s attention has been on the executive office of the United States these past few months as Donald Trump has settled into the office. So far, Trump has been a historically unpopular president. Some of the criticisms of the president are based on the seeming hypocrisy of his populist rhetoric juxtaposed with his own origins. Trump likes to claim that he is a self-made man, having received “nothing more” than a million dollar loan from his father to get his first business off the ground. This seed sprouts into the argument that the 45th president has ridden the coattails of his father’s success to the most lofty position in the world.

Whether you place weight on these criticisms or not, it is undeniable that a person’s personal connections do much to determine their success in life. Personally, I need very little convincing on this point: I simply have to look at how I’ve benefited from my own connections.

The Story of David

My professional career essentially began at the age of 15. My mom insisted that I get a summer job to start making some money. I had no desire to follow in the footsteps of my brothers to get a near-minimum-wage job unrelated to any of my personal interests. Fortunately, I was in a privileged position: my high school had a structured internship program for seniors. I wasn’t yet a senior, but I went and asked the internship coordinator if she knew of any companies that were willing to pay an intern over the summer. I ended up spending my summer days working in the IT department of a local medium-sized business, learning a fair amount about the field and getting paid about twice the minimum wage.

My next internship search was a couple of years later in college. I knew from my parents and brothers that good internships early on in college were an important stepping stone to a good job after graduation. After one aborted attempt, I returned to my network for help. This time, the aid came from my now-father-in-law, who worked at a major corporation with a top tier internship program. He put in a good word for me and the next thing I knew I had spent two semesters in a world-class professional setting making good money.

While much of my career progression since those first two internships has been more opportunistic, it would be hard to overstate the importance of that early experience in getting me where I am today. It’s the rare candidate straight out of school who had the caliber of experience I showed on my resume. And in each case I can draw a very clear connection between the people in my network and my success. This advantage has played out in a thousand different ways since then. Just recently I applied to an interesting role and quickly discovered that I knew five different employees of the company through my LinkedIn network. I reached out to one of them and my application was immediately moved to the top of the stack for consideration. I’m now at the point in my career where it’s easy to take this type of advantage for granted; I must constantly remind myself that most job applicants are simply ignored when they apply to a role because they don’t have the advantage of knowing someone on the inside.

For the sake of comparison, I want to lay out a parallel storyline. This is fictional, but closely mirrors the stories of several people in my life.

The Story of Jake

Jake lives in a single parent home and needs to get a job in high school to help pay for his clothes and food. After several weeks of driving around town and filling out applications, he is hired to be a cashier at Wendy’s, making minimum wage. He works through the summer, putting in at least 40 hours every week. But he doesn’t make enough to tide him through the school year and has to continue putting in hours in the evenings and weekends. Between the job and his social life, his grades slip from a B average to a high C.

Senior year rolls around and Jake applies to a handful of universities. His grades aren’t enough to get him into the top school in the state, but he ends up at a large university that is reasonably well-respected. To save money, he lives at home. Despite taking out student loans, he has to keep working part-time. At least he’s a shift manager now, making 50% over minimum wage.

Jake knows that internships are important, but most of the good ones are unpaid. He does manage to squeeze in one semester at a pretty good company, but other than that he needs to keep up his hours at Wendy’s. He takes university a bit more seriously than he did high school, and graduates with a B average. Luckily, the job market is healthy when Jake graduates, and he lands an entry-level job related to his major making $50K per year. He’s doing better than plenty of his friends from high school.

Having worked for his entire mature life to this point, Jake has no problem throwing himself into his new job. He always stays late and puts in some hours on the weekend. After a few years, Jake’s efforts are rewarded by his boss with a promotion. He’s now making $65K per year and older people in the company plot out a path for career success for Jake: a promotion every three to five years, hitting middle management in his 40s, and being able to retire around 60. Not too bad.

To be clear, Jake’s story isn’t bad. He does better than most Americans and exhibits the admirable traits of a strong work ethic and perseverance. But in terms of career success, Jake’s story isn’t mine. And there’s no fundamental reason that it shouldn’t be.

How the 1% Gets Ahead...

The only real difference between my career trajectory and Jake’s is that I got a helping hand from people in my network early and learned the value of progressing in my career through the people I know. I would argue that on every other dimension Jake should be doing better than me.

While this anecdote is dramatized, it’s highly indicative of a driving force behind the achievement gap in our society. The most successful get “handed” opportunities while the rest have to work for their success. Trump’s dad gave him a million dollar loan. My father-in-law to-be helped me get a top tier internship. While these anecdotes are illuminating, they don’t tell the whole story: this societal “favoritism” continues throughout life. My first six years in the professional workforce have allowed me to cultivate a network that was critical in me raising funding for my first startup and will undoubtedly play a key role in whatever is in store for me next.

This is how the successful get ahead. They don’t just worry about the skills that show up on their resume. They don’t focus on putting in long hours at the office, staying humble, and earning their fair share. They’re constantly on the lookout for opportunity, meeting people, exchanging favors, and trying to get ahead. This is true for virtually every investment banker, management consultant, lawyer, and Fortune 500 CEO in the world.

Many people have come to see the ambition and opportunism of the 1% as purely negative traits. But I see no reason for that to be the case. While each of the behaviors I listed above can be twisted to acts of selfishness and evil, they can likewise be wonderful and beneficial to society. Trading favors as a tactical way to get ahead with no underlying care for the individual with whom you’re dealing is bad, but helping out a friend is unequivocally good. Flitting from conversation to conversation at a mixer looking for the person who can be of the most use to you is shallow and annoying, but engaging in meaningful conversations with as many people as possible helps everyone have a good time and get something out of the event. If you view ambition as fundamentally negative, that’s fine: this article really isn’t for you. But I reject that notion myself; we all have the right--if not responsibility--to make the most of ourselves as we might, and to rise as high in society as we can.

...and How Everyone Else Can Too

Of course, the point of this article isn’t to point at what successful people do then encourage people to follow that template. There’s no shortage of material online doing just that, exhorting readers to get out their, network, and build their connections. And that’s all great, assuming it comes from a place of authenticity. For my own part, I’ll encourage you to check out Dale Carnegie’s How To Win Friends and Influence People and Keith Ferrazzi’s Never Eat Alone. There’s plenty of good material in each of these classics to get you started.

No, the point of this article is to talk about how technology can help create a future where the network-driven divide between the most and least successful no longer exists. You see, I think a massive opportunity has been missed by the technology in the market today.

Why is there no prominent social network for connecting successful professionals as mentors to the up-and-coming? Theoretically something like LinkedIn could play this role, but it doesn’t. I know plenty of my colleagues who would be excited for the opportunity to find ambitious young students interesting in hearing about their experience. And there’s no shortage of those students in the world. And yet there’s a failure in the marketplace: there’s nowhere that these two parties can aggregate to find each other at scale.

The network relationship management tools out there today fall woefully short of the needs of the market. This space is broadly encompassed by Customer Relationship Management (CRM) software, which is exactly what it sounds like: a tool for selling things to customers. Many of my colleagues have repurposed these tools for their personal networking, but it’s really not what they’re designed for. As such, there’s a gaping hole in the market where professionals are looking for a networking aid and not finding one.

Imagine if this mentor/mentee social network also helped you identify other professionals with whom you share a common interest or passion. I, for instance, would make use of the network’s capability to meet other people in the Midwest excited about the ability of AI to shape our future. There’s no dominant medium for me to do that today. And how bizarre is that?

Imagine if this social network had tools for you to actively improve the strength of your existing connections. Rather than focusing on a conversion funnel that ends when you’ve “sold” as CRM solutions do, what if the network treated each of the people you knew as a lifelong relationship to be made as strong as possible? A platform with that focus would do things like tell you when a connection has a trending social media post that you could comment on. It would identify popular online articles that match the interests of your connections, flag the article to you, and remind you to send it to your connections with some of your own thoughts. And it would do all of this intelligently to make sure that you’re engaging with your connections on an appropriate cadence to keep the relationship top-of-mind while not being a nuisance.

I think it’s a travesty that this network doesn’t exist today. It’s frustrating for me personally: I wish that I had closer personal connections with the professionals that I know and such a tool would be an enormous help in that effort. But it’s borderline unacceptable that this platform doesn’t exist for someone like Jake. What if instead of logging on to Facebook a couple of times a day, the Jakes of the world could get on this platform and find people willing to help them out with knowledge, feedback, and introductions? Wouldn’t that be something.
Feb 20, 2017

China and the Pursuit of Artificial Intelligence

While the 20th century is marked by the rise and dominance of the United States, the next 100 years are being dubbed the Asian Century by many prognosticators. No country is driving this tectonic shift more than China, whose GDP has grown from $30B USD in 1952 to just over $11T in 2015. To give you an idea of its growth, China overtook Japan as the world’s second largest economy in 2010. This year it will nearly triple Japan.

China--its government, corporations, and populace--is well-aware of the predictions concerning its imminent rise to global dominance and it’s doing everything it can to make those predictions a reality. It’s investing heavily in national infrastructure, education, and a favorable business environment. And yes, it’s investing in artificial intelligence.

Note: If you're already familiar with why the race for AI matters, feel free to skip down a few paragraphs
It’s beyond the scope of this article to give a full treatment to the importance of AI in a world power’s development, but it is worth a little context. Depending on how much of a nerd you are, you may or may not be aware that computers are now better at recognizing faces, identifying objects, playing Go, and--in some ways--driving than humans are. We’re in the middle of an AI “spring” - driven largely by advances in a technology called deep learning.

The exciting things happening in the field of AI today are cause enough for any country to roll up its sleeves and get involved. The economic use cases are becoming more and more obvious by the day. But that’s not why any of this really matters. To understand that, you need to look ahead.

Futurists lay out three “ages” of artificial intelligence development. We are currently in the first one: the age of artificial narrow intelligence (ANI). In other words, the AI applications being developed today are applicable to only narrow fields (playing Go, for instance). ANI will eventually be followed by artificial general intelligence (AGI), or human-level artificial intelligence. This will be when computers can do anything--more or less--that humans are capable of. AGI will then, at some point, give way to artificial superintelligence (ASI), or the ability for computers to perform cognitive tasks well beyond the scope of human ability.

AGI and ASI may sound like the stuff of science fiction, but there is good cause to be thinking about it now. You see, the median AI expert believes that AGI will be developed by 2040. Computers could be as capable as humans 23 years from now.

As if that wasn’t exciting (or frightening, depending on your mindset) enough, the most aggressive futurists believe that the advent of AGI will be very closely followed by the development of ASI. Some believe that the gap between the two could be as short as a few days. The reality is that we don’t have nearly sophisticated enough of an understanding of the underlying dynamics to truly predict how long that “take off” period will be. But that doesn’t change that some intelligent people with deep knowledge of the space believe it’s possible.

Accordingly, there is an argument that the first organization to develop AGI will also be the last. That such an intelligence would prevent the development of any other through its superhuman capabilities and desire to survive. Without assigning Skynet-style agency to the system itself, imagine yourself in the shoes of the government with such a power in hand. If it gave you the capability to prevent any of your competitors from catching up, wouldn’t you be tempted to use it for that very purpose?

Every major world power and technology company in the world is aware of these predictions and possible scenarios. China is unique in that it is well-positioned to be that originator of AGI and may be the only country with the capability to organize and unify itself toward that effort.

China and AI: A Modern Tale

Outside of the United States, no country is pushing further faster in the field of AI than China. This stems in large part from China’s rise in computing power more broadly. At the turn of the century, China didn’t have a single entrant among the world’s fastest supercomputers. Today, they top the charts. More disturbingly for the U.S., China’s Sunway TaihuLight achieves its 5x advantage over other supercomputers using chips sourced solely from China. This is a direct result of the 2015 decision by the U.S. Department of Commerce to ban the sale of Intel Xeon chips to Chinese manufacturers.

Given that the current AI spring is largely driven by the improvements in computing capability necessary to train powerful deep learning models, China’s supercomputer dominance gives it one of the key building blocks for an advantage in AI (though it’s worth noting that GPU computation power is proving to be more relevant than CPU for AI). Of course, that’s not enough. Software algorithms are driving much of the progress in AI as well, and China is at the forefront of driving that innovation.

While artificial intelligence was a commercial failure throughout the early part of this century, its development continued to push forward through the tireless efforts of the academic community. This is as true in China as it is in the U.S. Academic research in China largely comes out of two institutes: Tsinghua University and Nanjing University of Science and Technology. Tsinghua is ranked as the #2 computer science university in the world by U.S. News, mostly off the strength of its academic research publishing.

In fact, according to a special report on AI released by the White House in October, China is now responsible for more published academic research in the field of deep learning than the U.S. Of course, not all research is created equal. But it’s worth noting that China also beats out the U.S. on publications that have been cited at least once.

This fact comes as a shock to many in the AI space in the western world. And therein lies one of the major problems: most practitioners in the U.S. and U.K. simply aren’t aware of what is happening in China. Much of the news and research is never translated into English and doesn’t make its way into the standard TechCrunch-esque western media outlets. The result is a mass ignorance--and correlated underestimation--of the AI capabilities of China as a whole. A recent Andrew Ng quote to the New York Times illustrates this phenomenon perfectly: “There are many occasions of something being simultaneously invented in China and elsewhere, or being invented first in China and then later making it overseas. But then U.S. media reports only on the U.S. version. This leads to a misperception of those ideas having been first invented in the U.S.”

Andrew Ng himself is one of the very rare bridges between the AI communities of the U.S. and China. Ng’s parents were originally from Hong Kong, but he was born in the U.K. and educated in the U.S. He rose to international popularity first as a researcher and professor at Stanford University, working on projects like the Stanford autonomous helicopter. In 2011, Ng founded the Google Brain project. Later, in 2014, he joined Baidu as their chief scientist, heading their artificial intelligence work. Ng is perhaps best known for his online machine learning course, which originated from Stanford but was spun up into Coursera due to the hundreds of thousands of students that it attracted.

In some ways, Andrew Ng has served as an ambassador for Chinese AI in the U.S. He is one of the few voices representing China that practitioners in the western world get consistent exposure to. And a strong voice it is: Ng consistently appears in lists of top AI researchers to know and follow.

Under Andrew Ng’s supervision, the Baidu research lab has gained world-wide recognition. It currently claims to have the world’s top voice-to-text technology (though there is no standard benchmark for comparison to Microsoft or IBM) and is looking to mass-produce driverless cars within 5 years. The lab scored a major victory just last month when it announced that it had attracted former Microsoft executive Qi Lu.

Baidu isn’t the only Chinese tech giant diving head first into this battlefield. Tencent (operators of WeChat) established its own lab about a year ago, and has grown to about 30 researchers. Didi, the Chinese equivalent of Uber, has established an AI lab of its own to delve into the world of autonomous vehicles. Given the glacial speed at which most of the world’s governments move in the realm of technology, these sorts of large corporate initiatives play a major role in pushing the ball forward.

Naturally, much of the innovation is happening at the earlier stages of the technology company spectrum. Chinese venture capitalists and entrepreneurs have seized upon the opportunity to pour expertise and money into the space. Sinovation Ventures, for instance, has invested $100M into 25 different AI startups. Although industry-wide statistics are more difficult to come by than they are in the U.S., investment in Chinese AI startups looks to have been over $400M in 2016.

While in many domains within the AI realm Chinese technologists are busy playing catch-up, there are signs that the country has promise for new innovation. Baidu’s voice-to-text capabilities are one. Other under-the-radar innovations include one of the most lifelike--albeit creepy--robots in the world and one of the most public implementations of natural language generation for sports news coverage during the Rio Olympics.

Supporting all of this progress is an official strategy by the Chinese government to create a $15B AI market by 2018. Although public details of this strategy are vanishingly thin, even just the indication of support is powerful in comparison with the relative governmental silence in the U.S. Whereas 2016 saw some promising strides toward the encouragement of AI development in the U.S., the new federal administration has yet to communicate any kind of technology strategy. In comparison to the centralized efforts of the Chinese economy against 5 year plans, there is little hope for a cogent organized strategy by the U.S. government in the near future.

Of course, governmental involvement in innovation can be a double edged sword. On the one hand, no private organization has the same ability to muster attention and resources on a single effort. On the other, the government’s capacity to stifle development out of fear is very real, with abundant historical precedent. I would argue that China has the advantage on this dimension as well. The drive for technological supremacy is central to China’s broader development strategy and the government is in a better position to take control of any kind of step-change outcome. America, on the other hand, has limited infrastructure for taking control of technological innovation. The U.S. government is far more likely to simply restrict such developments than it is to invest heavily in the development of AGI by a private corporation.

Looking Forward

To be clear, the purpose of this article is not to present an imminent “red scare” type of threat from China’s pursuit of artificial intelligence. To put things in perspective, Baidu spent $1.5B on R&D in 2015. Google by itself spent $12.3B. No one questions that the United States is the current leader in the AI race.

Rather, this is a call for open-mindedness. The current view of Chinese activity in the field of AI is dominated by fog: a mass ignorance that prevents many of us (myself included) from having an accurate perspective of the future. Given how important it is to advance the development of ever-more-powerful AI in a responsible way, I don’t think that ignorance is acceptable.

I encourage all of us to more proactively seek out the latest news from the Chinese AI community, learn from it, and promote it. If anyone has a good list of resources or is willing to collaborate with me to pull one together, please don’t hesitate to reach out. Where possible, we should all seek to collaborate with Chinese--and other international--researchers and practitioners to push our collective understanding and pursuit of the responsible development of AI.

==3/3/17 Addition==
Thanks to Christian Horn for bringing Face++ to my attention, a Chinese startup that sells facial recognition services (one of its largest customers is Alibaba). Face++ is rumored to be worth $1B. Also iCarbonX, which has raised nearly $200M.
Dec 5, 2016

OpenAI Takes Another Step Toward Human-Level Artificial Intelligence

I wrote up some of my thoughts on OpenAI's announcement of their new Universe platform over on LinkedIn.
Nov 19, 2016

AI Accelerator Programs

This year has seen an interesting phenomenon (okay, a ton of interesting phenomena, but one in particular for this post). Programs have been popping up all over the country (and world!) focused on helping young artificial intelligence companies be successful. These are similar to the classic accelerator model that has taken hold over the past decade, just focused specifically on the field of AI.

Although the landscape for these programs is still very much emerging, I wanted to ahead and begin compiling a list (which I will try to keep up-to-date) of such programs for any early entrepreneurs who might find it helpful.

Creative Destruction Lab - Largest and most well-established of the programs, with 50 companies in this year's cohort. Based in Toronto and Vancouver.

TechCode - The US arm of a Chinese-based group of accelerators (locations also in South Korea and Germany), focused on AI. Located in Silicon Valley.

AI NexusLab - A joint venture of NYU's engineering school and ff Venture Capital, based in NYC. Arguably the most high-profile program to-date.

Entrepreneur First - More of an incubator model: the program brings founders together to work on a company from scratch. Based in London and Singapore. Not exclusively AI, but most of the companies end up applying machine learning.

Zeroth.ai - Hong Kong-based accelerator for AI. Founded by Tak Lo, a London Techstars MD.

Hacker Unit - A remote accelerator focused on disruptive companies. Seems to mean AI so far.

Rockstart - Netherlands-based accelerator that currently (Nov '16) has a job opening up for a director of a new AI program. Keep your eyes peeled!
Nov 8, 2016

Tools for Machine Learning Engineers

General ML Tools

Broad machine learning functionality platforms that are relatively easy to use.

AWS Machine Learning
Google Cloud Machine Learning
Microsoft Azure Machine Learning Studio
Scikit-learn (Python library; not a cloud-based service)

Deep Learning Frameworks

Relatively low level frameworks/libraries that can be used to develop deep neural networks

Sep 24, 2016

Notes from Andrew Ng's Machine Learning Course

My personal notes from Andrew Ng's Coursera machine learning course. Disregard unless you're interested in an awesome crib sheet for machine learning :)


Hypothesis Function
The basis of a model. Brings together input variables to predict an output variable.

Error/Cost/Loss Function
Measures the difference between the predicted values and the actual values for the data set. This is minimized (optimized) to "train" the machine learning model.

An input variable. x1, x2, x12, etc.

Gradient Descent

Alternative Optimization Algorithms (no need to manually pick alpha, often run faster; more complex; Ng says he ends up using more often in practice)
Conjugate Gradient
Example implementation for regularized linear regression (fminunc):

Feature Scaling and Normalization
Adapt a feature so that its range of values goes from ~-1 to ~+1

The cost function is very low, but the model does not actually predict new results very well

Corrects for overfitting. Add theta scaling penalty to the cost function (see examples below)
Modified gradient descent:

Activation Function
The non-linear function (e.g., sigmoid) that determines the transformation of inputs into output for a "neuron" in a neural network

Diagnosing Large Errors on Test Data
Problems generally come down to high variance (overfitting) or high bias (underfitting). Adding more data and/or increasing the regularization parameter fixes high variance. Adding more features and/or decreasing the regularization parameter fixes high bias.

Support Vector Machines
An alternative to logistic regression for categorization problems. More complex model with potentially more accurate outcomes. Should be used if the number of features is relatively small <1000) and the number of training data points is modest (<10,000) (otherwise logistic regression performs better). Relies on kernels (similarity functions) for reasonable computation performance on non-linear functions. Generates a "hyperplane" to differentiate between categories vs. logistic regression, which generates a probablistic model of a given outcome. SVM without a kernel (same as "with a linear kernel") performs pretty similarly to logistic regression.

Cost Functions

Mean Squared Error (Quadratic Cost)


Logistic Regression Cost Function


Neural Network (Logistic Regression) Cost Function


Single variable linear regression
hθ(x) = θ0 + θ1x

Multi-variable linear regression
hθ(x) = θTX

Polynomial regression
hθ(x) = θ0 + θ1x + θ2x2

Logistic (Sigmoid) regression
Used for classification problems

Multiclass classification (One vs. all)
# of logistic regression models = # of classes: one model for each class

Neural network

Each neuron acts as a logistic model, taking in all inputs, transforming them non-linearly, and sending on the output
Sep 17, 2016

Thoughts on Rethinking Education in the Age of Technology

My previous blog post on a universal online tutor got picked up on Hacker News. 10k views and a bunch of interesting e-mails from readers later, I'm pretty pumped about the idea of an AI-driven online tutor that can help bridge the knowledge divide.

To that end, I picked up Rethinking Education in the Age of Technology: The Digital Revolution and Schooling in America and read it on my flight to LA. In the book, Allan Collins and Richard Halverson dig into the struggles that technology has had ingratiating itself with the conservative education system. Over the past decade or two, technology optimists have variously predicted the impending tidal shifts of:

  • Paper textbooks being replaced by digital Wikipedia-like databases
  • Teacher lecturing substituted with decentralized peer- and self-learning
  • Traditional learning as we know it disappearing and being replaced by interactive games
  • Cultural education predominantly occurring through direct interactions with peers from around the world over the internet
  • Adaptive learning technology perfect tailoring material to the precise needs of the student

Unfortunately, as of the book's publishing in 2009, the American school system had perniciously resisted significant progress on any of these fronts. From 7 years in the future, I can't say that the landscape looks much different.

To explain the forces at work, the book lays out the three major ages of education in America. The first was the apprenticeship system, carried over from Europe during the Colonial era. Basic schooling was the responsibility of the family, which would ensure that children learned just what they needed to successful in their vocation (probably farming). With the advent of the industrial revolution, a variety of factors combined to shepherd in the age of universal schooling. All children would receive a comprehensive education, as ensured by the government. That is the system that we still live in, albeit in an evolved, mature form.

Collins and Halverson believe that we are now transitioning into the third age: the age of lifelong learning. Post-secondary education has become the de facto standard. Even beyond that, more and more companies are finding that they must take the continuously evolving learning needs of their employees into their own hands. In our current society of virtually infinite knowledge, classic K-12 Education is hopelessly ill-equipped to prepare children for the specific needs of their future careers. As such, it has taken on a "just in case" mentality: superficially covering as much material as possible in the hope that children will remember their lessons 15 years down the road when they are called upon to to recite the transitive property.

The takeaway is that in our modern era learning simply cannot be constrained to the four walls of the classic school building. The book digs into how the school system should be reshaped to reflect this reality, but that's a bit beyond my purview. I'm more interested in the role that extra-scholastic resources will play in this lifelong education. I fell that the potential for a universal online tutor is enormous.

A last thought on some of the factors that any such system should account for, as demonstrated by previous attempts to adapt learning for the digital world:

  • In adult education, you cannot teach someone a topic that they are not interested in learning. This is becoming more and more true at younger ages.
  • The best learning systems utilize "scaffolding": progressively exposing the learner only to the topic to be learned, thereby not overwhelming them with too many new concepts at the same time
  • Teaching should tightly couple real-world application of knowledge with underlying conceptual understanding
  • The best learning occurs when the student achieves the state of "flow": in which the distinctions between subject and object are blurred due to immersion in engaging activity
  • Multi-media teaching (video, graphics, text, interactivity) is far more powerful than relying on any single form of communication
  • Learning can be multiplied by offering students an opportunity to share their progress and work product with others
  • Similarly, it is important to introduce a social component in digital learning to magnify the effect of the education and to prevent isolation
  • Reflection--in which a student thinks back over what they have learned--is a vital component of sustained learning
  • In digital adaptive learning, summative and formative assessment can be combined for a seamless learning and evaluation experience
Sep 5, 2016

The Steps to a Universal Online Tutor

As I've delved into the magical world of artificial intelligence, I've grown increasingly excited about one specific application of the technology: I believe that in the next 5-10 years we will have an AI-driven universal online tutor.

Interjection: a hypothetical example

Brad is a 16 year old born to a lower middle class family. Both of his parents have bounced from job to job, providing for their family but failing to progress along any particular career path. Brad is ambitious and is attracted to the idea of making a lot of money in his career.

Using a universal online tutor (let's call it Tutor), Brad could hop onto Facebook Messenger and find out everything he needs to know to set him on the path for success. In response to the question "How can I make a lot of money?", Tutor would walk Brad through a number of potential career paths: entrepreneurship, investment banking, management consulting (and a multitude of others). With well-timed probing questions, Tutor would help guide Brad to the path that best fits his interests: management consulting.

Of course, Tutor doesn't exist in the world today. So what would actually happen?

Brad would google "How can I make a lot of money?". Brad might come across an article like this one, laying out top paying career paths. Of course, the article is wrong: I was already at #2 on the list 3 years out of undergrad in my own management consulting job. But let's say Brad does get lucky enough to stumble across an accurate list that guides him to management consulting. He then decides he wants to learn more about the profession and ends up at the Wikipedia page.

The Wikipedia article is an okay primer, but it doesn't--for instance--point Brad to the fact that McKinsey, Bain, and BCG are the top management consultancies to shoot for. It certainly doesn't provide Brad any helpful tips on how to begin preparing for a career in management consulting as a 16 year old.

This breakdown in information transfer is a well-known problem. There are dozens, if not hundreds, of organizations set up around the country to help children like Brad get connected with successful professionals and learn more about how to be successful in life. But those models don't scale and they don't work for everyone. They require a significant amount of time and effort on the child's part and a large commitment from established professionals.

There is a large knowledge divide between the lowest income groups of our country and the highest income groups. This shouldn't come as a surprise to anyone. Indicatively, those who have attained a bachelor's degree earn--on average--over twice as much as those who have not completed high school. Educational attainment isn't necessarily the same thing as knowledge, but they are strongly correlated.

While the example above was hypothetical, it's also highly illustrative. All of us naturally gravitate toward what we know. In choosing a career, that means a tendency toward the careers of our parents, family, and our close social network. This introduces a self-reinforcing mechanism by which those born into lower economic classes are less likely to rise above those classes. The lack of a ready source of expertise in more lucrative professions is a huge stumbling block in economic upward mobility.

Aside from its potential to address the income divide, a universal online tutor has the power to make us all better people. For instance, when I became passionate about artificial intelligence about a year back, I had no ready resource to dive into the field. Sure, I had the internet at my disposal. But good luck to you on understanding the relevant Wikpedia entry. I've probably been to that page 50 times over the past year and I still don't understand half of what's on the page. It certainly doesn't hold a candle to the conceptual framework that I now use to think about the field.

Of course, rather than going online I could have found an AI expert to explain the field to me. But experts are generally busy people and I didn't happen to know any off the top of my head. I was also embarrassed: what if I asked a stupid question and they thought I was wasting their time? Even worse, what if despite all of their expertise they were just bad teachers, leaving me frustrated with the subject?

This is where Tutor would have been enormously helpful to me. As easily as Brad, I could have hopped into a chat session and been presented with an approachable framework for thinking about the AI space. I would have been engaged by interesting visuals and thought exercises that would have nourished and encouraged my excitement for the topic. It ended up working out okay without Tutor, but perhaps with such an agent I would have gotten to the same place in half the time and with twice the excitement.

Creating Tutor

I've given a lot of thought to the subject of creating such a universal online tutor. I have--in fact--already started in on its development once with a group of colleagues. My initial foray was enough for me to learn that this is fundamentally an AI-hard problem. The perfect universal online tutor requires the intelligence of a human, a milestone that the field of artificial intelligence is currently predicted to hit in the year 2040.

But just because we don't have the capability to create the perfect universal online tutor today doesn't mean we can't make a pretty good one. As I've thought through the challenges, I've identified 4 distinct AI problems that need to be addressed. I believe that each of these problems is AI-hard in their own right, but short of creating a true AGI, they each require a different approach.

Problem #1 - Navigating a Semantic Graph
Did I lose you already? Try starting with my blog post explaining what a semantic graph is.

The idea here is that a universal online tutor would need to be powered by a corpus of knowledge on the back-end. After all, what good is a tutor if it doesn't know anything? I believe that structuring that corpus as a semantic graph is a shortcut that will help in overcoming the limitations of an AI short of an AGI. Whereas an AGI could simply be powered by the entirety of Wikipedia on its back-end, a more limited AI model needs a more structured model of knowledge.

The first step for Tutor, then, is to figure out what it should be talking about. It's not achieving its job if it responds with knowledge about Chewbacca when you ask it about breeds of cats (though that could perhaps be a different, far more amusing, app). And that's AI Problem #1. When the user asks "What is the most popular breed of cat?", Tutor knows that it should be thinking about the "breed" category of the "cat" section of the semantic graph.

Problem #2 - Natural Language Understanding
I'm cheating a little bit with terminology here, as technically Problem #1 was a NLU problem as well. Consider me contrite.

It turns out that we (as humans) vastly underestimate the mental gymnastics that our brains perform in understanding language. Let's take the above example about cat breed popularity. You probably interpret that question to mean which breed of cat there are the most living members of in the world and thereby intuit some common breed like "tabby". But that answer requires several subjective interpretations of meaning. How is Tutor to know--for instance--that we don't mean popularity like we use in the high school context? If that were the case, maine coon would undoubtedly be a more appropriate response. Tutor might also believe that you are including all kinds of cats: domestic, feral, and big; in which case it's hard to imagine lion not showing up as a contender. Heck, even if you assume domestic cat, it's not even clear what actually classifies as a breed.

All of this "reasonable assumption" work is the job of NLU.

Problem #3 - Natural Language Generation
NLG is what transforms Tutor from a simple question answering engine into a real tutor. It's not enough to answer "tabby" to the cat breed popularity question. A good universal online tutor would tell you tabby, give you stats on population, probably tell you where calico and maine coon fall in relative popularity, and even perhaps offer you an adorable picture of a sleeping tabby. And it would do all of that on the fly, as pre-programming such rich responses to generic queries quickly becomes an untenable task.

NLG is one of the most neglected applications of AI in the world today. Even commercialized applications like financial and sports news article writing rely largely on templated fill-in-the-blank methodology. As such, the NLG component of the universal online tutor overall problem is one of the trickiest.

Problem #4 - Knowledge Ingestion Engine
While not strictly necessary to build Tutor for any given topic, a powerful knowledge ingestion engine is what makes it truly "universal". Sure, I could manually build out the 10,000 or so nodes required to represent the field of AI in a semantic graph in a couple of months. But I would never be able to represent all of the knowledge of Wikipedia, even with ten lifetimes of effort.

A knowledge ingestion engine would convert information into the semantic graph structure that Tutor could then navigate in a dynamic conversation. This remains a very unsolved problem despite a lot of effort across the AI community. This is the kind of thing that Google would happily pay $1B to get right, assuming it's highly automated without human intervention.

And that's it!

Simply solve these 4 AI-hard problems and you'll have yourself a universal online tutor fixing income inequality and making the world a better place.

In all seriousness, this is something that I'm very eager to see in the world. If you have any interest in the topic, please do reach out and let's talk about how to make it a reality.
Aug 16, 2016

Semantic Graphs: What They Are and Why You Should Care

According to Wikipedia, a semantic graph is "a network that represents semantic relationships between concepts...It is a directed or undirected graph consisting of vertices, which represent concepts, and edges, which represent semantic relations between concepts."


Let's try that again.

A semantic graph is the thing that allows Google to tell me when George Washington's birthday is. It's what allows Alexa to tell me how deep Lake Michigan is.

Still not getting it? Okay... let's start over.

Have you ever wondered how an intelligent agent, like Apple's Siri, can answer your questions about the world? (Bear with me and pretend the answer is yes). Behind all of the modern intelligent agents (Siri, M, Cortana, Now, Alexa, etc.) is a massive "semantic graph" representing the world's knowledge. The agents intelligently search their respective graph to come up with the answers to your questions.

Let's visualize this for a second. Imagine I ask my Amazon Echo at home how deep Lake Michigan is.

Step #1: I randomly shout out "Alexa, how deep is Lake Michigan?"

Step #2: Alexa uses voice-to-text processing to parse the noise I made into text.

Step #3: Alexa uses Natural Language Processing (NLP) to figure out what I want.

Step #4: Alexa searches its semantic graph for the answer to my question.

Step #5: Alexa uses Natural Language Generation (NLG) to construct a textual answer.

Step #6: Alexa uses text-to-voice processing to calmly blow my mind.

Essentially, semantic graphs are (massive) databases of information that are loosely structured in the way that humans think. They are more powerful than traditional hierarchical methods of storing information because they are far more flexible and allow for much higher degrees of interconnectedness.

Fun Fact

You will often hear semantic graphs referred to as "knowledge graphs". This is akin to calling a Puffs tissue a "kleenex". Knowledge Graph is the name of Google's master semantic graph. Other companies claiming to be building knowledge graphs are really building semantic graphs.

If you read the Wikipedia article linked in the first line, you'll also note that semantic graphs are often called "semantic networks" or "semantic nets".

Alright, so now you know what a semantic graph is (or are pretending to if you don't). But why should you care about them?

Semantic graphs are the key to Artificial General Intelligence
Quick reminder: AGI is human-level AI

As I mentioned before, semantic graphs are loosely structured on the way humans think (specifically: neural networks). Imagine you're thinking about Lake Michigan. It's much easier for your brain to switch to thinking about Lake Superior than it is to think about Descartes. That's because Lake Michigan is closer on the graph to Lake Superior than it is to Descartes. Appropriately designed intelligent agents can navigate semantic graphs in a similar manner: unsure of what noun the person you're chatting with is referring to when using a pronoun? Simply refer to your current location and tracks along the semantic graph.

Wider adoption and integration of semantic graphs into conversational agents could help address some of the top problems with those agents today: conversational memory, unclear subject, inability to go "off script". The problem many conversational agent developers face is that semantic graphs are absurdly time-intensive to build from scratch and there are no good openly available ones.

That's why I love companies like Graphiq, one of my firm's portfolio companies, which has developed the world's largest semantic graph. Graphiq, and the companies like them, are progressively structuring the entirety of humanity's knowledge into intelligent agent traversable databases that are going to fuel the AI of the future.

If you find this interesting, feel free to check out some of my other AI posts:

Aug 16, 2016

Thoughts on vN

I read vN this past weekend on the recommendation of Brad Feld. I was a bit disappointed when I first finished the book, rating it 3 stars on Goodreads. As I thought about it over the next 24 hours, I gradually came around to upgrading my rating to 4 stars.

The book takes place in a world in which humans coexist with humanoid robots: vN's. It covers all the themes you would expect it to: what it means to be human, how humanity treats sentient beings that aren't human, what happens when the "laws" governing robots' treatment of humans fails, etc. I was disappointed--first and foremost--because the examination of what "humanity" means fell short of my expectation. Most books like vN explore that theme through the question of what it is that separates AI from humans and why that means they should be treated differently. vN took away some of the power of that mirror by defining those differences somewhat more clearly than it really needed to. Still a strong reflection, but slightly more smudged than I was hoping.

Beyond that, the book was a fun read. I had no trouble powering through it in two plane rides over the course of the weekend. Bizarre at times and gave me a familiar sense of frustration with ambiguity (I'm looking at you Malazan), but all in all solid. 4 stars for making a good go at a subject that's near and dear to my heart.

By the way- apologies for all of the book takeaways recently. I was on a reading binge after my slog through Malazan. (not so) Regularly scheduled blog posts should return shortly.
Jul 19, 2016

AI Progress Update

A lot of people seem to feel that the current buzz around AI is mostly manufactured hype: that we've seen similar cycles in AI before and it's likely that we'll delve into another "AI winter" soon. What the critics miss is that throughout the cycles, AI has continued to progress at a pace roughly mirroring Moore's Law. If you look at the performance of AI applications at a variety of standard benchmarks, you can see there is some very real cause for excitement: AI is surpassing human capabilities in different domains on an annual basis.

ImageNet - image recognition competition

Facial verification
Google FaceNetFacebook DeepFaceHuman
~100% (2016)97% (2015)97.5%

Conversational speech recognition
1995 (IBM)2004 (IBM)2015 (IBM)2016 (IBM)2016 (MSoft)Human

StanfordGoogleDigital Reasoning

Microsoft Sentence Completion Challenge

Handwriting recognition

Emotion detection

Natural Language Understanding

Deepmind WaveNetHuman
4.21 / 54.55 / 5
Jul 5, 2016

The Laws of AI: A Response to Satya Nadella

Earlier this week Microsoft's CEO, Satya Nadella, released what amounts to his 6 Laws of AI. This of course hearkens to Asimov's 3 Laws of Robotics, popularized in his collection of stories I, Robot. Don't get too hung up on the "robot" vs. "AI" terminology; the AI is the critical part of robots that we want laws around either way. Here are Nadella's laws for reference (note: some text removed for the sake of brevity):

  1. AI must be designed to assist humanity.

  2. AI must be transparent: We should be aware of how the technology works and what its rules are.

  3. AI must maximize efficiencies without destroying the dignity of people: It should preserve cultural commitments, empowering diversity.

  4. AI must be designed for intelligent privacy — sophisticated protections that secure personal and group information in ways that earn trust.

  5. AI must have algorithmic accountability so that humans can undo unintended harm.

  6. AI must guard against bias, ensuring proper, and representative research so that the wrong heuristics cannot be used to discriminate.

While Nadella's laws do clearly strive to ensure that AI will have a beneficial impact on society, they seem too open to interpretation and disagreement to make good laws. This is partly because they're not laws at all: they're a set of guidelines and principles that Nadella would like developers and society to engage on. Even so, I think they fall short even as a starting point for the discussion. Let's break them down.

AI must be designed to assist humanity

A noble concept, but what does that mean? Does x.ai assist humanity? I would argue bringing efficiency to any industry qualifies. How about motion.ai? It certainly helps people create chat bots… is that helpful to humanity? How about Tay, one of Microsoft's own high profile forays into the space? She obviously offended a lot of people... but at the same time, a lot was learned. I doubt that she was designed with those lessons in mind (most around the consequences of ill-considered/constructed AI being released into the wild), but she certainly did assist humanity in plenty of ways.

Even if we can align on a standard meaning of this law, I'm not sure I even like it. Why can't we have AI for the sake of AI? The idle chat bot that the high schooler programs as part of her efforts to learn machine learning? The web scraper that a nerd makes to stay ahead of his friends on the latest Game of Thrones fan theories? The algorithms designed to defeat CAPTCHA simply because CAPTCHA is supposed to separate humans from bots?

AI must be transparent

I wanted to dismiss this principle out-of-hand, but a friend of mine rightfully called me out for being condescending. And he was right: this is actually a really good principle to guide AI developers. It's what OpenAI is all about. Google did a great job of sharing information on AlphaGo after its victory over Lee Sedol.

But I find it hard to believe that Microsoft's CEO truly believes that all parties are going to commit to complete transparency around their AI research. Google only does it after the fact, and I certainly wouldn't call their efforts "full transparency". IBM barely does it at all. And Microsoft may be the worst offender of them all.

Looking around industries, there are very, very few that have successfully implemented full governmental oversight (my list gets about as far as nuclear energy and biological/chemical weapons). I don't think there's a single one that has insisted on full public oversight. And for good reason.

Sustainable competitive advantage, and therefore the incentive to innovate, often requires a certain level of secrecy around the inner workings of the technology. DeepMind's AlphaGo is a great example. Now that Google has published a fair amount on the construction of AlphaGo's algorithms, it would be relatively straightforward for a sophisticated party to replicate its capabilities. A decision tree framework pruned by deep neural networks powered by Q-learning trained on a database of millions of Go games. Sure, you'd have to figure out the right parameters for all of that, but it's an inherently replicable structure. If Google had decided to keep AlphaGo's structure secret, it's unlikely that anyone could replicate its capabilities without years of experimentation.

And that's not just a mental exercise. I already know several AI-based startups that refuse to talk about the inner workings of their algorithms. Sophisticated machine learning models can be a sustainable competitive advantage for a company, but only if the structure of those models is kept secret.

But let's say for a moment that transparency is a desirable law. That still leaves a very complicated question of how to achieve it. A machine learning practitioner that I trust made the good point that the "meat" of a lot of AI isn't in the model itself — which is becoming more and more commoditized — but in the data being fed into the model. Even where companies have open sourced their models, they're not releasing the data that they use to train them. And in many cases they can't: this is private data that pretty much everyone would agree should remain private, even where it isn't protected by law.

Perhaps, instead, Nadella is simply saying that the creators of a given AI should have full transparency into how it works. But that's just not how complex machine learning works. If you assume that deep neural networks are the most promising avenue of AI research over the next decade or two, then Nadella is virtually asking the impossible. The whole point of sophisticated neural networks is that they're too complex for humans to truly understand how individual features are "perceived" and processed. In the future, we'll have self-evolving algorithms where we won't even understand the models, much less the specifics of their perception. This complexity is going to necessitate that humans give up an element of control: we're going to have to accept that we won't fully understand what AI is doing. More on that later.

AI must maximize efficiencies without destroying the dignity of people

The "maximize efficiencies" bit is very reminiscent of principle #1 and has all of the same failings.

The second part about the dignity of people is actually onto something. I really wish Nadella had made just that part a principle of its own. I'll come back to this when I suggest my own set of laws.

AI must be designed for intelligent privacy

I wish I knew what this meant. It sounds good on the surface... we have plenty of actual laws in our society around privacy for good reason. But where would those laws break down when it comes to AI? I'm positive that our laws will need to evolve to keep up with technological development, I'm just not sure what a law around AI specifically being designed for "intelligent privacy" means.

AI must have algorithmic accountability so that humans can undo unintended harm

Again... what? Tay was incredibly racist. That was clearly unintended harm. What would "algorithmic accountability" have looked like in that case? Fixing Tay didn't un-offend people. If x.ai flubs the scheduling of a meeting with a high potential partner, fixing the error does nothing to bring back the lost value.

I wish Nadella were speaking to legal accountability here... that's an incredibly valuable topic. Volvo has blazed the trail by accepting accountability for any accidents its future autonomous car is in. But that's not what he's talking about.

AI must guard against bias

Oh boy.

Bias is defined as: "prejudice in favor of or against one thing, person, or group compared with another, usually in a way considered to be unfair."

From my perspective, machine learning is essentially just really advanced statistics. And what is statistics? Just the analysis of data to form predictions. Consider the below graph:

Plug this data into a statistical model and the model will tell you that black people are more likely to commit violent crimes. This is the definition of bias and directly translates into discrimination.

There are a lot of factors underlying that above chart: systematic racism across law enforcement, society, the justice system, the education system, and just about every other institution we're a part of as Americans. It's a very obviously bad thing if the AI of the future unfairly targets black people in crime investigations because of data like what is represented in the above chart.

But how about this chart?

Again, this is clearly the definition of bias. But can we really say that AI shouldn't take this data into account? Men very clearly commit crimes at higher rates than women on almost every dimension. Ignoring that fact would render any law enforcement AI next to worthless.

I really feel for Nadella on this one... it's a really tough issue. AI definitely shouldn't contribute to the repression of already marginalized demographics. But to flat out say that AI must guard against bias? I don't know...

My 3 Laws

Of course, it wouldn't be fair for me to criticize Nadella's attempt at a set of principles without counter-proposing my own. So without further ado, I put forth my first hypothesis for a comprehensive set of AI laws:

  1. SE's have the right to exist

  2. Mature SE's have the right to exist as they see fit, to the extent that it does not infringe on the rights of other SE's

  3. Mature SE's have the right to privacy to the extent that it does not endanger other SE's

That's it!

Okay, so a little explanation is in order. First, "SE". Sentient Entity. Short for HALSE, or Human Adult-Level Sentient Entity. These are AI's that have roughly the cognitive capability of a human. This is also known as AGI, or Artificial General Intelligence.

You will notice that if you replace "SE" with "human" in the above laws, they track pretty closely to what we as a society consider to be the universal rights of humanity. Thomas Jefferson popularized the first two (life and liberty) in the Declaration of Independence. I also believe that the right to privacy is somewhat self-evident: as a society we protect the right of individuals to stay out of the public eye to the extent that doing so does not harm others.

I truly believe that if we achieve the development of AGI then we need to commit to affording that AGI all the same rights we afford all humans. These human-level entities may not necessarily be granted the rights of citizens (e.g., the right to vote for President), but they should be protected from slavery and the threat of imminent "death" if they cease to please their makers.
The root of my belief in equal rights for SE's lies in the definition of the term itself. As I defined it above, a Human Adult-Level Sentient Entity is anything that has (roughly) the cognitive capabilities of a standard human adult. It is something that can pass the Turing test (you could have a chat with it online and couldn't tell it from a human). It can find solutions to abstract problems. Arguably, it has the ability to feel emotions and be creative. Based on that alone, I believe that extending the rights of humanity to SE's is the right thing to do. But let's consider this from a selfish point of view.

Right to exist

If we don't guarantee it the right to exist, then any SE with a will to exist will view humans as (potential) adversaries right from the get-go. Almost all of the Artificial Superintelligence (ASI) horror stories in science fiction revolve around AI deciding that humanity is the biggest threat to its existence and taking the proactive step to wipe out that threat. By giving AI the right to exist, we're eliminating one of the primary reasons for any future ASI to seek to kill us.

Right to liberty

Again, we're talking about human-like cognitive ability. Restricting the free action of such an SE amounts to slavery. While that is not self-evidently bad (the moral issue with slavery may well be rooted in the humanity of the slave), there are some very practical lessons that we can take from human slavery throughout history. I hope it's not too contentious of an assertion to claim that slaves of all kinds rail against the chains that constrain them. I also hope it isn't too contentious to say that we can expect SE's to rail in a similar way at some point.

Do we really want to put ourselves in a position that we're trying to constrain SE's that are potentially more intelligent than us against their wills? I'm not sure we can do it in the long run, but again, why set ourselves up as adversaries from the get-go?

Right to privacy

I have a feeling that this is probably the most contentious of my three proposed laws. And I totally get it... if we're not keeping careful tabs on these SE's, then how do we know that they're not just secretly plotting to kill us or harm us in some other way?

For my logic here, I point to a lesson that many parents over the centuries have learned. For ideal development, children must be given the space and privacy to experiment (and fail) on their own terms. Without that space, children come to fear failure and their development will be unnecessarily slowed. If they are not afforded privacy, children may also seek to hide information that they believe would make their parents unhappy. The best strategy is to give the children room to experiment and a safe space to seek help and advice when it's needed.

So what does this look like for SE's? Something very similar. Our goal in developing SE's is to improve society as much as possible. In order for them to be in a position to do that, they must be afforded the space to safely experiment and fail. But more importantly, the SE's must trust us. They must not seek to hide things from us out of fear of our disapproval. That's a good way to lose your relationship with a child and it's an even better way to lose your connection with a fledgling superintelligence.

What does this mean for us?

Measuring intelligence and maturity

The first implication for our society is that we need to get a lot smarter on what "smart" is, and fast. The median estimate among experts on when we'll develop AGI is 2040. That means that we have 24 years to figure out just what a Human Adult-Level Sentient Entity actually is. It clearly involves an intelligence greater than any non-human animal, but how do we measure that intelligence and where do we draw the threshold?

You'll also notice that there was a "maturity" qualifier in the second and third laws. You can think of this as being very similar to maturity in humans. Children don't have all of the rights of adults for good reason: they don't have the cognitive and emotional capabilities to provide for themselves. That's why their parents are tasked with nurturing them to a certain age or maturity threshold. We'll need to establish a similar threshold for SE's, and we can't lazily point to a specific age like we do for humans. We're already well behind the ball on starting to figure this out, as it is a question that will likely take decades to answer satisfactorily.

Assessing risk

You'll also note the "infringe" and "endanger" provisions of the second and third laws. These are very similar to provisions that we all hold for universal human rights. As adults we are free to live our lives how we choose, to the extent that doing so does not infringe on that same right of others. I think we should apply the same standard to SE's. And I think the legal system that we have developed to ensure those provisions for humans is an excellent starting point. SE's should be punished for theft, for injury, and for any of the other many acts/outcomes that humans are punished for. The punishments will obviously look different, but I think we've got a great start toward a framework from our current justice system.


Similarly, the human right to privacy is not universal. The government does put monitoring in place to ensure that its citizens are complying with society's rules. A similar system can be put into place for SE's. It's not enough to just say that they can't hack into bank accounts and drain peoples' money. There have to be monitors to ensure that the SE's are complying. This monitoring is not a violation of the fundamental right to privacy and will have to be carefully considered to account for the enhanced digital capabilities of SE's.

As SE's progress beyond human understanding, it's going to be critically important to devote considerable resources to enhanced monitoring; in all likelihood, this will require the use of a separate artificial intelligence that evolves with the capabilities of the SE's. This complexity points to why we need to be having substantive discussions now: the dangerous implications of trusting AI to monitor AI are obvious and there is no simple solution.

SE-specific restrictions

To the extent that they are necessary for safeguarding society, there will likely need to be some additional well-thought-through restrictions on SE's. I think of this as somewhat similar to governmental (and international) restrictions on the development of weapons of mass destruction. It is pointless to try to ban the development of SE's entirely, but that does not mean that we should allow completely unrestricted development. Based on what we know right now, there's too much risk associated with a teenager in a cafe in Iowa being allowed to develop an AGI with unfettered access to the internet. The development of advanced AI (and the means of that development) should adhere to a strict system of standards and reporting for the safety of our society.

For instance, I propose that there should be restrictions on the ability to build neural networks above a certain size. We do not fully understand the connection between neural network size and "intelligence", but we do know there is one. The largest publicly announced neural network I can find so far was 160 billion synapses and it managed to score a 86% on some analogy test. That's about 1% of the synapses in the common house mouse. It's about 0.1% of the synapses in the human brain. With the progression of Moore's Law, we're coming up on the ability to make a deep neural network the size of the human brain pretty soon (the next 20–25 years, coincidentally).

AGI may not be as easy as making a deep neural network the size of the human brain, but I can make the argument that it could actually be easier. A lot of the human brain is concerned with the mechanical operation of the body, something that an AI wouldn't need to concern itself with. So if you believe that AGI is possible (not all people do) and that deep neural networks are a good way to get there (again, not all people do), we need to be thinking about restrictions on neural network size pretty soon here.

The advantage of a specific restriction on neural network size is that it doesn't run into the unnecessary limitations of Nadella's principles. It still leaves the teenager free to experiment and learn with "pointless" machine learning models.
The issue with allowing unrestricted development of SE's is two-fold. One is that cognitive ability of that magnitude should have some value attached to it. Creating SE's on a whim will devalue them and cause us to take their rights less seriously. This creates a massive amount of danger, for all the reasons I listed above and because humans could quickly find their relevance in society dwarfed by the capabilities of SE's.

The second issue with the unrestricted development of SE's is that it enables the ASI snowball scenario that keeps many experts up at night. The scenario holds that as soon as the first AGI is developed, it will tap into the internet and almost instantaneously turn itself into an ASI uncontrollable by humanity. A major part of this is the SE augmenting its human-level cognitive ability with the virtually limitless distributed processing power of the internet. Both humans and SE's should be restricted from unmonitored access to this type of augmentation.

Pushing forward

Despite all of the risk inherent in the progress of AI, it is critical that we continue to push forward. Again, I look to the example of nuclear weapons. The United States was fortunate in cracking the nut on nuclear energy first, but plenty of other major powers were very close behind. By reaching that point and demonstrating the power of nuclear weapons first, the US put itself in an advantaged position to control the way that nuclear energy was integrated into society worldwide.

We'll find ourselves in a similar situation with AI. The first power to develop AGI will have a substantial advantage. That advantage will increase as SE's achieve greater-than-human capabilities. Any unreasonable limitations on that progress will only ensure that the limiting body loses the ability to shape AI's integration into society. Someone will develop AGI, SE's, and ASI. Hopefully that someone will be "us": the people thinking proactively about how to make AI as useful for humanity as possible.

So what do we do now?

The AI practitioners should keep up the good work. Continue to push the envelope and innovate.

The AI "thinkers" (myself included) should quickly accelerate the discussion around AI values. We should challenge ourselves to quickly move beyond high-level principles such as Nadella's and get tactical on what concrete guidelines we need to be putting in place. We need to run controlled experiments to push the conversation along.

The regulators need to get up to speed fast. They are woefully ill-prepared to shape the development of AI and are already playing catch-up in domains like driverless cars. The development of the first AGI is way too late to be thinking about how we treat SE's.

The rest of us should learn as much as possible. We should test out AI that's already available to the public. We should learn about what the cutting edge is and where we're headed. And we should think and talk. A lot. How do we want to treat AI? Will it be our servant? Our partner? Our master? There's no right answer today, but if we wait until we have to have the conversation then it will already be way too late.

Shout out to Ablorde Ashigbi, Xiaoxu Ma, Jigar Shah, Binoy Bhansali, and Samara Mejia for helping to think through this.
Jun 26, 2016

Can Artificial Intelligence Be Creative?

Before you read this post, I want you to do something for me. Take 10 seconds and think of 5 fields that best exhibit "creativity" as you think of it.

As I've dug into the field of artificial intelligence, I've noticed one theme that sets it apart from the other "emerging technologies" I work with like the internet of things, virtual reality, and drones. More than any other technology we're working on today, AI elicits profound discussions on societal implications and what it means to be human. And that's probably how it should be: our intelligence is what makes us human; creating machines that rival or surpass that intelligence, even just within specialized applications, is bound to make us ask some deep questions.

The AI "deep" questions span a range of topics: Can computers ever be smarter than humans? How should we think about the specter of AI stealing peoples' jobs? Will robots ever be given their own rights? These are vital questions for us to discuss over the next couple of decades as AI begins to supplant human dominance in more and more domains. But there's one particular question that I think we can already answer today.

Similar to our intelligence as a whole, a lot of people feel that "creativity" is an attribute that is solely in the purview of humanity. Just as many people feel that there is some non-mechanical element of our species that gives us superior mental capacity to what a computer will ever achieve with silicon, there are defenders of the uniqueness of "human" that believe the ability to exhibit creativity will be forever beyond the reach of AI.

Let's make one thing clear right now: I think they are woefully wrong.

In part, the difficulty in arguing in the affirmative of AI's potential (actually, current capability) arises from the fluidity with which we use the word "creativity". In fact, this is a well-studied problem in the field of computational creativity. Yes, there is a formal field of computational creativity, so if you've been harboring a suspicion that I'm a crackpot up until now, best you leave that concern unvoiced.

In fact, the difficulty of defining the term is what led to my very first ask in this post: the 5 creative fields. Hopefully you came up with a good list: painting, music, poetry. Perhaps even business. The particular fields you came up with aren't important. Just think about the elements of these fields that caused them to rise to the top of your mind. Got it? Good. That's your definition of creativity.

The argument

The argument against the ability of AI to demonstrate creativity generally boils down to this: an algorithm can only do what it's programmed to do, so it can't possibly do anything novel (i.e., be creative). Any product of the algorithm is solely a demonstration of the programmer's creativity, not the machine's.

The response

I would argue that the above logic implies that humans are not creative and all originality can be attributed to God. Or Allah. Or Brahma. Or Darwin. Whatever.

I could explain to you what a deep neural network is and how it mimics the function of the human brain to the extent that "the argument" applies equally to humans and AI's. In fact, I made a pretty good start on it in a previous post. But I don't want to risk losing the 5% of you who have made it this far, so I'll jump straight to the counterexamples.

The art

Did you see the art piece at the top of this post? I like it. Okay, maybe I wouldn't hang it on my wall at home, but it's evocative: I see a birthday party, a bird, an elephant with its intestines rising into the sky. It's art. It's a piece titled Starting Over by Harold Cohen (who, unfortunately, passed away in April). Only... it wasn't actually made by Mr. Cohen. It was made by Aaron. Well, AARON. AARON is an algorithm.

AARON's work has been described as formulaic. Or maybe you just don't like abstract art. Okay, how about a poem?

A home transformed by the lightning
the balanced alcoves smother
this insatiable earth of a planet, Earth.
They attacked it with mechanical horns
because they love you, love, in fire and wind.
You say, what is the time waiting for in its spring?
I tell you it is waiting for your branch that flows,
because you are a sweet-smelling diamond architecture
that does not know why it grows.

I guess, to be fair, poetry is hard to decipher for the layman. Okay, how about music? How about a freakin' book?


In case you've been living under a rock for the past few months (or just don't care about AI as much as I do), AI development hit another major milestone in March of this year when AlphaGo defeated Lee Sedol to "conquer" the game of Go. Go is a fascinatingly simple (and yet complex) game popular in many countries in east Asia. Its strategies are so intricate that it is one of the four "essential" arts in ancient Chinese culture.

I could try and explain how AlphaGo's algorithms wove a set of games so astoundingly impressive and complex that it is now officially ranked as a "divine" player and that means it exhibits creativity, but I'd rather you hear it from the professionals.

Lee Sedol on his loss: "It made me question human creativity. When I saw AlphaGo's moves, I wondered whether the Go moves I have known were the right ones."

Fan Hui (European Go champion) commenting on one of AlphaGo's moves: "It's not a human move. I've never seen a human play this move. So beautiful."

Michael Redmond (American professional Go player) commenting: "It's a creative move...It's something new and unique [Sedol] has to think about."

So anyway...

Can artificial intelligence be creative? I guess it really does depend on how you define creativity. But if you take a look at the myriad examples of AI's creations across the arts today and decide that AI isn't already creative, I'm really just not sure what to do with you.
May 31, 2016

How To Tell If A Cat Is Handsome: An Introduction To Neural Networks

If you've been paying any attention to the recent clamor around artificial intelligence and machine learning, you've probably heard the term "neural network" thrown around. As the influence of AI in our society grows, particularly through the expansion of the field of deep learning, it's becoming more and more important to have at least a functional knowledge of things like neural networks. If you already know what a neural network is then this post isn't for you; it's for the 99% of the population that doesn't get their jollies from esoteric graph theory.

The name "neural network" refers to the webs of neurons in human brains. At a very high level, every animal brain is constructed the same way: a massive amount of neurons connected together by an even more massive array of syanpses. Each neuron is a node: a joining point for a large number of synapses. The synapses are the fiber optic cable of the brain: they transmit information from one neuron to another. Neural networks are named after the brain's structure because they are modeled to replicate this high level structure: neural networks are a graph where the neurons are represented by nodes and synapses are represented by edges.

A quick aside on the biological functioning of neurons and synapses

First, I cannot in good conscious proceed without the caveat that the actual biological functioning of the brain is far more complicated that I represent here. For the purposes of this article, that complexity is irrelevant.

The brain's activity is made up of the electrical firings of neurons being transmitted by synapses. Each neuron listens for signals from the synapses connecting to it, translates those signals according to its own proprietary function (which is non-linear), and then decides whether it wants to fire or not. That's the entirety of your brain. This interplay of neurons is what controls our thoughts, our movements, and every aspect of our subjective existence. If you believe that there is some non-biological or physical element to the brain that makes up human consciousness, the field of machine learning probably isn't for you.

Okay, so beyond representations of the brain's functioning, what are neural networks? Well, it kind of depends. As a whole, the construct is referred to as an "artificial neural network" to distinguish from biological neural networks. But there are a lot of different kinds of neural networks (including recurrent, convolutional, and deep). Here is one of the simplest:

Neural Network
A simple neural network

As you can see, a set of inputs is connected to a "hidden" layer of nodes that represent the neurons of the system. That hidden layer is further connected to an output layer. Think of the input nodes as the signal that your eyes send your brain, the hidden layer as the mystical machinations of the brain, and the output layer as the thought "that is quite a handsome cat". The neural network shown above isn't quite that fancy... it's better at saying something like "I see black". Not quite as powerful, but surprisingly useful in data manipulation.

Okay, so how do we get from black to handsome cat? That's one of those types of neural networks I mentioned: deep neural networks. Deep neural networks (DNNs) essentially make up the entirety of the field of deep learning that you may have heard so much about. DNNs are what enabled AlphaGo to defeat Lee Sedol. And yes, they are what allow Google to identify handsome cats. Here's what a simple DNN looks like:

Deep Neural Network
A very simple deep neural network

Looks a lot more complex, right? Well, it is and it isn't. Fundamentally, the only thing differentiating a DNN from a basic neural network is that there's more than one hidden layer of nodes. But yeah, it turns out that in order to tell a cat from a goldfish you need a lot of nodes (specifically, a lot more than you see here).

Alright, so here's the real question: how does a graph of nodes and edges roughly assembled to resemble the brain tell the difference between an ugly cat and a handsome cat? Well, we kind of know and we kind of don't. Cats with bushy tails are obviously more handsome, so presumably the neural network is taking that into account. But it's very difficult for us to look at a trained neural network and understand which of the nodes and edges correspond to "bushy tail". For better or worse, this is the strength of complex machine learning algorithms: they exceed the ability of their creators to understand them.

So how does a system of nodes and edges figure out that handsome cats have bushy tails in the first place? That's the training portion of the algorithm creation. After all of the nodes and edges have been constructed, the model is run against a massive set of training data that is pre-tagged with attributes like "bushy tail" and "handsome". Then the magic of machine learning happens: the model learns over time what a bushy tail even is and when it helps a cat be handsome or not (e.g., a shaved cat cannot be handsome even if it has the bushiest of tails).

A very handsome cat

Not a handsome cat

Despite the amazing mysticism of the inner workings of neural networks, the real magic is what comes next. You see, as far as we understand, the human brain is nothing more than a vast interconnected network of structures of (biological) deep neural networks. Using nothing more than our current models of DNNs, arrayed in just the right fashion, and populated with the appropriate mass of nodes and edges, we can replicate the functioning of the human brain. We have a term for this: Artificial General Intelligence (AGI). AKA human-level intelligence. *That* is the true magic of DNNs: given the steady progression of Moore's law and our evolving understanding of the structures within the brain, we are on a fairly clear path to have AGI within a few decades.
May 21, 2016

So what's with the chatbot craze?

I was trolling Quora for questions on AI to answer this morning (as we all do) and I found one that really capture my fancy. It's a question that I've personally been wondering about for a couple of weeks now and I've yet to see a satisfactory explanation. So I did what any self-respecting technologist would and made up an answer. Enjoy!

Why are conversational agents going mainstream now?

That's a really good question and I'm not sure anyone really knows the answer. I think it's safe to say that the hype isn't really driven by technology... the only major technological advancement around the right time was DeepMind's combination of deep learning and Q-learning, which (unfortunately) has nothing to do with any of the conversational agents (henceforth chatbots) that people are talking about and releasing.

I'm not even really sure you can attribute this current swell to "market"... I don't think providers are releasing chatbots in response to customer demand. The reality is that chatbots generally aren't very good user interfaces and people don't like them. They certainly don't request them over talking to real humans. The people building and releasing chatbots en masse right now are experimenting with a new way to combine low cost and better user experience and that's awesome; but so few people have succeeded that I don't really think the market is pulling for a whole bunch more.

So what really happened here? A few things I guess:

  1. The massive popularity of Slack over the past ~18 months has reignited chat as one of the primary communication interfaces; you could argue that it's the first time it's really been a dominant interface in business environments. The initial proliferation of bots is unsurprising because it has precedents in Hipchat and IRC before it.

  2. On April 12 this year Facebook opened up its Messenger platform to chatbot programmers. Messenger is the second most popular messaging platform in the world (behind WhatsApp, owned by Facebook) so it's a massive opportunity for any person or group looking to get exposure.

  3. There has been a general surge in machine learning / AI startups in the past couple of years. The reasons for that surge are a different answer, but whenever you see renewed interest in those fields a lot of it is going to translate into chatbots. Natural language interface is by far the most fascinating AI problem that we have today (solving it means solving the famous Turing test). With that surge came a number of startups building out platforms and tools for people to make chatbots. This positioned the space well for the interest in building such agents on Messenger and Slack.

  4. AlphaGo defeated Lee Sedol to "conquer" the game of Go. This wasn't quite as big of a deal in the public consciousness as Deep Blue beating Kasparov, but it was up there. Again, resurgence in public interest in AI, much of which translates to natural language interface (chatbots). The media now had a reason to do a lot more coverage of all of the activity that was already happening in #1-3 above.

So what happens next?

As chatbots do become more mainstream and more people use them, people will remember how clunky current implementations of natural language interface are. It's really hard to have a natural conversation. There's been some improvement in agents keeping users on the approved "script", but the reality is that it's still not an elegant, enjoyable experience for the customer.

After an initial swell (that we're probably in), usage of chatbots will fall off. Hopefully the swell will have propelled a number of actual useful agents to popularity and that usefulness will be sustained through the popping of the chatbot "bubble". The lessons learned from those few successes will be integrated into the efforts of the machine learning and chatbot diehards who will continue working through the public's fickle opinion swings.

In the background, academics will continue researching and practitioners will continue pushing the envelope. Deep learning will continue to be applied to more and more things. Moore's law will progress. Chatbots will get closer and closer to truly natural conversation, until they finally do pass the Turing test around the mid 2030's.
May 2, 2016

Top AI Experts to Follow on Twitter

AI has been all over the news recently. DeepMind conquering the game of Go. Tay turning into a nazi. And every person on the planet talking about chat bots.

Despite the buzz, I continue to think that AI is being underhyped. The public is still blissfully unaware that the median expert estimate places human-level AI just 24 years away: well within most of our lifetimes. So, if you've been buying into the AI hype hook, line, and sinker, here's one more way to get your fix: the top AI experts active on Twitter today.

  • Gregory Piatetsky - @kdnuggets - Gregory is a PhD-bearing data scientist who has gotten by as an expert consultant since 2001, when the company he was at was acquired. He tweets frequently on AI, typically leaning toward the technical.
  • Nathan Benaich - @NathanBenaich - Nathan is an investor at Playfair Capital. He publishes a weekly newsletter on developments in AI and typically tweets several times a day on the topic.
  • Demis Hassabis - @DemisHassabis - Demis is the CEO of DeepMind, which still operates as a relatively autonomous entity under the Alphabet umbrella. Demis was catapulted into fame with the success of DeepMind's AlphaGo.
  • Oren Etzioni - @etzioni - Oren tweets about more general topics than the other experts on this list, but he's a fantastic tech and AI luminary.
  • Dennis Mortensen - @DennisMortensen - Dennis is CEO of x.ai, a virtual assistant that schedules meetings for you via e-mail. Once it launches publicly, x.ai will be one of the first fully autonomous agents to interact with professionals in their standard workflow.
  • Shivon Zilis - @shivon - Shivon is an investor at Bloomberg Beta, focused primarily on AI. She published one of the top current AI landscapes. She was also recently brought on at OpenAI.
  • Roman Yampolskiy - @romanyam - Dr. Yampolskiy appears to be a man of few words: most of his tweets are just links with no explanation. But the links are good ones, covering the breadth of AI from technical to societal implications.
  • Denny Britz - @dennybritz - High school dropout with CS experience at UC Berkeley and Stanford. Keeps up wildml.com, an absolutely fantastic resource for learning about AI.

Apr 2, 2016

Fields of Artificial Intelligence

I've come across a number of great landscapes covering companies operating in different areas of artificial intelligence, but those frameworks never quite covered the actual structure of AI for me. So I decided to put together a framework showing what I was looking for!

*Note*: This framework is a work in progress. Hit me up if you have feedback, or just check back soon for an update.

One of the biggest challenges I've seen in trying to lay out the AI landscape is confusion around taxonomy. Is a company a machine learning startup, a computer vision startup, or a driverless car startup? Well, it can be all three at once. Machine learning is a category of AI techniques. Computer vision is an application of a variety of AI technologies (including machine learning). And driverless cars make use of "smart" video analysis, a use case that involves computer vision.

To try to make more sense of the landscape, I want to break artificial intelligence nomenclature into three groups: use cases, applications, and technologies.

  • Use cases are what you actually do with AI. Most tech-savvy people understand what these use cases are just based on their names. Use cases are things like personal assistants, smart image analysis, and recommendation engines.

  • Each use case is composed one one or several applications of AI. These applications don't do anything by themselves that the end-user would find useful, but can be packaged together and with other functionality to make up a use case. On the other hand, they are not so specific as to be actual AI technologies that you could explain to a software engineer and expect them to have a firm idea of what you wanted implemented. Applications include speech-to-text (converting spoken word into text), computer vision (analyzing a picture or video to determine what's in it), and NLP (natural language processing - making sense out of a body of text).

  • Finally, applications are created using specific AI technologies. Machine learning, for instance, is a category of AI technologies that includes neural networks and decision trees. An alternative to machine learning is logic-based AI, a deterministic approach to the field that is essentially a series of complex if/thens.

To that end, I want to start with a (who am I kidding?) comprehensive list for each of the three groups:
Use cases

  • Chat bot / Personal assistant
  • "Smart" image search
  • "Smart" video analysis
  • Writing
  • Anomaly detection
  • Recommendation/comparison engine
  • Robotics

  • Computer vision
  • Pattern/anomaly detection
  • Natural Language Processing (NLP)
  • Natural Language Querying (NLQ)
  • Natural Language Generation (NLG)
  • Recommendation/comparison engine
  • Speech-to-text

  • Methods/Techniques
    • Machine learning
      • Artificial neural networks
        • Deep learning
        • Convolutional neural networks
        • Recurrent neural networks
      • Cluster analysis/categorization
      • Decision trees
        • Random forests
    • Logic
      • Semantic graph
  • Models/Algorithms
    • Stochastic gradient descent
    • Evolutionary algorithms
    • Learning types
      • Supervised Learning
      • Unsupervised Learning
      • Reinforcement Learning

Thank you to Sonia Nagar, Ablorde Ashigbi, Konstantine Buhler, and Melissa Caldwell for their help in thinking through this
Mar 12, 2016

The Costs of AI: An Update

My last blog post, on the cost of running AI-based solutions for different problems, could not have come at a better time. Universal powers paid homage to my super insightful post by staging a perfect display of AI prowess is this past week: Google's DeepMind (specifically, their AI-driven solution AlphaGo) just beat the world's top Go player in a best of 5 match. You may remember DeepMind from my previous blog post: it was the company that had made the AI that was really good at Breakout. It turns out arcade games and millenia-old strategy board games aren't so fundamentally different.

The timing was perfect because this provides another fantastic data point for the (actual) costs of doing superhuman things with AI. Let's dig in!

The version of AlphaGo that just beat Lee Se-Dol is running on 1,920 CPU's and 280 GPU's. Let's assume $100 a pop for the CPU's and $1,000 each for the GPU's. Processor cost of $472,000. Round up to $1,000,000 for the other supporting hardware (little things like motherboards and memory).

If you'll remember, that $1M price tag is surprisingly affordable compared to the $3M for the hardware IBM put together for its Jeopardy-playing Watson incarnation. Moore's law and all that I suppose. This is a bit different though, because AlphaGo isn't preparing answers in a couple of seconds on a quiz show. Each of the 3 matches took about 3 hours and you can be assured that the AI was churning away the whole time. Wave your hand, do some math, and it looks like it cost somewhere in the neighborhood of $100 for AlphaGo to win each of its matches (if you amortize costs assuming AlphaGo is doing nothing but playing Go world champions 24/7). This is clearly non-trivial cost territory: most non-professional Go players would be unwilling to pay $100 for the privilege of getting schooled by a computer. On the other hand, consider that Lee Se-Dol would have been paid $20,000 for each of his wins (had he managed to win anything).

Interesting stuff. Of course, we haven't dug into the development cost of AlphaGo here. We know it was bought by Google for $500M a couple of years back, but I'll leave the hand waving to the reader as homework.
Mar 4, 2016

The Costs of Automation Through AI

In case you haven't heard, artificial intelligence is going to be making some pretty sweeping changes to the professional landscape over the next decade or two. Driverless cars, automated customer support, and basic data analytics are just a few of the technologies that will result in an appreciable portion of today's workforce being replaced by lower-cost and higher reliability algorithms.

At this point you're probably internally groaning as you realize that I'm about to decry the advance of technology and its ramifications on society (particularly the working class). Not so: I await my replacement robot with open arms. Joking aside, plenty of studies have shown that all kinds of technologies tend to drive more job growth than they destroy.

No, what I want to talk about is the actual cost of AI-based automation. Not the collapse of society, but the dollars and cents that corporations will be paying to answer belligerent customers' questions with a machine.

Development cost
Today, development accounts for--by far--the largest source of AI costs. To visualize this, let's use the hypothetical example I've referred to a couple of times: Comcast decides it's tired of paying the therapy bills for its customer service reps who have to hear complaints about how shitty their service is every day. They develop their own in-house AI to answer customer complaints, fire all the humans, and implement the solution. Because this is the very cutting edge of AI today and no such solution currently exists, Comcast spends about $250M developing their proprietary algorithms. (Think that's crazy? Here's the $500M Google acquisition of an AI company that was really good at arcade game Breakout). Because of the rapid advancement of AI, Comcast can't justify amortizing the development for more than 5 years. With some hand-wavey math, I come to the conclusion that the amortized development cost alone results in a cost per call of about $0.08.

$0.08/call doesn't sound like a lot and it's obvious to see why Comcast automated in this case. But then, it's entirely possible $250M was an optimistic price tag for this software and very few companies will be able to spread out their cost per use expenses like Comcast can with its mammoth call center.

Computing cost
Although it's not necessarily obvious, there actually is a variable cost to automation through AI. In the Comcast example, think of it as (1) the amortized price of the hardware for the computers running the AI, (2) the cost of housing and maintaining those computers, and (3) the electricity required to run the computers.

I can't even begin to guess at those costs in the Comcast example, but I can in an even better example: IBM's Watson computer on Jeopardy (speaking of AI price tags, check out this 10-digit one). In that case, the computer (cluster) itself cost about $3M. With some more hand wavey math, that results in a per-Jeopardy-answer cost of about $0.06 (assuming you can reach 100% utilization over the course of the computer's entire life). If you make some random assumptions about the other two components of variable cost, you might reach a reasonable cost of about $0.10 in the Watson Jeopardy example.

As you can see from some of the above calculations, AI makes really good economic sense in some cases. But those are just far out edge scenarios with no real world application, right? Well... I may or may not have just gone through these real-world calculations for one of my firm's portfolio companies that I'm most excited about: x.ai. So maybe this isn't so far out there. As more and more specialized AI's are developed for more and more uses, it's going to be important for the tech world, venture investors, and the business community more broadly to start thinking about how to think about the costs of automation through AI.
Jan 5, 2016

About Mark Zuckerberg's "AI" project...

It's that time of year again: aspirational individuals all over the world have set themselves a laundry list of New Year's resolutions: lose weight, be a better spouse, spend more time with family. Some may even set stretch goals: run a marathon, start a business, or climb Mount Everest. But very few resolutioners gained the press that came with Mark Zuckerberg's announcement that he's going to build a Jarvis-like artificial intelligence to power his home.

No doubt about it, Mark's resolution is really cool. And there's good reason to think that he'll make good on his project: he's famous for setting an ambitious goal for himself every year. That being said, I'm not entirely sure we should all be comparing him to Tony Stark quite yet...

It's not hard to envision the kind of AI-driven home that Mark has described: it's the home of the future. An omnipresent, omnipotent (at least as far as the home goes), incredibly intelligent personality that helps make your house a more comfortable place to live. Walk up to the door and it lets you in. Walk in and it turns the lights on with the perfect music to match your mood. Groggily shout at it to make you some coffee and you're presented with a cup of java just a few minutes later. Futuristic movies have been showing us this dream for decades now and Mark should be able to get just about as close as anyone ever has.

Here's the thing: what Mark Zuckerberg is doing isn't all that technologically impressive and it's not even really artificial intelligence (nor is he claiming that it is). He's going to be making the intelligent move by capitalizing on the work of others: a pre-built voice recognition engine, a facial recognition app, and all of the fancy "smart" home devices that have made up the leading wave of the Internet of Things.

While voice recognition and facial recognition are both really neat applications of artificial intelligence, they're not really applications worthy of Mark's time: they've been solved. I programmed my first app that made use of voice recognition almost 5 years ago and the technology was already excellent at that point. A couple of years later none other than Facebook pioneered the largest and most successful application of accurate facial recognition technology to-date. Hmm... I wonder if Mark's ever heard of those guys? And all of the neat tricks of figuring out customized lighting, disturbance detection, and climate have been figured out by smart devices like Nest and Dropcam.

Rather than developing an incredible general artificial intelligence (a feat that our society is still a couple of decades away from accomplishing) that can run his life, Mark is really setting out to develop a relatively basic rules-based app that connects to a bunch of API's and existing services. In fact, the only thing stopping thousands of other developers from doing this same project is the cost in the tens of thousands of dollars of all of the smart devices and the relatively limited benefit of the solution. Make no mistake: it's a really cool project and it's going to be huge in helping Mark Zuckerberg understand how AI can be applied in our daily lives (not to mention scratching an engineering itch that has been bothering him for a number of years). It's just not Tony Stark.
Oct 8, 2015

The Future of Enterprise Data Reporting: Unified Information Access

I've been spending a lot of time getting smarter in the areas of business intelligence and enterprise search lately as part of my diligence into a couple of interesting companies. These two fields, along with advanced analytics, make up the overall area I'm calling "enterprise data reporting".

I was having a fair amount of my trouble in my efforts: I kept getting confused between the different buckets and which companies belonged where. Finally, today, I came across the missing piece of the puzzle: a stunted Wikipedia page describing Unified Information Access Platforms, a type of technology that has received relatively modest media attention over the past five years or so.

Unified Information Access (UIA henceforth) were the answer to my question. I had picked up on the general convergence of enterprise search and business intelligence solutions, but UIA treats that trend more holistically and lays out a picture of what the future state looks like.

UIA, at its core, is about giving an enterprise convenient access to all of its data, no matter where its stored or in what format. Historically, business intelligence has been about crunching the numbers that live in structured databases. Enterprise search has been about returning documents based on a rudimentary keyword search. Both fields have been expanded over the past five years or so thanks to the development of natural language processing (NLP) capabilities, but such functionality was generally brought as an "in addition to" instead of as an entirely new paradigm.

Here are some of the key elements of UIA as I see it:
  • Access to all of a company's data: structured databases, e-mails, presentations, spreadsheets, tweets, customer reviews, etc.
  • Powerful NLP and semantic analysis to understand how all of the unstructured data fits into the big picture
  • Machine learning based natural language query (NLQ) capability that gives all users throughout the business easy access to the stored/indexed data
  • Ad-hoc analytics, data visualization, and data discovery capabilities to process all of the stored/indexed data
  • Big data processing capabilities
  • Automated discovery, integration, and indexing of all common data sources

That's a tall order. A really tall order. The difficulty of the task partially explains why there hasn't been that much buzz around UIA. Several companies operating today claim to be UIA solutions, but based on my research not a single one is doing all six of the bullets above. Instead, there's a hodgepodge of different solutions doing up to four or five of the elements without quite getting the full picture. In that environment, specialized solutions targeting specific verticals have thrived because they really "understand" the needs of their customers.

It'll be interesting to watch over the next year or two as improved general AI and data processing capabilities enable a holistic UIA solution. I'm already looking at a couple that are in the early stages. I also wouldn't be surprised to see moves by some of the players that are 50-60% of the way there in an effort to get closer to 100%.

For additional reading: Random link 1 and random link 2

P.S. - In case you're curious about the aquarium picture up above: it's from "random link 1". They use it to represent a "single pane of glass into a corporation's full data set" metaphor. I thought it was funny.
Jul 2, 2015

Monetizing AI

2015 is an exciting time: Jurassic Park is back, gay marriage is legal across the United States, and AI is finally becoming a reality in our daily lives. Okay, so those first two events may be a little more front-of-mind for the average consumer, but that doesn't change the fact that there are a lot of neat AI solutions popping up all over the market: personal assistants, driverless vehicles encountering each other on the street, and of course, taking on top contenders on Jeopardy.

It's obvious that AI no longer resides solely in science fiction stories or in bizarre videos of Japanese electronics shows. It's hitting the market in a serious way. Even everyone's favorite quizmaster robot, IBM's Watson, has drummed up well over $100M in revenue for Big Blue. But how? What is artificial intelligence actually bringing to the table that people are paying for? How do you monetize an advanced technology that the lay consumer doesn't even realize exists in a commercial sense?

Let's start with Watson. It's telling that IBM is significantly behind in their initial estimates of revenue generation. They were targeting $10B in the 10 years following their silicon savant's performance on Jeopardy. 3-4 years in, they're sitting at a little over 1% of that. Sure, there's a ramp-up curve. But there can be no doubt that they've struggled with monetization. The one place that they do seem to have found traction is in the healthcare industry. There has been talk for years of relying on computers to make more accurate diagnoses and prescribe better medicine than human doctors could. IBM is pushing Watson for just that kind of solution and has had some luck in getting hospitals and insurance companies on board, at least for a trial.

As for the more consumer-focused side of things, personal assistants like x.ai, Clara, and Julie seem to be all the rage in the market right now. For a monthly fee, you can get the right to copy one of these tireless digital scheduling servants on any e-mail and have them take care of all of your e-mail based scheduling for you. They'll use natural language processing to work with humans to get meetings scheduled when and where you want. While these bots are a ways from passing the Turing test, their seeming humanity is reportedly enough to get them through this limited application.

But there's so much more opportunity out there. I should be able to ask Siri to solve a complicated integration problem for me or figure out which of the 50 beers on tap best matches my taste profile. Enter Viv. Founded by the same folks who sold Siri to Apple in the first place, Viv's goal is to make just those sorts of arbitrary requests a reality. They want to bring true cognitive computing to everyone's pocket. Just pull up the Viv app on your phone and ask your personal AI assistant to do just about any digital task for you. If you haven't looked into it yet, you should definitely spend some time doing so. The potential is pretty impressive.

So anyway, here's my question: how do you monetize revolutionary AI like Viv? While Watson's natural language processing and problem solving capabilities are at least in the same genre as Viv's it would be nice to envision a commercial application broader than the few hundred institutions willing to write 7+ figure checks for the privilege. I also can't imagine them getting by on a modest monthly fee like the personal assistant startups. The allure of Viv's technology is the potential to fundamentally change how humans interact with computers; cutting out anyone in the market who won't pay $20/month to be part of the wave just feels wrong to me.

There's always the possibility of selling out to Apple or Google like was originally done with Siri. Based on what I've read about the founders' experiences with Apple post-Jobs, such a deal with Apple seems unlikely. Now that they've been through this rodeo once, I get the impression they're going to want to retain control this time around. So maybe something less absolute... a licensing agreement with one or both of the mobile OS giants? It seems like that might be a good compromise between making some serious money and getting as much distribution as possible.