Learning From Machine Intelligence





It’s a very exciting time for technology. In just the past 5 years we’ve surrounded ourselves with connected mobile devices, moved our businesses to the cloud, and watched as every aspect of the content industry has been radically disrupted. Now, just as we’re starting to accept all these changes, another big shift is getting underway. Software is starting to become intelligent and autonomous.

Research into deep learning, statistical modeling, and neural networks has been advancing in the halls of academia for decades but suddenly there’s a Cambrian explosion of solutions impacting every sector. These tools finally have enough computation and enough data to really perform – and a large amount of talent and capital is pouring into the fast-moving sector.

Now, software is shifting from being programmatic – it does only what we tell it – to being proactive: it makes decisions autonomously, based on continuously-updated models of the world. Machine intelligence delivers classification of unstructured data – it can read the news and identify us in camera streams. It can graph relationships between objects, data, and people and then make context-aware recommendations. And it can run sophisticated models that show us the probable outcomes of decisions.

Machine intelligence is being democratized, moving into SaaS platforms and cloud services, mobile applications and connected devices. Microsoft Office 365 uses deep learning to support collaboration workflows. Hospitals and financial planners can address Watson through an API. Mattel has announced a connected Barbie doll that learns how to converse with children.

What we see is that wherever there is data, there are computational resources moving to evaluate it. And increasingly, wherever there is data and computation, there are machine intelligences arriving to grapple with the complexity, bring structure to the streams, and act accordingly in the blink of an eye.

This is a story about a new wave of disruption and the great opportunities it brings. It is also a story about speed: the pace of change, the flows of information, and the unique ability of machine intelligence to process huge volumes of information at gigahertz rates. But mostly, it is a story about how we as people are starting to make tools that mimic our own ability to understand and make decisions about the world, and how this shift is helping enterprises work better.

Georges Nahon CEO,

Orange Silicon Valley

  • Software is undergoing a fundamental shift from doing only what we tell it, to using context and probability for autonomous decision-making.
  • Machine intelligence is the solution to Big Data. 
  • IT infrastructure is shifting to address real-time stream computing.
  • The commercializing of machine intelligence is stirring up a new wave of digital transformation.
  • Businesses are now experimenting with machine intelligence across their value chains, to enable high-resolution sensing, model complex systems, and engineer enterprise-scale learning at machine speeds.
  • Predictive modeling moves businesses squarely into the rapidly-unfolding present with a sharper eye to the future.

In the vast expanses of dusty red Australian outback, one of the world’s largest iron ore mines is being taken over by autonomous systems. Enormous yellow haulage trucks, 7 meters tall and with no human occupants, pick their way across the engineered landscape moving materials from drill sites to crushers and out to trains a mile long. In 2015, the drills and trains will also become autonomous in an ongoing effort by Rio Tinto — one of the world’s largest mining concerns — to remove human operators from dangerous conditions, increase the efficiency and scale of operations, and to enable much more rapid response to changes.

Rio Tinto’s “Mine of the Future” at Pilbara, Australia hosts over 50 autonomous trucks managed from a central command station 1500 km away in Perth. The trucks have sensors, GPS, and algorithms that communicate continuously with central command, responding to directives and waypoints. The Mine Automation System integrates data from equipment, geologic sensors, and control algorithms into real-time models that adapt to the dynamic conditions of their global demand chain. In 2013, the Pilbara site generated record production levels.

The Mine of the Future was “designed to respond rapidly to change.” This is perhaps the most animating feature of how the enterprise is currently evolving — and why machine intelligence is moving into business operations. We’re now inundated with the details of change at unprecedented scales. Big Data has become huge and continuous, overwhelming the human capacity to consume and process it all. Many human systems have grown too complex to manage directly. From global mining operations and supply chains to networks and enterprise tools, algorithms are proving to be much more capable at integrating complexity, responding to change, and optimizing productivity.

Indeed, software is undergoing a fundamental shift from doing only what we ask of it, to doing things that we may not expect. It is now starting to read the news, watch videos, and listen to conversations.

With near-infinite computation and waves of information, older tools like convolutional networks, back propagation, and natural language processing are now making rapid advances. The best minds in the field have been migrating from academia into corporate product lines, lured by the likes of Facebook, Baidu, and Google.

Companies are now developing intelligent systems for their operations, and helping democratize machine learning as a service. IBM, Microsoft, and Google offer addressable machine intelligences for third party integration, suggesting that the rewards of AI willaccrue to central players while the features will be made available to all. It’s important to watch these big platform players: they have large SaaS channels that are already distributing machine intelligence services to their subscribers. They also have large checkbooks and have begun shopping for acquisitions among a growing host of pure-play algorithm companies.

For the enterprise, the impacts will be far-reaching. Employees and assets can be graphed and classified, then connected. These tools bring inference and contextual assistance directly to the workforce. They also offer modeling and predictive analytics, helping companies shift away from mining historic data towards actively monitoring the present and modeling the future.

A greater promise of machine intelligence lies in the ability to integrate many isolated-but-adjacent processes into orchestrated meta-systems. How might an organization perform when sales, market intelligence, business intelligence, ERP systems, HR, IT, DevOps, and product innovation groups are all networked and overseen by sophisticated algorithms able to make immediate decisions without human oversight? If computation has given organizations more brainpower, and API’s and sensors are creating a nervous system, then machine intelligence offers a new layer of fast reasoning on top of it all.

It is in this context that businesses are evolving, straining towards the forms and functions of sensing and adapting living systems. In times of seeming acceleration and continuous change, the enterprise is starting to behave more like biology, with feedback loops, fast iteration, and flexible, adaptive structures that learn and evolve. In doing so, the enterprise is driving a virtuous cycle of funding, R&D, productization, and service integration.

If machine intelligence spreads into the gaps between nearly-ubiquitous computation and sensing, the disruptive impacts will be tremendous and the challenges and opportunities will fundamentally re-shape our world. At some point soon in the balance between human agency and machine assistance, we may find ourselves asking, “Who’s really running our business?”




Many of the algorithmic tools we often refer to as artificial intelligence or machine learning have been around for decades. Up until recently, these tools had been starved and under-powered. There wasn’t enough data or computation for them to really perform.

Now, 15 years into the 21st century, we have near-infinite, cheap computational resources and far more data than we can handle. Moore’s Law has continued to reward us with greater transistor density and lower cost. New solutions like distributed computing and GPU’s offer massive parallelization of processing. Nvidia, a world leader in graphics processing units (GPU’s), is undergoing a pivot from solely serving the market for graphics rendering pipelines to commercializing GPU computers designed to support machine intelligence applications. Now, somewhat-dormant deep learning solutions are being scaled across clusters and are branching into many new algorithmic typologies, feasting on the relentless streams of data that are spinning off the digitalized world.

Analysts at Gartner estimate that in 2015 there will be 4.9 billion connected things sending data across networks and into registries. Banks, for example, can have literally billions of transactions to comb through in order to identify fraud, which highlights another condition of the current context: most enterprises have embraced Big Data but are now trying to figure out how to make it meaningful. They face extremely large data sets and real-time streams from operations and logistics, IT and security, human resources, customer relationships, sales, social media, and any number of embedded sensors. For enterprise-scale organizations, business itself has grown exponentially complex.






The environment in which businesses operate has become increasingly accelerated, ephemeral, and programmable by a host of competitors, disruptors, and inter-connected systems. Data, computation, and the IT infrastructure it relies upon are shifting towards real-time processing to keep up with the pace of information. When change is continuous it no longer makes sense for organizations to invest in static structures.

The twin shocks of 9/11 and the global economic contraction of 2008 have set the tone for our young century. We are now in a world of fast-moving change, sudden discontinuity, and growing environmental pressures. At the same time, powerful digital platforms have emerged that offer rich fields of innovation, opportunity, and competition. Business is increasingly driven to to pay closer attention, learn quickly, and be able to shift its footing when necessary.

In the quest to be more responsive, the contemporary enterprise is experimenting with tools that help it become more aware of its physiology and environment, and that enable it to easily evolve and adapt by constantly mining information to reveal insights and model outcomes.

There is urgency for market leaders to leverage their historic gains and market position in a way that propels them successfully into the future without dragging them back into the past. This is a risky situation and underlines one of the biggest challenges in navigating the landscape of machine intelligence. What solutions will add real value and increase performance, and how do we measure that success?.




Google has an estimated 900,000 servers distributed in 13 data centers around the world. They live in huge warehouses with massive cooling towers, often situated near waterways that offer opportunities to generate hydro-electric power. The data centers are deeply complex, with feedback loops that can drive non-linear changes in power usage. Even ambient weather conditions shape the energy footprint.

A Google engineer named Jim Gao wanted to increase the energy efficiency of their data centers. He developed a neural network that analyzes the data center for runtime parameters that are used to model power usage effectiveness (PUE) – the energy efficiency of the data center. This enables operators to view performance against expected metrics, and to model changes in controller settings before committing them to the data center. They can even model and test proposed physical changes before spending the time and money to reconfigure the hardware.

“By creating an accurate, robust PUE model, you’ve essentially created a virtual environment where you can simulate changes to your data center without actually putting your plant at risk,” says Gao. They’ve applied the method to several data centers with considerable success.

The PUE learning model integrates the details of all the controllers and setpoints with the broader goals of energy efficiency. “At the high end, we achieved a 25% overhead energy reduction,” Gao says of the machine learning effort. When an average Google data center uses upwards of 300 million watts of power, minor optimizations can yield large outcomes in cost performance.

Humans have proven the ability to create complex systems that then grow beyond our ability to effectively manage them. But if we instrument them with networked sensors and controllers, then machine intelligences can be deployed to nudge those systems back into equilibrium with the broader needs of people, organizations, and the planet itself.

Deloitte has projected that the current $1 billion market for cognitive computing will expand to $50 billion within 3 years. There is now a virtuous cycle of funding, research, and commercialization propelling advancements in machine intelligence and its applications.




In the 1980’s Yann LeCun was the first to move convolutional neural networks – one of many tools referred to now as “deep learning” – out of the lab and into the market. At Bell Labs he developed early machine vision algorithms still used by many ATM’s to recognize hand-written signatures on the backs of checks. In 2014 he was hired by Facebook as their Director of AI Research to help them better understand their users, their advertisers, and the content they post so that it can more effectively deliver personally relevant information to each of them. Facebook machine intelligences must reckon with the torrent of images, video, and text posted by its roughly 900 million active daily users. LeCun’s present goal is to remove the training and supervision required by Facebook’s neural nets – an effort that could lead to truly autonomous, self-directed machine intelligence.

LeCun’s approach began with an emulation of the human brain but soon diverged to use whatever techniques yielded the most accurate results. At Cambridge, Geoffrey Hinton was pursuing a similar effort to computationally mimic our thinking – or at least, to mimic what we know of how the brain seems to function. In 2004, Hinton founded the Neural Computation and Adaptive Perception program, regarded as one of the key enablers pushing deep learning into the mainstream. In 2013 he was hired by Google to advance their existing deep learning program and help make better sense of the search giant’s massive hoard of data.

"We ceased to be the lunatic fringe. We’re now the lunatic core."

Geoffrey Hinton

Hinton’s work in Google has picked up where Andrew Ng left off after he was hired away by another web giant, China’s Baidu. Having earned degrees from Carnegie Mellon, MIT, and Berkeley, Ng took a professorship at Stanford in 2002 where he paired machine learning and robotics to develop autonomous systems such as the Stanford Autonomous Helicopter project. In 2011, Ng joined Google part-time to develop Google Brain which he then successfully scaled across 16,000 CPU cores. Google’s machine intelligence made headlines when it was able to successfully understand and identify the higher-level concept of “cat” after watching millions of YouTube videos. It was never actually told what a cat is.






As the tech industry lures the wizards of deep learning away from academia, they’re also busy acquiring companies focused on bringing machine intelligence to the broader marketplace. The past year has seen a flurry of activity as market leaders work to consolidate their position on the next wave of digital transformation.

Google bought DeepMind for an estimated $500 million. MIT’s Technology Review noted that DeepMind has “one of the biggest concentrations of researchers anywhere working on deep learning”, echoed by the managing director of Google UK, Eileen Naughton, who recently declared that DeepMind is “the most advanced artificial intelligence think tank in the world.” Their expertise in “reinforcement learning” recently produced an algorithm that was able to learn how to play – and beat – several classic Atari games simply by watching the screen.

Facebook bought Wit.ai for its natural voice processing technology and their network of 6,000 third-party developers – human brains helping further productize the platform. With Wit.ai the social network is trying to better understand written language in its stream of posts. Notably, prior to the acquisition, Wit.ai grabbed $3 million in a seed round led by Andreessen Horowitz.

Indeed, in the past year most large players have been buying machine intelligence companies. Salesforce acquired RelateIQ and their “relationship intelligence platform” for $38 million. Microsoft spent $200 million on Equivio for text analysis. Infosysy grabbed Panaya, a SaaS/ERP risk analysis tool, for $200 million. Twitter bought MadBits for their visual search solution. IBM bought text analysis and computer vision company AlchemyAPI. Yahoo snapped up Lookflow to bring machine learning and image analysis to Flickr. These are big bets on the trajectory of data and intelligence.





While M&A activity in the young sector has been growing rapidly, so too has private investment in fledgling companies. Industry data analysts CB Insights report that in 2014 VC investments in “artificial intelligence startups” grew by an impressive 302% from the previous year. At the end of 2014 Shivon Zilis published an infographic about the “Machine Intelligence Landscape,” showing over 200 companies shaping the fast-moving ecosystem.

In part, hype and the bandwagon effect are contributing to the jump, but the sudden Cambrian explosion in ML companies (and the attendant dealflow) was likely inspired by the scale of investments in machine intelligence made by industry giants in the past few years.

"A breakthrough in machine learning would be worth ten Microsofts."

Bill Gates

"Enterprise applications — in every use case imaginable — should and will become inherently more intelligent as the machine implicitly learns patterns in the data and derives insights. It will be like having an intelligent, experienced human assistant in everything we do."

Peter Levine,

partner at Andreessen Horowitz



Many major players in the tech sector have in-house machine intelligence programs. Some, like Amazon, have been using deep learning tools across their value chain for years. In 2011, Amazon CEO Jeff Bezos spoke with his shareholders about the company’s commitment to re-investing in its core technologies, including machine intelligence. “We use high-performance transactions systems, complex rendering and object caching, workflow and queuing systems, business intelligence and data analytics, machine learning and pattern recognition, neural networks and probabilistic decision making, and a wide variety of other techniques,” he declared. “In many cases, advanced machine learning techniques provide more accurate classification and can self-heal to adapt to changing conditions.” Four years later, the market is catching on to the value of these tools.

Others players, like IBM, have underscored their substantial pivots away from hardware and into data-driven services by aggressively pursuing intelligent machines. Big Blue recently committed $1 billion to the commercialization of Watson on the path towards democratized and centralized cognitive systems.

"Everything we do now is influenced, one way or another, by machine learning."

Peter Lee,

Microsoft Director of Research

The platform giants are now shaping the market to receive machine intelligence - and they are distributing these tools through their massive SaaS channels.

Google’s AI & machine learning research group has impacted Search, Maps, Google Now, and even Google Sheets. Google’s Prediction API is the customer-facing solution for Google Storage users to detect customer sentiment, predict churn, and identify unwanted activity. The search giant also runs Sibyl, a massively-parallel machine learning platform that sits behind YouTube recommendations generating click-rate probabilities. Launched several years ago, it’s been so successful at predicting what YouTube users might want to see next that Google has since used it to raise the quality of advertising, achieve greater efficiency in monetization, and detect spam in email.

Microsoft has been pursuing machine learning for years and integrating it into its product line. Infoworld recently spoke with Microsoft’s Director of Research, Peter Lee, who was at DARPA prior to joining. “I honestly can’t think of any recent product development that Microsoft has been involved in that hasn’t involved machine learning,” he says. 

Nvidia’s 2015 GPU product line features hardware designed specifically to support deep learning and machine intelligence computation. The Titan GPUs and Digits Devbox kit respond directly to the 10x growth in GPU computing, powering hungry algorithms with parallel processing.





One of the more salient impacts of these efforts is that they’ve kicked off an arms race between top-tier competitors that will drive more money and talent into advancing machine intelligence. Just after Facebook published record-shattering benchmarks from its machine vision component DeepFace, Google announced that its own FaceNet solution had scored even better. Perhaps even more notably, there is a steady re-architecting of the IT stack to move away from historical analytics – querying a hard drive to evaluate past performance – towards in-memory real-time processing of data streams. This is a fundamental shift in computing.

Just as machine intelligence couldn’t bloom without enough data, our ability to use Big Data effectively is increasingly dependent on machine intelligence for help. IBM has stated that in 2012 there were 2.5 exabytes of data generated every day. That’s 2.5 billion gigs. About 75% of it was coming from unstructured text, voice, and video — ie not directly machine-readable. EMC estimates that in 2013 there were 4.4 zettabytes of data in the “digital universe,” projected to grow to 44 zettabytes by 2020. IDC claims that the world’s data doubles every year.

Which is simply to say that there is an incomprehensibly large amount of data being generated — and it’s getting larger every day.

Much of this data arises as the byproduct of our interactions with the digital world but, increasingly, more innovation is data-native, designed intentionally to capture and deliver operational metrics. This trend represents a significant feedback loop: computation generates data which asks for more computation.






John Deere provides farm machinery to producers across the world. They came to understand that both farm vehicles and the fields themselves are rich data centers. By instrumenting and connecting their machines they can monitor performance, detect errors, and provide remote maintenance. Some tractors can even use machine vision and learning to autonomously work the fields. The tractors can survey the crops and soil and transmit their properties to information dashboards. This enables farmers to understand historic patterns in productivity with great detail, rapidly attend to shifting conditions, and develop predictive models for future yields.

What’s important about John Deere’s effort is the shift in mindset to see valuable infor mation in what was previously considered opaque. Precision agriculture and agronomics is enabled by the digitization of nature — the farm becomes connected, sensing, and much more visible by bridging the gap into digital systems.

Similarly, Microsoft has been evolving Office 365 to acknowledge the fundamental relationship at its core: people and information. Microsoft understands that its productivity tools are also significant data centers that can be mined for insights and understood as touch points for professional productivity and collaboration. At the center of this is Microsoft’s Office Graph, a machine learning core component that sees people as behavioral nodes than can be algorithmically coupled to the information in their documents, spreadsheets, and to each other.

Office Graph “has mapped over 6 billion actions and interactions within Office 365

in just the last six months,” says Julia White, general manager for Office 365. Office Graph is learning from the behaviors of Office365 users, analyzing communications, filtering them for validity, and matching them to appropriate human resources. It’s like a smart, autonomous routing system that connects relevant human resources around the evolving needs of a project. By giving employees an active network identity, relevancy graphs and machine intelligence can break down the silos that harden business units against each-other.

This ability to process enormous amounts of data and then find patterns that show meaningful correlations is one of the ways that machine intelligence is starting to amplify human knowledge about the world.

Enlitic is a young startup that uses deep learning to unlock the data within medical visualizations. X-Ray scans, CT scans, and MRI’s rely on doctors to accurately parse the information in these complex images. With neural networks, doctors can have a cognitive assistant that can spot patterns that they themselves might miss. Perhaps more importantly to public health, the algorithms can look through huge volumes of these images and correlate them to medical records and data stores, highlighting patterns and connections between them, and even surfacing environmental factors.

Enlitic founder and CEO Jeremy Howard says that recent breakthroughs in machine intelligence represent “the biggest opportunity for positive impact using data that I’ve seen in my 20+ years in the field.”

Algorithms are no longer just scripts that require structured data types and direct instructions. They are becoming active personal assistants.






Google’s data centers generate massive amounts of operational data. Jim Gao’s work is a response to this condition, using machine intelligence to manage power consumption. “Machine learning is simply the framework that makes this data useful,” he says. “Machine learning is a tool to turn information into knowledge.”

These examples illustrate the power of coupling data to machine intelligence. For many businesses that have already begun down this path, the challenge is in making the data useful and in linking those insights into organizational performance. Machine intelligence is the solution to data overload, and a path to a new level of enterprise optimization and productivity.

Businesses must continuously learn about themselves, their markets, and the broader landscape in which they operate. The evolution of digital tools is beginning to engineer these capabilities directly into the infrastructure of business, enabling the rise of the digital learning enterprise. Much has been said about the need for employees to learn continuously but what happens when smart IT systems are learning more quickly than the workers?

If you ask data scientists, you’re likely to get very conservative responses about these tools and their limitations. If you talk with product managers and marketers, you may be transported to a world of transformational artificial intelligence. The metrics for success are still being discovered but the reality lies somewhere in between the two groups.

Machine intelligence enables software to process natural language, identify visual patterns, make connections and inferences, and create models and probabilities about the future. When we read about “machine learning”, the authors are usually referring to pragmatic methods like back propagation and convolutional neural networks that classify things and generate probabilities about what they are, but in ways that diverge from what we know about how human intelligence works. For this reason, it may be more accurate to think of it as machine intelligence rather than a simlution of human intelligence.

"There’s a big difference between intelligence and sentience. Our software is becoming more intelligent, but that does not imply it is about to become sentient."

Andrew Ng,

Chief Scientist at Baidu

Watson, for example, notoriously beat a human at Jeopardy — but it had absolutely no knowledge of why it was right or wrong. It’s mostly a very fast encyclopedia reader with a set of tools it uses to build a basic map of how concepts relate. Which, admittedly, is something like how we do it, but Watson can’t premeditate or celebrate after. It has neither intention nor self-reflection.

The nature of deep learning algorithms is that they do indeed learn — their models continually update and become more accurate - and their capabilities increase with advances in networked computation. Classifying, modeling, testing, and learning is at the core of how machine intelligence is assuming roles that are impossible for traditional software.

"We know the ultimate answer is unsupervised learning, but we don’t have the answer yet."

Yann LeCun

One of the challenges to advancing these tools is the leap from supervised neural networks to unsupervised, where the algorithms essentially learn how to learn on their own. Another is the ability to scale cognitive systems beyond a single core so that Google, for example, could distribute a machine intelligence instance across its global data centers.

trading algorithms

read the news

High-frequency trading algorithms were developed to read the structured stock and company information published by the public markets, like market cap, stock price, P/E, etc... As these indices update continuously each day, the algorithms read them and execute trades almost immediately. Speed became such an advantage that HFT companies moved their servers as close as they could to the fiber hubs in their cities.

Now companies like Sentient.ai have trained natural language processing algorithms to look beyond the structured stock ticker and to read unstructured data. They watch world news feeds to learn about events impacting markets and then bring these into their models. Analysts can then ask questions like “what happens to US homebuilder stocks if a category three hurricane makes landfall?” With this information they move past rote execution and are now feeding sophisticated models of global markets to predict the impacts of change.

DeepMind masters video games

Researchers at Google DeepMind have developed a cognitive system that was able to learn how to play — and in some cases master — classic video games. “The only information we gave the system was the raw pixels on the screen and the idea that it had to get a high score. And everything else it had to figure out by itself,” says Dr. Demis Hassabis, DeepMind’s vice president of engineering. Of the 40 video games it learned to play, it was better than a human in 29 of them. Sadly, the young machine intelligence struggled with Pac-Man.

the Turing Test

Developed by Alan Turing in 1950, the Turing Test is presented as a way to identify artificial intelligence — but it’s a much better illustration of how we perceive human intelligence. Typically, the test is performed through a chat window with a human on one side and, on the other, a machine intelligence disguised as a human. If, through their conversation, the human is convinced that they are chatting with another human, then the machine has passed the Turing Test. It doesn’t prove that the machine is intelligent but rather that its behavior appears close enough that we grant it humanness.

In 2014, a program called Eugene Goodman was the first ever to pass the Turing Test, sort of, by convincing 10 of 30 judges of its humanness within a reasonable degree of certainty. But it turns out Eugene was cheating a bit by classifying itself as a 13-year-old Ukranian boy that didn’t speak English very well, thus providing an excuse for its inelegant conversation.

What the Turing Test shows is not the presence of intelligence but, rather, the subjectivity of intelligence itself and our willingness to see intelligence where it may not actually exist. It both defends the uniqueness of human consciousness while establishing a place in our world for non-human intelligence.





Most current machine intelligences must be trained — so called “supervised learning.” They are given a pre-configured set of data and are assisted by humans that reinforce the right answers. The training establishes the model that the machine intelligence will refer to — a history of known classifications like “cats within a set of mammals.” One of the most popular approaches is the use of convolutional neural networks, often referred to as deep learning.

Deep learning is a layered approach that isolates progressively greater and greater details within the sample set. For example, an image of a cat in a field may be initially sorted as the simple boundary between the different textures, colors, or shading. This is the first layer. It then may look for the edges within the first layer to establish the second layer. Then it may isolate discrete features before finally “seeing” a clear picture of the animal. At each step there are references to its trained model — the way it was taught to identify cats — that look to classify the probability of what it is “seeing” as animal, mammal, four-legged, and pointy ears. The net result is that the software often performs poorly at first, but then gets better and better over time. It’s a radical shift from how software has behaved historically.

"Enterprise software used to be about making existing work more efficient. Now, the opportunity for software is to transform the work itself."

Aaron Levie, CEO Box

All the money and talent, all the compute and data and modeling... It’s all driving machine intelligence out of labs and internal deployments, and into the hands of the marketplace.






Just as Amazon took its expertise in deploying and running its own cloud and then turned it into a new business line with EC2, big players are doing the same with machine intelligence. Google has opened its Predict API, and Microsoft has launched machine learning on Azure. Not surprisingly, after 2 years of internal use, Amazon announced its own APIs on EC2. Deep learning and predictive services are now being offered by major enterprise service platforms as up-sells for cloud customers.

IBM’s Watson is probably the most famous machine intelligence, often invoked as the inscrutable face of AI. Watson has a toolbox of classification, regression, reduction, and clustering algorithms that it chooses from to solve a query. They recently launched the Watson Developer Cloud to open Watson to the public as an intelligent service.

Ashley Hathaway is a product manager for the group. Her work includes “empowering developers to tell us what can be done with the suite of cognitive APIs.” She sees the young program as a way to get developers “coupling and combining APIs to create cognitive recipes.” There are now 14 APIs that developers can use to address Watson.

The potential for this kind of augmentation to human capabilities is illustrated by IBM’s interest in fighting cancer. Watson for Oncology is a partnership with the Memorial Sloan Kettering Cancer Center (MSK) that aims to personalize cancer treatment. MSK oncolo- gists train Watson to match patient information against a large database of cancer case studies, medical journals, and oncological research. Watson then makes personalized recommendations to physicians about the best ways to treat a patient. Not only does this assist and amplify the capabilities of oncologists, it also enables a new degree of personalization that was previously impossible.

Cancer is so complex and there are so many variations, each relating in nuanced ways to the individual life of the patient. It takes machine intelligence to sift through it all in fine detail and surface the patterns.

"Machine learning creates knowledge. It creates links between concepts."

Mike Haley, Autodesk



Mike Haley is the Senior Director of Emerging Products and Technology at Autodesk, one of the world’s largest providers of computer-aided design (CAD) software. Their customers have tremendous amounts of data in every single part of every model but, in many cases, they don’t know how to leverage that information. For Autodesk, machine intelligence is a way to help designers be more creative and efficient by turning their data into context-aware insights.

“We realized that if we could understand the components and the design information, then we could infer knowledge of the projects and of the people designing them,” says Haley. “With machine learning and cloud we can look at data anonymously, in aggregate, and derive understanding.”

A person working in one of Autodesk’s future 3D design applications might be modeling a bicycle crank. The application’s algorithms can look at the geometry using machine vision (3D geometry, not 2D images) and then classify it: this is a bicycle crank. It can then identify specific properties to infer that the crank is for a mountain bike. Then it can recommend additional part geometries that it knows are common to mountain bikes, as well as potential fabricators and suppliers that can best meet the needs of the design project. It can even make suggestions about alternative designs that the designer may not have considered.

The users of Autodesk’s 3D design applications will also become an important source of feedback that helps continuously train the algorithms. If a user dismisses a part sugges- tion, the machine intelligence registers that input and updates its user model accordingly. The entire user community then becomes a source of feedback and knowledge genera- tion for the learning algorithms, which can deliver similar knowledge services to any application and customer across the Autodesk product portfolio.

For Autodesk, machine intelligence offers designers more time by assisting them with the design process. In a sense, they’re making a cloud-scale, context-aware recommendation system with visibility into the rich strata of data underneath their models, workflows, and collaborations.

“Machine learning creates knowledge,” says Haley. “It creates links between concepts.”



Enterprise examples are important to show how much money and intention are behind the shift from automated workflows to autonomous systems. The large actors are actively shaping the market and pushing machine intelligence solutions into their product pipelines. The impacts broaden significantly, however, when these tools move into consumer products. Mattel, one of the world’s largest toy makers, is experimenting with dolls that can have a natural conversation and learn how to relate to children as individuals.

"The number one request we hear from girls around the world is that they want to have a conversation with Barbie. Now, for the first time ever, Barbie can have a two-way conversation,"

said a spokeswoman for Mattel,

about the new Hello Barbie machine intelligence.

What are the implications of toys that act as if they are alive? How will humans socialize with an emerging landscape of machines that pay attention to us, respond to our words, and offer to help us with our tasks? Imagine Siri and Cortana, the iOS and Android artificial mobile assistants, in 10 years’ time, trained to watch the world on our behalf. Global consulting firm Deloitte refers to these tools as “Cognitive Systems,” which is effectively what we are creating: outsourced thinking that can be scaled, distributed, and attached to anything that has computation and network connectivity.




Information may in fact want to be free, as the saying goes. But the ability to process that information into meaningful insights and predictive models is usually not free, and this fact is forming new markets for computational intelligence. The promise is bold – and deeply evolutionary. If computers could reckon like humans but at inconceivable speeds over enormous data sets, what would that mean for human endeavors? If the arc of civilization is coupled to the progressive development of tools and knowledge, what happens when we deputize machine intelligences to look at it all and help us figure out the next steps?

The democratization of machine intelligence means that these capabilities are starting to work their way into not only every part of the enterprise, but every part of the world.

the mind of a young drone

Skydio is a young startup funded by Andreessen Horowitz that is developing an artificial intelligence for drones. Small UAVs are having large impacts on media, agriculture, and monitoring of industrial facilities. Part of their strength is in their ability to autonomously navigate along waypoints. Founded by MIT grads and ex-Googlers, Skydio is working to give them better vision so they can map the world around them, carefully navigate through complex structures, and identify things they see. The drone vision startup complements another Andreessen Horowitz company, Airware, who are building an operating system for drones.

toys that learn who you are

The inflection point for the spread of machine intelligence may have already been reached. In February of 2015, a New York startup called CogniToys (whose COO is the former president of Mattel) used Kickstarter to successfully fund a toy dinosaur that uses Watson to converse with and learn from children. It’s a green plastic dinosaur that listens to your child’s voice and learns about them so it can converse and respond. The toy gets its brains from IBM’s Watson by connecting to wifi and consulting the cloud-born machine intelligence. With obvious con- cerns about corporations listening to your child in their private moments, the Kickstarter nevertheless passed their funding goal by 500%. Not to be outdone, Mattel has announced Hello Barbie, a version of their iconic Platonic woman endowed with conversational machine intelligence.

There’s a Japanese tradition of Shinto animism that sees a spirit of personality in the most mundane and inanimate things. The distribution of machine intelligence is sort of like having these little helpers all around us, working on our behalf, helping us to find things, and occasionally stealing our jobs... which may seem a bit more like Mickey’s fantastic assistants in Disney’s Fantasia.

Whether intentionally through R&D or indirectly via SaaS subscriptions, such enterprises are getting more visibility into their data and operations; they’re fine-tuning performance and productivity; and are becoming more directly attuned to information flows, environmental sensing, customer experience and sentiment, and the emerging capabilities of predictive modeling and testing.

The democratization of machine intelligence is driving another wave of digital transformation that is further redrawing the lines between automation and humanity, and redefining what we think can and cannot be accomplished through computation. Ultimately, the trajectory is heading towards an enterprise that is programmable and self-correcting, with learning and decision-making integrated across the chain of operations.




When Oracle completed its hostile takeover of PeopleSoft, both the CEO and Chief Strategist of the acquired ERP giant left to found Workday. The company has since become a $15.9B SaaS business by re-defining human capital and financial management. In November of 2014, the company launched Workday Insight Applications, a suite of predictive analytics services based on data and machine learning. At the core of Insight is SYMAN, a relevancy and predictive engine that will “automatically map, classify, and normalize similar definitions across multiple sources and structures of data.” This means that by studying operational data Insight can make smart recommendations for specific situations across the business, evaluating the possible outcomes of decisions and learning from new data.

“Our customers want an experience similar to consumer applications that understand user activity over time to deliver more relevant, personalized recommendations on what book to buy, what movie to watch, or what route to take,” said Aneel Bhusri, Workday’s co-founder and CEO, indirectly acknowledging the influence of Amazon, Netflix, and Google Maps on how we expect services to meet our needs by helping us beyond what we already know. 

“Think of something as simple as looking at a map application on your smartphone to figure out the fastest way home,” says Workday VP of Technology Dan Beck. ”The application describes the routes and predicts how long each route will take based on historical patterns of data. It then makes new, faster route recommendations given heavy traffic on your normal route. Cracking that code for the enterprise has been difficult, but we’re doing it.”










Insight’s coretech, SYMAN, came from Workday’s 2014 acquisition of Identified, a machine learning startup focused on finding and retaining talent. By analyzing responses to employee surveys, Insight Applications can spot which of their top performers are most likely to leave the company. They can model organizational changes to that employee’s work assignment and then predict how likely the move might contribute to attrition or retention.

It’s truly a fascinating development if an algorithm is better than a manager at under- standing employee sentiment. But what does it really mean when such systems are recommending courses of action for how the business will evolve and grow in a com- petitive landscape? In a time of continuous change and information overload, businesses can no longer live by their past. Machine intelligence and predictive modeling offer to move business squarely into the rapidly-unfolding present with a sharper eye to the future.






Sri Ambati is the founder and CEO of H20, a prediction engine for enterprise data. He’s helping companies make the jump from correlation in data to causation. “Enterprise is still trying to learn from the last quarter,” he says, referring to the way that, for example, inventory projections are based on the past 3 months of sales. With H20, customers can build and score models continuously and much more rapidly, bringing streaming data from the present, comparing it against historic performance, and then casting it into predictive models of future demand. Ambati’s clients are data scientists and engineers in large enterprises like PayPal and Cisco who are tasked with identifying anomalies, anticipating customer needs, and surfacing relevant and timely insights. For them, “it’s always about prediction,” he says. “Prediction is the new Search.”

Again, this shows how software is evolving beyond merely executing on explicit commands. It is now starting to use data to build models of possible outcomes and to make decisions based on those models.



Not only do such platforms offer assistance to users, they also offer a lens for companies to observe and learn from every aspect of their business. As data becomes relentless and instantaneous, spinning off of millions (potentially billions) of end points, there is both a challenge and an opportunity for organizations to wrap computational intelligence around the streams and plug them into living models of the business landscape. This shift is beginning to re-architect IT infrastructure and shape the next platforms for scaled compute. Hadoop is facing Apache Spark. NoSQL is evolving towards memSQL. Reading hard drives is evolving towards computing on fast data in memory. The old stack cannot keep up with the volume of real-time data.

When machine intelligence is penetrating businesses through IT, SaaS, and mobile devices, learning becomes part of the structure of the company — and the IT department becomes its primary caretaker. How they evaluate procurements, educate employees, and manage information flows and the machine intelligences that operate on them will play a large role in the adoption and evolution of these emerging toolsets.

This notion that data and machine intelligence can assist businesses in their operations while simultaneously helping them learn from the results of that assistance may be the genie in the bottle of artificial intelligence. It engineers learning directly into the soft and hard infrastructure of the enterprise. Machine intelligence can create a powerful feedback loop with the potential to radically scale businesses into massive data processors and learning engines.



There are not enough data scientists to support the growing needs of the enterprise. Solutions must address lay people who know nothing about statistical regression or random forest algorithms. By moving the functionality directly, but invisibly, into user workflows they can adopt the functionality without even recognizing it as artificial intelligence. Like how we don’t recognize Amazon or Netflix recommendations as algorithms that are watching us and learning our behaviors.

We are already being augmented with cognitive systems that amplify our intelligence and productivity. Information is being assembled around our identities and our context to better meet and anticpate our needs. The way that this relationship unfolds will continue to be determined by the friendliness and unobtrusiveness of the user experience. 












Recall that Autodesk uses one set of machine intelligence algorithms to train another which then acts on data streams to derive insights. In a connected world where information flows are getting bigger and faster, we seem to be heading towards a point where machine intelligences can train and educate themselves and each other at Internet scales and computational speeds. As members of the same networks, we’re also helping them learn. This is perhaps why so many still approach it with caution, and why the performance and operations of our businesses.

Enterprise leadership may soon find that operational decision-making has been sub-contracted to hyper-connected machine intelligences.

In the march of digital transformation our tools are starting to ask us questions, learn from our behaviors, and predict the next engagement — which is a very smart thing for a tool to do, and one that should give us pause as we confer more and more agency to autonomous systems.





One fo the things that really makes us unique is our large forebrain and its ability to model the world and predict the probable outcomes of actions we may take. We plan, we act, and then observe the results, updating our models accordingly. This is how we innovate, test, learn, and create knowledge. And this is essentially what machine intelligence is starting to do, but through statistical pathways and on the substrate of networks and silicon.

The results of human and machine intelligence are looking more alike but they come from very different sources. One is emergent and self-aware; the other is statistical and unconscious. It is unlikely that there is any link between statistical methods and sentience. Even autonomous systems require our hand in design, and we have almost zero understanding of how consciousness arises or how we’ve become self-aware. As Facebook’s deep learning wizard Yann Lecun notes,

“how this is done in the brain is pretty much completely unknown. Synapses adjust themselves, but we don’t have a clear picture for wht the algorithm of the cortex is.”

In spite of all the hype and sensationalism, it’s evident that we won’t get sentient AI any time soon - but we are poised for another wave of digital transformation rising up on machine intelligence.




Machine intelligence offers assistance and augmentation to human users but it also makes it much easier to have autonomous decision-making. This is beyond automation which is coded and directly programmatic. Autonomy is agency and the ability to make decisions and act on them independently. This radically amplifies the ability of software to both assist us and to further displace human labor. We are increasingly confronted with the need to define our unique contributions to economic productivity, at a time when economies are increasingly programmable.

The evolution of digital tools will likely continue to shift more power into the hands of those with the biggest compute resources and the greatest access to data. The rest of us will be able to invoke these tools but not at the scale of heavily-funded corporate and government clouds running powerful cognitive systems across vast networks. Just as information asymmetry has risen between those with the data and those without, a greater gap may open beneath those with the strongest machine intelligence.

Companies want control of their data and they will likely want control over their machine intelligences as well.

Perhaps most importantly is the question of how well these tools really work. It’s still early and there are not many public long-term case studies. We may not even understand the metrics of success. For service providers it’s just subscription numbers and revenue but for the rest of us, inviting young machine intelligences into our lives is uncharted territory clouded in fear and uncertainty. 

As these tools inevitably move further into architecture and infrastructure, gaining access to CCTV’s, listening devices, embedded systems, and networks, the opportunities for assistance and insight are blended with bureaucracy, governance, and control. Our identity and status on networks will determine how much the connected world enables us or contains us. Automobile ignitions, toll passes, and connected locks can turn against us under the watch of some networked machine intelligence flagging our status. Driverless cars will test our willingness to give up control and our acceptance of 3rd-party intervention in our most cherished individuality.

When do our cybernetic dreams of control unleash the thing we can no longer manage? Remember Mickey’s magical servants in Fantasia. The success or failure of the current wave of talent and capital behind these assistive tools will depend in large part on our willingness as people to accept them. 



Questions of agency and labor may be difficult to overcome but if the tools integrate easily and the user experience is lightweight and barely visible, adoption may be quick. For the enterprise, development and deployment cannot be wholly dependent on expensive data scientists. This is a strong case for turn- key SaaS solutions and IT procurements. But this raises another challenge: are businesses willing to let 3rd party machine intelligences walk through their data, extracting insights and modeling their future? This same dynamic is now playing out with the cloud. As more enterprises adopt it, the demand for on-premise hosting becomes greater.




This report has focused on the impact of machine intelligence in the enterprise but these tools are already spreading far beyond, reminding us again that wherever there is computation and data and the need for rapid response to complex situations, machine intelligence is coming online to engage, model, and act.

It’s not just that businesses will feel the impacts of these tools... It’s that the world is shifting to enable machine intelligences to become part of the broader global landscape. It may become pervasive. We don’t yet know how effective these tools will be but the trajectories are strongly positive, and there is more talent and money and computation and data feeding them than ever before. If it works, the world may be as transformed by machine intelligence as it has been by globalized telecommunications and the Internet. We may even be led to confront our own cherished ideas about intelligence and humanity.

the future is here:

swarming autonomous systems

In August of 2014 the Office of Naval Research demonstrated its swarming autonomous Naval response. 13 small, unmanned littoral combat boats fitted with the CARACaS (Control Architecture for Robotic Agent Command and Sensing) technology escorted a Navy ship into a simulated attack by an enemy boat. The escort craft autonomously broke formation, synchronized into a swarm response, choose their own routes and attacked the enemy.

“This demonstration was a cost-effective way to integrate many small, cheap, and autonomous capabilities that can significantly improve our warfighting advantage,” said Adm. Jonathan Greenert, chief of naval operations. For the armed services, autonomous weapons systems offer greater flexibility while removing warfighters from the field of operations.

“This multiplies combat power by allowing CARACaS-enabled boats to do some of the dangerous work,” said Dr. Robert Brizzolara, program manager at ONR. “It will remove our Sailors and Marines from many dangerous situations. If an adversary were to fire on the USVs, no humans would be at risk.”

But isn’t war supposed to be human? Shouldn’t we feel its viscerality in hopes that it will deter us from its propagation?

This is not science fiction, and there are numerous examples of machine intelligences colonizing and animating robotic bodies. The Navy’s CARACaS effort is also an example of another emerging component of machine intelligence: the ability to coordinate and swarm. Using simple rules derived from ants and bees and birds, autonomous systems can communicate and coordinate around objectives. The Laws of Nature are moving into silicon. Machine-to-machine relationships are being formed, which may further disintermediate our own agency and control as we increasingly socialize with machine intelligences. 



Orange Silicon Valley is the San Francisco Bay Area division of Orange – one of the world’s largest telecommunications operators, serving 231 million customers. Orange Silicon Valley participates in the disruptive innovations that are changing the way we communicate, which is the core business of Orange in the 220 countries where we operate. We contribute to and engage with the Silicon Valley ecosystem continuously through outreach, partnerships, and rich engagements.

At Orange Silicon Valley, we thrive on collaboration, seek out change and novelty, and anticipate that the future is closer than we think. We foster innovators through our Orange Fab accelerator program and Orange Digital Ventures, and we promote discussion of the keys issues and trends shaping the sphere of technology and human behavior. We have an abiding commitment to be objective, insightful, and inclusive in everything we do.





This report was inspired by the fast-moving evolution of the many technologies collectively referred to as “machine learning,” and the attendant hype and sensationalism surrounding the notion of artificial intelligence. With an abiding interest in the field, we wanted to understand how these solutions are spreading and moving into the enterprise, how capitol and talent are organizing to drive development, and what it means for business when we adopt these tools into our systems and workflows. We also wanted to debunk some of the myths and clarify the language but this report deliberately avoids confronting the specter of “artificial intelligence” and the many dystopic scenarios that follow such think- ing. That aside, we’d like to thank several contributors for their help in navigating these seas: Autodesk Senior Director of Emerging Technologies and Products, Mike Haley; Product manager Ashley Hathaway, on the IBM Watson team; Sri Ambati, the CEO & Founder of H20; Google Mechanical Engineer, Jim Gao; Startup.ml Founder Arshak Navruzyan; and Lee Gomes at IEEE Spectrum for guidance and perspective. We’d also like to thank the Internet, house music, and artisanal coffees for their unyielding support.



_ Mark Plakias, OSV VP Strategy

_ Gabriel Sidhom, OSV VP Technology Development

_ Georges Nahon, OSV CEO Orange Silicon Valley

_ Chris Arkenberg, OSV Principal Researcher & Analyst

_ Xavier Quintuna, OSV Big Data Principal

_ Gregory Veran, OSV Business Analyst

_ Carolyn Ma, OSV Graphic Design Lead

other publications from OSV

Learning From Machine Intelligence

[[ preview ]]
Invalid: Tell us your email.This is not a valid email.

Learning From Machine Intelligence

Get free access to the full report now

Invalid: Tell us your email.This is not a valid email.