Trends in Artificial Intelligence

From a business perspective, what counts is how Artificial Intelligence can make you money. At this point in a book, I hope I’ve conveyed the general framework that AI is here to help you automate and optimize business processes with which you deal daily. It’s supposed to make you faster, more creative, and more efficient in what you do, and by that, give you some time and money back.

In this chapter, I will look at various business domains and go through how AI is already used in practice. From the most striking applications in retail to healthcare and video games, we’ll go over use cases that made headlines in recent years. Moreover, I’ll speculate about the general directions of AI applications and how your organisation can benefit from them.

Regardless of your geographical location or your budget, you’ll be able to use AI in your organisation thanks to democratisation trend within the community. Open-source code available on Github allows you to pick recent research, tune it quickly with your data, and start using it to solve your problems. The whole process is fast, especially if you know a bit of Python yourself.

Before diving in more specific domain-trends, I’d like to point you to general meta-trends that appear in applied AI. Those are:

1. Autonomous machines
2. Automated processes
3. Optimized processes with human supervision

Autonomous machines trend is about making real-world robots, vehicles, or otherwise constructed things into independent self-sufficient (to a large extent at least) beings that perform one particular task very well. The simplest example is iRobot Roomba, an autonomous vacuum cleaner, which is already widely used in American households. Or think about self-service checkout. In this category, we will have autonomous cars (e.g. Tesla) or autonomous sorting machines operating in Amazon warehouses.

Automated processes trend is to take repetitive processes and use algorithms to make them fully automatic. The oldest example of this is a calculator, which led to more advanced computing machines, and the automation of computations. More advanced uses are, for example, macros in spreadsheets or connecting various applications thanks to Zapier or similar solutions. The current work done in this space often goes under the name Robotic Process Automation, and some notable players are Automation Anywhere, UIPath, and Blue Prism. On the other hand, automated processes also encompass healthcare problems: from diagnostics to drug design.

Optimized processes is a category where machines are supervised by humans, and eventually, it is up to a supervisor to choose the optimal answer from insights provided. Great examples can be found in marketing when one wants to optimize ad spend and effectiveness, by automatic A/B testing, looking for patterns and insights in marketing data. Nevertheless, optimization occurs everywhere, from marketing to real estate, from retail to manufacturing, as we will see in this section. Optimization is the most common reason why AI is employed at all. Machine learning builds upon statistics using most cutting edge hardware to deliver outstanding optimization performance across a variety of fields.

All in all machine learning is driving changes at three levels:

• tasks and occupations, for example, using computer vision to free up radiologists’ time in detecting potential cancer cells.
• business processes, for example, Amazon warehouses introducing robots and optimization algorithms to their workflow.
• business models, for instance, going from pricing for an individual product (song, movie) to a subscription model based on personalised recommendations a’la Netflix or Spotify.

We’re still early in the process, and hence machine learning systems do not replace entire jobs or processes but rather boost and complement human workers. AI is applied whenever a task is tedious and repetitive and can be automated, hence freeing up human time. For example, chatbots used commercially in customer service are not meant to replace employees entirely but make them focus only on the most unusual cases where a personal touch is necessary.

This approach at the current stage of research is also the only one feasible in production, as building a technology that would commit no error is either too hard or too costly right now. Nevertheless, you should keep in mind that machine learning is an active research field, with thousands of researchers constantly trying to push the boundaries and breakthroughs are announced every month.

Thus building a hybrid workflow of human-machine interactions is the most effective way to progress and build a great company in the age of AI.

Let’s now focus on particular use cases of AI in various business domains. This list is far from comprehensive, but I just wanted to show what’s possible. The chances are that some applications are already outdated. Our level of progress is so fast that by the moment you read these words, many more applications will emerge, which are even more amazing.

AI in retail

The sale of goods is a prototype of all human economic endeavors. AI plays already a prominent role in boosting sales of virtually any item, be it online or offline. Just think about the following use cases:

• classifying customers based on their purchase history;
• personalization of offer for each potential customer;
• analysing demographics and proposing an individual message for each customer;
• automation of sales via in-store technology;
• customer service done by chatbots;
• analysing competitors live 24/7 and reacting to changes fast;
• optimising ad spend and marketing campaigns;

to name just a few more basic ones. They are all related to smart analytics, and while they existed before the technological revolution, AI allowed retailers to boost analytics in a tremendous way. It gave companies tools to analyse and act upon insights 24/7, perform hyper-personalization, and considerably lower the cost of operating a retail business.

In January of 2018, Amazon opened its first high-tech grocery store14 that does not require a traditional checkout. Amazon Go is an experimental grocery in Seattle, which allows shoppers to take goods off shelves and just leave without checkout. Computer vision identifies them as they enter the store, then links them with products taken from shelves. When shoppers leave, the system deducts the cost of the items in their bag from their Amazon accounts and sends an email receipt.

Enhanced user experience is the area that offers probably the most futuristic perspectives for AI in retail. Deep learning and computer vision technologies also will help store owners compete with the one-click convenience of online retailers by eliminating checkout altogether. Cashless shops around the world, especially in China are becoming more popular.

Navigating a hardware store can be difficult, and that’s why in 2016, Lowe’s introduced LoweBot,15 an autonomous retail service robot, in Lowe’s Stores in the San Francisco Bay area. LoweBot was able to find products in multiple languages and help customers effectively navigate the store.

As LoweBot helps customers with simple questions, it enables employees to spend more time offering their expertise and specialty knowledge to customers. LoweBot also assisted with inventory monitoring in real-time, which helped detect patterns that might guide future business decisions

Sephora16 has installed an AI-powered mirror in its store in Madrid to come up with recommendations for customers. The mirror detects data about the shopper looking into it, including gender, age, look, and clothing, and uses those points to recommend makeup, skincare, and fragrance offerings that best match the shopper’s needs.

Airlines like KLM started introducing chatbots to help with orders. You can meet KLM’s BB17 on Messenger or through the Google Assistant to find a destination and book a flight.

Similarly, companies in other niches experiment with ordering through a chatbot for a couple of years already. Pizza Hut was one of the first companies to introduce online ordering through Facebook.18

The Olay Skin Advisor19 offers a web-based skin analyst application and advisor tool that uses artificial intelligence to create a personalized beauty experience once a user uploads a picture of their face (a selfie). The app evaluates skin health and makes recommendations for problem areas with personalised skincare regimen recommendations.

A great example of how AI is applied in business is done by Starbucks. Starbucks’ CEO Kevin Johnson talked about Deep Brew, an in-house AI solution20:

“Deep Brew will increasingly power our personalisation engine, optimise store labor allocations, and drive inventory management in our stores. We plan to leverage Deep Brew in ways that free up our partners so that they can spend more time connecting with customers. Deep Brew is a key differentiator for the future, as we continue our quest to build world-class AI capabilities, to better support partners.”

This Deep Brew technology is used in automatic espresso machines, to make the best coffee possible. Starbucks has also implemented chatbots in their My Barista App.

Another example of an enterprise applying AI comes from H&M,21 which created its AI department in 2018 and has been since working on using AI in different segments. Most profoundly, it changed supply-chain. Forecasting demand is crucial for H&M to lower costs of operations.

Summing up, AI can be used in retail to evaluate demand at restaurants or other shops based on time or weather. Smart analytics can improve customer experience and suggest changes, increasing profitability. In-store virtual assistants can identify returning customers using facial recognition, analyze their shopping history, and make suggestions via a chat.

Online retailers are going for personal recommendations to customers, the more, the better. They are ahead in targeted marketing thanks to data gathered online. However, also traditional retailers have started to collect data from physical stores and then analyse it using data science solutions.

The future of retail is in hyperpersonalisation and data-driven decision making.

Manufacturing

Manufacturing produces 16% of global GDP, and for that reason, AI can hugely benefit the whole economy if it can boost manufacturing. And it seems like this is the case. I will divide manufacturing into:

• predictive maintenance,
• forecasting demand,
• factories (also partially covered in chapter about robots),

and discuss each use case below. The term Industry 4.0 emerged in 2011 to address the trend of total automatization of manufacturing. As a business vertical manufacturing is very susceptible to automation - many factories have already implemented automation to a large degree. Due to the digitization of manufacturing, it’s easy to enter with AI tools and try to optimize processes. However, the most difficult challenge lies in using AI in the way which won’t make errors in scenarios where errors would cost too much (human lives).

Predictive maintenance

Maintenance is a critical area that can drive significant cost savings and production value around the world. The cost of machine downtime is high: according to the International Society of Automation, $647 billion is lost each year globally.

Predictive maintenance is about detecting anomalies. Deep learning, thanks to its capabilities to analyse vast volumes of data, can take existing preventive maintenance systems to a new level. AI’s ability to predict failures and allow planned interventions can be used to reduce downtime and operating costs while improving production yield.

This kind of solution is, for example, offered by Uptake or Dataiku. Both companies specialise in industrial applications of AI.

Forecasting demand

Hedge funds have been using machine learning to forecast demand for quite some time in order to predict demand for a given commodity. With more information available digitally, it becomes possible to use machine learning to forecast demand beyond finance and hence enhance any business by reducing waste and increasing profitability. For example, grid scale electricity is currently hard to store. This creates a substantial economic and environmental cost for both underestimating demand (blackouts) and overestimating demand (wasted energy).

DeepMind, a subsidiary of Google, managed to reduce wasted energy in Google’s data centers by 40% by using machine learning algorithms.22 Smart grids, combining AI with expert knowledge, are still in development, with many challenges ahead23.

On the other hand predicting demand for retail purposes, like, for example, H&M we mentioned, is much easier to implement. In general, the more digital the process, the easier it is to optimise it using AI.

You can find many tools powered by AI, which are suitable for forecasting demand like FuturePlanning or Pipedrive. Often a good CRM with AI capabilities can significantly increase your productivity.

Factories

Factories are at the core of the industrial revolution, and it’s the same for Industry 4.0 movement. The future of intelligent manufacturing lies in using machine learning together with experts’ knowledge.

At Siemens’ Electronic Works Amberg,24 about 50 million items of processes and product data need to be evaluated and used for optimization for production to run smoothly. With Edge Computing, data can be immediately processed where it’s generated, right at the plant or machine. People manage and control the production of programmable logic circuits through a virtual factory that replicates the factory floor.

The Nanjing25 factory is part of Ericsson’s global supply chain set up. State-of-the-art cellular IoT technologies in the Nanjing factory enable an automated alert system for the immediate attention of critical issues and faults. Implemented at 45 work stations, it allows increased efficiency and speed of the production system.

China is also investing massively in AI factories. MEGVII26 is one of the very few companies in the world that have developed proprietary deep learning frameworks in this space. Their solution Brain++ functions as a unified underlying architecture and provides critical support for algorithm training and model improvement processes. Brain++ enables a customer to build a semi-automatic algorithm production line that is continuously self-improving and becoming more automatic over time.

This is just the tip of an iceberg as factories throughout the world are massively experimenting with AI-powered production lines, achieving significant increases in productivity.

Logistics

Machine learning algorithms can learn how to optimally allocate resources, like fleets of vehicles, to address dynamically changing demand (e.g. passenger requests) while maximising resource utilisation. Thus it’s no surprise that logistics is another domain with AI adoption in progress.

In general, AI can help logistics in:

• demand forecasting,
• assisting last-mile delivery (from chatbots to autonomous drones),
• real-time decision making,
• creating contingency plans,
• tracking movement.

Rolls-Royce27 is working with Intel to develop self-driving ships. Rolls-Royce released the Intelligence Awareness system in 2018, a system that can classify all the nearby objects under the water. It can also monitor the engine condition and recommend the best routes.

DHL28 has developed a machine learning-based tool to predict transit time delays of air freight to enable proactive mitigation. By analyzing 58 different parameters of internal data, the machine learning model can predict if the average daily transit time for a given lane is expected to rise or fall up to a week in advance. Furthermore, this solution can identify the top factors influencing shipment delays, including temporal factors like departure day or operational factors such as airline on-time performance. This can help air freight forwarders plan ahead by removing subjective guesswork around when or with which airline their shipments should fly.

UPS saves 10 million gallons of fuel per year by optimizing routes with surprising ‘don’t turn left’ strategy.29

Satellite imagery company DigitalGlobe delivers high-resolution pictures of the planet’s surface to Uber. These images provide rich input sources for the development of advanced mapping tools to increase the precision of pick up, navigation, and drop off between its drivers and riders. DigitalGlobe’s satellites can decipher new road-surface markings, lane information, and street-scale changes to traffic patterns before a city adds them to its official vector map. This level of detail from satellite imagery can provide valuable new insights to planning and navigating routes not only for the transport of people but for shipments as well.

Startups like Transmetrics and ClearMetal are offering their AI tools for logistics to Fortune 500 companies and beyond, enabling supply chain organizations to optimize logistics and provide their customers with easy access to trusted, live information about their shipments.

Robotics and Autonomous Vehicles

Media and science fiction movies love stories about autonomous conscious robots taking control of the world. This vision is far from reality, and in this chapter, I will cover commercial use cases of robotics.

The most vivid imagery for robots is created by Boston Dynamics. Boston Dynamics has made tremendous progress in the last ten years, from barely walking robots to parkour performing athletic robots able to walk and run on any terrain. Each year they present innovation, and then there’s a bit of public concern about the potential use of those in military missions.

However, the reality for robots is usually more boring as they are widely used for warehousing and logistics tasks.

Warehouse robots

Tractica Research30 estimates that the worldwide sales of warehousing and logistics robots will reach $22.4 billion by the end of 2021. Robots are locating, tracking, and moving inventory inside warehouses; they are conveying and sorting oversized packages at ground distribution hubs.

Autonomous workers in Amazon Warehouses is one example. Amazon employs over 200,000 robots in its warehouses today, which is doubling from year to year. These robots service fulfillment and sorting, completing these jobs faster and better than human workers would. Amazon builds whole technology around helping robots navigate simultaneously in a cramped warehouse environment with QR codes and blocking sunlight.

Moreover, Amazon works on autonomous drones that would deliver parcels to secluded locations. In 2016 Amazon announced a partnership with the UK government, and it also works with the US government.31 Amazon Prime Air plans to use the aircraft to establish a package delivery operation in the United States.

Ocado, a UK online supermarket, often grabs media attention with new implementations of AI and robotics at its core operations. Machine learning algorithms steer thousands of products over a maze of conveyor belts and deliver them to humans just in time to fill shopping bags. When fully operational, Ocado’s hive of robots will be processing 3.5 million items or around 65,000 orders every week. The tasks being undertaken by Ocado’s robots are very basic, and they can be expressed in one word — “lifting,” “moving,” “sorting.” Thus we can expect to see the application of these sorts of robots also in other industries.32

Autonomous robots also work alongside people to increase productivity and reduce injuries. Swisslog’s robot-based solutions combine KUKA robots and Swisslog’s intralogistics know-how. They are designed to reduce operational costs and improve warehouse efficiency.33

DHL constantly tries to collaborate with various companies and push the research boundaries themselves.34 In 2016 they unleashed a pair of fully automated trolleys that follow pickers through the warehouse and relieve them of physical work. Similarly, LocusBots from U.S.-based Locus Robotics are self-guiding vehicles that navigate autonomously around a warehouse carrying standard plastic tote bins. A display on the bot tells nearby warehouse associates what to pick for each bin, and a scanner confirms each item as it is loaded. By allowing workers to spend more time picking and less walking the aisles, Locus says that its system can double worker productivity.

Chuck35, created by 6 River Systems, uses machine learning to help associates work faster. Chuck leads warehouse associates through their work zones to help them minimize walking, stay on task, and work more efficiently. It can be used in all put-away, picking, counting, replenishment, and sorting tasks.

Brain Corp. signed a deal with Walmart to scale up from an initial 360 robotic floor cleaner trial and add 1,500 more robots.36 BrainOS, built by Brain Corp., is a cloud-connected operating system for commercial autonomous robots. Robots powered by BrainOS navigate autonomously, avoid obstacles, adapt to changing environments, manage data, generate reports, and seamlessly interact with end-users and other robots.

Berkshire Grey combines AI and robotics to automate omnichannel fulfillment for retailers, e-commerce, and logistics enterprises. In 2020 they announced securing a $263 million funding in series B led by SoftBank.37

China is heavily investing in AI, especially when it comes to manufacturing, logistics, and robotics. JD.com, the logistics giant from China, unveiled in 2018 a warehouse that can handle 200,000 orders a day but employs just four people – with their jobs centered around servicing the robots that run the place.

This all shows that the robotization trend is growing: 35,880 robots were added to U.S. factories in 2018, 7% more than in 2017.38 And as we can see, not only warehouses employ robots. Logistic companies are also investing in implementing robots in their tasks.

We can expect an even larger increase of automated factories and warehouses as technology progresses and cuts down some of the costs. Within 10 years 90% of warehouses and factories may be entirely automated with a couple of human staff per building to oversee robots.

Autonomous cars

Autonomous cars are another hot topic when it comes to automation and autonomy. Many companies are receiving permits to ride their passengers in autonomous vehicles.39 Tesla, Uber, and Lyft are already making tests on autonomous cars driving around California in 2019 and 2020.

California was the first state in the US with autonomous vehicle testing regulations. Ten states have authorized the full deployment of autonomous vehicles without a human operator, including Nevada, Arizona, or Texas. Some states like South Carolina, Kentucky, and Mississippi, already regulate truck platooning.

In 2018 California licensed testing autonomous vehicles for over 50 companies and more than 500 autonomous vehicles (AV) - which summarily drove over 2 million miles. In 2018 AVs in California had 46 crashes noted as being in the autonomous mode when the collision occurred.

In general, the California Public Utilities Commission gives permits as a part of the state’s Autonomous Vehicle Passenger Service pilot. As part of the program, companies must provide data and reports to the CPUC regarding any incidents, the number of passenger miles traveled, and passenger safety protocols. Companies must also have a safety driver behind the wheel and not charge passengers for rides.

The number of companies that were accepted into a program is growing.

Similar tests are performed around the world. The first driverless bus is already cruising in Singapore.40 China, Japan, and Singapore are especially keen on pushing autonomous vehicles. Europe is catching up with Scania and Volvo, each producing their own autonomous buses and deploying them in their respective countries.

Moreover, there are more tests of autonomous trucks. The headlines were made by Otto, which were later bought by Uber. However, it still seems early for commercial use at scale.

The crucial thing when it comes to training autonomous vehicles is mileage. After “State of AI report 2019” we can cite that: Waymo drove more than 1 million miles in 2018, 2.8 times next best (GM Cruise), and 16 times third best (Apple). The average Californian drives 14,435 miles per year. Only 11 out of 63 companies with DMV approval drove more than this in 2018. Self-driving mileage accrual in California is still microscopic when compared to all drivers. In total self-driving car companies racked up 0.00066% of the miles driven by humans in California in 2018.

This all means there’s still a way to go for autonomous vehicles, and we shall see it unfold in this decade. When it comes to AVs and their use in commerce - be it passenger rides or transportation - the largest obstacles lie in creating suitable legislation and a proper environment for deploying autonomous cars on our streets.

Robotic Process Automation

A report by McKinsey Global Institute called “A Future That Works: Automation, Employment, and Productivity”41 predicts that nearly half of work tasks will be performed by some form of a robot by the year 2055. AI agents will automate any kind of job that is repetitive on various levels.

I genuinely believe that. Not only repetitive tasks will be automated but also creative ones, as we can judge by recent breakthroughs in text understanding.

Robotic Process Automation (RPA) software is a fundamental piece of automation. RPA, at its core, is just an approach to automating business processes through the deployment of bots or AI. It dates back to the 90s, and back then, it was a purely software engineering task of breaking down a process into smaller pieces and connecting various APIs to replace humans. For example, having to copy a particular text from one document to another and then sending it through email. Though useful, it was primarily concerned with boring, repetitive tasks that could be automated by writing ‘if-then’ instructions.

Everything changed with the advent of AI. Now software is able to deal with anomalies, previously unseen cases, and make a decision for itself on the spot. Enhanced with text understanding algorithms RPA software is capable of taking upon more complex business process and don’t need to be always guided by human labeling of every possible scenario. We’re entering a new age of automation. We should expect even more disruption once reinforcement learning methods will be implemented within RPA software, which would allow for automating entire office jobs.

Let’s review the most popular RPA solutions on the market right now.42 Most of the companies have been around for over 10 years and became quite mature organisations.

Automation Anywhere empowers people to focus on the work that makes their companies great with robotic process automation that automates virtually anything. Automation Anywhere operates a Bot Store – a marketplace for pre-existent bots suited for different roles. These “Digital Workers” are given job titles by the company, for example, “Digital Employee Onboarding Specialist” with tasks such as identify, shortlist, and onboard candidates. Originally founded in 2003 as Tethys Solutions, the company acquired its current name in 2010, emphasising its focus on robotic process automation.

Blue Prism is a UK RPA platform with customers like eBay, the NHS, and Walgreens. Its intelligent RPA platform comes in both on-premise and SaaS varieties (as most mature RPA solutions), working with the public sector, manufacturing, financial services, and beyond. Blue Prism has a drag-and-drop interface built around connectable objects containing actions and events, with a documented history of processes. The company was founded in 2001.

The digital workforce is built by the operational teams or accredited Blue Prism partners using their robotic process automation technology to rapidly build and deploy automations through leveraging the presentation layer of existing enterprise applications. The automations are configured and managed within an IT governed framework and operating model, which has been iteratively developed through numerous large scale and complex deployments.

UiPath is an RPA company that offers a platform for automating repetitive manual tasks. Thanks to the ease of use for its automation designer, UiPath’s robots can operate with or without human supervision (help desks and call centers). Now based in New York, the company was founded in Bucharest, Romania, in 2005. The company raised $568 million in its latest Series D funding round and is one of the RPA giants with over $7 billion valuation.

SAP Intelligent Robotic Process Automation is a complete automation suite where software robots are designed to mimic humans by replacing manual clicks, interpreting text-heavy communications, or making process suggestions to end users for definable and repeatable business processes. SAP’s offering incorporates machine learning and conversational AI alongside RPA.

Other startups worth observing in RPA domain are:

• WorkFusion,
• Pegasystems,
• Cognizant.

All of them are working on their proprietary software and heavily investing in AI. Large tech companies are also slowly entering the market, some of them like IBM collaborates with existing RPA providers (Automation Anywhere), others like Microsoft, try to develop their own set of tools.

Even though RPA companies are relatively big, there is still a place for newcomers. Startups looking to disrupt the RPA market should look towards reinforcement learning and recent breakthroughs in text processing. The space of RPA tools is big enough to welcome new players without stealing the business from established companies. Another potential direction for startups is to narrow down automation tools only to a particular business niche like law, logistics, or finance.

Having said that, RPA is yet to grow immensely, thanks to progress in machine learning. Most RPA systems are legacy solutions compared to what is currently possible with machine learning.

Image generation

Deepfakes are hyper-realistic AI-generated images and videos. They entered the mainstream, making real and fake media indiscernible. This shows another side of democratising AI: an easy availability for malicious use in disinformation and malware.

Media companies are the first to engage in using image generation.

At the end of December 2019, Snapchat acquired AI Factory, a Ukraine-based startup developing computer vision products, for $166M43. Snap had previously worked with AI Factory to power Cameos, a feature that enables users to insert their selfies into GIFs to create animated deepfakes. Snapchat Cameos are an alternative to Bitmoji for quickly conveying an emotion, reaction, or silly situation in Snapchat messages.

TikTok, owned by ByteDance, is working on a similar feature: it has built technology to let you insert your face into videos starring someone else.44

Samsung engineers have developed realistic talking heads that can be generated from a single image, so AI can even put words in the mouth of the Mona Lisa.45

Hollywood also is betting on AI.

The cast of “The Irishman,” Robert De Niro and Al Pacino, was digitally de-aged in the film using AI. Hollywood is heading towards “digitally resurrecting” actors from the ’50s and ’60s in films: supposedly James Dean is going to appear in one of the movies thanks to AI.46

Warners Bros. started a collaboration with Cinelytic47 to use their comprehensive data and predictive analytics to guide decision-making at the greenlight stage. The integrated online platform can assess the value of a star and how much a film is expected to make in theaters and on other platforms. This means that AI has a say in what movies will be produced. A similar example is Scriptbook, a startup developing AI tools to analyse a script and try to predict whether it will be a hit or a flop. Results seem promising.48 To finish with the movie industry, Netflix estimates its recommendation engine to be crucial for its existence and worth over $1 billion.49

Fashion is also heavily investing in AI.

UK-based startup Superpersonal has created an app that will allow users to try on clothes virtually. Users feed the app basic information, including gender, height, and weight. The app then records the user’s head movements. From this data, the app creates a virtual version of the user modeling clothes: great both for personal and commercial use.

The Echo Look is Amazon’s “style assistant” that takes a photo of your outfit and makes fashion recommendations that are conveniently available for sale on Amazon.

Zao, a free deepfake face-swapping app that’s able to place your likeness into scenes from hundreds of movies and TV shows after uploading just a single photograph, has gone viral in China.

On the other hand, solutions are needed to counter deepfakes used for malicious purposes. Two such examples are Sherlock AI and TruePic, which were created to detect deepfakes and verify content. Facebook, in collaboration with Microsoft, started Deepfake Detection Challenge to create benchmarks for evaluating deepfakes.

All in all, image generation will be an important route for the entertainment and fashion industries. Both are already experimenting heavily with deepfakes and image generation. Next come e-commerce and retail businesses, which will use image generation to improve customer experience, and go into the direction of hyperpersonalisation. Also, the media will be influenced by image generation, especially TV anchors - the world’s first AI news anchor has already debuted in China.

Text generation and Chatbots

Text generation had experienced the most significant breakthrough in 2019 when OpenAI announced GPT-2. This Transformer-based model was able to generate coherent pieces of text on a large scale.

GPT-250 is a large transformer-based language model with 1.5 billion parameters, trained on a dataset of 8 million web pages. GPT-2 is trained with a simple objective: predict the next word, given all of the previous words within some text. The diversity of the dataset causes this simple goal to contain naturally occurring demonstrations of many tasks across diverse domains. GPT-2 is a direct scale-up of GPT, with more than 10X the parameters and trained on more than 10X the amount of data.

The whole 2019 was full of surprises when it comes to text generation models with Megatron from NVIDIA being 5 times larger than GPT-2 and finally Turing-NLG from Microsoft being 10 times larger than GPT-2 (released in February 2020).

We are just beginning to experience the consequences of these breakthroughs:

• chatbots are getting much better,
• voice assistants are on the rise,
• tools for language understanding are getting better.

One example of a super-powered chatbot is Meena from Google.51 Meena is an end-to-end, neural conversational model that learns to respond sensibly to a given conversational context.

In 2018 we saw Google Duplex, an assistant from Google, call a restaurant and book a table.52 Similar tools are already in use in China, developed by Baidu and Alibaba.

Text generation tools are already used in customer service and support. In the form of chatbots, these tools can quickly assess a question, answer simpler ones and direct more complex to a human worker. This saves a lot of time by allowing human workers to focus on more involved and interesting cases, rather than constantly answering the same type of questions.

There is plenty of solutions available on the market which are easy to implement and don’t require coding.

On the other hand, the most complex ones are custom-made powered by machine learning and trained on proprietary internal data. The cost of entrance to text generation is still relatively high, especially compared to other domains of machine learning. That’s not surprising as text understanding is one of the hardest areas of machine learning.

In general, applications of Natural Language Generation (NLG) vary from sales and marketing to market research and customer service. To name a few bigger startups in this domain:

Conversica’s AI platform converts leads to sales opportunities via natural, two-way email conversations. It creates engaging conversations to boost your sales.

Persado Marketing Language Cloud delivers AI-generated language that resonates the most with your audience.

Acrolinx is in the content marketing and advertising sector. It aligns your content with your guidelines and uses automation for shortening your editorial process.

Narrative Science interprets and transforms your enterprise data into intelligent narratives like content for your website or analytic reports.

Automated Insights develops technology that automatically generates narratives on a massive scale that sound like a person crafted each one.

OneSpot generates personalized content after viewing a website user’s history on the internet.

The opportunities for a content generation will only grow with time as this technology will become more mature. Together with Image Generation methods, companies will be able to automate their content production or at least radically save time on producing content.

AI-powered education

Education is also transformed by AI. Most of the innovations so far were on the side of running massive online classes. Companies like Coursera or edX are leading the way of online education, enrolling millions of students into courses from higher education institutions.

Nevertheless, more edtech companies are investing in machine learning solutions to track the progress of students and personalise their learning experience. AI promises global access to personalised education for anyone.

Advances in speech and text understanding allow AI to answer questions from students instantly, guiding them through the process along the way.

Reports from EdTechXGlobal and Ibis capital estimated that schools spent nearly $160 billion on education technology, or edtech, in 2016, and forecast spending to grow 17 percent annually through 2020. Also, private investments in educational technology increased by 32% in the last couple of years.

Coursera’s online classes use machine learning to alert teachers when a large number of students make similar errors on an assignment, suggesting possible gaps in the teachers’ lectures or course materials.

Another application of AI in education will come in the form of profiling students to divide them into groups according to their skills and the pace of learning. Collaboration.ai uses artificial intelligence to process data on each student’s experience, knowledge, and capabilities and recommend group formations best suited for the learning objective. Machine learning can identify complementary skills that would maximize critical thinking and test students’ capacity to adapt and collaborate.

Also, traditional universities are experimenting with Artificial Intelligence to improve student retention. Some schools are testing advanced analytics to identify students in trouble and offer them support before they drop out; for example, this is done by Civitas Learning and Salesforce.53 The Salesforce tools use machine learning to recommend engagement strategies to improve graduation rates and minimise churn.

AI also helps grade students’ assessments. Gradescope helps seamlessly administer and grade all of the assessments, whether online or in-class. Teachers save time grading and get a clear picture of how students are doing.

Microsoft also does a lot to promote access to education: collaborating with schools, individual educators, and other companies. They are particularly keen on bringing AI to schools. Microsoft commissioned a report on education from IDC, which covered 509 higher education institutions in the US, and found that 99.4% of respondents say AI will be instrumental to their institution’s competitiveness in the next three years. Furthermore, 15% called AI a “game-changer,” and 54% have started to experiment with AI, while 38% have adopted AI as a core part of their business strategy.

At the University of New South Wales in Sydney, Australia, David Kellermann has built a question bot capable of answering questions and delivering past video lectures. The bot can also flag student questions for teaching assistants to follow up. What’s more, it keeps getting better at its job as it’s exposed to more and different questions over time.

Duolingo uses AI in its gamified lessons. The company reaches over 300 million users with more than 32 language courses—from French and Tamil to endangered languages such as Hawaiian and Navajo. You learn by completing short lessons, repeating those where you failed. The algorithms learn your speed of learning and suggest appropriate tasks to boost your learning.

China is also heavily investing in AI education. Adaptive learning is an education technology that can respond to a student’s interactions in real-time by automatically providing the student with individual support. SquirelAI is the first AI-powered adaptive education provider in China. They provide personalized and high-quality K-12 after-school tutoring at an affordable price, addressing:

• lack of personalized attention in traditional classrooms,
• unequal distribution of educational opportunities.

HolonIQ54 predicts that AI adoption in education will explode over the next five years and is expected to reach a global expenditure of $6b by 2025. Much of the growth will come from China, followed by the USA, together accounting for over half of global AI education spend.

This is good news for anyone interested in the education space. The 2020 remote work environment caused by lockdowns around the world is also changing how students engage with teachers. We have to wait to see how education will change, but definitely, in the upcoming years, we will witness a disruption of standard models of education that were dominant for the past hundreds of years. Thanks to AI and online learning, education will be democratised and freely available to anyone.

AI in Healthcare

Artificial Intelligence can already review health records and medical data with more speed and accuracy than humans. Thus AI in healthcare can significantly increase the accuracy and reduce the likelihood of human error in:

• diagnostics,
• treatment plans,
• overall patient care.

In the next years, we will see more and more doctors working closely with software, which will boost largely available help for patients, not only in developed countries but even in the most remote regions.

A good example here is Bosch Vivascope,55 which is a cell-analysis platform using artificial intelligence to detect anomalies in biosamples. There are many regions of the world, where laboratory medicine is scarce. Sometimes there’s only one pathologist to 1.5 million people in a region. There’s often no one to examine blood for diseases and make a diagnosis. Two-thirds of the examinations are still carried out with a microscope, which is time-consuming. Vivascope can help those people in rural regions, preventing the spreading of diseases and boosting their overall health.

Cancer diagnoses

Artificial Intelligence is already being used to detect diseases more accurately than ever before. The AI program reliably interprets mammograms and translates patient data into diagnostic information 30 times faster than a human doctor, with 99% accuracy.56

A high proportion of mammograms yield false results, leading to 1 in 2 healthy women being told they have cancer. The use of AI helps reduce the need for unnecessary biopsies.

There are other examples of using AI in detecting cancer. Automated classification of skin lesions using images is a challenging task owing to the fine-grained variability in the appearance of skin lesions. In a 2017 Nature article, Esteva et al.57 describe an AI system trained on a data set of 129,450 clinical images of 2,032 different diseases and compare its diagnostic performance against 21 board-certified dermatologists. The AI system could classify skin cancer at a level comparable to the dermatologists.

Google has used machine learning to improve the detectability of prostate cancer. It has achieved an overall accuracy of 70% when grading prostate cancer in prostatectomy specimens. The average accuracy achieved by US board-certified general pathologists in the study was 61%.58

MaxQ.AI is building towards AI-augmented healthcare through medical diagnostic solutions that help empower physicians around the world to prioritize better and identify life-threatening conditions such as stroke and traumatic brain injury in acute care settings.

Turbine.AI models how cancer works on the molecular level and tests millions of potential drugs with artificial intelligence.

There are plenty more startups in healthcare space diagnosing cancer thanks to progress in image recognition algorithms and access to vast data of scans. We’re bound to see more applications of AI in cancer diagnosing and fighting it in the first phases.

General screening and general care

AI will be applied to the design and development of clinical trials, which is a labor-intensive and manual process. With a market size expected to hit $68.9 billion by 2026, optimizing the clinical trial workflows could help to reduce spending, lower costs, optimize processes. Already, AI-automated trial matching can integrate data from electronic health records, medical literature, and eligibility criteria from legislative bodies and learn how to interpret the trial requirements based on patient cases. In one IBM Watson Health™ study, AI-based identification cut the time required to screen patients for clinical trial eligibility by 78%.59

I have already mentioned Vivascope above. Boosting pathologists with AI is also a goal of PathAI startup.60 PathAI is developing technology that assists pathologists in making rapid and accurate diagnoses for every patient, every time.

Screening goes beyond just diagnosing cancer, and the list grows longer with every day. Machine learning models were used to:

• Detect and classify cardiac arrhythmia using ECGs, achieving cardiologist-level performance.
• Reconstructing speech from neural activity in the auditory cortex.
• Prevent or manage diseases.
• Address unhealthy behaviors before people become patients (tracking diets, fitness routines, etc.).

DeepMind analyzed over 1 million anonymous eye scans to train itself to be able to identify the early signs of eye disease. For that, Google’s DeepMind collaborated with NHS working with London’s Moorfields Eye Hospital to develop a machine learning system that will detect the early signs of degenerative eye conditions that humans might miss. 61

Another use case is detecting diabetic retinopathy from eye scans. This condition happens when high levels of blood sugar lead to damage in the blood vessels of the retina. Autonomous AI that instantly detects disease was created by IDX.62

With more data sources coming from our smartphones, smartwatches, laptops, camera, IoT devices, and so on, global organisations can track our behaviors better. Apart from privacy concerns you might have, the benefit would be an improvement in our health, thanks to monitoring early symptoms and suggesting behaviors that we should correct. You can already download various apps from Samsung or Apple, which monitor your health based on data coming from your smartphone and smartwatch. With the more sophisticated technology, we will be able to prevent many strokes, heart failures, and other diseases.

More general solutions already exist. Babylon Health has been designed around a doctor’s brain to provide accessible healthcare for millions. Their AI solution can understand and recognise the unique way that humans express their symptoms. Using this knowledge, combined with a patient’s medical history and current symptoms, AI provides information on possible medical conditions and common treatments.63

A big segment of the health industry is related to care for elderly people. There are a lot of experiments using robots or VR with the idea of reducing loneliness and add a point of contact in emergency cases.

A different use case is applying machine learning to help people visually impaired navigate. AI is now able to identify objects and read a text, convert handwriting or printed text to digital text and read it aloud. Microsoft’s Seeing AI used those techniques to bring a solution available in 70 countries free of charge. A similar solution is OrCam MyEye camera, which is mounted on standard glasses and converts what is seen into spoken output.

Thus AI can be a great tool to improve access for impaired or elderly people. We’re yet to see this direction of research to develop, especially that going from a research phase to production can take time due to regulations that are required to be met.

Research and development

The path from the research lab to the end-user, which is a patient, is long and costly. California Biomedical Research Association estimates it takes an average of 12 years for a drug to go from the research lab to the patient and around $359M. Furthermore, only five in 5,000 of the drugs that begin preclinical testing ever make it to human testing, and just one of these five is ever approved for human usage.64

The greatest challenge today in pre-clinical drug discovery and development is identifying a drug candidate that is both effective and safe. It is a problem faced by researchers in every pharmaceutical company, whether small or large and in thousands of research institutions across the world.

Drug research and discovery is one of the more recent applications for AI. By directing the latest advances in AI to streamline the drug discovery and drug repurposing processes, there is the potential to significantly cut both the time to market for new drugs and their costs.

Genomic modeling is one of the domains ripe for disruption with AI. DeepMind has built Alphafold65 to understand protein folding and determine the 3D structure of proteins. Their solution borrows concepts from natural language processing to predict distance and angle between amino acids. Understanding their structure and how they fold presents the opportunity to develop drugs for previously unknown targets.

This leap forward in drug discovery allows Relay Therapeutics to leverage the relationship between protein motion and function, creating opportunities to develop more effective therapies for multiple diseases.

Atomwise develops technology that uses a statistical approach that extracts the insights from millions of experimental affinity measurements and thousands of protein structures to predict the binding of small molecules to proteins. This tool makes it possible for chemists to pursue discovery, lead optimization, and toxicity predictions with precision and accuracy.

Exscientia is another startup using AI to discover new drugs. Exscientia’s Centaur Chemist™ approach transforms drug discovery into a formalized set of moves and a system that learns strategy from human experts. Then AI algorithms can outperform expert human drug designers in the search for optimised drug compounds.

LabGenius is a biopharmaceutical company developing next-generation protein therapeutics using a machine learning-driven evolution engine, integrating machine learning, synthetic biology, and robotics. They use deep learning to explore protein fitness landscapes and improve multiple drug properties simultaneously.

At Insitro, which launched with $100M in VC financing, computational experts and biologists work together to create lab experiments and produce massive custom data sets. Then machine learning models find patterns to suggest new tests and potential therapies. Automated pipetting machines and other robotics tools reduce human error. Thanks to that Insitro can do experiments in a matter of weeks instead of years.66

In general, pharma is among the most willing to invest in AI, as drug research is expensive to carry on successfully. The costs of hiring a data science team are relatively low compared to the costs of developing even a single drug available commercially. Definitely, in the upcoming years, we will see breakthroughs in drug discovery coming from AI and other cutting edge technology, especially quantum computing is promising when it comes to applications in pharma.

Cybersecurity powered by AI

When it comes to our security in the digital world, AI is transforming both sides - it’s both defending and attacking us. Its malicious uses can be tracked to hackers trying to get into our bank accounts, stealing precious information from corporates and governments, or simply cracking our social media accounts.

On the other hand, we have better and more reliable defense systems which, through ultra-personalization, allow us to identify whether the agent is really what he claims to be, think about face-unlocking on phones, or determining who’s typing by the speed of typing.

Yet, on the other hand, ultra-personalisation goes together with a lack of privacy and surveillance capitalism. It’s often privacy versus security. Is there a way out?

More Cybersecurity AI startups are raising funds to defend us against hackers and malicious use of the software. Darktrace, a global machine learning company specialized in cyber defense, raised over $230M in total. It uses behavioral analytics to detect abnormal behavior in organizations automatically.

Some other notable cybersecurity startups using AI include:

• Cylance applies AI algorithms to predict, identify, and stop malware.
• ThetaRay uses AI to provide real-time detection of anomalies and unknown threats, especially in the financial sector and industry.

The malicious use of AI gets more sophisticated too. Criminals used artificial intelligence to impersonate a chief executive’s voice and demand a fraudulent transfer of €220,000 in March 201967. There were other cases like that.

There’s also a potential for using deepfakes to spread misinformation and influence public opinion. Fake news were largely used in political campaigns in the US in 2016 with massive reach thanks to Facebook (Cambridge Analytica scandal).

With lowering entry barriers, AI will be more often used by hackers in the upcoming years. This means that to protect ourselves and our companies, we will need to use AI-based solutions as well. The global cyberwar is yet to begin.

Climate Change

Climate is becoming more of an issue each year, with the weather becoming more extreme in various parts of the world. We’ve already passed a point where restrictions will suffice, and we need to proactively change the way we manufacture, consume, and live. AI will play a role in our fight for the climate. Not only algorithms already provide better analytics and actionable insights, but paired with advances in robotics AI will be able to help us transit into renewable energies, reducing waste and emissions of greenhouse gas.

The most visible effect of pollution is plastic floating in the oceans. It ends up in animals’ stomachs leading to their death. This interrupts the food chain, influencing directly all other animals and humans alike and causing damage to the marine industry. Here comes Clear Blue Sea with FRED – the Floating Robot for Eliminating Debris. FRED is a solar-powered marine vessel capable of harvesting floating marine debris. Another example is a marine drone called the WasteShark cleaning up plastic waste off the coast of Devon in the United Kingdom. Unmanned, autonomous vehicles can help with cleaning tasks where humans don’t want or can’t go.

Plastic in oceans is, of course, just the tip of an iceberg when it comes to climate. A group of researchers from various US institutions has prepared an in-depth guide on how machine learning can help tackle climate change.68 The whole paper is over 100 pages long and is worth reading for anyone interested in how they can leverage AI to tackle climate change. The following summarises69 some of the aspects the authors have mentioned throughout the paper, which is divided into the following sections:

Electricity Systems: Forecasting supply and demand, Improving scheduling and flexible demand, Accelerating materials science, Managing existing technologies, Accelerating fusion science, Reducing life-cycle fossil fuel emissions, Reducing system waste, Modeling emissions, Improving clean energy access, Approaching low-data settings.

Transportation: Understanding transportation data, Modeling demand, Shared mobility, Freight routing and consolidation, Alternatives to transport, Designing for efficiency, Autonomous vehicles, Electric vehicles, Alternative fuels, Passenger preferences, Enabling low-carbon options.

Buildings and Cities: Modeling building energy, Smart buildings, Modeling energy use across buildings, Gathering infrastructure data, Data for smart cities, Low-emissions infrastructure.

Industry: Reducing overproduction, Recommender systems, Reducing food waste, Climate-friendly construction, Climate-friendly chemicals, Adaptive control, Predictive maintenance, Using cleaner electricity.

Farms and Forests: Remote sensing of emissions, Precision agriculture, Monitoring peatlands, Estimating carbon stock, Automating afforestation, Managing forest fires, Reducing deforestation.

CO2 Removal: Direct air capture, Sequestering CO2, Understanding personal carbon footprint.

Climate Prediction: Data for climate models, Clouds and aerosols, Ice sheets and sea-level rise, Working with climate models, Storm tracking, Local forecasts.

Societal Impacts: Monitoring ecosystems, Monitoring biodiversity, Designing infrastructure, Maintaining infrastructure, Food security, Resilient livelihoods, Supporting displaced people, Assessing health risks, Managing epidemics, Disaster response.

Solar Geoengineering: Understanding and improving aerosols, Engineering a planetary control system, Modeling impacts.

As you can see, the list is pretty comprehensive. For each of those points, authors discuss what AI can and can’t do, and what are current opportunities, what is being done and what’s missing. If you’re interested in any of these aspects of climate change, I highly suggest you go to the original paper.

On the other hand, looking from a general perspective, what AI enables in most of these cases is a superb capability to monitor data and predict/forecast events, even rare ones. We can better understand what our actions are doing to the climate, what we can improve, and more importantly how. Machine learning allows us to monitor vast amounts of data in real-time, from carbon footprint, gas emissions to plastic. Monitoring can then be used to engineer solutions that would minimise our bad influence on climate: like improving the efficiency of cities, predicting cyclones, making agriculture smart. These kinds of changes are already applied across every domain.

However, the problem with climate is often more political than economic. It boils down to how to implement changes in large organisations and countries which are far from the lean mindset of constantly changing startups. Definitely, competition and monetary incentives can play here a role, as AI will boost organisations which are ready to implement it and not only make them faster but also lower their overall costs considerably. We should make sure that this change will happen sooner than irreversible climate changes.

Games and Reinforcement Learning

Video games are an excellent simulation environment for training machine learning algorithms. Because of games’ increasing complexity, games can be viewed as a model for our own reality. Learning how to play games is the first step to learn how to operate in real life. We value play as a way to learn for our children, and play is equally good for machines.

Reinforcement Learning (RL) techniques seem particularly well suited for games. The main focus of RL is to reward an algorithm when it completes a sub-task or moves in a good direction, and give it a penalty when it doesn’t. This is the closest to raising a child with a set of rules on which it builds its world-view model. Reinforcement Learning agents learn tasks by trial and error. They must balance exploration (new behaviors) with exploitation (repeating past behaviors).

Experiments in Reinforcement Learning within games like Go, DOTA 2, or Quake III Capture the Flag show that without an explanation of rules, algorithms can figure out the game and learn complex strategies through self-play.

Over nearly 3 weeks in 201770, AI system Libratus played 120,000 hands of Texas hold’em against the human professionals and defeated four top human specialist professionals in heads-up no-limit Texas hold’em. In a study completed in December 2016 and involving 44,000 hands of poker, another AI system DeepStack defeated 11 professional poker players with only one outside the margin of statistical significance.71

A DeepMind agent reached human-level performance in a modified version of

Quake III Arena Capture the Flag, which is a complex, multi-agent environment and one of the canonical 3D first-person multiplayer games. The agents successfully cooperate with both artificial and human teammates and demonstrate high performance even when trained with reaction times comparable to human players.72

An interesting part of Deepmind’s research, later also explored by OpenAI in their Hide&Seek project, is related to training multiple agents simultaneously. Deepmind researchers write that agents must learn from scratch to see, act, cooperate, and compete in new environments, all from a single signal per match: whether their team won or not. This is a challenging learning problem, and its solution is based on three general ideas in reinforcement learning:

• Instead of training a single agent, they trained a population of agents, which learn by playing with each other.
• Each agent in the trained population learns its own reward signal, which transfers to their own internal goals like capturing a flag. Then a two-tier optimisation process optimises agents’ internal rewards for winning and uses RL on the internal rewards to deduce the agents’ policies.
• Agents operate at two timescales, fast and slow, which boost their ability to use memory and generate consistent sequences of actions.

These agents trained by Deepmind showed human-like behaviors such as navigating, following, or defending. They have also exceeded the win-rate of strong human players both as teammates and as opponents.

OpenAI has trained AI systems on Dota 273. Their OpenAI Five agents learned by playing over 10,000 years of games against itself. AI showed the ability to achieve pro-level performance, learn to cooperate with humans, and operate at scale. This was a great achievement as Dota 2 is an example of a multiplayer collaborative game with fast-paced action.

Another achievement by Deepmind came with AlphaStar. AlphaStar is the first AI to reach the top league of StarCraft II without any game restrictions.74 However these applications come at a cost, and it is estimated that training AlphaStar required $26M in computing resources.75

Probably the best known use case of Reinforcement Learning so far is AlphaGo Zero, which learns how to play Go from scratch to the level of top professionals, and beyond. The whole story is shown in a documentary called AlphaGo (highly recommended!).

This application was surprising because Go is a much more complex game than chess, equally renowned and old, with a whole ecosystem of fans, professional players, and broadcast, especially in Asia. Loss of Lee Sedol to AlphaGo came as a shock to the Go community, and it was a moment that made the Chinese government invest heavily in AI, as Go is also very popular in China. In a sense, this was the Sputnik moment for the Chinese AI industry.

Reinforcement learning is also coming to the real world. OpenAI trained76 a pair of neural networks to solve the Rubik’s Cube with a human-like robot hand. The interesting part was that the neural networks were entirely trained in simulation, using the same reinforcement learning code as OpenAI Five with some improvements. The system can handle situations it never saw during training.

RL is often applied in autonomous vehicles, drones being a good example. Azure Drones is a civilian drone maker specializing in data capture and processing services for commercial applications. Skydio is a drone used to automatically follow you and record, great to record extreme sports’ stunts or just a social gathering. A team of researchers at Stanford has also successfully deployed an autonomous helicopter.77

All in all, reinforcement learning is still mainly in the development phase. It seems that RL, together with text understanding, is the hardest subdomain in machine learning when it comes to the level of architecture complexity, amount of training needed, and expertise to deploy it. That’s why we still have to wait to find scalable commercial applications of RL beyond games. But even within the game industry, there’s a big potential when it comes to building AI agents, which could be good sparring partners for human players and guarantee an enjoyable game at any level.

Hardware and beyond

In this final section on Artificial Intelligence trends, I want to talk about hardware and physical devices related to AI: AI chips, IoT, smart cities, quantum computing. Let’s have a look behind these buzzwords and see actual applications and opportunities.

It’s worth noting that machine learning by itself, that is, as a set of learning algorithms, is not useful until you’re able to provide it with enough computing power. That’s why advances in computing power and general computation techniques are influencing AI progress. In 2019 and 2020, we could see that through Transformers models with billion parameters trained on millions of texts (GPT-2, Megatron, Turing-NLG). AlphaGo and AlphaStar from Deepmind needed millions of dollars in cutting edge computing power to achieve human-level performance in Go and StarCraft II, respectively. We can expect new amazing applications of deep learning will require even more computing power. That’s why big tech companies are heavily investing in new AI chips and hardware, from TPU to quantum computing. Anything which has a chance to boost your AI algorithms and give you an advantage over your competitors is worth investing in.

On the other hand, computing power is not enough if you don’t have data to feed to your algorithms. We are fairly good at collecting data online. Enterprises monitor their processes, and these are noted in CRMs, spreadsheets, and other tools that allow further extraction and processing of data. However, we only learn how and what to collect from the real world. Manufacturers invest in IoT to collect data within factories, municipalities install cameras and sensors to monitor human flow within cities, governments, and private companies look at satellite data, and the list goes on. In the future, we might expect centralized systems for monitoring data both in commercial and government use. Setting privacy issues aside, gathering data through monitoring will allow us to make our cities and organisations more smooth: less traffic, faster and more effective production. Of course, our privacy is a big issue, and the whole ethical side of implementing these solutions need to be thoroughly discussed.

IoT and Smart Cities

A smart city is a broad concept encompassing surveillance, mobility, and data infrastructure used together as data sources for machine learning algorithms. The goal of a smart city is to make living in a city a smoother, frictionless experience, just like buying online is nowadays.

Toronto Sidewalk Labs by Alphabet was probably the most comprehensive smart city project in the Western world. This Alphabet company aimed to build an entire neighborhood from the ground up, making it ‘smart’ in every aspect along the way. However, it was put on hold in 2020. As Sidewalk’s CEO writes: “it has become too difficult to make the 12-acre project financially viable without sacrificing core parts of the plan we had developed together with Waterfront Toronto to build a truly inclusive, sustainable community.”78

There are, however, a growing number of smart city projects around the world with ambitious plans, working by adding new sensors and new software to already existing infrastructure or building one from scratch:

• Toyota Woven City in Japan,
• ReGen Villages in California,
• Singapore as a whole.

China is also investing massively in smart cities. Surveillance is of key importance, and the government collaborates with private companies to attain its goals. Utilizing comprehensive real-time city data, Alibaba’s City Brain79 holistically optimizes urban public resources by instantly correcting defects in urban operations. This has led to numerous breakthroughs in urban government models, service models, and industrial development.

Reimagining cities for the future is a must with a growing population worldwide, congestion, pollution, and other civilization problems on the rise. Reimagining workplaces is also crucial for more effective and pleasant work. I have described how AI is applied in various domains in previous chapters to that end. What’s left to be mentioned is that various sensors and devices which capture data are entering more workplaces to monitor and analyse workflows, and eventually improve it. Industry IoT sensors are the largest part of this market, of course, because semi- and fully automated factories need access to real-time data to operate. However, more sensors are entering standard offices. Just think about motion sensors, occupancy sensors, cameras, and beacons. We’re entering a new, interconnected world.

5G, Satellites and Artificial Intelligence

The vast amount of data coming from the physical and online world needs to be fed to machine learning algorithms to be useful for applications. And for that, we need a faster connection. Here comes 5G, a new generation of the industry standard. For comparison, 4G has a theoretical 100 megabits per second (Mbps) maximum speed, while 5G could achieve up to 10 gigabits per second (Gbps). Thus 5G is up to a hundred times faster than 4G. Moreover, it should also offer a more stable information transmission.

When it comes to 5G, China is far ahead of everybody else. However, the US and EU are actively pushing towards 5G. Huawei80 has the most extensive declared 5G portfolio of patents, followed by Samsung, LG, and Nokia. Qualcomm and Intel are the largest US companies with 5G patents, while Sharp and NTT DOCOMO are the largest Japanese 5G patents declaring companies.

A patent advantage can position Huawei as the critical player in the upcoming 5G era, which would also include the whole ecosystem of network providers, device makers, and app developers. In Europe, the UK and Germany are using Huawei hardware, whereas the US looks for domestic alternatives.

Another contender for fast broadband connection is the Starlink project run by SpaceX and Elon Musk. SpaceX leverages its rocket building skills to deploy the world’s most advanced broadband internet system powered by dozens of thousands of satellites (an initial plan mentioned 12,000 satellites, which was raised in 2019 to 42,000). The goal is to provide high-speed internet access across the globe. That will be a game-changer, especially for remote locations, where no Internet was available before. Hence Starlink is complementary to 5G, which will be mostly deployed in high population density areas.

Next-generation connectivity will allow us to gather data from all available sources and use it in machine learning applications.

AI Chips

Once the data is gathered and you have it ready for your machine learning algorithms, what’s missing is computing power to handle the training of models. GPUs came as dominant hardware pieces for AI applications, with NVIDIA leading the race. However, since the new boom for AI, more hardware companies are entering the market of AI chips intending to speed up deep learning training.

Google came with TPU, a tensor processing unit, in 2016. TPUs are custom application-specific integrated circuits (ASIC) tailored for machine learning workloads on TensorFlow and potentially 15-30 times faster than GPUs (depending on to which model we compare them).

NVIDIA is not staying behind with constant innovation in GPUs and delivering commercial supercomputers for the office use like NVIDIA DGX station, which you can combine into DGX SuperPod.

Snapdragon 855 from Qualcomm is dedicated hardware and software designed to accelerate on-device AI. You can squeeze up to 7 teraOPS (7 trillion calculations per second) out of AI processing on Snapdragon 855.

Graphcore is a startup producing IPU (intelligent processing unit). Graphcore C2 IPU-Processor PCIe card achieves 3.7x higher throughput at 10x lower latency compared to a leading alternative processor. The company is backed by Microsoft.

It’s no surprise that all tech giants either build their own AI chips or invest in startups building AI chips: Intel, Microsoft, Baidu, IBM, Samsung, Apple, LG, Amazon, Facebook. All are in this space, though for different reasons and different applications.

The most active startups in this space are Arm, Groq, SambaNova, and Wave Computing, among others. We’re yet to see whether there will be a new standard for AI computations that would replace GPUs.

Quantum computing

Quantum computing is a new, revolutionary approach to computing with a goal to overcome Moore’s Law and hugely boost computing speed. Quantum computers are based on qubits, which not only can be ‘0’ or ‘1’ like classical bits but can also be a superposition of those. A quantum phenomenon called entanglement allows interactions between any qubits in a quantum computer, whereas classical computers are limited to linear communication between neighboring bits. That roughly explains an exponential increase in power with a growing number of qubits in a quantum computer, compared to an only linear increase in power of classical computers.

Nevertheless, we are still early in the process of building and harnessing quantum physics for computing. Currently built quantum machines are not stable, have too much noise, and in general, are not ready to tackle commercial problems. However, a lot of research is being done in this space with investments coming from all tech giants: Google, Microsoft, IBM, and Fujitsu, among others. The most advanced company seems to be D-Wave when it comes to producing and selling actual quantum computers.

In recent years more and more startups are joining the race, building software solutions on already existing platforms. The hope is that with quantum computing, AI algorithms will gain enough power to identify patterns and derive insights from systems so complex that there just has not been enough classical computing power to model them. This includes problems like modeling chemical reactions, identifying compounds with similar chemical properties, and drug research and development. Startups tend to focus on particular niches and develop their own algorithms within the space, partnering with larger organisations to access to quantum computing.

Hybrid models look promising. They combine classical machine learning algorithms with quantum AI and has great potential for commercial applications. Interesting tests were done by Volkswagen together with D-Wave in the optimization space81. A Canadian startup Xanadu82 used hybrid models to approach various machine learning problems, and their results also look promising. Worth noting is also Rigetti, a Californian startup both devising their own quantum chips and software ecosystem.

Quantum cloud computing war is yet to start once the technology is mature enough to be available commercially through the cloud. We should see significant investments and startup acquisitions coming from the likes of Google, Microsoft, IBM, and Amazon.

---