The artificial-intelligence industry is often compared to the oil industry: once mined and refined, data, like oil, can be a highly lucrative commodity. Now it seems the metaphor may extend even further. Like its fossil-fuel counterpart, the process of deep learning has an outsize environmental impact.
In a new paper, researchers at the University of Massachusetts, Amherst, performed a life cycle assessment for training several common large AI models. They found that the process can emit more than 626,000 pounds of carbon dioxide equivalent—nearly five times the lifetime emissions of the average American car (and that includes manufacture of the car itself).
It’s a jarring quantification of something AI researchers have suspected for a long time. “While probably many of us have thought of this in an abstract, vague level, the figures really show the magnitude of the problem,” says Carlos Gómez-Rodríguez, a computer scientist at the University of A Coruña in Spain, who was not involved in the research. “Neither I nor other researchers I’ve discussed them with thought the environmental impact was that substantial.”
The carbon footprint of natural-language processing
The paper specifically examines the model training process for natural-language processing (NLP), the subfield of AI that focuses on teaching machines to handle human language. In the last two years, the NLP community has reached several noteworthy performance milestones in machine translation, sentence completion, and other standard benchmarking tasks. OpenAI’s infamous GPT-2 model, as one example, excelled at writing convincing fake news articles.
But such advances have required training ever larger models on sprawling data sets of sentences scraped from the internet. The approach is computationally expensive—and highly energy intensive.
The researchers looked at four models in the field that have been responsible for the biggest leaps in performance: the Transformer, ELMo, BERT, and GPT-2. They trained each on a single GPU for up to a day to measure its power draw. They then used the number of training hours listed in the model’s original papers to calculate the total energy consumed over the complete training process. That number was converted into pounds of carbon dioxide equivalent based on the average energy mix in the US, which closely matches the energy mix used by Amazon’s AWS, the largest cloud services provider.
They found that the computational and environmental costs of training grew proportionally to model size and then exploded when additional tuning steps were used to increase the model’s final accuracy. In particular, they found that a tuning process known as neural architecture search, which tries to optimize a model by incrementally tweaking a neural network’s design through exhaustive trial and error, had extraordinarily high associated costs for little performance benefit. Without it, the most costly model, BERT, had a carbon footprint of roughly 1,400 pounds of carbon dioxide equivalent, close to a round-trip trans-America flight for one person.
What’s more, the researchers note that the figures should only be considered as baselines. “Training a single model is the minimum amount of work you can do,” says Emma Strubell, a PhD candidate at the University of Massachusetts, Amherst, and the lead author of the paper. In practice, it’s much more likely that AI researchers would develop a new model from scratch or adapt an existing model to a new data set, either of which can require many more rounds of training and tuning.
To get a better handle on what the full development pipeline might look like in terms of carbon footprint, Strubell and her colleagues used a model they’d produced in a previous paper as a case study. They found that the process of building and testing a final paper-worthy model required training 4,789 models over a six-month period. Converted to CO2 equivalent, it emitted more than 78,000 pounds and is likely representative of typical work in the field.
The significance of those figures is colossal—especially when considering the current trends in AI research. “In general, much of the latest research in AI neglects efficiency, as very large neural networks have been found to be useful for a variety of tasks, and companies and institutions that have abundant access to computational resources can leverage this to obtain a competitive advantage,” Gómez-Rodríguez says. “This kind of analysis needed to be done to raise awareness about the resources being spent […] and will spark a debate.”
“What probably many of us did not comprehend is the scale of it until we saw these comparisons,” echoed Siva Reddy, a postdoc at Stanford University who was not involved in the research.
The privatization of AI research
The results underscore another growing problem in AI, too: the sheer intensity of resources now required to produce paper-worthy results has made it increasingly challenging for people working in academia to continue contributing to research.
“This trend toward training huge models on tons of data is not feasible for academics—grad students especially, because we don’t have the computational resources,” says Strubell. “So there’s an issue of equitable access between researchers in academia versus researchers in industry.”
Strubell and her coauthors hope that their colleagues will heed the paper’s findings and help level the playing field by investing in developing more efficient hardware and algorithms.
Reddy agrees. “Human brains can do amazing things with little power consumption,” he says. “The bigger question is how can we build such machines.”
After spending a few years cutting through Big Tech’s B.S. about AI ethics, one of our AI writers created a glossary to help you decode what all of their favorite terms actually mean.
AI researchers often say good machine learning is really more art than science. The same could be said for effective public relations. Selecting the right words to strike a positive tone or reframe the conversation about AI is a delicate task: done well, it can strengthen one’s brand image, but done poorly, it can trigger an even greater backlash.
The tech giants would know. Over the last few years, they’ve had to learn this art quickly as they’ve faced increasing public distrust of their actions and intensifying criticism about their AI research and technologies.
Now they’ve developed a new vocabulary to use when they want to assure the public that they care deeply about developing AI responsibly—but want to make sure they don’t invite too much scrutiny. Here’s an insider’s guide to decoding their language and challenging the assumptions and values baked in.
accuracy (n) – Technical correctness. The most important measure of success in evaluating an AI model’s performance. See validation.
adversary (n) – A lone engineer capable of disrupting your powerful revenue-generating AI system. See robustness, security.
alignment (n) – The challenge of designing AI systems that do what we tell them to and value what we value. Purposely abstract. Avoid using real examples of harmful unintended consequences. See safety.
artificial general intelligence (phrase) – A hypothetical AI god that’s probably far off in the future but also maybe imminent. Can be really good or really bad whichever is more rhetorically useful. Obviously you’re building the good one. Which is expensive. Therefore, you need more money. See long-term risks.
audit (n) – A review that you pay someone else to do of your company or AI system so that you appear more transparent without needing to change anything. See impact assessment.
augment (v) – To increase the productivity of white-collar workers. Side effect: automating away blue-collar jobs. Sad but inevitable.
beneficial (adj) – A blanket descriptorfor what you are trying to build. Conveniently ill-defined. See value.
by design (ph) – As in “fairness by design” or “accountability by design.” A phrase to signal that you are thinking hard about important things from the beginning.
compliance (n) – The act of following the law. Anything that isn’t illegal goes.
data labelers (ph) – The people who allegedly exist behind Amazon’s Mechanical Turk interface to do data cleaning work for cheap. Unsure who they are. Never met them.
democratize (v) – To scale a technology at all costs. A justification for concentrating resources. See scale.
diversity, equity, and inclusion (ph) – The act of hiring engineers and researchers from marginalized groups so you can parade them around to the public. If they challenge the status quo, fire them.
efficiency (n) – The use of less data, memory, staff, or energy to build an AI system.
ethics board (ph) – A group of advisors without real power, convened to create the appearance that your company is actively listening. Examples: Google’s AI ethics board (canceled), Facebook’s Oversight Board (still standing).
ethics principles (ph) – A set of truisms used to signal your good intentions. Keep it high-level. The vaguer the language, the better. See responsible AI.
explainable (adj) – For describing an AI system that you, the developer, and the user can understand. Much harder to achieve for the people it’s used on. Probably not worth the effort. See interpretable.
fairness (n) – A complicated notion of impartiality used to describe unbiased algorithms. Can be defined in dozens of ways based on your preference.
for good (ph) – As in “AI for good” or “data for good.” An initiative completely tangential to your core business that helps you generate good publicity.
foresight (n) – The ability to peer into the future. Basically impossible: thus, a perfectly reasonable explanation for why you can’t rid your AI system of unintended consequences.
framework (n) – A set of guidelines for making decisions. A good way to appear thoughtful and measured while delaying actual decision-making.
generalizable (adj) – The sign of a good AI model. One that continues to work under changing conditions. See real world.
governance (n) – Bureaucracy.
human-centered design (ph) – A process that involves using “personas” to imagine what an average user might want from your AI system. May involve soliciting feedback from actual users. Only if there’s time. See stakeholders.
human in the loop (ph) – Any person that is part of an AI system. Responsibilities range from faking the system’s capabilities to warding off accusations of automation.
impact assessment (ph) – A review that you do yourself of your company or AI system to show your willingness to consider its downsides without changing anything. See audit.
interpretable (adj) – Description of an AI system whose computation you, the developer, can follow step by step to understand how it arrived at its answer. Actually probably just linear regression. AI sounds better.
integrity (n) – Issues that undermine the technical performance of your model or your company’s ability to scale. Not to be confused with issues that are bad for society. Not to be confused with honesty.
interdisciplinary (adj) – Term used of any team or project involving people who do not code: user researchers, product managers, moral philosophers. Especially moral philosophers.
long-term risks (n) – Bad things that could have catastrophic effects in the far-off future. Probably will never happen, but more important to study and avoid than the immediate harms of existing AI systems.
partners (n) – Other elite groups who share your worldview and can work with you to maintain the status quo. See stakeholders.
privacy trade-off (ph) – The noble sacrifice of individual control over personal information for group benefits like AI-driven health-care advancements, which also happen to be highly profitable.
progress (n) – Scientific and technological advancement. An inherent good.
real world (ph) – The opposite of the simulated world. A dynamic physical environment filled with unexpected surprises that AI models are trained to survive. Not to be confused with humans and society.
regulation (n) – What you call for to shift the responsibility for mitigating harmful AI onto policymakers. Not to be confused with policies that would hinder your growth.
responsible AI (n)- A moniker for any work at your company that could be construed by the public as a sincere effort to mitigate the harms of your AI systems.
robustness (n) – The ability of an AI model to function consistently and accurately under nefarious attempts to feed it corrupted data.
safety (n)- The challenge of building AI systems that don’t go rogue from the designer’s intentions. Not to be confused with building AI systems that don’t fail. See alignment.
scale (n)- The de facto end state that any good AI system should strive to achieve.
security (n) – The act of protecting valuable or sensitive data and AI models from being breached by bad actors. See adversary.
stakeholders (n) – Shareholders, regulators, users. The people in power you want to keep happy.
transparency (n) – Revealing your data and code. Bad for proprietary and sensitive information. Thus really hard; quite frankly, even impossible. Not to be confused with clear communication about how your system actually works.
trustworthy (adj) – An assessment of an AI system that can be manufactured with enough coordinated publicity.
universal basic income (ph) – The idea that paying everyone a fixed salary will solve the massive economic upheaval caused when automation leads to widespread job loss. Popularized by 2020 presidential candidate Andrew Yang. See wealth redistribution.
validation (n) – The process of testing an AI model on data other than the data it was trained on, to check that it is still accurate.
value (n) – An intangible benefit rendered to your users that makes you a lot of money.
values (n) – You have them. Remind people.
wealth redistribution (ph) – A useful idea to dangle around when people scrutinize you for using way too many resources and making way too much money. How would wealth redistribution work? Universal basic income, of course. Also not something you could figure out yourself. Would require regulation. See regulation.
withhold publication (ph) – The benevolent act of choosing not to open-source your code because it could fall into the hands of a bad actor. Better to limit access to partners who can afford it.2 free stories remaining Sign in Subscribe now
When Xiaomi entered the fiercely competitive smartphone market in 2010, it did so without even offering a real phone. The company only offered a free Android-based operating system (OS). Yet, within seven years, Xiaomi became one of the world’s largest smartphone makers, reaching $15 billion in revenue. Accelerating its growth rate, Xiaomi transformed into the world’s largest consumer IoT (Internet of Things) firm by 2020, with its revenue surpassing $37 billion and more than 210 million IoT devices (excluding smartphones and laptops) sold across more than 90 countries. How was Xiaomi able to grow so explosively and what lessons can other companies learn from Xiaomi’s rise?
We sought answers via an in-depth, multi-year study of the firm, including extensive interviews with 12 top executives (including cofounders, chairman, CEO, president, senior VPs, and executives leading R&D, distribution, and marketing), as well as the founder and CEO of Smartmi, Xiaomi’s largest ecosystem partner. Our research also involved analyzing more than 100 hours of conversations and reviewing more than 5,000 Xiaomi documents (from 2010–2020) as well as 470 external reports and data sets.
We learned that the secret to Xiaomi’s growth lies in what we term as “strategic coalescence.”The word “coalesce” originates from the Latin words co (“together”) and alescere (“to grow”). Strategic coalescence thus refers to a process through which a firm intimately connects with demand and supply-side stakeholders, bolsters tangible benefits for all, and triggers exponential market growth. Let’s first understand the key aspects of strategic coalescence at Xiaomi.
Coalescence with consumers
Xiaomi entered its first market — China — by offering a smartphone OS, called MIUI, for free. At the time, there were several strong domestic (e.g., Huawei, Lenovo) and international players (e.g., Apple, Samsung) battling over every tier of the market, from economical to premium. Most Chinese manufacturers simply smacked the Chinese version of Android on their smartphones, with little customization.
Instead of competing head on, Xiaomi courted tech-savvy smartphone users by offering them free software and building a fully-fledged online community to engage with them and understand which features they craved and which they disliked. This segment of consumers loved the unprecedented attention from a tech firm and were highly motivated to interact and contribute suggestions.
Xiaomi released a new OS version for download every Friday afternoon, as its tech-savvy consumers were heading home for the weekend. Its engineers followed up on user suggestions as soon as they were received, often corresponding with users to resolve issues together. This co-development process enhanced Xiaomi’s brand awareness and likability and prepared a segment of potential consumers for the entry of Xiaomi’s phones, without spending money on traditional advertising.
When it introduced its first phone in August 2011, Xiaomi positioned itself as offering “quality technology at an affordable price.” It sold directly to consumers, through its own website, at a margin of below 5% — the thinnest margin in the industry. Because of its direct engagement with tech-savvy consumers, Xiaomi was able to trim out all intermediaries — the many tiers of national, regional and local wholesalers and retailers, each of which charged a markup. Its direct-to-consumer approach created a significant cost advantage — the phone’s feature to price ratio was far more favorable than anything else on the market — and increased the speed at which Xiaomi could reach its consumers. Target consumers responded: Demand outpaced production so much that the firm could only open its e-commerce site one day per week and stocks sometimes sold out within minutes. The constant and instant sell-outs led to social media storms, spreading the brand to an ever-wider audience, stimulating further demand.
Coalescing operations around the core value proposition
After gaining a foothold in the tech-savvy, value-conscious segment in the top cities, Xiaomi began to expand into other segments — consumers who were less tech savvy, as well as those residing in smaller cities. Many of these consumers preferred an offline shopping experience, wanting to discuss their needs with a staff member or get a demonstration.
To serve these new customers, Xiaomi built an offline retail infrastructure, setting up hundreds of stores spanning major metros and small cities. Unlike other smartphone makers, who co-located their stores at the “telecom street” (an area dedicated to telecom stores), Xiaomi set up its stores in locations with high foot traffic, like malls, where its new target consumers were likely to shop. Importantly, Xiaomi chose malls where existing “high value at a reasonable price” anchor stores could help reinforce its own positioning. It also started offering different sub-brands (Redmi as an economical product line and Mi MIX for more advanced tech seekers), always ensuring that the features-to-price ratio of each new phone was more appealing than competing products.
In sum, during its initial phase, Xiaomi focused on quickly building a large smartphone consumer base across value-seeking consumer segments along with an appropriate on- and offline distribution infrastructure, always keeping its promised low margin on hardware. This enabled it to achieve massive volumes. Xiaomi expanded the share of wallet of this huge and growing customer base, with higher margin post-purchase services (commissions on music, videos, or game purchases), to help attain profitability. These laid the foundation for Xiaomi’s subsequent IoT endeavors.
Leveraging coalescing synergies
Xiaomi’s expansion into the IoT sphere was further empowered by four coalescing synergies.
In-Home IoT Synergy
Xiaomi leveraged its smartphones as an “omni-remote control” and began launching products that could be linked to and controlled by its phones (such as, TVs, air conditioners, air purifiers, smart lamps). In addition to developing its own products, Xiaomi sought partners who could help the firm quickly expand the range of its IoT offerings. The products from the partners were easily integrated into Xiaomi’s in-home system as they were built on its IoT protocol. This meant that once consumers acquired their first Xiaomi IoT product, they were more likely to seek out other products from Xiaomi. In other words, it became progressively harder for competitors to lure customers away in an IoT category.
Design Aesthetics Synergy
To further strengthen the bond between Xiaomi and its customers, the firm ensured that all Xiaomi-branded IoT products, including those manufactured by ecosystem partners, followed similar design aesthetics. So, if a consumer purchased another Xiaomi product, that item would be more aesthetically congruent with the Xiaomi products they already owned, creating synergy through design congruency and visual gestalt.
Product Portfolio Synergy
A key challenge associated with offline distribution is the high and ever-increasing square footage cost, especially at prime locations. Non-smartphone products (including those by partner firms) could yield much higher margins than smartphones, making the opening and running of offline stores more financially viable. Also, selling a variety of products in the stores attracted consumers who were not specifically looking for smartphones, creating opportunities to promote Xiaomi smartphones and, more generally, to cross-sell its entire portfolio. Furthermore, a broader product portfolio encompassing items with shorter replacement cycles (such as fitness bands and smart light bulbs) created higher foot traffic, leading to additional unplanned purchases and cross-selling opportunities in store.
Multi-Channel Synergy
To maximize returns on its brick-and-mortar stores, Xiaomi leveraged online sales data, using analytics to inform which products to sell offline and how to optimize the product mix at the store level. Offline stores were leveraged to offer potential consumers demonstrations for more experiential products (such as vacuum cleaner robots or AI speakers), moving those potential customers along the decision process whereby the demonstration could either seal an immediate purchase or nudge the consumer towards a later online purchase. The latter provided an additional multichannel synergy, from offline to online.
These four synergies coalesced together, amplifying each other’s effect. Consequently, Xiaomi was able to attract a growing number of potential customers to visit its stores (as opposed to smartphone competitors’ stores) and enhance the likelihood that they make purchases within Xiaomi’s ecosystem. This propelled rapid adoption of Xiaomi’s IoT products, with customers frequently visiting Xiaomi stores and purchasing multiple items.
Coalescence with partners
To effectively expand into categories outside Xiaomi’s expertise and bolster the four synergies, Xiaomi implemented a unique process for identifying and developing partnerships. These yielded some advantages:
Partners were hand-picked by Xiaomi cofounders and top executives through their personal networks. Because of the close personal connections, executives at Xiaomi had in-depth knowledge about each partner. They understood the capabilities and values of the management team, enabling Xiaomi to better assess the likelihood of collaboration success.
Leveraging personal networks meant that Xiaomi executives were also well connected to the social network of each partner. If some partners performed poorly or violated the partnership agreement, there would be an immediate and direct reputational cost to them, making it harder for them to leverage their social network for further business endeavors, a crucial success factor, particularly in the Chinese business context. This social cost complemented the economic incentive for partnering, bolstering the likelihood of successful collaboration. Certainly, this cherry-picking approach also had drawbacks — it limited the number of potential partners from which Xiaomi could select. However, in the case of Xiaomi’s IoT transformation, its executives believed the pros outweighed the cons.
Xiaomi invested in the partner firms but did not acquire controlling shares. While the investment incurred a risk, it created significant benefits. The “co-owner” relationship facilitated communication and increased trust. Xiaomi was able to gain access to information about each partner’s cost structure and operations, as well as participate in their business decisions. Because partners retained majority shares, they were motivated to develop and sell successful products. As a shareholder, Xiaomi benefited from the growth of its partner firms and the profits they made. Simply put, this form of co-ownership created a win-win outcome for both Xiaomi and its partners.
Xiaomi purposefully selected firms that were small or startups, so that partnering with Xiaomi offered significant value to them. These firms typically focused on a single category of products, and this specialization meant higher likelihood of producing great products. One important benefit Xiaomi offered to its partners was “incubation”: It assisted them with R&D by sending in teams of its own engineers, and it helped its partners identify key suppliers and negotiate contracts. Xiaomi’s investment and operational involvement brought brand awareness and prestige, making suppliers more willing to offer favorable terms to partner firms (compared to a “nobody” startup). Importantly, by ensuring that partner firms had access to solid designs and used quality inputs at reasonable cost, Xiaomi safeguarded the quality and price attractiveness of their final products.
These approaches enabled Xiaomi to effectively manage the partner network and offer an ever-growing portfolio of products consistent with the Xiaomi brand in design, aesthetics, quality, and technology/price ratio. Xiaomi’s coalescence with partners laid another foundation for the firm to become a global IoT giant.
Xiaomi’s growth path differs from conventional strategic thinking. While we are often taught that a firm’s strategy should be based on either cost leadership or differentiation and must serve either a few needs of a broad segment or broad needs of a narrow segment, Xiaomi is clearly an outlier. It differentiated on multiple frontiers and at the same time attained cost leadership. It achieved these through strategic coalescence — by coalescing with consumers and partners, which erected and continuously fortified barriers to entry on both the demand and supply sides. This resulting sustainable competitive advantage catapulted Xiaomi forward at warp speed.
Haiyang Yang is an associate professor at the Johns Hopkins Carey Business School, Johns Hopkins University. His research focuses on decision-making. His work has appeared in premier journals such as the Journal of Marketing Research, Journal of Consumer Research, Journal of Consumer Psychology, and Psychological Science.
Jingjing Ma is an assistant professor at the National School of Development, Peking University. Her research focuses on marketing. Her work has appeared in premier journals such as the Journal of Marketing Research, Journal of Consumer Research, and Journal of Consumer Psychology.
Amitava Chattopadhyay is the GlaxoSmithKline Chaired Professor of Corporate Innovation at INSEAD. He is co-author of The New Emerging Market Multinationals: Four Strategies for Disrupting Markets and Building Brands and has published extensively in premier journals such as the Journal of Marketing, Journal of Marketing Research, Journal of Consumer Research, Marketing Science, and Management Science. You can follow him on Twitter @AmitavaChats.
This list marks 20 years since we began compiling an annual selection of the year’s most important technologies. Some, such as mRNA vaccines, are already changing our lives, while others are still a few years off. Below, you’ll find a brief description along with a link to a feature article that probes each technology in detail. We hope you’ll enjoy and explore—taken together, we believe this list represents a glimpse into our collective future.
We got very lucky. The two most effective vaccines against the coronavirus are based on messenger RNA, a technology that has been in the works for 20 years. When the covid-19 pandemic began last January, scientists at several biotech companies were quick to turn to mRNA as a way to create potential vaccines; in late December 2020, at a time when more than 1.5 million had died from covid-19 worldwide, the vaccines were approved in the US, marking the beginning of the end of the pandemic.
The new covid vaccines are based on a technology never before used in therapeutics, and it could transform medicine, leading to vaccines against various infectious diseases, including malaria. And if this coronavirus keeps mutating, mRNA vaccines can be easily and quickly modified. Messenger RNA also holds great promise as the basis for cheap gene fixes to sickle-cell disease and HIV. Also in the works: using mRNA to help the body fight off cancers. Antonio Regalado explains the history and medical potential of the exciting new science of messenger RNA.2021
Large natural-language computer models that learn to write and speak are a big step toward AI that can better understand and interact with the world. GPT-3 is by far the largest—and most literate—to date. Trained on the text of thousands of books and most of the internet, GPT-3 can mimic human-written text with uncanny—and at times bizarre—realism, making it the most impressive language model yet produced using machine learning.
But GPT-3 doesn’t understand what it’s writing, so sometimes the results are garbled and nonsensical. It takes an enormous amount of computation power, data, and money to train, creating a large carbon footprint and restricting the development of similar models to those labs with extraordinary resources. And since it is trained on text from the internet, which is filled with misinformation and prejudice, it often produces similarly biased passages. Will Douglas Heaven shows off a sample of GPT-3’s clever writing and explains why some are ambivalent about its achievements.
Since its launch in China in 2016, TikTok has become one of the world’s fastest-growing social networks. It’s been downloaded billions of times and attracted hundreds of millions of users. Why? Because the algorithms that power TikTok’s “For You” feed have changed the way people become famous online.
While other platforms are geared more toward highlighting content with mass appeal, TikTok’s algorithms seem just as likely to pluck a new creator out of obscurity as they are to feature a known star. And they’re particularly adept at feeding relevant content to niche communities of users who share a particular interest or identity.
The ability of new creators to get a lot of views very quickly—and the ease with which users can discover so many kinds of content—have contributed to the app’s stunning growth. Other social media companies are now scrambling to reproduce these features on their own apps. Abby Ohlheiser profiles a TikTok creator who was surprised by her own success on the platform.
Electric vehicles come with a tough sales pitch; they’re relatively expensive, and you can drive them only a few hundred miles before they need to recharge—which takes far longer than stopping for gas. All these drawbacks have to do with the limitations of lithium-ion batteries. A well-funded Silicon Valley startup now says it has a battery that will make electric vehicles far more palatable for the mass consumer.
It’s called a lithium-metal battery and is being developed by QuantumScape. According to early test results, the battery could boost the range of an EV by 80% and can be rapidly recharged. The startup has a deal with VW, which says it will be selling EVs with the new type of battery by 2025.
The battery is still just a prototype that’s much smaller than one needed for a car. But if QuantumScape and others working on lithium-metal batteries succeed, it could finally make EVs attractive to millions of consumers. James Temple describes how a lithium-metal battery works, and why scientists are so excited by recent results.Advertisement
Technology companies have proven to be poor stewards of our personal data. Our information has been leaked, hacked, and sold and resold more times than most of us can count. Maybe the problem isn’t with us, but with the model of privacy to which we’ve long adhered—one in which we, as individuals, are primarily responsible for managing and protecting our own privacy.
Data trusts offer one alternative approach that some governments are starting to explore. A data trust is a legal entity that collects and manages people’s personal data on their behalf. Though the structure and function of these trusts are still being defined, and many questions remain, data trusts are notable for offering a potential solution to long-standing problems in privacy and security. Anouk Ruhaak describes the powerful potential of this model and a few early examples that show its promise.
Hydrogen has always been an intriguing possible replacement for fossil fuels. It burns cleanly, emitting no carbon dioxide; it’s energy dense, so it’s a good way to store power from on-and-off renewable sources; and you can make liquid synthetic fuels that are drop-in replacements for gasoline or diesel. But most hydrogen up to now has been made from natural gas; the process is dirty and energy intensive.
The rapidly dropping cost of solar and wind power means green hydrogen is now cheap enough to be practical. Simply zap water with electricity, and presto, you’ve got hydrogen. Europe is leading the way, beginning to build the needed infrastructure. Peter Fairley argues that such projects are just a first step to an envisioned global network of electrolysis plants that run on solar and wind power, churning out clean hydrogen.2021
As the coronavirus began to spread around the world, it felt at first as if digital contact tracing might help us. Smartphone apps could use GPS or Bluetooth to create a log of people who had recently crossed paths. If one of them later tested positive for covid, that person could enter the result into the app, and it would alert others who might have been exposed.
But digital contact tracing largely failed to make much impact on the virus’s spread. Apple and Google quickly pushed out features like exposure notifications to many smartphones, but public health officials struggled to persuade residents to use them. The lessons we learn from this pandemic could not only help us prepare for the next pandemic but also carry over to other areas of health care. Lindsay Muscato explores why digital contact tracing failed to slow covid-19 and offers ways we can do better next time.
Sign up for The Coronavirus Tech Report
– A weekly newsletter about covid-19 and how it’s changing our world.Sign up
Stay updated on MIT Technology Review initiatives and events?
We all use GPS every day; it has transformed our lives and many of our businesses. But while today’s GPS is accurate to within 5 to 10 meters, new hyper-accurate positioning technologies have accuracies within a few centimeters or millimeters. That’s opening up new possibilities, from landslide warnings to delivery robots and self-driving cars that can safely navigate streets.
China’s BeiDou (Big Dipper) global navigation system was completed in June 2020 and is part of what’s making all this possible. It provides positioning accuracy of 1.5 to two meters to anyone in the world. Using ground-based augmentation, it can get down to millimeter-level accuracy. Meanwhile, GPS, which has been around since the early 1990s, is getting an upgrade: four new satellites for GPS III launched in November and more are expected in orbit by 2023. Ling Xin reports on how the greatly increased accuracy of these systems is already proving useful.
The covid pandemic forced the world to go remote. Getting that shift right has been especially critical in health care and education. Some places around the world have done a particularly good job at getting remote services in these two areas to work well for people.
Snapask, an online tutoring company, has more than 3.5 million users in nine Asian countries, and Byju’s, a learning app based in India, has seen the number of its users soar to nearly 70 million. Unfortunately, students in many other countries are still floundering with their online classes.
Meanwhile, telehealth efforts in Uganda and several other African countries have extended health care to millions during the pandemic. In a part of the world with a chronic lack of doctors, remote health care has been a life saver. Sandy Ong reports on the remarkable success of online learning in Asia and the spread of telemedicine in Africa.
Despite the immense progress in artificial intelligence in recent years, AI and robots are still dumb in many ways, especially when it comes to solving new problems or navigating unfamiliar environments. They lack the human ability, found even in young children, to learn how the world works and apply that general knowledge to new situations.
One promising approach to improving the skills of AI is to expand its senses; currently AI with computer vision or audio recognition can sense things but cannot “talk” about what it sees and hears using natural-language algorithms. But what if you combined these abilities in a single AI system? Might these systems begin to gain human-like intelligence? Might a robot that can see, feel, hear, and communicate be a more productive human assistant? Karen Hao explains how AIs with multiple senses will gain a greater understanding of the world around them, achieving a much more flexible intelligence.
For a look at what technologies made our 10 Breakthrough Technologies lists in previous years, check out this page, which starts with 2020’s list.
In an effort to help improve healthcare organizations’ resilience against ransomware, MITRE this week unveiled its new Ransomware Resource Center, offering an array of tools and strategies for IT and infosec professionals to better guard against the growing epidemic of costly malware.
WHY IT MATTERS The Ransomware Resource Center tailors its many offerings around the role of the healthcare professional who might be accessing them – whether business manager, technical manager or IT or cybersecurity practitioner – and also around the five stages of the National Institute of Standards and Technology (NIST) Cybersecurity Framework: Identify, protect, detect, respond and recover.
It also offers a well-stocked resource library that’s searchable and can be filtered for the materials that might be the most useful.
The tools are drawn from MITRE’s own expertise, from government sources and from provider best practices. The goal is to convene a variety of resources in a single accessible and intuitive location, say MITRE officials, and to help “network defenders, IT administrators and business managers better prepare for, respond to, and recover from ransomware attacks.”
In a Q&A on the MITRE website, Joanne Fitzpatrick, lead cybersecurity engineer in MITRE’s Cyber Solutions Innovation Center, said particular attention was paid to small and/or underfunded IT and security staff.
“There are two key considerations,” she explained.
“First, such organizations typically have smaller IT and security departments, with a handful of talented people wearing many hats, and each responsible for several major operational IT areas. Staff tend to be experienced in the operations of their own organization, but often have little access to growth/training/professional development on cybersecurity issues, such as threats and attacks. Lack of time or budget is usually the reason.”
Moreover, at under-resourced organizations, there’s often “little-to-no extra staff available to dedicate to specialty cyber topics, such as threat modeling or attack surface assessments,” said Fitzpatrick.
“Second, we recognize that both small and large healthcare organizations may be targets for adversaries. Size does not matter. We’ve witnessed successful attacks at all types of health organizations. Adversaries may even exploit a smaller hospital as part of their attack navigation to exploit a larger, partnering organization.”
THE LARGER TREND MITRE points to a recent report that showed 560 healthcare facilities suffered a successful ransomware attack in 2020 – and another that saw a 45% increase in exploitation attempts just in the past four months.
“We are currently fighting not only the COVID-19 pandemic, but also an epidemic that is spreading through cyberspace: ransomware,” said newly appointed Homeland Security Director Alejandro Mayorkas recently. “In addition to disrupting city governments, schools and companies, ransomware has also been disrupting hospitals and health care facilities, who are already strained, going above and beyond the call of duty during this ongoing crisis.
“Last October, CISA, together with other government agencies, warned of the growing threat of ransomware targeting the healthcare and public health sector,” he added. “Previous ransomware attacks illustrate the risk to COVID-19 vaccine deployment efforts that depend on key production and logistics facilities.”
ON THE RECORD “We hope the Ransomware Resource Center will make two key contributions,” said Fitzpatrick. “It will inform hospitals and healthcare organizations about how to prepare, respond to and recover from such an attack. It also will share freely with the broader community the unbiased guidance and best practices that MITRE cybersecurity and cyber resiliency professionals have provided for years to our many federal government sponsors.”
While the defense industrial base is healthy, there are single points of failure and dependencies on overseas suppliers that must be addressed, the undersecretary of defense for acquisition and sustainment said.
“Over a period of years, we have offshored many, many sources of supply,” Ellen M. Lord said during an online discussion Thursday with the Hudson Institute. “It’s not for one reason; it’s for a variety of reasons, whether it be regulations, whether it be labor costs, whether it be government support of different industries.”
The deindustrialization of the U.S. over the last 50 years, the end of the Cold War and the focus it gave the U.S. on defeating the Soviet Union, digital technology and the rise of China have all created challenges to national defense.
In the newly released Fiscal Year 2020 Industrial Capabilities Report to Congress, Lord said the department looked into those challenges and their effects on the defense industrial base and proposed key actions to address them.
“What we did in this report was try to really capture those risks, look at the opportunities and come up with some specific steps that we can really take to reform how we go about looking at that supply chain and, in the endgame, really get capability downrange to the warfighter as quickly and cost-effectively as possible,” she said.
First, Lord said, the U.S. must re-shore more of its industrial base — bring it back to the U.S. and U.S. allies.
“There are a couple [of] key areas there with shipbuilding, as well as microelectronics — fundamental to our capability,” she said.
Development of a modern manufacturing and engineering workforce along with a more robust research and development base is also critical. Declines in U.S. science, technology, engineering and mathematics education and industrial jobs hurt the ability of the defense industrial base to innovate, Lord said.
“We want to make sure that we have modern manufacturing and engineering expertise,” she said. “We do not have nearly the number of scientists and engineers as China has. We need to make sure that we develop our talent to be able to leverage on these critical areas.”
The department must also reform and modernize the defense acquisition process to better meet the realities of the 21st century, Lord said.
“We’ve started with a number of those, but there’s much further to go,” she said. “We want to make sure that our traditional defense industrial base is widened to get all of those creative, innovative companies. We know the small companies are where most of our innovation comes from, and the barriers to entry — sometimes to getting into the Department of Defense — are rather onerous.”
Lord said part of modernizing and reforming defense acquisition is the recently announced Trusted Capital Marketplace, which will match potential defense suppliers — many of them small companies that have never done business with DOD — with the investors they need to keep operating and innovating. The Trusted Capital Marketplace will vet investors to ensure foreign ownership, control and influence is nonexistent.
Finally, Lord said, the department must find new ways to partner private sector innovation with public sector resources and demand.
“We, as the government, I believe, need to work with industry to make sure that we diversify that industrial base and, also, that we much more quickly translate technological capability into features of current platforms and weapon systems, as well as incorporate it in new ones,” Lord said.
After Amazon’s three-week re:Invent conference, companies building AI applications may have the impression that AWS is the only game in town. Amazon announced improvements to SageMaker, its machine learning (ML) workflow service, and to Edge Manager — improving AWS’ ML capabilities on the edge at a time when serving the edge is considered increasingly critical for enterprises. Moreover, the company touted big customers like Lyft and Intuit.
But Mohammed Farooq believes there is a better alternative to the Amazon hegemon: an open AI platform that doesn’t have any hooks back to the Amazon cloud. Until earlier this year, Farooq led IBM’s Hybrid multi-cloud strategy, but he recently left to join the enterprise AI company Hypergiant.Ad: (2:07)Skip AdMicrosoft says hackers viewed source code, didn’t change it, and other top stories in technology from January 05, 2021.
Here is our Q&A with Farooq, who is Hypergiant’s chair, global chief technology officer, and general manager of products. He has skin in the game and makes an interesting argument for open AI.
VentureBeat: With Amazon’s momentum, isn’t it game over for any other company hoping to be a significant service provider of AI services, or at the least for any competitor not named Google or Microsoft?
Mohammed Farooq: On the one hand, for the last three to five-plus years, AWS has delivered outstanding capabilities with SageMaker (Autopilot, Data Wrangler) to enable accessible analytics and ML pipelines for technical and nontechnical users. Enterprises have built strong-performing AI models with these AWS capabilities.
On the other hand, the enterprise production throughput of performing AI models is very low. The low throughput is a result of the complexity of deployment and operations management of AI models within consuming production applications that are running on AWS and other cloud/datacenter and software platforms.
Enterprises have not established an operations management system — something referred to within the industry as ModelOps. ModelOps are required and should have things like lifecycle processes, best practices, and business management controls. These are necessary to evolve the AI models and data changes in the context of the underlying heterogeneous software and infrastructure stacks currently in operation.
AWS does a solid job of automating an AI ModelOps process within the AWS ecosystem. However, running enterprise ModelOps, as well as DevOps and DataOps, will need not only AWS, but multiple other cloud, network, and edge architectures. AWS is great as far as it goes, but what is required is seamless integration with enterprise ModelOps, hybrid/multi-cloud infrastructure architecture, and IT operations management system.
Failures in experimentation are the result of average time needed to create a model. Today, successful AI models that deliver value and that business leaders trust take 6-12 months to build. According to the Deloitte MLOps Industrialized AI Report (released in December 2020), an average AI team can build and deploy, at best, two AI models in a year. At this rate, industrializing and scaling AI in the enterprise will be a challenge. An enterprise ModelOps process integrated with the rest of enterprise IT is required to speed up and scale AI solutions in the enterprise.
I would argue that we are on the precipice of a new era in artificial intelligence — one where AI will not only predict but recommend and take autonomous actions. But machines are still taking actions based on AI models that are poorly experimented with and fail to meet defined business goals (key performance indicators).
VentureBeat: So what is it that holds the industry back? Or asked a different way, what is that holds Amazon back from doing this?
Farooq: To improve development and performance of AI models, I believe we must address three challenges that are slowing down the AI model development, deployment, and production management in the enterprise. Amazon and other big players haven’t been able to address these challenges yet. They are:
AI data: This is where everything starts and ends in performant AI models. Microsoft [Azure] Purview is a direct attempt to solve the data problems of the enterprise data governance umbrella. This will provide AI solution teams (consumers) valuable and trustworthy data.
AI operations processes: These are enabled for development and deployment in the cloud (AWS) and do not extend or connect to the enterprise DevOps, DataOps, and ITOps processes. AIOps processes to deploy, operate, manage, and govern need to be automated and integrated into enterprise IT processes. This will industrialize AI in the enterprise. It took DevOps 10 years to establish CI/CD processes and automation platforms. AI needs to leverage the assets in CI/CD and overlay the AI model lifecycle management on top of it.
AI architecture: Enterprises with native cloud and containers are accelerating on the path to hybrid and multi-cloud architectures. With edge adoption, we are moving to pure distributed architecture, which will connect the cloud and edge ecosystem. AI architecture will have to operate on distributed architectures across hybrid and multi-cloud infrastructure and data environments. AWS, Azure, Google, and VMWare are effectively moving towards that paradigm.
To develop the next phase of AI, which I am calling “industrialized AI in the enterprise,” we need to address all of these. They can only be met with an open AI platform that has an integrated operations management system.
VentureBeat: Explain what you mean by an “open“ AI platform.
Farooq: An open AI platform for ModelOps lets enterprise AI teams mix and match required AI stacks, data services, AI tools, and domain AI models for different providers. Doing so will result in powerful business solutions at speed and scale.
AWS, with all of its powerful cloud, AI, and edge offerings, has still not stitched together a ModelOps that can industrialize AI and cloud. Enterprises today are using a combination of ServiceNow, legacy systems management, DevOps tooling, and containers to bring this together. AI operations adds another layer of complexity to an already increasingly complex model.
An enterprise AI operations management system should be the master control point and system of record, intelligence, and security for all AI solutions in a federated model (AI models and data catalogs). AWS, Azure, or Google can provide data, process, and tech platforms and services to be consumed by enterprises.
But lock-in models, like those currently being offered, harm enterprise’s ability to develop core AI capabilities. Companies like Microsoft, Amazon, and Google are hampering our ability to build high-caliber solutions by constructing moats around their products and services. The path to the best technology solutions, in the service of both AI providers and consumers, is one where choice and openness is prized as a pathway to innovation.
You have seen companies articulate a prominent vision for the future of AI. But I believe they are limited because they are not going far enough to democratize AI access and usage with the current enterprise IT Ops and governance process. To move forward, we need an enterprise ModelOps process and an open AI services integration platform that industrializes AI development, deployment, operations, and governance.
Without these, enterprises will be forced to choose vertical solutions that fail to integrate with enterprise data technology architectures and IT operations management systems.
VentureBeat: Has anyone tried to build this open AI platform?
Farooq: Not really. To manage AI ModelOps, we need a more open and connected AI services ecosystem, and to get there, we need an AI services integration platform. This essentially means that we need cloud provider operations management integrated with enterprise AI operations processes and a reference architecture framework (led by CTO and IT operations).
There are two options for enterprise CIOs, CTOs, CEOs, and architects. One is vertical, and the other one is horizontal.
Dataiku, Databricks, Snowflake, C3.AI, Palantir, and many others are building these horizontal AI stack options for the enterprise. Their solutions operate on top of AWS, Google, and Azure AI. It’s a great start. However, C3.AI and Palantir are also moving towards lock-in options by using model-driven architectures.
VentureBeat: So how is the vision of what you’re building at Hypergiant different to these efforts?
Farooq: The choice is clear: We have to enable an enterprise AI stack, ModelOps tooling, and governance capabilities enabled by an open AI services integration platform. This will integrate and operate customer ModelOps and governance processes internally that can work for each business unit and AI project.
What we need is not another AI company, but rather an AI services integrator and operator layer that improves how these companies work together for enterprise business goals.
A customer should be able to use Azure solutions, MongoDB, and Amazon Aurora, depending on what best suits their needs, price points, and future agenda. What this requires is a mesh layer for AI solution providers.
VentureBeat: Can you further define this “mesh layer”? Your figure shows it is a horizontal layer, but how does it work in practice? Is it as simple as plugging in your AI solution on top, and then having access to any cloud data source underneath? And does it have to be owned by a single company? Can it be open-sourced, or somehow shared, or at least competitive?
Farooq: The data mesh layer is the core component, not only for executing the ModelOps processes across cloud, edge, and 5G, but it is also a core architectural component for building, operating, and managing autonomous distributed applications.
Currently we have cloud data lakes and data pipelines (batch or steaming) as an input to build and train AI models. However, in production, data needs to be dynamically orchestrated across datacenters, cloud, 5G, and edge end points. This will ensure that the AI models and the consuming apps at all times have the required data feeds in production to execute.
AI/cloud developers and ModelOps teams should have access to data orchestration rules and policy APIs as a single interface to design, build, and operate AI solutions across distributed environments. This API should hide the complexity of the underlying distributed environments (i.e., cloud, 5G, or edge).
In addition, we need packaging and container specs that will help DevOps and ModelOps professionals use the portability of Kubernetes to quickly deploy and operate AI solutions at scale.
These data mesh APIs and packaging technologies need to be open sourced to ensure that we establish an open AI and cloud stack architecture for enterprises and not walled gardens from big providers.
By analogy, look at what Twilio has done for communications: Twilio strengthened customer relationships across businesses by integrating many technologies in one easy-to-manage interface. Examples in other industries include HubSpot in marketing and Squarespace for website development. These companies work by providing infrastructure that simplifies the experience of the user across the tools of many different companies.
VentureBeat: When are you launching this?
Farooq: We are planning to launch a beta version of a first step of that roadmap early next year [Q1/2020].
VentureBeat: AWS has a reseller policy. Could it could crack down on any mesh layer if they wanted to?
Farooq: AWS could build and offer their own mesh layer that is tied to its cloud and that interfaces with 5G and edge platforms of its partners. But this will not help its enterprise customers accelerate the development, deployment, and management of AI and hybrid/multi-cloud solutions at speed and scale. However, collaborating with the other cloud and ISV providers, as it has done with Kubernetes (CNCF-led open source project), will benefit AWS significantly.
As further innovation on centralized cloud computing models have stalled (based on current functionality and incremental releases across AWS, Azure, and Google), the data mesh and edge native architectures is where innovation will need to happen, and a distributed (declarative and runtime) data mesh architecture is a great place for AWS to contribute and lead the industry.
The digital enterprise will be the biggest beneficiary of a distributed data mesh architecture, and this will help industrialize AI and digital platforms faster — thereby creating new economic opportunities and in return more spend on AWS and other cloud provider technologies.
VentureBeat: What impact would such a mesh-layer solution have on the leading cloud companies? I imagine it could influence user decisions on what underlying services to use. Could that middle mesh player reduce pricing for certain bundles, undercutting marketing efforts by the cloud players themselves?
Farooq: The data mesh layer will trigger massive innovation on the edge and 5G native (not cloud native) applications, middleware, and infra-architectures. This will drive the large providers to rethink their product roadmaps, architecture patterns, go-to-market offerings, partnerships, and investments.
VentureBeat: If the cloud companies see this coming, do you think they’ll be more inclined to move toward an open ecosystem more rapidly and squelch you?
Farooq: The big providers in a first or second cycle of evolution of a technology or business model will always want to build a moat and lock in enterprise clients. For example, AWS never accepted that hybrid or multi-cloud was needed. But in the second cycle of cloud adoption by VMWare clients, VMWare started to preach an enterprise-outward hybrid cloud strategy connecting to AWS, Azure, and Google.
This led AWS to launch a private cloud offering (called Outposts), which is a replica for the AWS footprint on a dedicated hardware stack that has the same offerings. AWS executes its API across AWS public and Outposts. In short, they came around.
The same will happen to edge, 5G, and distributed computing. Right now, AWS, Google, and Azure are building their distributed computing platforms. However, the power of the open source community and the innovation speed is so great, the distributed computing architecture in the next cycle and beyond will have to move to an open ecosystem.
VentureBeat: What about lock-in at the mesh-layer level? If I choose to go with Hypergiant so I can access services across clouds, and then a competing mesh player emerges that offers better prices, how easy is it to move?
Farooq: We at Hypergiant believe in an open ecosystem, and our go-to-market business model depends on being at the intersection of enterprise consumption and provider offerings. We drive consumption economics, not provider economics. This will require us to support multiple data mesh technologies and create a fabric for interoperation with a single interface to our clients. The final goal is to ensure an open ecosystem, developer, and operator ease, and value to enterprise clients so that they are able to accelerate their business and revenue strategies by leveraging the best value and the best breed of technologies. We are looking at this from the point of view of the benefits to the enterprise, not the provider.
VentureBeat’s mission is to be a digital townsquare for technical decision makers to gain knowledge about transformative technology and transact. Our site delivers essential information on data technologies and strategies to guide you as you lead your organizations. We invite you to become a member of our community, to access:
up-to-date information on the subjects of interest to you,
our newsletters
gated thought-leader content and discounted access to our prized events, such as Transform
SIERRA & LENNYby 
