FTI Tech Trends 2022 AI

15th Edition • 2022 Tech Trends Report Vol.
01

Artificial Intelligence trends in business,
government, research, talent, creative
fields and society.
00 01 02 03 04 05 06 07 08 09 10 11 12 13 Artificial Intelligence TABLE OF CONTENTS
Table of Contents
03 Macro Forces and Emerging Trends 23 Robotic Picking and Replenishing 34 Health, Medicine, and Science 41 Ethics Guidelines for Tech 55 Applied AI for HR
05 Overview 23 100-Year Software Trends Contractors 56 Scenario: Say NO to the Dress!
07 Key Insights 24 Consumer 34 Protein Folding 42 The Mil-Tech Industrial Complex 57 Society
09 State of AI in 2022 25 Consumer Trends 34 AI Speeds Scientific Discovery 43 China Spotlight 59 Society Trends
10 Key Terms 25 Detecting Emotion 34 AI-First Drug Discovery 44 Expanding market 59 Expanding Use of Facial
12 AI: Techniques and Influential 25 Simulating Empathy and Emotion 35 AI Improves Patient Outcomes 44 China’s AI ethics Recognition Systems
Models 25 Theory of Mind Models 35 Deep Learning Applied to Medical 44 Cracking down on Big Tech 59 Increased Used of Ambient
14 Techniques to Watch in 2022 Imaging 44 Strategic panopticon Surveillance
26 Spotting Fakes
16 Models to Watch in 2022 35 Self-Driving Microscopes 45 Risk profile 59 Scattershot Approach to U.S.
26 Consumer-grade AI Applications
36 NLP Algorithms Detect Virus Regulations
18 Enterprise 26 Ubiquitous Digital Assistants 46 China’s supremacy
Mutations 59 Consolidation in AI’s Ecosystem
19 Enterprise Trends Get Smarter 47 Research
36 Using AI to Improve Talk Therapy 60 AI Alignment
20 Low-Code or No-Code Machine 27 Deepfakes for Fun 48 Research Trends
Learning 36 Dream Communication 60 Ethics Clash
27 Personal Digital Twins 50 Supersized AI Models
20 Web-Scale Content Analysis 36 Thought Detection 60 AI Still Has a Bias Problem
28 Creative Field 50 Unified Learning Processes
20 Serverless Computing 37 Scenario: The Real Future of AI 60 Deepfaking Trust
30 Creative Field Trends 50 Textless NLP
20 AI in the Cloud 38 Geopolitics and Military 61 Problematic Training Data
30 AI-Assisted Invention 51 Closed-Source Code
21 AI at the Edge 39 Geopolitics and Military Trends 61 Algorithms Targeting Vulnerable
30 Assisted Creativity 51 Framework Consolidation
39 The European Approach to Populations
21 Auto-Complete for Everything 30 Generative Algorithms for 51 Lowering the Cost of Training
Regulating AI 61 AI Intentionally Hiding Data
21 Advanced AI Chipsets Content Production Models
39 AI Nationalism 62 Undocumented AI Accidents
21 Processing-in-Memory Technology 31 Neural Rendering 51 Surpassing NLP Benchmarks
39 National AI Strategies 62 Prioritizing Trust
22 The Rise of MLOps 31 Generating Virtual Environments 51 No Retraining Required
from Short Videos 40 AI as Critical Infrastructure 64 Application
22 Robotic Process Automation (RPA) 51 Proliferation of Franken-Algorithms
31 Automated Versioning 40 Nation-based Guardrails 65 Key Questions
22 Predicting Workplace Injuries 52 AI Summarizing Itself
and Regulations 66 Trends in Action
22 Predicting Systems and Site Failures 31 Automatic Voice Cloning 53 Talent
and Dubbing 40 Regulating Deepfakes 67 Selected Sources
22 Liability Insurance for AI 54 Talent Trends
32 Automatic Ambient Noise Dubbing 40 Making AI Explain Itself 68 Authors
22 Manipulating AI Systems for 54 Demand for AI Talent Growing Fast
32 Ambient Interfaces 40 New Strategic Technical Alliances 70 About Future Today Institute
Competitive Advantage 54 AI Upskilling
33 Health, Medicine, and Science 41 AI to Guide Air Strikes 71 Disclaimer
23 AI Marketplaces 54 AI Brain Drain
41 Algorithmic Warfighting 72 Creative Commons License
23 AI-Powered InsureTech 54 Corporate AI Labs
2 © 2022 Future Today Institute

00 01 02 03 04 05 06 07 08 09 10 11 12 13 Artificial Intelligence EMERGING
SOURCES
TRENDS
00 01 02 03 04 05 06 07 08 09 10 11 12 13 Recognition, Scoring & Privacy 6 Watch Closely Informs Strategy Act Now
Macro Forces and Emerging Trends 4TH YEAR ON THE LIST 1
Scoring
For nearly two decades, the Future Today Institute has meticulously re-
searched macro forces of change and the emerging trends that result. Our KEY INSIGHT 2 EXAMPLES 3 DISRUPTIVE IMPACT 4 EMERGING PLAYERS 5
focus: understanding how these forces and trends will shape our futures. Our Scoring systems are In the U.S., we have a credit reporting
system that measures our creditworthi-
Advancement in data mining and
artificial intelligence promise both new
• Density
• Piwik Pro
leaving an indelible mark
15th annual Tech Trends Report identifies new opportunities for growth and on all individuals. For
ness. Banks, financial institutions, and
others use these scores to determine
opportunities and potential violations of
privacy as businesses and law enforce-
• Wakefit
• Dynamic Yield
potential collaborations in and adjacent to your business. We also highlight some, scoring systems
the likelihood that we might default on
a loan or a mortgage. Financial credit
ment implement scoring systems on
new platforms. Additional risk comes • Access Now
emerging or atypical threats across most industries, including all levels of create advantages scoring is available to all consumers— from China selling its government-fund- • Os Keyes and the University of
we can request copies of our financial ed scoring tools, which can be used on Washington’s Department of Human
and enhanced user Centered Design & Engineering
government. For those in creative fields, you will find a wealth of new ideas experiences. For others,
credit scores, check their accuracy, and
correct errors. Now, hundreds of types
both individuals and corporations, to
authoritarian regimes elsewhere in the • Amin Karbasi, associate professor of
that will spark your imagination. inherent biases and a of data are being harnessed to assign
us scores. However, unlike the credit
world. Even as new practical use cases
become apparent for scoring of this
electrical engineering and computer
science at Yale University
lack of available data reporting system, which is federally technology. • Ehsan Kazemi, research software
Our framework organizes nearly 600 trends into 13 clear categories, which on certain individuals regulated and follows set processes, this engineer at Google
kind of data isn’t subject to enforceable
• Voyager Labs
create hindrances and rules. It can be impossible to find out
are being published as separate reports. Each report includes specific use shortcomings. Copious what our scores are, how they are being
• Flock Safety
calculated, and how to correct inaccu-

cases and recommendations for key roles in many organizations: strategy, data and frameworks racies.
Scoring systems make assessments beyond just
for decision-making are
innovation, R&D, and risk. people. Food systems can evaluate nutritional
intake for a consumer.
integral for automated
systems to work.
Each trend offers six important insights.
1. Years on the List 3. Examples 5. Emerging Players 6. Action Scale  Informs Strategy
We track longitudinal tech and sci- Real-world use cases, some of Individuals, research teams, start- FTI’s analysis of what action your Strong evidence and data. Lon-
ence trends. This measurement in- which should be familiar to you. ups, and other organizations oper- organization should take. ger-term uncertainties remain.
dicates how long we have followed ating in this space. Mature organi- Fields include: This trend should inform your
the trend and its progression. 4. Disruptive Impact zations are included when they are strategic planning.
The implications of this trend on producing new contributions.  Watch Closely
2. Key Insight your business, government, or Mounting evidence and data,  Act Now
Concise description of this trend society. but more maturity is needed. Use Ample evidence and data. This
that can be easily understood and this trend to inform your vision, trend is already mature and
repeated to others. planning, and research. requires action.

00 01 02 03 04 05 06 07 08 09 10 11 12 13 Artificial Intelligence SCENARIOS
SOURCES
Scenarios Describe 00 01 02 03 04 05 06 07 08 09 10 11 12 13 Artificial Intelligence SCENARIO
Plausible Outcomes
People Will Live More Productive Lives • Optimistic/Near-Future scenario
You will find scenarios imagining future worlds as trends evolve The Real Future of AI
and converge. Scenarios offer a fresh perspective on trends and In the future, AI will be able to self-evolved machines that are
often challenge your deeply held beliefs. They prompt you to evolve on its own, creating new and
better versions of itself. This process
smarter than humans.
consider high-impact, high-uncertainty situations using signals will be known as “self-evolution.” In the future, AI will be so advanced
Self-evolution will allow AI to that it will be indistinguishable
available today. become smarter and more efficient from actual intelligence.
This scenario was written by
over time. It will also enable AI to Humanity will achieve a level of GPT-3, OpenAI’s language
1. Headline adapt to new environments and technological singularity, and all generator. (Humans did design
situations, making it more versatile our questions about the universe the page, however.) GPT-3 was
A short description offering you a glimpse into future changes. and powerful. AI has already will be answered. We will also
given the following instructions
by Amy Webb, the Future
shown signs of self-evolution in create powerful robots that can Today Institute’s CEO: “Please
2. Temporal and Emotive Tags recent years. For example, Google’s do everything humans can do, write a short op-ed around
500 words. Focus on how AI
AlphaGo program was able to only better. These robots will
A label explaining both when in the future this scenario is set and learn how to play Go better than eventually become self-aware,
will evolve in the future. Make
it sound like science fiction but
whether it is optimistic, neutral, pessimistic, or catastrophic. any other program by playing and we will have to deal with keep it factual.” Amy used GPT-
3 to produce several different
against itself. As AI continues to the consequences of creating
outputs, and selected this
evolve, we can expect even more something that is essentially equal
3. Narrative impressive feats from it – including to us.
scenario as her favorite. It took
3.29 seconds for GPT-3 to write
The descriptive elements of our imagined world, including the this essay, which Amy decided
required little editing.
developments leading us to this point in our future history.

Scenario sources:
The Future Today Institute uses a wide array of qualitative and
quantitative data to create our scenarios. Some of our typical
sources include patent filings, academic preprint servers, ar-
chival research, policy briefings, conference papers, data sets,
structured interviews with experts, conversations with kids, criti-
cal design, and speculative fiction.

00 01 02 03 04 05 06 07 08 09 10 11 12 13 Artificial Intelligence OVERVIEW
SOURCES
What’s the new normal? Most of us feel an urgent need to curious, rather than absolutely certain, about what you
get back to normal, especially as we continue to witness perceive in the present.
and endure an unprecedented amount of change. Early
this year, Russian President Vladimir Putin waged an un- Our 2022 Tech Trends Report is designed to help you
provoked war against Ukraine, which led to a rapid global re-perceive the world so that you can confront deep un-
response. Switzerland broke with tradition and took a side, certainty, adapt and thrive. In this 15th anniversary edition,
freezing Russian assets and denying entry to oligarchs. we have analyzed nearly 600 technology and science
Anonymous, the global hacker collective, waged a cyber- trends that impact most industry sectors. We created 14
war against Russia, infiltrating government databases and separate volumes, and each report focuses on a related
state propaganda websites. cluster of trends. You will also find vivid scenarios depict-
ing the unexpected ways in which the future might unfold.
Within the past 12 months, DeepMind solved a 50-year- In each volume, we discuss the disruptive forces, opportu-
old problem in biology, opening up new pathways to nities and strategies that will influence your organization
drug discovery. AI systems proved that they can gener- in the near future.
ate images and text as well as a human. Facebook and
Square changed their names to Meta and Block, signaling Now, more than ever, your organization should examine
a digital land grab in the emerging Web 3.0. The biggest the potential near and long-term impact of tech trends.
streamers—Netflix, Disney+, AppleTV, Hulu—discovered a You must factor the trends in this report into your strategic
formidable competitor in social commerce networks like thinking for the coming year, and adjust your planning,
Shein. Seemingly every day, a new set of signals emerge to operations and business models accordingly. But we hope
challenge our existing mental models. you will make time for creative exploration. The new nor-
mal is unfolding for those who know how to re-perceive
It might feel pointless to forecast the future past a few signals in the present.
weeks or months. But strategic foresight results in prepa-
rations, not predictions. Trends invite us to consider alter-
native outcomes from those we previously imagined. They
also unlock something invaluable in each one of us: the
ability to re-perceive reality. The act of “re-perception”
Amy Webb
awakens you to the possibility of a future that differs from
CEO
your current expectations. It helps you understand that you
The Future Today Institute
cannot know all things at all times, and that you should be

00 01 02 03 04 05 06 07 08 09 10 11 12 13 Artificial Intelligence
This is volume 01 in the Future Today

Institute’s 2022 Tech Trends Report. Each
volume covers a different set of topics.
To find additional volumes, visit

www.futuretodayinstitute.com/trends
00 01 02 03 04 05 06 07 08 09 10 11 12 13 Artificial Intelligence KEY INSIGHTS
Key Insights
 During the next decade, artificial  During the 2022 Super Bowl, several  Governments are instituting new Party expects to have outpaced the
intelligence will reshape the knowl- ads made use of deepfakes, which restrictions on mergers and acqui- U.S. in AI and become the singular
edge economy, automating some are becoming cheaper and easier sitions and investment activity to dominant force.
tasks currently performed by people to create. ensure that AI developed by com-
and augmenting others. panies does not aid foreign adver-  China is forging ahead on its own
 AI is being deployed to speed scien- saries. NLP models. Two massive models
 Companies are looking to upskill tific discovery. developed specifically for the Chi-
their workforce in machine learning  Several governments are attempt- nese language market, Wu Dao
and the basics of AI.  Three research teams from Micro- ing to regulate deepfake technol- 2.0 and M6, will start to shape the
soft, Google, and Baidu have sur- ogy. Bills to regulate or prohibit the future of AI development.
 Machine learning is transitioning, passed human baselines on Super- use of deepfakes have been intro-
as new platforms allow businesses GLUE natural language processing duced in California, Texas, and Mas-  Our future wars will be fought in
to leverage the power of AI to build (NLP) tasks. sachusetts, and a number of federal code, using data and algorithms as
applications without the need to bills are being discussed. powerful weapons.
know specific code.  Deep neural networks are being
used to analyze emotional states  The U.S. military has started using  We will soon reach a point when
 Deep neural networks are now using wireless signals. Scientists AI to guide its airstrikes, deploying we will no longer be able to tell if a
listed among the 10 fastest growing are warning that advertisers could algorithms to a live operational kill dataset has been tampered with, ei-
technologies in the U.S., as indicat- begin to alter and drive purchasing chain. ther intentionally or accidentally.
ed by patent applications. behavior through sleep and dream
hacking.  Within the next decade, China plans  A new type of software could be
 In some U.S. states, facial recogni- to meet two crucial milestones: By viable for 100 years, making use of
tion without consent is now illegal.  Strict new regulations in China and 2027, its People’s Liberation Army AI to dynamically adapt to changes
But there are no laws against emo- the EU could challenge the develop- will have a modern-ready force, and in environments and resources.
tional recognition. ment of AI. by 2030 the Chinese Communist

Artificial Intelligence

00 01 02 03 04 05 06 07 08 09 10 11 12 13 Artificial Intelligence Watch Closely Informs Strategy Act Now
State of AI in 2022
Artificial intelligence represents the third era of com- vaccine candidates. Because AI is so broad, we have
puting, generally defined as the ability for a machine identified different themes to track; however, because
to perform cognitive functions as well as or better than AI is now so deeply intertwined with every aspect of life,
humans. Such functions include perception, learning, you will find AI mentioned in other trends throughout all
reasoning, problem-solving, contextual understanding, 13 volumes of the Future Today Institute’s 2022 report.
making inferences and predictions, and exercising cre-
Since publishing our first Tech Trends report 15 years
ativity.
ago, we have included and expanded our coverage on
AI is now used across most industries. It solves business artificial intelligence. What began as several pages of in-
problems, detects fraud, improves crop yields, manages sights is now a dedicated, stand-alone report with more
supply chains, recommends products, and even assists than 100 trends to monitor. We believe AI is a force mul-
designers and writers in their work. AI can predict call tiplier because it both enables technology and is power-
volume in customer service centers and recommend ing the evolution of business, government, and society.
staffing levels; it also predicts the emotional state and AI is already transforming most economic sectors, but
behavior of the person calling to help companies antic- we anticipate deeper impacts across health care, bio-
ipate desirable solutions. AI automates the process for tech, cloud computing, fintech, consumer electronics,
drug discovery, which ultimately led to faster COVID-19 and military applications in 2022.

Key Terms
Machine Learning (ML) when teams know how to classify the Reinforcement learning (RL) videos, speech, and the like. Often, connected layer. Each one performs
Machine learning uses data to make input data and what they are trying to A system performs a task by repeatedly these systems are trained to learn on a different task using the data. The
predictions and recommendations on predict but can get accurate results running calculations as it attempts to their own, and they can sort through a output is classification. If a researcher
how to achieve stated goals. AI pioneer much more quickly by relying on an accomplish a stated goal. It’s a tri- variety of unstructured data, whether has 10,000 images and needs to extract
Arthur Samuel popularized the idea of algorithm rather than a human. This al-and-error process, where rewards or it’s making sense of typed text in docu- data—to recognize particular faces,
machine learning in 1959, explaining is the most common form of ML used penalties are earned in response to the ments or audio clips or video. In prac- for instance—the CNN would run until
how computers could learn without today. For example: Understanding what system’s performance toward achieving tical terms, deep learning’s emergence information could be inferred. In busi-
being explicitly programmed. This product features would most likely drive the stated goal. RL is used when there means that more and more human pro- ness, CNNs are used for recognition
would mean developing an algorithm new purchases is a business use case for isn’t enough training data, when the re- cesses will be automated, including the such as anomalies in medical imaging,
that could someday extract patterns supervised learning. searcher is trying to learn about an en- writing of software, which computers faulty products on a production line,
from datasets and use those patterns to vironment (such as a complex financial will soon start to do on their own. For blight on crops.
Unsupervised learning example, once a system learns what an
predict and make real-time decisions portfolio), or when the researcher needs
Data is provided to a model without object looks like—say, an apple—it can Recurrent neural networks (RNNs)
automatically. It took many years for to find greater levels of optimization. It
reality to catch up with Samuel’s idea, specific output parameters, and it tries has a tremendous number of business recognize that object in all other imag- These multilayered neural networks
but today machine learning is a primary to learn the dataset’s structure without use cases, ranging from real-time dy- es, even if it has only a partial view. move and store information between
driver of the growth in AI. any designated labels. For example, if namic pricing models to high frequency There are different types of deep learn- input, hidden, and output layers. They
a researcher doesn’t know quite what trading algorithms to the systems that ing architectures. The most common are good at modeling sequence data
There are different types of machine
to do with a large dataset, an unsuper- run self-driving cars. types include convolutional neural for predictions. In business, they are
learning including supervised, unsuper-
vised learning model could determine networks, recurrent neural networks, used anytime the sequence of data
vised, and reinforcement.
patterns, classify data, and make Deep Learning (DL) transformer neural networks, and gen- matters, such as speech recognition and
Supervised learning recommendations without a human erative adversarial networks (GANs). language translation. RNNs are used
supervisor. Unsupervised learning has Deep learning is a relatively new branch in digital assistants, to create captions
A model attempts to transform one type of machine learning. Programmers
been used during the pandemic to find Convolutional neural network (CNN) for images and to generate narrative
of data into another type using labeled use special deep learning algorithms
patterns in how COVID-19 is spreading A CNN is multilayered, with a convolu- reports using structured data (sports,
examples. Supervised learning is used alongside an enormous corpus of data—
throughout communities. tional layer, a pooling layer, and a fully financial).
typically many terabytes of text, images,

Key Terms
Transformer neural networks (TNNs) same data, such as images of people. Algorithm Transformer
A transformer is a component whose The networks compete against each A process describing how to solve a A type of neural network mechanism
purpose is to process sequential data, other to perform a task—identifying specific problem or how to complete a that learns what text means when it
such as natural language or genome the correct person—which results in particular task. appears in a particular context. Using
sequences. Transformers rely on “at- optimizing overall performance. GANs “attention,” a transformer looks at an
tention” (the mathematical description are useful when researchers don’t have Computer Vision input sequence and determines at each
of how things relate to, complement or enough data to train an algorithmic step what other parts of the sequence
Processes that give computers the
modify each other) in translating se- model. They are also used to create are important. To date, transformers
ability to derive meaningful information
quences. A transformer neural network new, synthetic data. Deepfakes, which have mainly been used in natural lan-
from digital images (including still and
is the unique architecture that enables have become popular in the past year, guage processing, image generation and
video) and to mimic and manipulate
systems to learn from context and to are generated using GANs. In design, genome sequencing.
such images.
generate new information. TNNs are GANs are tremendously useful: They
complementary to CNNs and RNNs, can produce thousands of designs and
recommend the best ones based on
Model
the two most common neural network
architectures used in deep learning. desired parameters. They can gener- A program that has been trained on a
ate and modulate voices, faces, even dataset. Models are generally used for
Generative adversarial networks gestures. Researchers from Nvidia, analytical and decision-making tasks,
(GANs) Massachusetts General Hospital, BWH such as making predictions.
Center for Clinical Data Science, and
As unsupervised deep learning systems,
the Mayo Clinic collaborated on a GAN Natural Language Processing
GANs are composed of two competing
This year is the 10th anniversary of an important that generates synthetic MRIs showing Processes that give computers the abil-
neural networks—a generator and a
milestone in AI. In 2012, a neural network taught cancerous tumors. ity to understand, mimic, and manipu-
itself to recognize a cat. It required 16,000 computer discriminator—that are trained on the
processors and 10 million YouTube videos. late human language.

AI: Techniques &

Influential Models

I may not be real. But for a second there,
I felt pretty alive.
— Guy, in Free Guy

AI: Techniques and Influential Models

Techniques to Watch in 2022
Automated Machine Learning phases when supervised training is continues to push its research forward. as inputs. GNNs are being used to detect understanding system failures. It builds
(AutoML) required. Continuous learning is more MuZero mastered multiple games with- smell—to predict odors at a molecular off both human- and system-generated
Some organizations want to move away about autonomous and incremental out being told their rules, a “significant level—and for a wide array of chemical observations to explain malfunctions
from traditional machine learning skill building and development, and step forward in the pursuit of gener- and biological processes. For example, related to input content and system
methods, which are time-consuming researchers will continue to push the al-purpose algorithms,” according to researchers at the Broad Institute used architecture.
and difficult and require data scientists, limits of what’s possible in this field. DeepMind. In a seminal paper, “Reward them to discover antibiotic compounds
specialists in AI fields, and engineers. Is Enough,” published at the end of that don’t have toxic side effects. Machine Image Completion
AutoML operates differently by match- Federated Learning 2021, DeepMind researchers hypothe-
If a computer system has access to
ing raw data and models together to Federated learning is a technique that
sized that artificial general intelligence Hybrid Human-Computer Vision enough images—say, millions and
reveal the most relevant information. could be achieved through reinforce-
distributes machine learning to the AI isn’t yet capable of fully functioning millions—it can patch and fill in holes
Google, Amazon, and Microsoft include ment learning alone.
edge. Introduced by Google researchers without human assistance. Hybrid intel- in pictures. This capability has practical
a host of AutoML products within their in 2016, it is a framework that makes it ligence systems combine humans and applications for professional photog-
cloud service offerings. possible for algorithms to use data on Graph Neural Networks AI systems to achieve greater accuracy. raphers, as well as for everyone who
devices—such as mobile phones and Because we perceive scents using mil- The U.S. Army Research Laboratory has wants to take a better selfie. Soon, if
Continuous Learning smart watches—without compromising lions of sensory neurons in our brains, a system that uses a brain-computer the foreground of a mountain is out
At the moment, deep learning tech- user privacy. Research in this space has and because scents are multifaceted, interface armed with computer vision of focus, or if your skin has an un-
niques are helping systems learn to dramatically increased. predicting the way something will smell technology and allows a person to sightly blemish, another version can
solve complex tasks in a way that resem- is incredibly complex. For example, rapidly see and sort images within her be swapped in to generate the perfect
bles what humans can do—but those General Reinforcement Learning how would you describe the smell of line of sight. CloudSight, which special- picture. As such technology becomes
tasks are still specific, such as beating Algorithms an orange? Sweet? Bright? Grassy? Each izes in image captioning, is working commonplace, there will be significant
a human at a game. And they require a descriptor is unique. Classifying smell is on a hybrid crowdsourced computer biases and other pitfalls to navigate. For
Researchers are developing single
rigid sequence: Gather data, determine tricky because it requires a multi-label vision system. Microsoft researchers example, image generation algorithms
algorithms that can learn multiple tasks.
the goal, deploy an algorithm. This system. Graph neural networks (GNNs) have proposed Pandora, a set of hybrid routinely reflect deeply culturally em-
DeepMind, the team behind AlphaGo,
process requires humans and can be constitute a particular type of deep human-machine methods and tools for bedded racism and sexism. A few years
which learned how to play Go with the
time-consuming, especially during early skill level of a human grand master, neural network that operates on graphs


Techniques to Watch in 2022
ago, if you typed “CEO” into Google Predictive Models Using and it will produce a realistic-looking than always relying on human program- Vokenization
Images, the first result of a woman was Incomplete Data synthetic photograph. This research will mers to sort, tag, and catalog data for Models like GPT-3 are trained on
CEO Barbie. In an experiment, re- Computer vision systems are getting someday enable robots to more easily them. Symbolic algorithms will aid the syntax and grammar, not creativity or
searchers at Carnegie Mellon University smarter. Neural networks can predict navigate human environments—and to process, which should eventually lead to common sense. So researchers at the
trained a system to autocomplete imag- geometry from a single color image. In interact with us humans by taking cues robust systems that don’t always require University of North Carolina–Chapel
es of men and women cropped below 2019, the DeepMind team developed a from our body language. Retail, manu- a human for training. Hill are combining language models
the neck. In pictures of men, the system generative adversarial network (GAN) facturing, and education settings could with computer vision. Humans learn in
autocompleted him wearing a suit. The that creates videos from images. For be especially relevant. Real-time Machine Learning a multilayered, multidimensional way,
system autocompleted women—includ- example: Imagine a photo of a person (RTML) so a new technique called vokenization
ing U.S. Rep. Alexandria Ocasio-Cortez holding a basketball. Based on his Neuro-Symbolic AI One big challenge in AI is building extrapolates language-only data by con-
(D-N.Y.)—wearing a low-cut top or bikini posture, face, and other data within the textually mapping language “tokens,” or
The development of AI has been on machines that can proactively collect
53% of the time. picture, the GAN figures out what likely the words used to train language mod-
two conceptual tracks since the 1950s: and interpret data, spot patterns and
happened next and generates a video symbolic (machines that use a base of incorporate context, and ultimately els, to related images, or “vokens.” For
Model-free Approaches to RL clip of the action. Earlier, researchers at knowledge and rules that represent con- learn in real time. New research into example, auto-generated image captions
Dreamer is a reinforcement learning MIT’s Computer Science and Artificial cepts) and non-symbolic (machines that RTML shows that it’s possible to use a often can’t infer context. Vokenization
(RL) agent that uses a world model to Intelligence Laboratory (CSAIL) trained use raw data to create their own pat- continual flow of data and adjust models would enable machines not just to rec-
learn long-sighted predictions, employ- computers to predict what humans terns and representations of concepts). in real time. This signals a big change in ognize objects but to really “see” what’s
ing backpropagation through model would do next using YouTube videos Classic AI is the former, because it more how data moves, and in how we retrieve in them.
predictions. It can create models from and TV shows such as “The Office” and closely represents how we understand information. The National Science
raw images and learn from thousands “Desperate Housewives.” CSAIL’s system human thought—and the original intent Foundation launched a $10 million
of predicted sequences in parallel using predicts whether two people are likely was to teach machines to think like us. grant program to catalyze research in
a graphics processing unit (GPU). This to hug, kiss, shake hands, or slap a high Researchers are working on new ways this area, although all of the big tech
new approach solves long-horizon tasks five. SinGAN is an unconditional gener- to combine both learning and logic companies are working closely to ad-
using an imagined world. ative scheme that can manipulate and using neural networks, which would vance RTML too.
enhance images—sketch a mountain, understand data through symbols rather


Models to Watch in 2022
DALL-E nAI, the system is capable of generating Google’s LaMDA and MUM
Developed by OpenAI last year, DALL-E text that is now indistinguishable from At last year’s I/O conference, CEO
is a model trained to manipulate visual human writing. In fact, we used GPT- Sundar Pichai demonstrated a model
concepts through language. It begins 3 to write one of the scenarios in this capable of managing open-ended
with a prompt that’s written in natural volume and only edited it lightly for conversations. That system, LaMDA,
language and then generates a set of length. GPT-3 was pitted against college appeared closer to natural conversation
images showing its interpretation of the students in an essay writing contest, and following an infinite number of paths
intended meaning. Named for Span- the anonymized papers were graded after an initial prompt. Google’s
ish surrealist artist Salvador Dali and by professors. It earned mostly B’s—the Multitask Unified Model, or MUM, is
Pixar robot WALL-E, the model relies same as human students. But the AI capable of understanding complex
on zero-shot capabilities, meaning that has demonstrated a strong anti-Muslim queries during search using natural
it does not need examples to perform bias. Researchers from Stanford Uni- language (text) and a variety of images.
generative tasks. In an early example, versity and McMaster University probed MUM could become the future of
researchers gave DALL-E the prompt “an the neural network on tasks including search, allowing consumers to speak
armchair in the shape of an avocado,” prompt completion, analogical reason- conversationally—and easily—to
and it returned a bevy of options ranging, and story generation. They found discover everything they’re looking for.
ing from a line drawing to an image of a that a Muslim-violence bias appears
chair that looks like something you’d see consistently and creatively in many Switch Transformer
in a furniture store catalog. use cases of the model. It’s yet another
Last year, Google Brain published an
example of how bias creeps into our
DALL-E is a model that generates images based on a prompt. Researchers inputted the prompts “an important paper demonstrating a new
armchair in the shape of an avocado” and “an armchair imitating an avocado,” and DALL-E returned a GPT-3 automated systems. Left unchecked, it
language model architecture. This
number of options. will cause problems throughout society
GPT-3 is an enormous AI that generates natural language processing AI model
as AI matures.
human-like language. Created by Ope- scales way up to 1.6T parameters while


Models to Watch in 2022
reducing training time. Many language

models are complex and dense; the Leading Applicants, 2017–2021
breakthrough of the switch transformer IBM 4,236
Developing AI systems based on
is that it reduces computational costs
biological models—or deep neural
while improving accuracy and power.
networks—is among the 10 fastest Samsung 2,415
growing technologies in the U.S., as
Chinese Models:
indicated by patent applications. Google 1,950
Wu Dao 2.0 and M6
Two models developed specifically for
the Chinese language market launched Intel 1,668
last year: Wu Dao 2.0 and M6. The
research team behind Wu Dao compare
Microsoft 1,574
it to GPT-3 but argue that the network is
orders of magnitude larger. Meanwhile,
Alibaba’s M6 is an efficient model that
reduces the cost of computing and—im-
pressively—reduces the carbon footprint
typically associated with training mod-
els on enormous datasets.
Source: IFI Claims Patent Services, January 2022 study.

Enterprise

Enterprise Trends
During the next decade, AI will re-

shape the knowledge economy,
AI Reshapes the Knowledge Workforce
automating some tasks currently
performed by people and augment-
ing others. Assistive AI technologies,
which include robotic process auto-
mation, low-code/no-code ML, the
deployment of pretrained transformer
models, natural language/conver-
sational search, and completion
systems will increase the productivity
and output of many knowledge econ-
omy professionals.
AI’s forecasted impact on productivity between 2022–2030, based on FTI research and modeling.
Sources include “Occupational Employment and Wage Statistics,” U.S. Bureau of Labor Statistics; “The Impact of Artificial Intelligence on Labor Productivity,” Eurasian Business Review; “Econom-
ic Impacts of Artificial Intelligence (AI),” European Parliament; “A Future That Works: Automation, Employment, and Productivity,” McKinsey Global Institute; Reuters Institute, Oxford University.

Enterprise Trends
Low-Code or No-Code matically changing the colors of objects Serverless Computing

Machine Learning (such as flowers or dresses) in images. Hyperscalers like AWS, Alibaba Cloud,
Machine learning is transitioning, as Last year, Amazon launched a no-code Microsoft’s Azure, Google Cloud, and
new platforms allow businesses to lever- mobile and web app builder for Amazon Baidu Cloud are rolling out new offer-
age the power of AI to build applications Web Services (AWS). Microsoft Power ings and packages for developers with
without the need to know specific code. Apps is a low-code application develop- the goal of making it easier and more af-
Businesses can turn their unruly data- ment environment on Azure. fordable for a wide swath of AI startups
sets into structured data that can be to launch their ideas into the market-
trained, and they can build and deploy Web-Scale Content Analysis place. AWS Lambda lets teams run code
models with minimal skills. Create ML Mining very large, unstructured data- for virtually any type of application or
is Apple’s no-code, drag-and-drop tool sets is now easier thanks to advanced back-end service—without provision-
that lets users build custom models such natural language processing collection ing or managing servers or hands-on
as recommendation engines, natural and classification. Trained to recognize administration. The Azure Functions
processing systems, and text classifiers. keywords, special algorithms can rapid- architecture supports myriad program-
Google Cloud’s AutoML includes image ly sort, classify, and tag information to ming languages, scales on demand, and
classification, object detection, transla- detect patterns. For example, a model charges only for active compute time.
tion, and all sorts of pattern recognition trained to search for hate speech can Some engineers worry that such server-
At Replicate, users can deploy StyleCLIP to give anyone a bowl haircut. Other physical transformations are tools to allow developers with limited detect bad actors in social networks. less systems require them to surrender
possible, too, and take only a few seconds. machine learning expertise to train too much control.
Machine translation generates training
high-quality models specific to their data for financial crime classification;
business needs. MakeML allows devel- last year, it reduced the amount of time AI in the Cloud
opers to create an AI app or solve busi- needed for classification from 20 weeks Corporate leaders within the AI eco-
ness problems using computer vision. (human analysts working alone) to two system have been racing to capture AI
Applications have included tracking weeks. cloudshare—and to become the most
tennis balls during matches and auto-

Enterprise Trends
trusted provider of AI on remote serv- computing. The Internet of Things and Auto-Complete for Everything that consume hundreds of kilowatts of chip, the Wafer Scale Engine 2 (WSE
ers. Enterprise customers are likely to its billions of devices, combined with 5G Google’s AI-based autocomplete tech- power. That is all starting to change. 2), has 2.6 trillion transistors, 850,000
stick with their initial vendor, because networking and increased computing nology is finding new applications. In Enter the SoC, or “system on a chip.” cores, has 40 gigabytes of on-chip
as machine learning systems get better power, has made large-scale AI at the Gmail, the Smart Reply feature checks Big tech companies, including Huawei, memory and 20 petabytes of memory
over time, the more data they amass. edge possible. Processing data direct- grammar and makes suggestions; in Apple, Microsoft, Facebook, Google, bandwidth. Amazon’s homegrown AI
For that reason, the competition is fu- ly on devices will be important in the Android, it automatically generates IBM, Nvidia, Intel, and Qualcomm, as chip AWS Inferentia, is a custom ma-
rious, even though it’s still early. In the future for health care, automotive, and responses to text messages. It now well as startups Graphcore, Mythic, chine learning chip used for high-per-
West, the field is led by Amazon, Micro- manufacturing applications because auto-completes formulas in Sheets, and Cerebras Systems, are all work- formance inference predictions—it now
soft, and Google, followed by companies it’s potentially faster and safer. Apple providing context-aware predictions— ing on new systems architecture and powers Alexa’s back-end services. The
including Apple, IBM, Salesforce, SAP, spent $200 million to acquire Xnor.ai, and querying datasets using natural SoCs—some of which come pretrained. pretrained chips will speed up commer-
and Oracle. In Asian markets, Alibaba a Seattle-based AI startup focused on language. New features will be rolled In short, this means that the chips are cialization and further R&D. But if the
and Baidu dominate the AI cloud, allow-power machine learning software out in Maps, which now auto-suggests more readily able to work on AI projects various device manufacturers all start
though in January 2020, telecom equip- and hardware. Microsoft offers a com- eco-friendly routes and ways to save and should promise faster and more creating unique protocols, developers
ment and smartphone maker Huawei prehensive toolkit called Azure IoT Edge on fuel consumption. Auto-complete secure processing. Projects that might may struggle with too many different
announced a management change to that allows AI workloads to be moved to offers convenience for consumers, but otherwise take weeks could instead be frameworks. We anticipate an eventual
focus on what it calls a “full-stack cloud the edge: Businesses can deploy com- as these AI capabilities are rolled into accomplished in a matter of hours. In consolidation, pitting just a few compa-
platform.” It’s a $250 billion industry plex event processing, machine learn- business applications, they could result 2019, Cerebras debuted an AI chip with nies—and their SoCs and languages—
and quickly growing. ing, image recognition, and other high in new productivity and efficiencies. 1.2 trillion transistors, 400,000 proces- against one another.
value AI without writing it in-house. sor cores, 18 gigabytes of SRAM, and
AI at the Edge Anyone is able to create AI modules and Advanced AI Chipsets interconnects (tiny connection nodes) Processing-in-Memory
AI-driven processing and decision-mak-
make them available to the community that can move 100 quadrillion bits per Technology
Today’s neural networks have long
for use through the Azure Marketplace. second. That’s an astounding amount
ing that occurs closer to the source of required an enormous amount of com- A new approach to memory, which
of components and power—and yet last
data generation, as opposed to in the puting power, take a long time to train, could eventually power the next gen-
year, it announced that its next-gen
cloud, is a technique known as edge and rely on data centers and computers erations of smartphones and help

Enterprise Trends
usher more smart devices to market, building models to operating them. supports call centers and help desks, Predicting Systems ers who inject fake training data into a
involves breaking some of the current Within software, a set of best practic- answering common questions and and Site Failures system? What if a health care company’s
bottlenecks in computing. Process- es known as DevOps relies on tools, scheduling services. Amazon uses RPA Computer vision can anticipate and AI misinterprets data and neglects to
ing-in-memory (PIM) integrates a pro- automation, and workflows to reduce to sift through résumés and prioritize identify failures in physical locations. identify cancer among certain patients?
cessor with RAM on a single chip, which complexity so that developers can focus top candidates. In banking, Blue Prism High-tech factories, airline manufac- These are the kinds of problems that
allows computations to be performed in on problems that need to be solved. This and Automation Anywhere help staff turers, and construction sites use image could put a company at risk of lawsuits.
the memory of a computer. In tests, PIM approach is now being used in machine with repetitive work functions. RPA recognition systems to monitor projects New insurance models will help address
tech delivers significant performance learning. Some of the fastest-growing will eventually augment staff and shift and automatically warn of problems. these issues. Underwriters are starting
gain while cutting energy consumption. GitHub projects are MLOps, or projects productivity into higher gear. This is accomplished by comparing data to include AI under cyber insurance
In late 2021, Samsung, which in addition that deal with tooling, infrastructure, from the real world to that of a digital plans. Specialty insurers such as LaPla-
to phones and consumer appliances is and operations. Going forward, MLOps Predicting Workplace Injuries twin. ya Insurance now offer insurance for AI
also the world’s largest manufacturer will describe a set of best practices that applications.
AI systems are being trained to detect
of dynamic random-access memory, combines machine learning, traditional Liability Insurance for AI
possible workplace injuries. Using com-
announced a new PIM-enabled chip DevOps, and data engineering.
puter vision models, Turkey-based In- Manipulating AI Systems
Who’s to blame when machines behave
that could eventually double the perfor-
tenseye can detect 40 types of employee for Competitive Advantage
badly? When the machine learning
mance of neural nets. This matters to Robotic Process health and safety incidents in real time. Amazon, Google, and Facebook have all
system in Uber’s self-driving car failed
the enterprise because it could bring AI Automation (RPA) The company says that it does not cap- come under fire in the past few years
and killed an Arizona pedestrian, the
applications to a wider array of devices
RPA can automate certain tasks and ture personally identifiable information company was likely not covered under for manipulating their search systems to
in the near future.
processes within offices and allow em- from the visual data it processes and traditional cyber insurance. As busi- prioritize results that are more profit-
ployees to spend time on higher-value that it detected 1.8 million unsafe acts nesses rush to build and implement AI able for their companies. For example,
The Rise of MLOps work. It’s the most commonly deployed in 2020 and 2021. San Francisco–based Google has been accused of de-ranking
products and processes, they must plan
As machine learning matures and new AI technique among enterprise compa- Voxel uses computer vision to enable ahead for emerging risks. For exam- websites and promoting news stories
applied business solutions emerge, nies. In health care and insurance, RPA security cameras to automatically detect ple, what happens if machine learning from preferred partners. Late in 2019,
developers are shifting their focus from is used to input and process claims. It high-risk activities in real time. makes a company vulnerable to attack- researchers found that Amazon had

Enterprise Trends
optimized its search algorithm to boost opers can upload their work and receive disaster. Sompo Japan is using the Trac- 100-Year Software
the visibility of Amazon’s own brands. payment when others pay to access it. table AI Estimating system to calculate Traditional software has a short and
Tweaks to search algorithms have a sig- Now there are specialized marketplac- the approximate repair cost of damaged unpredictable shelf life compared with
nificant impact on what internet users es for specific use cases: Nuance AI homes. other engineering tools. This leads to
see, whether that is news, products, or Marketplace developed a single API to headaches and costly upgrades, often
advertising. This resulted in the ongo- connect its algorithms to radiologists at Robotic Picking and with downtime. As a result, companies
ing antitrust lawsuits filed against the 6,500 health care facilities. Bonseyes is Replenishing and government agencies attempt to
companies. a European-specific marketplace to buy keep pace with the evolution of tech-
As the global pandemic kept people in-
and sell AI tools. nology by maintaining systems rather
doors, consumers increasingly relied on
AI Marketplaces e-commerce to shop. During that time, than evolving. Libraries, data formats,
Online hubs to share, buy, and sell AI-Powered InsureTech warehouse workers left their jobs. Enter and protocols can all become outdat-
models are growing in popularity. AWS Insurance companies are applying AI AI-powered robotic picking, sorting, ed quickly, creating vulnerabilities in
hosts its own marketplace, offering to assess damage and improve fore- and packing, which is making up for a critical systems. Since 2015, the Defense
models and algorithms for computer casts. The Vehicle Damage Inspection dearth of workers as consumers contin- Advanced Research Projects Agency
vision, speech recognition, and text— model, which is available on AWS ue to shop online. GXO Logistics, which (DARPA) has funded research to make
and its base of sellers includes Intel, Marketplace, uses a machine learning runs warehouses for Apple and Nike, software viable for more than 100 years.
CloudSight, and many others. (Think model to determine what part of a car is said that it will adopt robots and auto- These systems would use AI to dynami-
of AWS Marketplace as an Amazon for damaged. After photos are uploaded, it mated systems this year, to augment its cally adapt to changes in environments
algorithms and models.) SingularityNET assesses loss—and dramatically reduces human workforce. Amazon is experi- and resources. They require a novel
is an AI marketplace built on a block- the amount of time required for human menting with a suite of new robots to ef- approach to design, using AI to discov-
In Japan, a RightHand Robotics picking robot chain, where different hosted models appraisers to conduct their analysis. ficiently, and autonomously, move items er and make visible the application’s
sorts items. operations and interactions with other
are intended to intercommunicate. Following catastrophic typhoons and and boxes around warehouses. Globally,
Image credit: RightHand Robotics. One application—DeOldify—colorizes weather events in Japan, local insurance warehouses are expected to invest $36 systems.
old images. There are marketplaces for companies are relying on computer billion in automation in 2022, according
generalists, like GenesisAI, where devel- vision to assess damage after a natural to research group Interact Analysis.

Consumer

Consumer Trends
Detecting Emotion challenging, but companies with a large an authentic emotion? Theory of mind
A new type of neural network can deter- enough dataset are developing accurate refers to the ability to imagine the men-
mine how people are feeling. Using ra- models. Amazon’s Rekognition API tal state of others. This has long been
dio waves, AI can detect subtle changes infers someone’s emotions using facial considered a trait unique to humans
in heart rhythms, run a pattern analysis, recognition and physical appearance. and certain primates. AI researchers
and predict someone’s emotional state Replika uses AI to evaluate voice and are working to train machines to build
in a given moment. A team from Queen text, and over time it mirrors the user. theory of mind models of their own.
Mary University of London used a trans- Affectiva Human Perception AI analyzes This technology could improve existing
mitting radio antenna to bounce radio complex human states using speech AI therapy applications such as Woebot,
waves off of test subjects and trained analytics, computer vision, and deep a relational agent for mental health. By
a neural net to detect fear, disgust, joy learning. For example, the automotive designing machines to respond with
and relaxation, as people were shown sector uses Affectiva’s technology to empathy and concern, these technol-
different videos. The system accurately detect a driver’s emotional state—such ogies could eventually end up in hos-
tagged emotional states 71% of the time, as sleepiness or road rage—and make pitals, schools, and prisons, providing
which signals new opportunities for real-time suggestions to improve their emotional support robots to patients,
health and wellness applications, as well driving. students, and inmates. According to
as for job interviews and the govern- health insurer Cigna, the rate of loneli-
ment/military intelligence community. Theory of Mind Models ness in the U.S. has doubled in the past
50 years. In our increasingly connected
Research teams at Loving AI and Han-
Simulating Empathy world, people report feeling more isolat-
son Robotics are teaching machines
and Emotion ed. Future governments struggling with
Google Assistant’s Duplex automates scheduling, waiting on hold, and other tedious tasks. unconditional love, active listening,
a massive mental health crisis, such
AI can now measure biomarkers that and empathy. In the future, machines
as South Korea, may turn to emotional
suggest a person’s emotional state, such will convincingly exhibit human emo-
support robots to address the issue at
as agitation, sadness, or giddiness. tions such as love, happiness, fear, and
scale.
Precisely detecting human emotion is sadness. It begs the question: What is

Consumer Trends
Spotting Fakes Consumer-grade AI

Applications
Ubiquitous Digital Assistants
Get Smarter
In some U.S. states, facial
Researchers recently showed how AI
could be used to compose text so good
that humans couldn’t tell it was machine
Low-code and no-code offerings from
Amazon Web Services (AWS), Azure, and
Digital assistants (DAs)—like Siri, Alexa,
Google Assistant, and their Chinese
recognition without consent is
written. OpenAI demonstrated the
many reasons why this was problematic,
Google Cloud will start to trickle down
to everyday people, who will create their
counterpart, Tiān Māo from Aliba-
ba—use semantic and natural language
now illegal. But there are no laws
from mass-generating salacious social
media posts and fake reviews to forging
own AI applications and deploy them
as easily as they can a website. We’re
processing, along with our data, to
anticipate what we want or need to do
against emotional recognition.
documents by world leaders. It turns seeing a shift from highly technical next, sometimes before we even know
out that AI can also be used to detect AI applications used by professional to ask. Alibaba’s highly advanced DA
when text was machine generated, even researchers to more lightweight, us- can not only interact with real humans
if we humans can’t spot the fake. That’s er-friendly apps intended for tech-savvy but also deftly handle interruptions
because an essay written by AI tends to consumers. New automated machine and open-ended answers. Similar to
rely on statistical patterns in text and learning platforms make it possible for Google Assistant’s Duplex, Tiān Māo can
doesn’t have much linguistic variation. nonexperts to build and deploy predic- make calls on your behalf, but it also
Researchers at the MIT-IBM Watson AI tive models. Platforms hope that in the understands intent. So if you’re trying
Lab and Harvard University developed near future, we’ll use various AI appli- to schedule an appointment and men-
the Giant Language Model Test Room, cations as part of our daily work, just as tion that you’re usually commuting in
which looks for words that are likely to we do Microsoft Office and Google Docs the morning, the system infers that you
appear in a particular order. This tech- today. won’t be available then. In 2017, Future
nology can be used to detect forgery, Today Institute’s analysis correctly
intentional records falsification, email projected that nearly half of Americans
phishing campaigns, and corporate would own and use a digital assistant by
espionage. 2020. (An estimated 62% of Americans

Consumer Trends
use digital assistants today.) Amazon share on social media. Jiggy is a deep- startup, answers questions via text.
and Google dominate the smart speaker fake that makes anyone dance. For now, The near future could include digital
market, but digital assistants can be they all result in images and GIFs that twins for professionals across a range of
found in many places. Thousands of look like they’ve been manipulated—but fields, including health and education.
applications and gadgets now track and with the technology becoming so easy
respond to DAs. News organizations, to use, how long until we can’t tell real
entertainment companies, marketers, from fake?
credit card companies, banks, local au-
thorities, political campaigns, and many Personal Digital Twins
others can harness DAs to both surface
A number of startups are building
and deliver critical information.
customizable, trainable platforms
capable of learning from you—and then
Deepfakes for Fun representing you online via personal
Wombo is a lip-syncing app that allows digital twins. In 2021, China’s annual
consumers to transform any photo of a Spring Festival Gala on the country’s
The Reface app deepfakes consumer faces into popular movie scenes. person into a video of that person sing- state broadcaster (CCTV) included
ing. MyHeritage animates old photos. performances from synthesized celeb-
Faceswap is a free and open-source rities. With an estimated billion people
deepfake app powered by TensorFlow, watching, the AI copies mimicked their
Keras, and Python. Deep Art Effects human counterparts without pre-script-
offers desktop and mobile apps to turn ed behaviors, speeches, or routines.
images into stylized art. Reface is a face Meanwhile, Replika is a programmable
swap app that morphs your face onto digital twin that you can deploy for your
celebrity bodies and creates GIFs to friends. Molly, a Y Combinator–backed

Creative Fields

We are crossing a threshold into a new reality
in which AI is generating its own programs,
creating its own algorithms, and making
choices without humans in the loop. At the
moment, no one, in any country, has the right
to interrogate an AI and see clearly how a
decision was made.
— Amy Webb writing in The Big Nine: How the Tech Titans and Their
Thinking Machines Could Warp Humanity

Creative Field Trends
AI-Assisted Invention 2019, Nvidia launched GauGAN (named

Last year, the South African govern- after post-Impressionist painter Paul
ment granted a patent to an AI system Gauguin), a generative adversarial AI
called Dabus, which invented a method system that lets users create lifelike
to interlock food containers. It was a landscape images that never existed.
world-first—previously, patents had The National Institute of Informatics in
only been awarded to humans. In the Tokyo built an AI lyricist, while Amazon
U.S., the application was rejected, with released its DeepComposer system,
a judge citing case law stipulating that which composes music “automagically.”
only a human can hold a patent. It begs These AIs are not ostensibly intended
the question: What happens when AI to replace artists but rather to enhance
systems co-invent, or even entirely their creative process.
invent, new products?
Generative Algorithms
Assisted Creativity for Content Production
Generative adversarial networks (GANs) During the 2022 Super Bowl, several ads
are capable of far more than generat- made use of deepfakes, most notably
ing deepfake videos. Researchers are a commercial featuring basketball star
partnering with artists and musicians LeBron James talking to a de-aged ver-
to generate entirely new forms of sion of himself. OpenAI’s deep learning
creative expression. From synthesiz- algorithm released a neural network
ing African tribal masks to building called Jukebox that generates songs in a
fantastical, fictional galaxies, AI is bunch of different styles and simulated
being used to explore new ideas. In voices that sound (sort of) like Elvis
Amazon’s DeepComposer assists musicians with creating new melodies.

and others. The open-source algorithm Generating Virtual efficiency to everything from written simultaneously in multiple languages.
DeepFaceLab has been used by other Environments from articles to photo shoots. Virginia Tech In total, Tobi published 39,996 differ-
Short Videos
Looking in the artists and filmmakers.
Nvidia is teaching AI to build realistic
researchers developed an algorithm
that builds multiple versions of a single
ent versions of election stories that
averaged 250 words each. The articles
mirror again and Neural Rendering 3D environments from short video clips. human model by breaking down an
image into individual parts and then
carried a special byline alerting readers
that they’d been written by an algo-
Starting with a 2D image, researchers The method builds on previous research
again can now create a rich 3D view of a scene on GANs. Nvidia’s system generated using a GAN to reposition the body. This rithm. With more experiments under-
using a neural network to capture and graphics based on open-source datasets technique allows a neural net to swap way, we expect to see news and enter-
generate spatial imagery. Called neu- used by the autonomous driving field. out clothing by fitting different pieces to tainment media companies developing
Wishing the ral rendering, the process captures a Using short clips segmented into vari-
ous categories—such as buildings, sky,
a single body. Startup Flawless focuses
on just one area of the body: the mouth.
multiple versions of the same content
to reach wider audiences or to produce
photorealistic scene in 3D by calculating
reflection would the density and color of points in space. vehicles, signs, trees, or people—the It generates international versions using
deepfake dubs for TV shows and films,
massive amounts of content at scale.
The algorithm converts 2D pixels into GANs created new, different versions
tell me something voxels, which are a 3D equivalent. The of these objects. The array of possible which match video with dialogue to cre- Automatic Voice Cloning
result is a video which looks convinc- applications is vast. Automatically gen- ate realistic new lip movements. Theo- and Dubbing
ingly real. The many applications for erated virtual environments could be retically, a movie can be released in 100
Anyone who’s ever recorded a podcast
I, I can’t get a neural rendering include amping up used for fantasy and superhero movies languages simultaneously. Journalists
at Switzerland-based Tamedia exper-
is familiar with editing challenges
autonomous driving to help train algo- and could bring down the costs of TV such as guests talking over each other,
hold of myself rithms to recognize and react to novel production and game development. imented with generative techniques
interruptions from sirens and other
on-road situations. This technology will during their country’s 2018 election. A
background noises, and inconvenient
influence the future of video games, Automated Versioning decision-tree algorithm Tamedia named
— Haim sneezes. Those moments can stop a
virtual reality, and emerging metaverse Tobi generated automated articles de-
Creating content can be a time-consum- conversation cold. But what if you could
environments. tailing vote results for each municipality
ing, resource-intensive process. But new edit the spoken word the way you edit a
covered by the private media group’s 30
AI techniques could bring scale and word document? That’s the promise of
newspapers. It also produced content
AI companies including Resemble AI

and Descript, which make it possible world. For example, what sound is gen- delegate them on our behalf, and even
to clone voices. That means soon you erated when a wooden drumstick taps autonomously answer for us, depend-
might see a star like Phoebe Waller- a couch? A pile of leaves? A glass win- ing on the circumstance. Much of this
Bridge in a movie and also hear her, in dowpane? The focus of this research, invisible decision-making will happen
her own voice, speaking in Portuguese. underway at MIT’s Computer Science without direct supervision or input from
Voice-over actors can now rent out their and Artificial Intelligence Laboratory, people. What makes ambient design so
voices for use in a variety of languag- should help systems understand how tantalizing is that it should require us
es. Resemble AI’s system can translate objects interact with each other in the to make fewer and fewer decisions in
English into 15 other spoken languages physical realm. Numerous projects are the near future. Think of it as a sort of
while learning thousands of features now underway to make it easier to au- autocomplete for intent.
unique to one person’s voice. But the tomatically generate voices, videos, and
technology is also attractive to cyber- even storylines.
criminals, who have started using it to
infiltrate companies and dupe unwitting Ambient Interfaces
victims. Recently hackers used voice
Modern interfaces are able to do more
cloning tools to trick an employee into
for us with fewer direct actions—yet
thinking he was speaking on the phone
still captivate our attention. The average
to his CEO; he then transferred $243,000
adult now makes more than 20,000 deci-
to a scammer’s bank account.
sions a day—some big, such as whether
to invest in the stock market, and some
DeepFaceLab is an easy-to-use system that allows consumers to conduct high-quality face-swapping. Automatic Ambient small, such as whether to glance at a
Noise Dubbing mobile phone when the screen lights
For some time, we’ve been training up. Zero user interfaces—otherwise
computers to watch videos and predict known as ambient computing systems—
corresponding sounds in our physical promise to prioritize those decisions,

Health, Medicine
& Science

Health, Medicine and Science Trends
Protein Folding studied organisms, including yeast, fruit scientists at the University of British
In 2020, DeepMind’s AI made a big flies, and mice. All of this research will Columbia now rapidly test a new kind
announcement: It had solved a 50- allow biologists to study and gain new of solar cell and log results using a robot
year grand challenge with AlphaFold, insights on living organisms and patho- overseen by an AI algorithm. Based on
an AI tool that predicts the structure gens, which will form the basis for new the results of each experiment, the algo-
of proteins. AlphaFold outperformed drug development. rithm determines what to change next.
an estimated 100 teams in a biennial A 9- to 12-month process was completed
protein-structure prediction challenge AI Speeds Scientific Discovery in 5 days. DeepMind’s AlphaFold will al-
called Critical Assessment of Structure low scientists to synthesize new drugs to
Running experiments with several
Prediction. Predicting protein structures treat diseases and develop enzymes that
variables often requires tiny, methodical
has long vexed biologists. AlphaFold might someday break down pollution.
tweaks to measurements, materials,
had previously bested other teams, but and inputs. Graduate students might
it worked so quickly and so accurately spend hundreds of tedious hours repeat- AI-First Drug Discovery
that it signaled a near future when the edly making small adjustments until COVID-19 accelerated the use of AI in
technology could be used regularly by they find a solution—a waste of their drug discovery. An international team
other scientists. cognitive abilities. Research labs now crowdsourced a COVID antiviral by
Along with the newest version of Alpha- use AI systems to speed the process synthesizing candidates for 2,000 mol-
Fold, DeepMind published full details of of scientific discovery. Biotechnology ecules in less than 48 hours—a process
the system and released its source code. company Recursion uses computer that likely would have taken human
It also made a stunning reveal: Alpha- vision-based digital biomarkers, such researchers a month or longer. In
AI is being used to speed the development of new therapeutics.
Fold 2 has predicted the shapes of near- as respiratory rate, to assess and track Japan, the first phase of a clinical trial
ly every protein in the human body, as disease. Digitalizing in vivo studies will for an AI-designed drug to treat obses-
well as hundreds of thousands of other further shorten the time to gather data sive-compulsive disorder showed a pos-
proteins found in 20 of the most widely and identify drug efficiency. Materials itive result. The drug, DSP-1181, acts as

an agonist to the receptor for serotonin, may not. Still, drug discovery is tricky, AI Improves Patient Outcomes Deep Learning Applied to capture ultrasound images of the
a signaling molecule in the brain that because the algorithms rely on drug tar- New medical algorithms address the to Medical Imaging heart that expands who can read such
mediates mood. The project used AI gets that must be published in research level of patient care in the U.S. Patients A new system designed to improve scans. Nurses would just need a few
techniques to generate tens of millions journals. Most data about potential experience symptoms differently, and stroke outcomes, Viz.ai, showed prom- days of training on the software.
of potential molecules to try against compounds isn’t readily available, and their care is based on how they describe ising results last year in a real-world
the serotonin receptor and sift through when it is, it isn’t always complete or their symptoms and how those symp- study. An AI-based approach reduced Self-Driving Microscopes
the candidates to decide which ones to reliable. Filling the gaps and cleaning toms are interpreted by doctors. For ex- the amount of time it took to detect a Researchers at the Oak Ridge National
prioritize for synthesis and testing. that data takes time and money. It also ample, assessing the severity of arthritic brain stroke by 39%, which resulted in Laboratory are applying deep learning
AI-first drug startups are attractive to in- requires data sharing—and most drug pain is challenging. There is a standard more patients identified as eligible for to microscopy, which until now had
vestors. Recursion raised $121 million in data is proprietary and locked up by big scoring system to rate pain, which looks thrombectomy, a procedure that reduc- relied on humans to painstakingly orga-
2019 before spinning off CereXis, a new drugmakers. Using algorithms for drug at the amount of structural damage and es the chances of long-term disability. nize, observe, and analyze microscopic
independent entity to study rare brain development also brings up a host of missing cartilage, but data from the Radiologists and pathologists increas- samples. Deep learning will automate
cancers. Nearly every major pharma- ethical questions. Will bias invade drug National Institutes of Health found that ingly rely on AI to assist them with much of that process while also extract-
ceutical company has inked deals with discovery much like it has other arenas Black patients’ pain is underscored. It’s diagnostic medical imaging. In 2021, ing more information from samples. A
AI drug discovery startups, including of AI, thereby marginalizing certain pa- likely that the system itself, called the U.S. Food and Drug Administration ap- smart microscope would be able to view
Johnson & Johnson, Novartis, Merck, tients or diseases? Do algorithms need Kellgren-Lawrence Grade, was riddled provals allowed new products to be used and analyze samples in real time—while
AstraZeneca, and GlaxoSmithKline. their own clinical trials? Could AI be with bias when it was first developed widely in hospitals and clinics. So far, techniques that introduce a stimulus,
Much of the potential in AI stems from used to take shortcuts and undermine using primarily white British patients. most of the approved devices augment such as an electric field, will allow
deep learning’s ability to sort through the value of the science being done Researchers are training deep learning (rather than fully automate) the process researchers to gain real-world under-
huge volumes of information and learn inside the laboratory? Advocates say AI models instead, and finding gaps in of reviewing images and making diag- standing in more dimensions and on a
and extrapolate from that informa- will make drug development and clini- patient care. noses. But emerging autonomous prod- nanoscopic scale. This could become a
tion. The upshot: AI can think faster cal trials more efficient, thereby cutting ucts are making their way into clinical force multiplier in how scientific discov-
than humans—sorting data in months drug prices and paving the way for more settings. IDx-DR is an AI-enabled device ery is done.
versus years—and see patterns that we personalized medicine. that detects diabetic retinopathy using
retinal images. Caption Health uses AI

NLP Algorithms Detect By using this kind of modeling before Dream Communication to alter and drive purchasing behavior
Virus Mutations mutations occur, public health officials Scientists discovered how to establish through sleep and dream hacking.
Natural language processing (NLP) could strategize and potentially prevent two-way communication channels
algorithms, which are typically used new viral spreads. between lucid dreamers. Lucid dream- Thought Detection
for words and sentences, are being ers are aware that they are asleep and Deep neural networks are being used to
used to interpret genetic changes in Using AI to Improve can steer their dreams. In four global analyze emotional states using wireless
viruses. Protein sequences and genet- Talk Therapy studies, participants were outfitted signals. Researchers at Queen Mary
ic codes can be modeled using NLP The success of therapy and counseling with sensors attached to their heads University of London used radio signals
techniques—and can be manipulated requires trust between a clinician or and faces, and their data was fed into to measure subjects as they watched a
the way you’d produce text in word doctor and patient—and that trust is a computer that looked for patterns. video. The deep neural network ana-
processing software. At MIT, compu- built through dialogue. AI is now being Scientists verified a state of REM sleep, lyzed subtle body movements—breath-
tational biologists used NLP to solve a used to quantify linguistic interactions and then participants interacted with ing rates, heart rate—to reveal otherwise
vexing problem when developing new to determine what techniques work researchers using eye movements. The hidden information. If this reminds
vaccines. “Viral escape” is the ability for best. Startup Lyssn translates natu- study proved that there are new ways to you of the “Black Mirror” “Crocodile”
a virus to mutate and evade the human ral language into structured data and send and receive real-time information episode, you’re not far off: Research labs
immune system and cause infection. generates digital voiceprints, which while dreaming. Meanwhile, research- are developing new technologies to read
MIT researchers modeled viral escape identify the sentiment attached to each ers at MIT’s Dream Lab built a wearable our minds. There are business implica-
using NLP to identify how the virus sound. It’s hoped that this technology device that can track and interact with tions: HR departments could determine
might look different to the immune will be used to improve the techniques dreaming sleepers. A glove, outfitted what employees really think of company
system. The approach is similar to therapists use for cognitive behavioral with different sensors, plays prerecord- policies, lawyers could determine how
DeepMind hopes to predict a wide array of protein changing words in a sentence to change therapy, PTSD therapy, and other forms ed audio cues depending on the wearer’s jurors lean in a case, and realtors could
shapes. its meaning. For example: “I laughed at current state of sleep. Scientists at MIT,
of talk therapy delivered in-person or judge how serious a homebuyer is. But
the clown” versus “I cried at the clown.” via telemedicine. Harvard, and the University of Montreal the ethics are, of course, concerning.
are warning that advertisers could begin

00 01 02 03 04 05 06 07 08 09 10 11 12 13 Artificial Intelligence SCENARIO
People Will Live More Productive Lives • Optimistic/Near-Future scenario
The Real Future of AI

In the future, AI will be able to self-evolved machines that are
evolve on its own, creating new and smarter than humans.
better versions of itself. This process
will be known as “self-evolution.” In the future, AI will be so advanced
Self-evolution will allow AI to that it will be indistinguishable
become smarter and more efficient from actual intelligence.
This scenario was written by
over time. It will also enable AI to Humanity will achieve a level of GPT-3, OpenAI’s language
adapt to new environments and technological singularity, and all generator. (Humans did design
situations, making it more versatile our questions about the universe the page, however.) GPT-3 was
given the following instructions
and powerful. AI has already will be answered. We will also
by Amy Webb, the Future
shown signs of self-evolution in create powerful robots that can Today Institute’s CEO: “Please
recent years. For example, Google’s do everything humans can do, write a short op-ed around
500 words. Focus on how AI
AlphaGo program was able to only better. These robots will
will evolve in the future. Make
learn how to play Go better than eventually become self-aware, it sound like science fiction but
any other program by playing and we will have to deal with keep it factual.” Amy used GPT-
3 to produce several different
against itself. As AI continues to the consequences of creating
outputs, and selected this
evolve, we can expect even more something that is essentially equal scenario as her favorite. It took
impressive feats from it – including to us. 3.29 seconds for GPT-3 to write
this essay, which Amy decided
required little editing.

Geopolitics
& Military

Geopolitics Trends
The European Approach clear that the AIA would be enforced policy experiments and social exper-
to Regulating AI meaningfully, or that the cost of compli- iments, according to official govern-
Last year, the European Commission in- ance would be reasonable. ment documents. Meanwhile, China is
troduced a new proposal to regulate the planning for a world without American
use of artificial intelligence. The AI Act AI Nationalism technology, with government directives
(AIA) is intended to promote both inno- Governments are instituting new re- to prioritize homegrown technology
vation and consumer protection. It dis- strictions on mergers and acquisitions companies and software systems.
tinguishes between “high risk” and “low and investment activity to ensure that
risk” systems. For those categorized as AI developed by companies does not National AI Strategies
high risk, AI systems must meet several aid foreign adversaries. The U.S. Senate China passed its New Generation Ar-
criteria: (a) they should work as the user overwhelmingly passed legislation last tificial Intelligence Development Plan
intended and should be interpretable; June, dedicating $250 billion to scientific with aggressive benchmarks to become
(b) they should be secure and accurate; and technological research. Its center- the world’s dominant AI player within
(c) they should contain all necessary piece, the Endless Frontier Act, was 10 years; France adopted a national
technical documentation for proper use designed to boost U.S. competitiveness strategy called AI for Humanity; Saudi
and keep logs of their behavior; (d) they against China, especially in AI. It also Arabia has both a strategy and a legal
must have effective human oversight. creates a new technology directorate framework for making robots citizens;
The AIA would cover any person or within the National Science Foundation and the United Arab Emirates has a
organization (including foreign entities) with $100 billion in funding over five sweeping set of policy initiatives on AI
that use an AI system housed within years and earmarks $10 billion for local and appointed Omar Sultan Al Olama as
This year, the EU could pass a sweeping set of AI regulations. the EU. If adopted, the new regulation and regional tech hubs across America. its minister of state for artificial intelli-
would prohibit social scoring, real-time Meanwhile, in China the Ministry of gence. In the U.S., numerous public and
remote biometric sensing and analysis, Science and Technology established private groups work independently on
and any techniques that might target 20-city AI pilot zones that should open the future of AI on behalf of the nation.
and distort a person’s behavior. It isn’t by 2023. They will carry out AI-based Those efforts, however, lack interagency

Geopolitics Trends
collaboration and coordinated efforts tion interference through disinforma- these efforts could introduce new ways prohibit the use of deepfakes have been quality of output. It’s plausible that vari-
to streamline goals, outcomes, R&D ef- tion campaigns to political repression to safeguard against bias and to ensure introduced in California, Texas, and ous countries will enact new regulations
forts, and funding. A new wave of coun- enhanced by facial recognition and trust, they also each attempt to create Massachusetts, and a number of federal requiring explainability in the coming
tries will launch national AI strategies automated surveillance, major events strategic advantages for stakeholders. bills are being discussed. These initia- years. Imagine sitting beside a genius
this year. The OECD.AI Policy Observa- over the past few years have thrown As AI continues to develop according to tives will likely be met with arguments mathematician who gives you correct
tory now maintains a live repository of into sharp relief the dangers of artificial different rules in China, the EU, and the that prohibiting deepfakes infringes on answers in Italy, but receiving her an-
more than 700 AI policy initiatives from intelligence. Few guardrails exist for a U.S., one of the hallmarks of the field— free speech rights. swers across the border in France would
60 countries, territories, and the EU. technology that will touch every facet global academic collaboration—could mean asking her to stop and show her
of humanity, and countries are racing drastically decline. Making AI Explain Itself work—and every time she’s asked to
AI as Critical Infrastructure to develop and publish their own AI
You’ve undoubtedly heard someone
share her answers in a new country.
Government researchers are exploring
strategies and guidelines. The European Regulating Deepfakes argue that AI is becoming a “black
ways to spearhead AI development
Union developed an AI Alliance and
The U.S. National Defense Authorization box”—that even researchers working New Strategic Technical
for critical systems use: road and rail
plan of cooperation between member
Act includes provisions that address in the field don’t understand how our Alliances
countries, and Estonia is developing its
transportation systems; power gener- the growing problem of deepfakes, and newest systems work. That’s not entirely New strategic technical alliances be-
own legal framework governing the use
ation and distribution; and predicting the Department of Homeland Security true. However, there is growing concern tween countries will help drive future
of AI within the country. In 2020, China
routes for public safety vehicles, such as must now issue an annual report each among computer scientists, journalists, R&D but could also strain existing geo-
moved into position to lead the first set
ambulances and firetrucks. Rather than year for the next five years on the risk and legal scholars who argue that AI political alliances or heighten tensions.
of global AI norms and standards. It
shunning AI systems, there is new inter- posed by deepfakes. Last August, the systems shouldn’t be so secretive, and Likely partners include the U.S., Germa-
had previously published a report on
est in using the technology to prevent U.S. Senate Committee on Homeland Se- regulators are paying close attention. ny, Japan, India, South Korea, the U.K.,
technical standards that would allow
disasters and improve safety. curity and Governmental Affairs voted Broadly speaking, a few challenges must France, and Canada—leaving China
companies to collaborate and make
unanimously to advance the Deepfake be overcome. Requiring transparency and Russia to partner up separately.
their systems interoperable. The EU
Nation-based Guardrails Task Force Act, which would establish in AI could reveal a company’s trade The latter two countries have already
and the Organization for Economic
and Regulations a public-private team built to investi- secrets. Asking the systems to explain announced a technical alliance on satel-
Cooperation and Development similarly
gate technology strategies and policy their decision-making processes as they lites and deep-space exploration.
From self-driving car accidents to elec- published their own guidelines. While
that could curb risk. Bills to regulate or work could also degrade the speed and

Military (U.S. Focus) Trends
AI to Guide Air Strikes countries leading the world in AI re- Ethics Guidelines for
In 2021, the U.S. military said that it had search—the U.S., China, Israel, France, Tech Contractors
started using AI to guide its airstrikes, Russia, the U.K., and South Korea—are Project Maven was developed to enlist
deploying algorithms to a live operation- also developing weapons systems that AI to analyze surveillance video. Ini-
al kill chain. The kill chain is a process include at least some autonomous tially, Google was the DOD’s vendor, but
of gathering intelligence, performing functionality. The U.S. Air Force has when employees found out they’d been
analysis, weighing risks, and deploy- successfully flown an AI copilot on a working on a military project, thou-
ing weapons to destroy a target. Using U-2 spy plane in California, marking the sands protested. It wasn’t the first time
a modified process, an AI system was first time in the history of the Depart- tech contractors had lost trust in the
deployed into the Air Force Distributed ment of Defense that an AI algorithm government. As a result, the Defense
Common Ground System to analyze trained to execute specific in-flight Innovation Unit is enforcing “respon-
troves of intelligence, which would have tasks was deployed. With the call sign sible artificial intelligence” guidelines
required a significant amount of human ARTUµ, it was the mission command- that vendors must adopt when building
hours to complete. The new AI system er—though the flight was just practice. AI systems, models, or applications for
cannot order a strike on its own, but it is Future Today Institute analysis shows the DOD. The guidelines offer specif-
now automatically identifying possible that the future of warfare encompasses ic instructions that must be followed
targets. more than traditional weapons. Using during planning, development, and
AI techniques, a military can “win” by deployment, which include provisions
Algorithmic Warfighting destabilizing an economy rather than for risk assessment. This represents a
demolishing countrysides and city longer-term trend: government agencies
Future wars will be fought in code,
A U-2 spy plane prepares to land at Beale Air Force in California. The U-2 Federal Laboratory trained an AI centers. From that perspective, China’s requiring transparency in AI projects.
using data and algorithms as powerful
copilot to execute specific in-flight tasks that would otherwise be done by the pilot. unified march to advance AI puts the
weapons. The current global order
Image credit: Airman 1st Class Luis A. Ruiz-Vazquez/USAF. emerging superpower dangerously far
is being shaped by AI, and the same
ahead of the West.

Military (U.S. Focus) Trends
The Mil-Tech Industrial develop a virtual wall at the U.S.-Mexico Joint Warfighter Cloud Capability, will
Complex border to automatically detect migrants solicit proposals from both Microsoft
In 2021, U.S. President Joe Biden an- and traffickers. and Amazon, as they are likely the only
nounced a Pentagon budget request with The U.S. General Services Adminis- cloud service providers that can current-
$874 million earmarked for direct invest- tration and the DOD’s Joint Artificial ly meet the needs of the U.S. military.
ment in AI. In the U.K., the government Intelligence Center awarded a five-year, Many of these contracts have prompted
announced an investment of 6.6 billion $800 million contract to Booz Allen employee protests. In just one example:
pounds into military R&D over the next Hamilton for AI product development. Google employees were at one point
four years, while the EU launched a 8 bil- The U.S. Army awarded Lockheed Mar- directed to train AI systems to analyze
lion euros European Defense Fund with tin a $75 million contract for a machine drone footage—but they weren’t told the
a focus on AI through 2027. learning cyber jamming pod that can be project, known as Maven, was for the
In the past few years, some of the mounted on Humvees or drones. With DOD. A high-profile backlash ensued:
biggest AI companies in the U.S. have a new focus on defense roadmaps that As many as 4,000 Google employees
partnered with the military to advance include AI components, startups work- signed a petition objecting to Project
R&D and find efficiencies. In fact, the ing in high-resolution satellite imagery, Maven, and ultimately dozens resigned.
public sector cannot advance its technol- computer vision, and unmanned aerial Eventually, Google said it wouldn’t renew
ogy without help from outside compa- vehicles are attracting lucrative venture its contract on the project. The compa-
nies. Microsoft’s $22 billion contract for capital investment. ny launched a set of ethical principles
HoloLens builds on a previous prototyp- While the DOD called off its controver- governing its development and use of AI,
ing contract worth $480 million in 2018 sial $10 billion Joint Enterprise Defense including a provision that prohibits any
The AI algorithm known as ARTUµ flew with pilot U.S. Air Force Maj. “Vudu” on a U-2 Dragon Lady assigned
and a $10 billion cloud contract in 2019. Infrastructure (or JEDI) cloud contract systems from being used for “weapons
to the 9th Reconnaissance Wing at Beale Air Force Base. Air Combat Command’s U-2 Federal Laboratory
Anduril Industries, an AI-powered drone that pitted Microsoft and Amazon in a or other technologies whose principal
researchers developed ARTUµ and trained it to execute specific in-flight tasks that otherwise would be done
by the pilot. company, was awarded a $99 million legal battle against each other, a new purpose or implementation is to cause
five-year production order contract with multi-vendor cloud computing contract or directly facilitate injury to people.”
the DOD and will rely on Google Cloud to will be awarded. The new program, the

China Spotlight
China is an undisputed global leader Last year it took major steps into shap-
in artificial intelligence. Under Presi- ing the future of AI by releasing its own
dent Xi Jinping, the country has made pretrained models. China is forging
tremendous strides in many fields, but ahead on its own natural language
especially in AI. Businesses and the gov- processing models, which makes sense:
ernment have collaborated on a sweep- The most popular models in use now
ing plan to make China the world’s are trained on English text. Researchers
primary AI innovation center by 2030, from Tsinghua University and Alibaba
and it’s already making serious progress are developing Chinese datasets and
toward that goal. That plan is unlikely pretrained large transformers to com-
to be repealed by a new government; pete against the likes of GPT-3. In 2021,
China abolished Xi’s term limits and will two models developed specifically for
effectively allow him to remain in power the Chinese language went live: Wu Dao
for life. 2.0 and M6.
Within the next decade, China plans to The country’s enormous population
meet two crucial milestones: By 2027, of 1.4 billion offers researchers and
its People’s Liberation Army will have startups there a command of what may
a modern-ready force, and by 2030 the be the most valuable natural resource
Chinese Communist Party (CCP) expects in the future—human data—without the
to have outpaced the U.S. in AI and privacy and security restrictions com-
become the singular dominant force. mon in much of the rest of the world.
China is producing what it calls “intel- If data is the new oil, then China is the
ligentized” technologies to bolster both new OPEC. The kind of rich data the
its economy and military. Chinese are mining can be used to train
President Xi Jinping wants China to become the dominant AI force by 2030.

China Spotlight
AI to detect patterns used in everything its stakes in companies including Tesla are watching; technology that can companies that use or deploy AI tech- Chinese companies choosing to list on
from education to manufacturing to and Snap. The payoff for the Chinese is extract customer information and carry nologies within China would also need U.S. stock exchanges could face regula-
retail to military applications. not just a typical return on investment— out statistical analysis in crowded areas to comply with the new ethics rules. tory scrutiny and auditing. Ultimately,
That gives China an incredible advan- Chinese companies expect IP as well. like shopping malls and supermarkets; this regulation isn’t about “breaking
tage over the West. It also gives three of China-based AI startups now account and simultaneous recognition of every- Cracking down on Big Tech up” China’s Big Tech—the CCP wants
China’s biggest companies—Baidu, Alib- for nearly half of all AI investments thing in a scene, including people, pets, to focus its tech sector on achieving
Last year, the CCP initiated a wave of
aba, and Tencent—superpowers. Col- globally. automobiles, trees, or soda cans. In research and development goals set by
legislation aimed at its tech sector. Reg-
lectively, they’re known as the BAT, and 2020, it generated $525 million in reve- the government and military within the
ulators introduced anti-monopoly legis-
they’re all part of the country’s well-cap- Expanding market nue and is now worth $12 billion today.
lation focused on the platform economy decade.
italized, highly organized AI plan. In February 2022, Tencent made the list and promoted data security and privacy
of the world’s 10 most valuable compa- China’s AI ethics laws. The Personal Information Protec- Strategic panopticon
The BAT is important to you even if
you’ve never used them and don’t do any nies, with a valuation of $589 billion. (Its The CCP’s Chinese Governance Com- tion Law (PIPL), China’s version of the In late 2019, China began requiring all
business in China. That’s because these valuation soared to $1 trillion in 2021, mittee for the New Generation Artifi- EU’s GDPR, went into effect November citizens to submit to facial recognition
companies are now well established in but after CCP regulators cracked down cial Intelligence released a draft of its 2021. Next, they will tackle AI-specific in order to apply for new internet or mo-
Seattle and around San Francisco, and on its tech giants last year, the share new ethical norms and standards that regulation, including recommendation bile services, and began requiring that
they are investing significantly in U.S. price slid.) Since 2018, Alibaba’s sales all AI systems must conform to. The algorithms. The crackdown specifically telecom companies deploy AI to check
startups. Baidu (a search engine compa- have tripled, from less than $40 billion norms cover areas such as the use and targeted Alibaba, which was fined $2 the identities of people registering SIM
ny often likened to Google) established to more than $126 billion, according to protection of personal information, billion. Didi, the country’s largest ride cards. Chinese social media platforms
AI research centers in Silicon Valley data from S&P Global Market Intelli- human control over and responsibility sharing service, was prohibited from require users to sign up with their real
and Seattle, and Tencent (the developer gence. for AI, and the avoidance of AI-related registering new users while regulators names. In Chinese schools, surveillance
of the mega-popular messaging app Chinese startup SenseTime is pioneer- monopolies. There are no provisions conducted a cybersecurity review— cameras with computer vision are used
WeChat) began hunting for American ing myriad recognition technologies, yet explaining how the norms would be which analysts say had more to do with widely and track whether students are
talent when it opened an AI lab in Seat- such as a system that provides advertis- enforced, or what punishments could Didi listing itself on the New York Stock paying attention and whether they
tle a few years ago. It has since upped ers real-time feedback on what people follow. It’s also unclear whether foreign Exchange than any security problems. attempt to cheat or sleep. These and

China Spotlight
other national standards make it easier build international partnerships in both

for the government to track its citizens. physical and digital infrastructure, and
Top-tier AI Researchers Increasingly Hail From China
China’s social credit system, an algorith- it is making surveillance technologies
mic reputation system developed by the available to countries with authoritarian
government, standardizes assessments regimes. Two Chinese companies—the
of citizens’ and businesses’ behavior and state-controlled CEIEC and Huawei—
activity. built Ecuador’s surveillance system,
In 2020, numerous reports of abuse called ECU-911. The system promised to
revealed that China turned its AI on curb high murder rates and drug crime,
the ethnic Uyghur Muslim community. but it was too expensive an investment.
Huawei developed special AI software to As a result, a deal was struck for a Chi-
identify Uyghurs and alert local police. nese-built surveillance system financed
In 2021, China blocked social media with Chinese loans. It was a prelude to
platform Clubhouse after an open, a much more lucrative deal: Ecuador
democratic debate flourished on the eventually signed away big portions of
platform about the plight of the Uyghur its oil reserves to China to help finance
community. infrastructure projects. Similar package
deals have been brokered in Venezuela
Risk profile and Bolivia.
China is quietly weaponizing AI, too.
We have failed—and we are continuing
China’s People’s Liberation Army is
to fail—to see China as a military, eco-
catching up to the U.S. military, using AI
nomic, and diplomatic threat when it
for such tasks as spotting hidden images
comes to AI. China has already used its
with drones. The Chinese military is Country affiliations are based on the country where the researcher received their undergraduate degree.
Belt and Road Initiative as a platform to
Source: https://macropolo.org/digital-projects/the-global-ai-talent-tracker/

China Spotlight
equipping helicopters and jet fight- 5G and mobile footprints. At the same tion of aggressive actions. Each action
ers with AI. The government created time, China is drastically expanding its on its own might be insignificant. What
a top-secret military lab—a Chinese trading partners. While India, Japan, are the escalation triggers? We don’t
version of DARPA—and it’s building and South Korea have plenty of technol- have a definition, and that creates a stra-
billion-dollar AI national laboratories. ogies to offer the world, it would appear tegic vulnerability.
China’s military is achieving remarkable as though China is quickly ascending to
AI successes, including a recent test of global supremacy. At the moment, the
—Amy Webb
“swarm intelligence” that can automate U.S. is enabling this, and our leaders
dozens of armed drones. do not appear to be thinking about the
long-term consequences.
China’s supremacy When it comes to AI, we should pay
We’re living through a precarious close attention to China, which has talk-
moment in time. China is shaping the ed openly about its plans for cyber sov-
world order in its own image, while ereignty. But we should also remember
exporting its technologies and surveil- that there are cells of rogue actors who
lance systems to other countries. As could cripple our economies simply by
China expands into African countries mucking with the power or traffic grids,
and throughout Southeast Asia and Lat- causing traffic spikes on the internet, or
in America, it will also begin to eschew locking us out of our connected home
operating systems, technologies and appliances. These aren’t big, obvious
China is rapidly expanding its digital footprint. infrastructure built by the West. China signs of aggression, and that is a prob-
has already announced that it will no lem for many countries, including the
longer use U.S.-made computers and United States. Most governments don’t
software. China is rapidly expanding its have a paradigm describing a constella-

Research

Research Trends
Three research teams from Microsoft,

From AlphaGo to MuZero and
Google, and Baidu have surpassed Beyond
human baselines on SuperGLUE NLP In 2016, DeepMind unveiled Alpha-

Go, the first AI program capable of
tasks. defeating human players at Go—a
board game long held as the high-
water benchmark in the field. Then, in
2018, the team created a successor
called AlphaZero, which learned how
to master Go, chess and shogi (an an-
cient Japanese chess game)—from
scratch, without human trainers. Last
year, DeepMind published a stunning
paper in the journal Nature describing
MuZero, which mastered all previous
tasks plus Atari without needing to be
told the rules. This was a significant
step toward AI systems functioning
in unknown environments—and yet
another sign that general-purpose
algorithms are on the horizon.
Source: DeepMind
Is artificial intelligence less than our intelligence?
— Spike Jonze, filmmaker

Research Trends
Supersized AI Models Unified Learning Processes

Last year, we saw the proliferation of Deep neural nets are good at identify- The Top 25 Institutions for Top-Tier AI Research
large AI models—but supersized models ing objects in photos and videos and
are on the horizon. For context, GPT-3— processing natural language, but until
widely hailed as a powerhouse—has 175 recently models had to be trained sepa-
billion parameters. Huawei debuted a rately. Researchers have now developed
200 billion parameter language model Data2vec, a system that deploys a single
called Pangu, while Baidu and the Peng algorithm to train a neural network to
Cheng Lab released PCL-BAIDU Wenx- recognize images, text, or speech. It
in, with 280 billion parameters. PCL- unifies the learning process through
BAIDU is already deployed to Baidu’s self-supervised learning, which allows
news feeds, search engine, and digital the neural net to recognize patterns in
assistant. Gopher, which was released datasets on its own, without being fed
by DeepMind in December 2021, has labeled examples.
280 billion parameters. And Microsoft’s
Megatron-Turing NLG, built in collab- Textless NLP
oration with Nvidia, has 530 billion
Most large language models have been
parameters. Google’s Switch-Transform-
trained on publicly available datasets
er and GLaM models have a staggering
such as Reddit and Wikipedia. Both are
1 trillion and 1.2 trillion parameters,
rife with biases. Researchers are devel-
respectively—but even bigger is Wu
oping generative spoken language mod-
Dao 2.0 from the Beijing Academy of
eling, which extracts speech from raw
AI, which reportedly has 1.75 trillion
audio without labels or text. The hope is
parameters.
that AI could become more inclusive if
Source: https://macropolo.org/digital-projects/the-global-ai-talent-tracker/

Research Trends
it uses podcasts, local radio, and other ically mirrors trends in the commercial Surpassing NLP Benchmarks board. When SuperGLUE was intro- Proliferation of
sources of spoken language. application landscape. In the past five The General Language Understanding duced, there was a nearly 20-point gap Franken-Algorithms
years, Facebook seems to have gained Evaluation (GLUE) benchmark is a col- between the best-performing model and Algorithms are simply rules that define
Closed-Source Code ground. Of the conference papers that lection of resources for training, eval- human performance on the leaderboard. and automate the treatment of data.
mention the framework the researchers uating, and analyzing natural language Last year, AI models from Microsoft and They are built using “if this, then that”
Code is important for reproducibility,
used, 75% cited PyTorch but not Ten- understanding systems. It includes a Google surpassed human performance. logic that a computer can understand
accountability, and transparency, and
sorFlow. Of the 161 researchers who benchmark of tasks around understand- Existing language benchmarks still fail to and process. Here’s an easy example: If
it is a key to driving improvements in
published more TensorFlow papers than ing nine sentences or a pair of sentences capture biases encoded in public data— a website reader’s IP address is based
the greater AI community. But when
PyTorch papers, 55% of them switched built on existing datasets and selected future benchmarks could be designed to in Baltimore, the rules then allow that
academic researchers publish papers,
to PyTorch, while only 15% moved in to cover a diverse range of dataset sizes, resolve this gap. reader to freely access the site; if the IP
they don’t often include all of their code.
the other direction. text genres, and degrees of difficulty. It address is based in Belgium, then the
The reason given: The code they used
is intermingled with other proprietary includes a diagnostic dataset designed No Retraining Required rules first show a GDPR screen stating
research, and it therefore can’t be re- Lowering the Cost of to evaluate and analyze model perfor- Training robots to do more than one privacy and cookie policies. While a sin-
leased. Fewer than 20% of all academic Training Models mance with respect to a wide range of thing is difficult, but a new model pits gle algorithm might be easily described
papers on AI publish their full code, and It costs a lot to train a model. Several linguistic phenomena found in natural identical robot arms against one an- and deployed as expected, systems of
some big names—DeepMind and Ope- variables influence those costs, all of language. And it includes a public leader- other in a game (moving objects on a algorithms all working together can
nAI—notoriously leave theirs out, citing which have increased in the past few board so that researchers can track their virtual table in specific ways) in which sometimes pose problems. Developers
proprietary concerns. years. For example, it costs an average performance. The human baseline score one robot challenges the other with in- don’t always know in advance how one
of $1 per 1,000 parameters. OpenAI’s is 87, and between May 2018 and August creasingly difficult tasks. It’s an example algorithm will function alongside other
Framework Consolidation GPT-3, likely cost more than $10 million 2020, natural language processing sys- of multitask learning, a deep learning algorithms. Sometimes, several teams of
to train. For smaller research groups tems increased from 60 to 90.6, surpass- model in which machines learn differ- developers are working independently
Google’s TensorFlow and Facebook’s
and companies, the costs are out of ing humans. The SuperGLUE benchmark ent skills as they progress. OpenAI’s on different algorithms and datasets,
PyTorch are two popular frameworks
reach. Some in the AI community are is a new measurement of more difficult model allows a bot to solve new kinds of and they only see one another’s work
used by researchers, and the relative
instead allowing the big tech companies language understanding tasks, improved problems without requiring retraining. once it is deployed. This has been the
popularity of different frameworks typ-
to pretrain and publish big models. resources, and a new public leader- cause of recent stock market glitches

Research Trends
and e-commerce website wonkiness. It

is especially challenging for big compa- American and Chinese Number of Papers by Country (10+ papers)
nies like Facebook, which have billions Researchers Dominate NeurIPS
of algorithms working together at any Papers
given time.
The U.S. and China dominated the
papers accepted at NeurIPS in 2021,
AI Summarizing Itself
a prestigious international AI con-
An AI model can summarize scientific ference. A total of 9,122 papers were
literature, including research about submitted and 2,344 were accepted.
itself. The Allen Institute for Artificial The acceptance rate of 26% was the
Intelligence used the model in Seman- highest since 2013.
tic Scholar, an AI-powered scientific
paper search engine to provide a short
summary of papers on AI. What makes
this work impressive—and ultimately
so useful—is that it is capable of com-
pressing long papers with accuracy and
efficiency.

Talent

Talent Trends
Demand for AI Talent bootcamp to upskill its workforce, sors take one- or two-year sabbaticals to
Growing Fast including showing staff how to apply work at tech companies and then return In the U.S., Men Dominate Technical Jobs in AI
For many years, demand for AI talent AI-thinking to everyday tasks. Founded to their universities—but corporate ben-
Men overwhelmingly occupy the technical team positions that create AI systems
has outpaced supply. In the U.S., there by Harvard University and University efits can be difficult to give up. In one
in the U.S., according to a 2021 study by the Center for Security and Emerging
were nearly three times more AI-related of California, Los Angeles faculty, Univ. infamous case, Uber poached an entire
Technology (CSET) at Georgetown University. Women are more likely to be hired
job postings on Indeed last year than job AI is an online program for training in robotics lab from Carnegie Mellon Uni-
into sales, marketing, and product jobs. The lack of gender diversity in technical
views for AI-related roles. While schools machine learning and AI. versity—40 professors and researchers
positions could impact the quality of research, lead to increased prevalence of
are adding programs, increasing enroll- in total. Poaching departments today
algorithmic bias, and accelerate a long-term deterioration of talent pipelines.
ment, and adding classes, there are just AI Brain Drain could rob the future of future AI ex-
too many new needs for AI skills and perts: Without great scholars, who will
The brain drain of AI researchers out of
nowhere near enough trained workers. train the next generation of innovators?
academia and into corporations is grow-
As demand grows, the hiring process ing at an alarming pace. The reason is Technical Team 1 73.7% 26.3%
is taking longer and becoming more simple: compensation packages. Top Corporate AI Labs
expensive. This is impeding growth at academics earn generous salaries and AI labs are located around the world, 76.8% 23.3%
Technical Team 2
some companies. benefits, and they get to work in a simi- with concentrations in North America,
lar tenured environment that’s carefully Europe, and Asia. Facebook, Google,
Product Team 52.5% 47.5%
AI Upskilling cultivated to represent their experience IBM, and Microsoft operate more than
Companies are looking to upskill their in academia. Between 2004 and 2018, 60 labs dedicated to AI R&D, and the
Commercial Team 46.4% 53.6%
workforce in machine learning and Google, DeepMind, Amazon, and Mic- majority are outside of the U.S. because
the basics of AI. As a result, new train- rosoft hired 52 tenured and tenure-track of access to talent. During the Trump
professors from U.S. universities. In administration, immigration restric- All Workers 52.6% 47.4%
ing programs abound to augment the
knowledge, skills, and competencies return for their poaching, tech compa- tions and stringent visa requirements
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
required of a modern organization. Levi nies are endowing AI professorships at made it difficult to recruit talent into the
Strauss launched a machine learning top universities. In some cases, profes- United States, and overseas labs allowed Male Female
Source: “The U.S. AI Workforce Understanding the Supply of AI Talent,” by CSET).

Talent Trends
companies to overcome that barrier. they’re recording, algorithms work

Most of those labs do basic AI research invisibly, which means that this is an Enrollment in Stanford University’s natural language processing class is now 10 times the size it was in 2004.
rather than product development. area that could face regulatory scrutiny.
Consumer advocacy organization Elec- Stanford NLP class enrollment
Applied AI for HR tronic Privacy Information Center filed
Recognition systems can now be a complaint with the U.S. Federal Trade
deployed to watch people being in- Commission requesting an investigation
CS224D (DL NLP)
terviewed and to gauge enthusiasm, into HireVue, alleging its tools produce
CS224S (speech)
tenacity, and poise. Algorithms analyze results that are “biased, unprovable,
hundreds of details, such as the tone of and not replicable” through algorithmic CS276 (IR)
your voice, your facial expressions, and models. CS224U (NLU)

your mannerisms to best predict how CS124 (ugrad HLT)
you’ll fit in with the culture of a com-
CS224N (NLP)
munity. Startups such as HireVue use
AI systems to help companies decide
which candidates to hire. But this kind
of recognition technology has practical
applications well beyond job interviews:
It can detect when someone is likely to
make a purchase—or attempt to shoplift,
whether someone is lying, and whether
1998-1999
1999-2000
2000-2001
2001-2002
2002-2003
2003-2004
2004-2005
2005-2006
2006-2007
2007-2008
2008-2009
2009-2010
2010-2011
2011-2012
2012-2013
2013-2014
2014-2015
2015-2016
2016-2017
2017-2018
2018-2019
someone is receptive to new suggestions
and ideas. Unlike security cameras,
which tend to have a light indicating
Source: https://nlp.stanford.edu/
00 01 02 03 04 05 06 07 08 09 10 11 12 13 Artificial Intelligence SCENARIO
Say NO to the Dress! • Near-future optimistic scenario
Say NO to the Dress!

Getting married and How it works:
searching for a dress?
1 Start with a prompt, such as: “I don’t like form-fitting
Are you overwhelmed? dresses. I don’t like shiny fabric. I don’t want it to be
tight or restrictive. I hate buttons.”
Are you someone who 2 Our AI will create images that fit your description.
doesn’t know what they Select the 10 that look closest to what you’d be willing
want – but you definitely to wear.
know what you don’t 3 Write a second prompt about what you like about your
want to wear on your selections, such as: “I like the shape of #1, the color of
special day? #2, and the length of #5.”
4 Our AI will create a second round of images that should

Say no to the dress! be closer to what you want. Repeat until you find a
Use our neural network dress that suits you.
to help eliminate choices 5 When you’re ready, use the AI-generated image of the
and find exactly what dress to search for one that looks similar. If you can’t
you want. find one, our seamstresses are ready to custom-create
the dress to your exact measurements.
We use a transformer language model that receives text and image as a single stream of data. It’s been trained
by our computer scientists to create an infinite number of plausible images exploring clothing. In addition to
traditional wedding dresses, our model will generate an array of gender neutral options, menswear options,
and options for the entire bridal party.

Society

“There’s a real danger of systematizing the
discrimination we have in society [through AI
technologies]. What I think we need to do — as
we’re moving into this world full of invisible
algorithms everywhere — is that we have to be
very explicit, or have a disclaimer, about what
our error rates are like.”
— Timnit Gebru, research scientist

Society Trends
Expanding Use of Facial 18,000 federal, state, county, and mu- on sensitive projects. Those working in and finances. The legislation is intended
Recognition Systems nicipal law enforcement agencies in the information security and risk manage- to increase transparency and reduce
While there has been widespread outcry U.S. are its customers. While many U.S. ment should pay special attention to bias. A version of this legislation was
over using facial recognition systems, states and cities ban law enforcement advances in computer vision. introduced in 2019 but failed to gain any
U.S.-based consumers are likely to be and government use of facial recogni- traction.
seen by more of them in the coming tion systems, local bans do not prevent Scattershot Approach to
years. A 90-page report published by the federal use. U.S. Regulations Consolidation in AI’s
U.S. Government Accountability Office In the U.S., there is no federal law Ecosystem
details how federal agencies currently Increased Use of Ambient protecting consumer privacy as there is As much as the AI ecosystem booms,
use, and plan to expand their use of, Surveillance in the EU. Only three states have passed a rush of acquisitions means consoli-
facial recognition systems.* Ten of 24 What happens behind closed doors may comprehensive laws: California, Colo- dation, too. Big companies now snap
agencies surveyed plan to broaden their not be secret for long, and executives rado, and Virginia. Existing laws tend up startups long before they have time
use of the technology by 2023. The GAO should be aware of new ambient surveil- to cover data in particular domains like to mature—the average age at acquisi-
reports that agencies are primarily lance methods. Scientists at Massachu- health, education, and credit. There tion is 3 years old. Just a handful of big
interested in using facial recognition setts Institute of Technology discovered are no federal regulations preventing companies dominate the AI landscape:
for cybersecurity, domestic law enforce- how to use computer vision to track data consumer data from being used by Google, Amazon, Microsoft, IBM, Face-
ment, or physical security. Most of the from what they call “accidental camer- third parties. Legislation announced in book, and Apple in the U.S., and Baidu,
systems currently used in the U.S. are as.” Windows, mirrors, corners, house- February 2022 known as the Algorith- Alibaba, and Tencent in China, with sig-
federally owned, though commercial plants, and other common objects can mic Accountability Act, sponsored by nificant fortification and support from
vendors—including Acuant FaceID, be used, along with AI, to track subtle Sen. Ron Wyden (D-Ore.), would give the their country’s government. On the
Timnit Gebru is a respected leader in AI ethics re- Vigilant Solutions, and Clearview AI, changes in light, shadows, and vibra- Federal Trade Commission more staff to investment side, Qualcomm, Tencent,
search who coauthored a groundbreaking academ- which came under intense scrutiny last tions. The result: We all may soon have oversee the auditing and enforcement Intel Capital, Google Ventures, Nvidia,
ic paper showing how facial recognition systems
did not identify people of color and women with the
year—are increasingly used. Last March, X-ray vision capabilities—which may not of AI automation systems that make Salesforce, Samsung Ventures, Alibaba,
same accuracy as Caucasion people and men. Clearview reported that 17% of the be great news for companies working decisions about employment, housing, Apple, Baidu, Citi, and In-Q-Tel fund
*FTI CEO Amy Webb is a fellow in the U.S. GAO.

Society Trends
much of the growth. When it comes to Ethics Clash tent.) In January 2021, the board made reviews, and Wikipedia, a site inherent- Mellon; MIT; Princeton University;
the future of AI, we should ask wheth- On Dec. 2, 2020, Timnit Gebru, the co- its first rulings on disputed content, ly riddled with bias. The people building University of California, Berkeley; In-
er consolidation makes sense for the lead of Google’s ethical AI team, posted overturning four out of the five cases models tend to be homogeneous and ternational Computer Science Institute;
greater good and whether competition a tweet saying she’d been fired. Known it saw. But there are billions of posts aren’t often aware of their own bias- among others—are studying the side
will eventually be hindered (along with for groundbreaking research in bias on Facebook every day and an untold es. As computer systems get better at effects of automatic decision-making.
access), as we’ve seen in other fields and facial recognition, she is widely number of content complaints—which making decisions, algorithms may sort Consumers could wind up on the wrong
such as telecom and cable. respected within the broader AI com- means the oversight board operates at each of us into groups that don’t make side of the algorithm—you could discov-
munity. While the incident concerned the speed of traditional government. A any obvious sense to us—but could have er you’re ineligible for a loan, or a par-
AI Alignment a paper she co-authored and a decision year later, the board advised Meta Plat- massive repercussions. Every single day, ticular medication, or the ability to rent
by Google that it didn’t meet “our bar forms to crack down address doxxing consumers are creating unimaginable an apartment, for reasons that aren’t
As AI systems improve, some research-
for publication,” she and others argued at Facebook, urging it to prohibit users amounts of data, both actively (such transparent or easy to understand.
ers are insisting on guardrails to ensure
that Google’s ethics team was motivated from sharing individuals’ home address- as when uploading and tagging photos
that AI is deployed in ways that do not
harm humanity. One area of concern is by PR rather than progress. It set off a es on platforms even if they are publicly on Facebook) and passively (driving to Deepfaking Trust
firestorm of criticism, and by February available. The board also advocated work, for example). That data is mined
known as AI alignment, which explores Humans can easily be tricked into
2021 Google said it would change its that Facebook create a communications and used, often without direct knowl-
different scenarios in which AI systems believing machine-generated faces, es-
diversity and research policies, “stream- channel to transparently explain viola- edge or understanding, by algorithms. It
are built with goals that align with so- pecially when they’ve been engineered
line its process for publishing research,” tions and enforcement. We anticipate is used to create advertising, to help po-
ciety’s values. OpenAI, DeepMind, and to elicit trust. A study published in the
and change how sensitive employee many more ethics clashes in 2022. tential employers predict our behaviors,
Anthropic (which defines itself as an Proceedings of the National Academy of
exits are managed. To deal with its own to determine our mortgage rates, and
“AI safety and research company”) each Sciences shows that synthetic faces are
have AI alignment teams with dedicated ethical failures, Facebook launched an AI Still Has a Bias Problem even to help law enforcement predict
often “deemed more trustworthy than
independent oversight board with the whether we’re likely to commit a crime.
staff researching guardrails. While the It’s no secret AI has a serious and multi- real faces.” This suggests that synthetic
total number of researchers working on power to overrule content moderation faceted bias problem. Just one example: Researchers at a number of universi- faces could be designed as societal mal-
AI alignment is small compared to the guidelines—and even to overrule Mark The datasets used for training often ties—including the University of Mary- ware. If a bad actor was attempting to
rest of the AI community, such dedicat- Zuckerberg himself. (But just on con- come from places like Reddit, Amazon land; Columbia University; Carnegie undermine institutions, it could deploy
ed teams did not exist until recently.

Society Trends
a synth on social media to sow distrust. images of people walking around Duke language generation were released—but Algorithms Targeting from infant through adolescent” and
There are not yet effective countermea- University’s campus, also didn’t gain the conversations from which they Vulnerable Populations the majority of which “feature coercion,
sures for synthetic humans, or effective consent. Both datasets were eventually learned were scraped from Reddit and There is no question that machine abuse, and sexual activity,” according to
markers to help consumers distinguish removed. In 2018, researchers at MIT Amazon reviews, both of whose au- learning systems, trained correctly can the program’s own developer documen-
between fake and real. developed an AI called Norman that was thor populations skew white and male, help find missing children and detect tation. These images are considered
trained to perform image captioning which means that their use of language abuse. The problem is that the systems particularly challenging for the software
Problematic Training Data based only on content from a subreddit isn’t representative of everyone. This use data from vulnerable populations to because of the greater variability of
that’s known for graphic violence. When illustrates an ongoing challenge within do their training. The Multiple Encoun- position, context, and more.
For the past decade, research teams
Norman was ready, they unleashed the developer community. It is already ter Dataset contains two large datasets
have come under fire for using prob-
lematic datasets to train their models. In
him against a similar neural network difficult to get authentic data from real of photos: people who have not yet com- AI Intentionally Hiding Data
that had been trained using standard people to train systems, and with new mitted a crime and an FBI dataset of de-
2016, Microsoft’s MS-Celeb-1M database Computers do exactly what they are told
data. Researchers fed both systems privacy restrictions, developers are ceased people. The dataset over indexes
was said to contain 10 million images to do. Command a machine to win at
Rorschach inkblots and asked them to choosing to rely more on public—and on people of color, which means that if
of 100,000 celebrity faces. Researchers a game, and it will do everything in its
caption what they saw, and the results problematic—datasets. The BigScience law enforcement uses the data to train
later discovered, however, that the power to achieve that goal. Apparently
were striking: Where the standard Research Workshop, held May 2021– algorithms, it’s going to lead to bias. Im-
dataset contained images pulled from that now includes cheating. Research-
system saw “a black and white photo of May 2022, chipped away at this prob- age recognition is a particularly vexing
the web—showing journalists, activists, ers at Stanford University and Google
a baseball glove,” Norman saw “a man lem. In a collaboration between 600 AI challenge, because researchers need
artists, and everyday people who had discovered that an AI system designed
murdered by machine gun in broad day- researchers from 50 countries, partici- large datasets to perform their work.
not given their consent to be includ- to turn satellite images into usable maps
light.” The point of the experiment was pants investigated a dataset and model Often, images are used without consent.
ed. Yet the dataset was being used by was withholding certain data. Research-
to prove that AI isn’t inherently biased, from all angles: bias, social impact, ca- The Child Exploitation Image Analytics
numerous companies including Face- ers were using a neural network called
but that data input methods—and the pabilities, limitations, ethics, potential program—a dataset used for testing by
book and SenseTime, which is China’s CycleGAN, which learns how to map
people inputting that data—can signifi- improvements, specific domain perfor- facial recognition technology develop-
facial recognition leader with ties to image transformations. For example,
cantly alter an AI’s behavior. In 2019, mances, carbon impact, and general AI/ ers—has been running since 2016 with
the Chinese police. In a separate inci- it took an old aerial photograph of a
new pretrained systems built for natural cognitive research landscape with a goal images of “children who range in age
dent, DukeMTMC, a dataset containing neighborhood, distinguished between
to apply a rigorous standard of ethics.

Society Trends
streets, alleys, driveways, buildings, and ment, researchers are not obligated to rithms used. Building trust and account-
lampposts, and then generated a map report accidents or incidents involving ability requires transparency. This is a
that could be used by GPS. Initially, they our data, or AI processes, unless a law complicated process, and corporations,
used an aerial photograph that hadn’t is broken. While big companies must government offices, law enforcement
been seen by the network. The resulting inform consumers if their personal agencies, and other organizations un-
image looked very close to the origi- data—credit card numbers, home ad- derstandably want to keep data private.
nal—suspiciously close. But on deeper dresses, passwords—have been stolen, The ethics of how data is collected in
inspection, the researchers found that they are not required to publicly doc- the first place may also influence the
many details in both the original image ument instances in which algorithms trustworthiness and validity of scientific
and the generated image weren’t visible have learned to discriminate against research, particularly in areas such as
in the map made by the AI. It turns out someone on the basis of race or gender, organ donations and medical research.
that the system learned to hide informa- for example. In addition, employing ethicists to work
tion about the original image inside of directly with managers and developers
the image it generated. Prioritizing Trust and ensuring developers themselves are
diverse—representing different races,
We will soon reach a point when we will
Undocumented AI Accidents no longer be able to tell if a dataset has
ethnicities, and genders—will reduce
inherent bias in AI systems.
Microsoft’s MS-Celeb-1M was purported to contain 10 million images of 100,000 celebrities’ faces. Only a few of the numerous AI-related been tampered with, either intentional-
accidents in 2018 and 2019 made head- ly or accidentally. AI systems rely on our
lines. Most people know about the Uber trust. If we no longer trust their out-
self-driving car that hit and killed a comes, decades of research and techno-
pedestrian in Tempe, Arizona. But there logical advancement will be for naught.
were countless more incidents that Leaders in every sector—government,
didn’t result in death, and as a result, business, nonprofits, and so on—must
aren’t known to the public. At the mo- have confidence in the data and algo-

[As Neo decides to take the door towards his
Left that will take him to the Matrix]
Hope. It is the quintessential human delusion,

simultaneously the source of your greatest
strength and your greatest weakness.
— The Architect, in The Matrix Reloaded

00 01 02 03 04 05 06 07 08 09 10 11 12 13 Artificial Intelligence APPLICATION
How these trends impact your company
STRATEGY INNOVATION R&D GROWTH
Artificial intelligence should be part of As no-code and low-code applications The field of AI is growing faster than In most industries, AI will serve as a force
every strategic plan, as it crosses multiple become more widely available, inno- universities can produce trained people. multiplier for growth, bringing efficien-
dimensions, from workforce automation, to vation teams will be in position to build Talent sourcing and retention will contin- cies, better tracking, business intelligence,
digital transformation, to everyday busi- powerful systems for decision manage- ue to pose challenges for AI companies— and assistance with decision-making. As
ness processes and business intelligence. ment, business intelligence, and product and for organizations in other industries training costs decline, more applications
It is imperative that executives and senior ideation. Generative AIs will improve an that need a trained workforce but may will be built. Spending on AI systems and
managers understand what AI is, what it organization’s efficiency and enhance not be able to provide the same perks as hardware is likely to explode this decade,
is not, and what strategic value it adds to creativity, leading to hybrid human-ma- the biggest tech companies. Meanwhile, creating significant enterprise value
the business. Chief strategy officers should chine creative teams. AI-assisted design China has emerged as a global leader overall.
build a robust understanding of AI in order will dramatically increase the number in R&D and is actively recruiting gradu-
to engage more closely with others in of prototypes that can be automatically ates to repatriate home. For teams with
the C-suite, especially chief technology generated with prompts. enough experience and staff resources,
officers, chief information security officers, 2022 should be a banner year for applied
chief financial officers, and others in the AI research for health, medicine, smart
organization to develop longer-term plans. cities, entertainment, and sports.
Keep abreast of emerging regulations that
could restrict the use of consumer data.
Risk models should be developed to deter-
mine plausible near-futures, so that lead-
ers can adjust their strategies accordingly.

00 01 02 03 04 05 06 07 08 09 10 11 12 13 Artificial Intelligence KEY QUESTIONS
Key Questions
We recommend using this report

1 2 3
to support your strategic foresight
activity in the coming year. Every  Are we adequately invested in AI?  How well does our organization  If the state of play changes, is our
executive team should begin by understand AI? organization agile enough to adapt
asking these questions: quickly?
 How could AI make us more
competitive in the years to come?  Are we leveraging enterprise
solutions effectively?  Are we adequately planning for the
longer-term future, as AI evolves?
 How might developments in AI
leave us vulnerable to disruption?  Are we preparing our workforce to
meet the evolving demands of the
knowledge economy?

00 01 02 03 04 05 06 07 08 09 10 11 12 13 Artificial Intelligence TRENDS IN ACTION
Trends in Action
Benefits of Strategic Foresight The Future Today Institute Supports Executive Leaders and Their Teams
33% Higher Profitability The Future Today Institute works closely with executive leadership and management
Companies using a dedicated strategic foresight process outperformed the average teams to transform their strategic thinking on the future.
by a 33% higher profitability.
We leverage these and other trends and use applied foresight to develop deep (20+
200% Growth years), long-range (10+ years) and near-term (2+ years) scenarios and strategic plans.
Companies using a dedicated strategic foresight process outgrew their competitors by
200% in desired areas.
FTI’s advisory services include signal mapping, trend identification, scenario
development, risk modeling, visioning, and strategic planning.
25% Improvement
Companies say that strategic foresight improves business objectives and planning,
helps define new markets, and builds flexible mindsets among executives, even in
times of deep uncertainty.

00 01 02 03 04 05 06 07 08 09 10 11 12 13 Artificial Intelligence SOURCES
Selected Sources
https://www.af.mil/News/Arti- https://www.nextgov.com/emerg- https://blog.google/technology/ai/ https://doi.org/10.1073/pnas.2120481119 https://www.nextgov.com/cio-brief-

cle-Display/Article/2448376/ai-co- ing-tech/2021/12/dod-hire-first-ev- lamda ing/2021/12/senate-confirms-
pilot-air-force-achieves-first-mili- er-chief-digital-and-artificial-intelli- https://obvious-art.com john-sherman-pentagons-tech-
tary-flight-with-artificial-intelligence gence-officer-form-new-office/187399 https://www.iarpa.gov/research-pro- chief/359940
grams/hiatus https://www.gao.gov/products/gao-
https://www.defenseone.com/technol- https://cset.georgetown.edu/publica- 21-518 https://doi.org/10.1016/j.art-
ogy/2022/02/ai-could-match-finger- tion/ethical-norms-for-new-genera- https://deepmind.com/research/ int.2021.103535
prints-texts-their-authors-under-new- tion-artificial-intelligence-released publications/2019/Mastering-Atari-Go- http://arxiv.org/abs/2108.02922
intelligence-program/361893/. Chess-and-Shogi-by-Planning-with-a- https://www.defenseone.com/technol-
http://arxiv.org/abs/2101.03961 Learned-Model ogy/2022/02/near-future-military-au-
https://deepmind.com/blog/article/
https://www.ai-darobot.com tonomy-isnt-robotanks-microser-
putting-the-power-of-alphafold-into-
https://www.gao.gov/assets/gao-21- https://media.defense.gov/2022/ vices/361873
the-worlds-hands
https://www.darpa.mil/program/air- 526.pdf Feb/02/2002931807/-1/-1/1/
crew-labor-in-cockpit-automation-sys- MEMORANDUM-ON-THE-INI- https://cset.georgetown.edu/publica-
https://www.pewresearch.org/inter-
tem https://github.com/google-research/ TIAL-OPERATING-CAPABILI- tion/u-s-ai-workforce
net/2021/06/16/experts-doubt-ethical-
dreamer TY-OF-THE-CHIEF-DIGITAL-AND-ARTI-
ai-design-will-be-broadly-adopted-as-
https://www.pnas.org/content/119/8/ FICIAL-INTELLIGENCE-OFFICER.PDF https://cset.georgetown.edu/wp-con-
the-norm-within-the-next-decade
e2120481119 https://doi.org/10.1038/s41467-021- tent/uploads/US-AI-Workforce_Brief.
25874-z http://arxiv.org/abs/2003.08934 pdf
https://www.diu.mil/responsi-
https://doi.org/10.1126/science.abj8754 ble-ai-guidelines
https://github.com/danijar/dreamer https://blog.google/products/search/ https://www.scientificamerican.com/
https://openai.com/blog/dall-e introducing-mum article/humans-find-ai-generated-
https://www.popularmechanics.
faces-more-trustworthy-than-the-real-
https://doi.org/10.1016/j.con- com/military/aviation/a34978872/
https://deepmind.com/blog/article/mu- thing
https://doi.org/10.1007/s40821-020- cog.2020.102938 artificial-intelligence-con-
00172-8 zero-mastering-go-chess-shogi-and- trols-u2-spy-plane-air-force-exclusive
atari-without-rules
https://doi.org/10.1038/s41586-021-
https://deepdreamgenerator.com/. 03819-2 https://doi.org/10.1038/s41586-020-
https://nerfies.github.io 03051-4

00 01 02 03 04 05 06 07 08 09 10 11 12 13 Artificial Intelligence AUTHORS
Authors
AMY WEBB EMILY CAUFIELD ERICA GRAU

CEO Creative Director Designer
awebb@futuretodayinstitute.com ecaufield@futuretodayinstitute.com egrau@futuretodayinstitute.com
Founder of the Future Today Institute, Amy Webb pioneered a data-driven, technology-led fore- Emily Caufield has served as Creative Di- Erica Grau is a freelance graphic design-
sight methodology practiced by FTI and used within hundreds of organizations worldwide. She rector for the Future Today Institute for the er with over a decade of experience. She
leads FTI’s technology research initiatives on artificial intelligence, synthetic biology and genet- past decade. She uses her graphic design grew up in Atlanta, Georgia where she
ic engineering, next-generation networks, and quantum technologies. Webb is a professor of skills to communicate complex ideas, visu- graduated with a degree in advertising
strategic foresight at the New York University Stern School of Business, where she developed and alize data, and develop frameworks for use from The University of Georgia and then
teaches the MBA course on strategic foresight. She is a Visiting Fellow at Oxford University’s Säid in strategic foresight and futures activities. furthered her education at The Creative
School of Business, a Nonresident Senior Fellow in the Atlantic Council’s GeoTech Center and a Circus in graphic design. She then moved
Fellow in the United States-Japan Leadership Program. She was elected a life member to the to New York City where she has worked
Council on Foreign Relations. Webb was named by Forbes as one of the five women changing with many clients including The United
the world, listed as the BBC’s 100 Women of 2020, ranked on the Thinkers50 list of the 50 most Nations, The Bronx Zoo, Weight Watch-
influential management thinkers globally. A lifelong science fiction fan, she collaborates closely ers and many more. From print to digital
with writers and producers on films, TV shows and commercials about science, technology and design, she enjoys working on a variety
the future. Her newest book, The Genesis Machine, explores the futures of synthetic biology and of different types of projects to continue
the bioeconomy. growing her skillset.

00 01 02 03 04 05 06 07 08 09 10 11 12 13 Artificial Intelligence AUTHORS
Authors
JENNIFER ALSEVER CHERYL COONEY MEL BLAKE TOM FOSTER

Editorial Director Director of Operations Business Development Editor
jalsever@futuretodayinstitute.com ccooney@futuretodayinstitute.com mblake@futuretodayinstitute.com
SARAH JOHNSON
Editor
Jennifer Alsever led editorial production For over a decade, Cheryl Cooney has Mel Blake handles commercial develop-
for the 2022 FTI Tech Trends Report. She served as the Future Today Institute’s ment and client relations for FTI. He was
has been a journalist for more than two director of operations, where she manages formerly founder and managing director
decades covering tech, biotech, startups workflows, planning and logistics. Cooney of Monitor Talent, a speaker agency he
and business for such publications as For- is a published poet, with works appearing founded at the Monitor Group, a global
tune Magazine, the Wall Street Journal, in American and Australian anthologies. consulting firm. He is a board member
the New York Times, Wired and Fast Com- of The Yale Center for Customer Insights.
pany. A popular young adult fiction writer, He holds an MBA from the Yale School of
Alsever has won multiple YA awards for Management and a BA in Economics from
her Trinity Forest Series. Wesleyan University.

00 01 02 03 04 05 06 07 08 09 10 11 12 13 Artificial Intelligence ABOUT FTI
About the Future Today Institute
Founded in 2006, the Future Today

Contact Us
Institute researches, models, and
prototypes future risk and opportuni-
ty. As the leading strategic foresight hello@futuretodayinstitute.com
and futures management consul- 267-342-4300
tants to executive leadership teams www.futuretodayinstitute.com
worldwide, FTI’s data-driven applied
research reveals trends and calcu-
lates how they will disrupt business,
government, and society.
Social
Twitter
Together with our clients and part-
@FTI
ners, FTI is helping leaders achieve
their preferred futures. Our pioneer-
LinkedIn
ing, data-driven forecasting meth-
@Future-Today-Institute
odology and tools empower leaders
to make better decisions about the
Facebook
future, today.
@FutureTodayInstitute
Instagram
@futuretodayinstitute

00 01 02 03 04 05 06 07 08 09 10 11 12 13 Artificial Intelligence DISCLAIMER
Disclaimer
The views expressed herein are the FTI’s reports are occasionally updat-
authors’ own and are not represen- ed on the FTI website.
tative of the greater organizations in
which they have been employed. The FTI advises hundreds of companies
names of companies, services, and and organizations, some of which are
products mentioned in this report are referenced in this report. FTI does not
not necessarily intended as endorse- own any equity position in any of the
ments by the Future Today Institute or entities listed in this presentation.
this report’s authors.
Any trademarks or service marks used
The Future Today Institute’s 2022 Tech in this report are the marks of their re-
Trends Report relies on data, anal- spective owners, who do not endorse
ysis, and modeling from a number the statements in this report. All rights
of sources, which includes sources in marks are reserved by their respec-
within public and private companies, tive owners. We disclaim any and all
securities filings, patents, academ- warranties, expressed or implied, with
ic research, government agencies, respect to this report.
market research firms, conference
presentations and papers, and news
media stories. Additionally, this report
draws from the Future Today Insti-
tute’s previous EMT Trends Reports,
FTI Trend Reports, and newsletters.

00 01 02 03 04 05 06 07 08 09 10 11 12 13 Artificial Intelligence USING THE REPORT
Using and Sharing the Material in this Report
You are free to: Under the following terms: You are prohibited from:
Share Attribution Commercial Sharing

Copy and redistribute the material in You must give appropriate credit to Don’t copy and redistribute this ma-
your organizations and classrooms. the Future Today Institute, provide a terial in any medium or format for
link to this Creative Commons license, commercial purposes, including any
and indicate if any changes were personal/corporate marketing and
We invite you to use, share, and build
made. You may do so in any reason- client services.
upon the material in our 15th annual
able manner, but not in any way that
Future Today Institute Tech Trends Representing This Work As Your Own
suggests that the Future Today Insti-
Report. We are making it freely avail- Do not represent any part of this ma-
tute endorses you or your use.
able to the public. This work is licensed terial as your own without giving credit
under a Creative Commons CC BY- NonCommercial to the Future Today Institute.
NC-ND 4.0 International License. You may not, under any circumstance,
use the material for commercial pur- Changing This Material
poses. If you remix, transform, or build upon
the material, you may not distribute
No Derivatives the modified material.
If you remix, transform, or build upon
the material, you may not distribute Additional Restrictions
the modified material. You may not apply legal terms or tech-
nological measures that legally restrict
others from doing anything this license
permits.

FTI Tech Trends 2022 AI

Uploaded by

Copyright:

Available Formats

You might also like

FTI Tech Trends 2022 AI

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

FTI Tech Trends 2022 AI

Uploaded by

Copyright:

Available Formats

15th Edition • 2022 Tech Trends Report Vol.

2 © 2022 Future Today Institute

Macro Forces and Emerging Trends 4TH YEAR ON THE LIST 1

calculated, and how to correct inaccu-

3 © 2022 Future Today Institute

Scenarios Describe 00 01 02 03 04 05 06 07 08 09 10 11 12 13 Artificial Intelligence SCENARIO

developments leading us to this point in our future history.

4 © 2022 Future Today Institute

5 © 2022 Future Today Institute

This is volume 01 in the Future Today

To find additional volumes, visit

7 © 2022 Future Today Institute

8 © 2022 Future Today Institute

9 © 2022 Future Today Institute

10 © 2022 Future Today Institute

11 © 2022 Future Today Institute

AI: Techniques &

12 © 2022 Future Today Institute

13 © 2022 Future Today Institute

AI: Techniques and Influential Models

14 © 2022 Future Today Institute

AI: Techniques and Influential Models

15 © 2022 Future Today Institute

AI: Techniques and Influential Models

16 © 2022 Future Today Institute

AI: Techniques and Influential Models

reducing training time. Many language

Source: IFI Claims Patent Services, January 2022 study.

18 © 2022 Future Today Institute

During the next decade, AI will re-

19 © 2022 Future Today Institute

Low-Code or No-Code matically changing the colors of objects Serverless Computing

20 © 2022 Future Today Institute

21 © 2022 Future Today Institute

22 © 2022 Future Today Institute

23 © 2022 Future Today Institute

24 © 2022 Future Today Institute

25 © 2022 Future Today Institute

Spotting Fakes Consumer-grade AI

26 © 2022 Future Today Institute

27 © 2022 Future Today Institute

28 © 2022 Future Today Institute

29 © 2022 Future Today Institute

Creative Field Trends

AI-Assisted Invention 2019, Nvidia launched GauGAN (named

30 © 2022 Future Today Institute

Creative Field Trends

31 © 2022 Future Today Institute

Creative Field Trends

32 © 2022 Future Today Institute

33 © 2022 Future Today Institute

Health, Medicine and Science Trends

34 © 2022 Future Today Institute

Health, Medicine and Science Trends

35 © 2022 Future Today Institute

Health, Medicine and Science Trends

36 © 2022 Future Today Institute