TECH Artificial Intelligence

Detailed analysis of AI, including key

techniques and influential models and
more than 100 AI trends related to the
enterprise, consumer ecosystem and
government.

2 © 2023 Future Today Institute

 TECH Artificial Intelligence

TABLE OF
CONTENTS 24 Adopting Unified Learning
Processes
24 Transitioning to Textless NLP
31
31
31
Adopting Serverless Computing
The Rise of MLOps
Leveraging AI in the Cloud
36
37

Self-Driving Microscopes
NLP Algorithms Detect
Virus Mutations
24 Amping Up Diffusion Models 37 Using AI to Improve Talk Therapy
32 Implementing AI at the Edge
24 Managing the Community- 38 Spotlight: Finance and Insurance
32 Low-Code or No-Code
05 Letter from FTI’s CEO 20 Graph Neural Networks Driven Open Sourcing of Code
Machine Learning 38 Mitigating Fraud
06 State of AI in 2023 21 Activity Vision 24 AI for Code Research
32 Consolidation in AI’s Ecosystem 38 Predicting Financial Risk
10 Key Insights 21 Visual Commonsense 24 Framework Consolidation
33 Hardware 38 Predicting Workplace Injuries
Reasoning (VCR)
11 Important Terms 24 Lowering the Cost of
33 Advanced AI Chipsets 38 Improving Damage Assessment
21 Machine Image Completion Training Models
13 Ones to Watch
33 Processing-in-Memory 38 Consumer-Facing Robo-Advisers
21 Predictive Models Using 25 Surpassing Benchmarks
15 Trends Technology
Incomplete Data 38 AI Claims Processing
25 Explainable AI (XAI)
15 Models, Techniques, & Research 34 Applied AI
22 Neuro-Symbolic AI 38 Liability Insurance for AI
26 Talent
16 AI Models 34 Applied AI for HR
22 Real-time Machine 40 Creativity & Design
27 Demand for AI Talent
20 Techniques Learning (RTML) 35 Spotlight: Health Care
Growing Fast 41 AI-Assisted Invention
20 Prompt Learning/Engineering 22 Vokenization 35 Protein Folding
27 Upskilling for AI 41 Assisted Creativity
20 Reinforcement Learning 22 Model-free Approaches to RL 35 AI to Speed Up
28 AI Brain Drain 41 Neural Rendering
With Human Feedback (RLHF) Scientific Discovery
23 Research
28 Corporate AI Labs 42 Generating Virtual
20 Imitation Learning 35 AI-First Drug Development
23 Building Supersized AI Models Environments from
28 Building a Silicon Valley
20 Automated Machine Learning 36 AI to Improve Patient Outcomes Short Videos
23 Creating Specialized of the Middle East
(AutoML)
Language Models 36 Anomaly Detection in 42 AI Voice Generators
30 Enterprise
20 Continuous Learning Medical Imaging
23 Developing Practical Tools 42 Automatic Ambient
31 Corporate Partnerships
20 Federated Learning for Language Models 36 Medical Deepfakes Noise Dubbing
31 Building In-House Solutions
20 General Reinforcement 23 Training Language Models 36 Generative Antibody Design 43 Society & Ethics
Learning Algorithms to Achieve Scale

3 © 2023 Future Today Institute

 TECH Artificial Intelligence

TABLE OF
CONTENTS 54 New Strategic Technical
Alliances
55 Defense
65


What if AI could help you
practice marriage before going
through with the ceremony?
66 What if AI could help you eat
55 Autonomous Weapons Policies
better and be healthier?
55 Simulating Warfare
68 How to Prepare
44 Detecting Emotion 48 Increased Used of
55 Unpiloted Military Vehicles
Ambient Surveillance 69 Key Questions
44 Simulating Empathy
55 AI Used to Guide Military Strikes
and Emotion 49 Trust 71 Selected Sources
55 Automated Target Recognition
44 Theory of Mind Models 49 AI Alignment 75 Author & Contributors
56 Automating Offensive Attacks
45 Consumer Applications 49 Addressing Bias 76 2023 Tech Trend Reports
Using AI
45 Generative AI for 49 Synthesizing and 78 About FTI
56 Algorithmic Warfighting
Personal Expression Generating Trust
79 Methodology
56 Mandating Ethics Guidelines
45 Turnkey Consumer-Grade 49 Worker Surveillance
for Tech Contractors 80 Disclaimer
Application Development
50 School Surveillance
57 Spotlight: China’s 81 Using and Sharing the Report
45 Responsive Recognition
50 Prioritizing Trust Long-Term AI Ambitions
Technology
52 Government & Defense 58 Expanding Market
45 Efforts to Thwart
Recognition Systems 53 Government 58 Breaking Up Big Tech
46 Biometric Scoring 53 Rise of AI Nationalism 58 Building a Strategic Panopticon
48 Safety 53 Countries Develop National 59 Deepening International Ties
AI Strategies
48 AI Generators and 62 Scenarios
Copyright Infringement 53 Europe Regulates AI
63 What if deepfakes attack our
48 Algorithms Targeting 53 Nation-based Guardrails public memory?
Vulnerable Populations and Regulations
64 What if AI can’t trust its users?
48 AI Intentionally Hiding Data 54 Regulating Deepfakes

4 © 2023 Future Today Institute

 TECH Artificial Intelligence

THIS YEAR REQUIRES FOCUS

In August 2017, a rare explosive event known Now more than ever, it’s important to carefully Trends on their own cannot predict the future.
as GW17817 took place in space. Two stars track new trends as they emerge. But that isn’t Rather, future-focused organizations use them to
collided, unleashing a blast energetic enough easy, given the rapid pace of change. For that deeply reflect on the tension between long-term
to form an incalculable number of new stellar reason, the theme of our 2023 Tech Trends report and short-term goals and to reduce uncertainty.
bits that continue to travel through interstellar is Focus. It is crucial to focus when new signals By understanding the trends and changes shap-
space. Over time, this stardust will combine into are forming because some may be lasting and ing the landscape, executives can make informed
small objects, evolve into large rocks, fuse with develop into impactful trends, while others might decisions and capitalize on new opportunities in
even more material, and form into planets. One burn out and fade away. In an increasingly com- the year ahead.
incredibly violent disruption will someday lead plex and fast-paced world, leaders who focus
We invite you to join us in observing how the star-
to the formation of a new corner of the universe. on the trends that matter and adapt to changing
dust settles into new signals and trends.
This is how our own sun and Earth, and all of circumstances make better decisions and see
Share your feedback with us at
human existence, came into being. improved outcomes. Trends enable them to antic-
2023trends@futuretodayinstitute.com.
ipate near-term change, understand the factors
Lately it’s as if we’ve been living through the
influencing their industries, and develop a point of
aftermath of cataclysmic explosions: the
view on the future.
release of generative artificial intelligence
systems like ChatGPT and Midjourney, a fusion Our research is presented in 14 in-depth reports
breakthrough that could someday generate that reveal the current state of play, a list of influ- Amy Webb
zero-carbon energy, Russia’s ongoing invasion encers to watch, key trends, detailed examples, Chief Executive Officer
into Ukraine, deep uncertainty about a global expert perspectives and recommendations de- Future Today Institute
recession, and AlphaFold’s protein-folding al- signed to help executives and their teams devel-
gorithms that predicted structures for nearly all op their strategic positioning. Some of the trends
cataloged proteins known to science, to name are new advancements on mature technologies,
a few. These and other forces of change are while others represent frontier technologies and
colliding, going supernovae, and resulting in an areas of science. When we look at them collec-
unfathomable amount of new signals—bits of tively, new centers of gravity come into focus, and
change that, over time, result in the trends that we can glimpse the impacts they will have on
shape society. every sector.

5 © 2023 Future Today Institute

 TECH Artificial Intelligence

ChatGPT, the viral chatbot from OpenAI that uses artificial intelligence to produce content for every-
day people, sparked imagination and alarm at the end of 2022. From our point of view, it is the most
powerful AI system ever released to the general public. Its ease of use, combined with increasingly
creative prompts, made it a global sensation. Created by the same company that developed GPT-3
and image generator DALL-E 2, the bot prompted school boards to hold emergency meetings, since
they worried the new tool would supercharge cheating on homework. It also spooked investors, who
wondered if this marked the end of traditional search, and therefore posed a credible threat to Google.
AI is a force multiplier on
technological progress because it To be sure, generative AI and other assistive technologies are bringing powerful capabilities to non-
is an enabler of other technologies technical users. But there is a lot more happening in the field of artificial intelligence.

and powers the evolution of business, In this section of our emerging tech trends report, we use artificial intelligence as an umbrella term to
government, and society. encompass the techniques, models, and frameworks that make up the field. Its aim is to create intel-
ligent machines that can sense, reason, act, and adapt like humans do, or in ways that go beyond our
capabilities. Today, cars can park themselves, while emerging platforms are capable of having seem-
ingly natural conversations. As the technology evolves to beyond-human capabilities, it might invent
new drugs, predict the real-time movement of wildfires, and autonomously design machine parts—or
even entire factories.

AI is transforming business, but it’s also changing programming. Wired magazine once explained
that “in traditional programming, an engineer writes explicit, step-by-step instructions for the com-
puter to follow. With machine learning, programmers don’t encode computers with instructions. They
train them.”

6 © 2023 Future Today Institute

 TECH Artificial Intelligence

STATE OF ARTIFICIAL INTELLIGENCE 2023

Likely Near-Term Developments Enterprise Automation Regulation and Geopolitics

• Top talent will start breaking away from the • AI coding assistants will grow in popularity. • Will this be the year the government breaks
General
largest players—Google, OpenAI, Meta—to OpenAI’s Codex, introduced in 2021, evolved up Big Tech? US lawmakers have begun
• General-purpose models will be commod- form their own startups, which will include from research to open commercialization by work to decouple Google’s ad business from
itized in the near-future. This will lead to large conversational agents, artificial general intel- the middle of last year. GitHub CoPilot is now the rest of the organization. In the European
language model (LLM) features being inte- ligence programs, AI-first biotech companies, available as a subscription ($10 per month). As Union, powerful new regulations of tech plat-
grated into every app. The first players to suc- and the like. of January, Amazon’s CodeWhisperer is avail- forms could go into effect. In China, regulators
ceed in the space will likely be creative apps, able in preview. Internally, Google is using a have eased up slightly on Big Tech crack-
and all of them will have LLMs as a backbone • Consolidation will continue to be a driving machine language–powered code completion downs—but the Chinese Communist Party is
to inspire and augment human creativity. theme in 2023. Here’s just one example: Mic- tool—it may at some point be made available intent on focusing new R&D toward servicing
rosoft plans to increase its investment in and to everyday users, possibly this year. the country’s long-term growth ambitions.
• Vertically integrated solutions will garner a leverage OpenAI for Bing to pull market share
higher transactional value. Some companies away from Google search. • We expect to see AI more deeply integrated
will win by providing “a refined/value-add- into health care and life sciences later this
ed LLM product” to the end consumer and • Google’s powerful ChatGPT rival, Bard, uses year. Generative AI will yield new proteins,
meeting the customer in desired distribution an AI model called LaMDA. Launched Feb- antibodies, and drugs, while biology and
channels, such as LLMs for health care, legal, ruary 2023, it could catalyze a new race for chemistry-specific models will result in faster
finance, and architecture. conversational search. discovery, further providing practical use cas-
es and leading to increased investment.
• In the next 12-18 months, watch for the pro- • We anticipate increased enterprise adoption
liferation of large-scale open-source models of AI. Leaders see AI as necessary for growth • Could 2023 mark the beginning of the end
and tools including Stable Diffusion and in the current macroeconomic environment, of human radiologists? Lithuania-based
HuggingFace. even as new developments make some job Oxipit, an AI-first medical imaging startup,
categories obsolete. built an autonomous system that reports on
chest X-rays that show no abnormalities. The
technology is good enough that it received
state-level certification to operate inde-
pendently, without a radiologist in the loop.

7 © 2023 Future Today Institute

 TECH Artificial Intelligence

STATE OF ARTIFICIAL INTELLIGENCE 2023

When will AI disrupt your in other trends throughout all volumes of Future Constraints on adoption: Even if a technolo- Public perception: How the public under-
Today Institute’s 2023 report. gy is maturing, constraints on its adoption can stands and responds to AI advancements will
organization? hinder its impact on an industry. For example, create or quell demand. This is especially true
Several factors are driving the momentum of AI a business may refuse to adopt an automated of generative AI and education/creativity/intel-
Advancements in AI will disrupt every industry
trends and the probable timing of an industry’s system because it challenges existing ortho- lectual property/misinformation, and the role
within the next decade, and Future Today In-
disruption: doxy or an existing successful strategy. This is assistive technologies will play in shaping the
stitute classifies it as a general-purpose tech-
nology. Like the steam engine and the internet Scaling: Enormous amounts of training data especially true in health care, insurance, and future workforce.
before it, AI has the potential to influence entire are still required for most AI models to learn. For financial services.
R&D developments: The pace of new research
economies and to alter society through political, example, recommender systems coupled with Regulations: Advances in technology typically breakthroughs can’t be scheduled to coincide
economic, and social structures. generative AI could lead to deep personalization outpace regulatory changes. This has benefitted with a board meeting or earnings report. Factors
for the hospitality and health care sectors—as AI, which until very recently was not targeted for like funding, quality and size of staff, and access
AI is now used across most industries, solving
long as data is made available. Historically, regulation. Additionally, whether local regula- to resources can improve the likelihood and
business problems, detecting fraud, improving
data is locked inside proprietary systems built tions are conflicting or complementary influenc- speed of new discoveries. We closely monitor
crop yields, managing supply chains, recom-
by third parties, and regulation often hinders es adoption in the marketplace. R&D developments but treat them as wild cards.
mending products, and assisting designers
access to certain forms of data.
and writers in creative work. AI can predict call
Media mentions: Increased awareness and
volume in customer service centers and recom- Investment: AI has passed through cycles of enthusiasm can influence the momentum of
mend staffing levels; it also predicts the emo- enthusiasm and disillusionment, leading to ei- a technology, even when there’s been no real
tional state of the person calling to help compa- ther too much or not enough capital being made breakthrough. Until OpenAI’s ChatGPT break-
nies anticipate desirable solutions. available. Investors prioritize commercialization through in late 2022, leaders weren’t talking
over basic R&D—though the latter yields bigger about the impact generative AI might have on
AI automates the process for drug discovery,
impact and often stronger returns. Whether their business. Media bursts related to AI will
which ultimately led to faster COVID-19 vaccine
investors are patient tends to influence progress drive momentum, especially if those stories
candidates. Because AI is so broad, we have
and pace of commercialization. are favorable, and more importantly, are easily
identified different themes to track; however,
because AI is now so deeply intertwined with understood by the public.
every aspect of life, you will find AI mentioned

8 © 2023 Future Today Institute

 TECH Artificial Intelligence

STATE OF ARTIFICIAL INTELLIGENCE 2023

Why AI trends matter to your

organization
We believe AI is a force multiplier on technolog-
ical progress because it is an enabler of other
technologies and powers the evolution of busi-
ness, government, and society. But new capa-
bilities of large language models—and ChatGPT
in particular—have deeply concerned some
in the professional and creative services. For
years, experts have warned about robots taking
away blue collar jobs; now we’re entering an era
where AI will erode white collar jobs.

Since publishing our first Tech Trends report 16

years ago, we have included and expanded our
coverage on artificial intelligence. What began
as several pages of insights is now a dedicated,
stand-alone report with more than 100 trends
to monitor. AI is already transforming most
economic sectors, but we anticipate deeper
impacts this year across insurance, finance,
entertainment, health care, biotechnology, and
cloud computing.

9 © 2023 Future Today Institute

 TECH Artificial Intelligence

Publicly available LLMs are often the

We have entered the “text-to-everything” era. Soon, we will use foundation for AI startups, but some
natural language to operate and interact with computers rather researchers and technologists are
than graphic user interfaces (GUIs). It’s possible for agents to learn the right questioning their defensibility when it
skills but the wrong objectives; an AI comes to capturing value. The moat is in
system can be asked to learn something data. Techniques and models will largely
If the 2010s were known for perception that then could be used for harmful get commoditized, and served via the
AI—systems that sensed signals such as purposes. Commercial AI products could infrastructure layer, where real value will

KEY
images and text—to understand the world inadvertently incentivize bad behavior. be realized.
around us, the 2020s will be known for
generative AI. These systems not only sense
and understand the world but can also Long-term sustainability
depends on network effects to

INSIGHTS
generate new content, concepts, and ideas
while communicating with us. gather enough user data. User-
generated data can be harnessed
to differentiate systems by
AI is emerging as an assistant for all knowledge workers. offering tuned models on top of
Within the next 18-24 months, we will see assistive foundational/commoditized LLMs,
technology developed for a variety of professions. Think: creating a flywheel effect. Longer
GitHub’s Copilot for financial analysts, commercial real term, we expect to see niche LLMs
estate developers, and lawyers. owned by a select few players,
We expect to see increased activity in
while general-purpose LLMs
the specialized AI accelerators space
become commoditized.
this year. Watch the UK’s Graphcore,
In the next 18-24 months, generative AI will be incorporated into consumer
which built a new type of processor
applications. Already, Canva, the popular online graphic design tool, has
for machine intelligence to accelerate At the infrastructure layer, we
integrated Stable Diffusion in its platform. Microsoft has incorporated OpenAI’s
machine learning and AI applications, anticipate a variety of generative
DALL-E 2 image-generating system into its Microsoft Designer and Image Creator
and Cerebras, which built one of the AI services built to better serve a
applications, which is a big upgrade to clip art. Google’s Bard includes an API
fastest AI accelerators based on the variety of applications. In addition
to encourage developers to start building new products. Notion, the freemium
largest processor in the industry, as well to OpenAI, specialized players are
productivity and note-taking web application, uses LLMs for assistive writing.
as AWS Inferentia accelerators. Will entering the market, including AI21,
Nvidia’s GPU/AI accelerator moat get Cohere, Snorkel.AI, and Scale.AI.
disrupted this year?

10 © 2023 Future Today Institute

 TECH Artificial Intelligence

IMPORTANT TERMS
Machine learning (ML) Unsupervised learning Deep learning (DL) recognize particular faces, for instance—the
Machine learning uses data to make predictions Data is provided to a model without specific Deep learning is a relatively new branch of ma- CNN would run until information could be
and recommendations on how to achieve stated output parameters, and the model tries to learn chine learning. Programmers use special deep inferred. In business, CNNs are used to identify
goals. AI pioneer Arthur Samuel popularized the the data set’s structure without any designat- learning algorithms alongside an enormous anomalies in medical imaging, faulty products
idea of machine learning in 1959, explaining how ed labels. For example, if a researcher doesn’t corpus of data—typically many terabytes of on a production line, blight on crops, and other
computers could learn without being explicitly know what to do with a large data set, an unsu- text, images, videos, speech, and the like. Often, irregularities.
programmed. This would mean developing an pervised learning model could determine pat- these systems are trained to learn on their own,
algorithm that could someday extract patterns terns, classify data, and make recommendations and they can sort through a variety of unstruc- Recurrent neural networks (RNNs)
from data sets and use those patterns to predict without a human supervisor. Researchers used tured data, whether it’s making sense of typed These multilayered neural networks move and
and make real-time decisions automatically. unsupervised learning during the pandemic to text in documents or audio clips or video. store information between input, hidden, and
It took many years for reality to catch up with find patterns in how COVID-19 spread through- output layers. They are good at modeling se-
Samuel’s idea, but today machine learning is a out communities. In practical terms, deep learning’s emergence quence data for predictions. In business, they are
primary driver of AI’s growth. means that more and more human processes used anytime the sequence of data matters, such
Reinforcement learning (RL) will be automated, including the writing of soft- as speech recognition and language translation.
There are different types of machine learning, A system performs a task by repeatedly run- ware, which computers will soon start to do on RNNs are used in digital assistants, to create
including supervised, unsupervised, and rein-
forcement. ning calculations as it attempts to accomplish a their own. Once a system learns what an object captions for images and to generate narrative
stated goal. It’s a trial-and-error process, where looks like—say, an apple—it then can recognize reports (sports, financial) using structured data.
Supervised learning rewards or penalties are earned in response to that object in all other images, even if it has only
A model that attempts to transform one type of the system’s performance toward achieving the a partial view. Transformers
data into another type using labeled examples. stated goal. RL is used when there isn’t enough A transformer is a component whose purpose
There are different types of deep learning archi- is to process sequential data, such as natural
Supervised learning is used when teams know training data, when the researcher is trying to tectures. These are the most common:
how to classify the input data and what they are learn about an environment (such as a com- language or genome sequences. Transformers
trying to predict but can get accurate results plex financial portfolio) or when the researcher rely on “attention” (the mathematical description
Convolutional neural network (CNN)
much more quickly by relying on an algorithm needs to find greater levels of optimization. of how things relate to, complement, or modify
A CNN is multilayered, with a convolutional
rather than a human. This is the most common It has a high number of business use cases, each other) in translating sequences. A trans-
layer, a pooling layer, and a fully connected lay-
form of ML used today. Understanding what prod- ranging from real-time dynamic pricing models former neural network is the unique architecture
er. Each one performs a different task with the
uct features would most likely drive new purchas- to high-frequency trading algorithms and the that enables systems to learn from context and
data. The output is classification. If a researcher
es is a business use case for supervised learning. systems that operate self-driving cars. to generate new information. Transformers are
has 10,000 images and needs to extract data—to

11 © 2023 Future Today Institute

 TECH Artificial Intelligence

IMPORTANT TERMS
complementary to CNNs and RNNs, the two Additional Terms Model
most common neural network architectures A program that has been trained on a data set.
used in deep learning. AI safety Models are generally used for analytical and
A field that studies and attempts to mitigate the decision-making tasks, such as making predic-
Generative adversarial networks (GANs) catastrophic risks that future AI could pose to tions.
As unsupervised deep learning systems, GANs humanity.
are composed of two competing neural net- Natural language processing
works—a generator and a discriminator—that Algorithm Processes that give computers the ability to
are trained on the same data, such as images A process describing how to solve a specific understand, mimic, and manipulate human
of people. The networks compete against each problem or how to complete a particular task. language.
other to perform a task, such as identifying the
Artificial general intelligence (AGI) Recommender systems
correct person, resulting in optimizing overall
A designation for systems that match and then A class of machine learning algorithms that use
performance. GANs are useful when research-
exceed the full range of human cognitive ability data to predict, narrow down, and find what
ers don’t have enough data to train an algorith-
across all economically valuable tasks. people are looking for among an exponentially
mic model, and are also used to create new,
synthetic data. growing number of options.
Automatic speech recognition
Algorithmic systems that give computers the Transformer
Deepfakes, which have become prevalent in the
ability to recognize and convert audio to human A type of neural network mechanism that learns
past year, are generated using GANs. In design,
readable language. what text means when it appears in a particular
GANs are tremendously useful: They can pro-
duce thousands of designs and recommend the context. Using “attention,” a transformer looks at
Computer vision
best ones based on pre-set parameters. They an input sequence and determines at each step
Processes that give computers the ability to
can generate and modulate voices, faces, even what other parts of the sequence are important.
derive meaningful information from digital imag-
gestures. Researchers from Nvidia, Massachu- To date, transformers have mainly been used in
es (including still and video) and to mimic and
setts General Hospital, BWH Center for Clinical natural language processing, image generation,
manipulate such images.
Data Science, and the Mayo Clinic collaborated and genome sequencing.
on a GAN that generates synthetic MRIs show-
ing cancerous tumors.

12 © 2023 Future Today Institute

 TECH Artificial Intelligence

ONES TO WATCH
Dr. Ali Madani, for leading the AI for protein Ashley Llorens, vice president, distinguished Dr. David Chalmers, professor of philosophy Dr. Ilya Sutskever, co-founder and chief
engineering moonshot (ProGen) at Salesforce scientist, and managing director, Microsoft and neural science at New York University, for scientist at OpenAI, and Sam Altman, CEO of
Research and developing large language Research Outreach, for leading inclusive, researching whether large language models are OpenAI, for pushing the frontiers of generative AI.
models used to design artificial proteins that collaborative research projects. sentient.
outperformed their naturally occurring peers. Dr. Irina Rish, professor of computer science
Dr. Cassie Kozyrkov, chief decision scientist at Dr. Dong Yu, vice general manager and and operations research at the University of
Allison Gardner, senior scientific adviser Google, for innovating in the combined fields of distinguished scientist at Tencent AI Lab, for his Montreal and a core member of the Mila—Quebec
at NICE, for her research in algorithmic bias, data science and behavioral science. work in natural language processing and speech AI Institute, for her research on sparse modeling
diversity, and inclusion. recognition. and probabilistic inference, dialogue generation,
Dr. Daniel Cohen-Or, professor at the biologically plausible reinforcement learning, and
Dr. Anima Anandkumar, director of ML Department of Computer Science and The Dr. Dou Shen, chief of Baidu AI Cloud, for dynamical systems approaches to brain imaging
research at Nvidia and Bren Professor at Caltech, Isaias Nizri Chair in Visual Computing at Tel Aviv leading the development of large Chinese analysis.
for her research in tensor-algebraic methods, University, along with Dr. Amit H. Bermano, language models and enterprise applications.
deep learning, and non-convex problems. assistant professor at Tel Aviv University; Dr. Dr. Jae Lew, director of computer vision,
Ron Mokady, computer vision research at Dr. Eric Boyd, corporate vice president of robotics, and AI, for developing innovative
Ari Kalfayan, global head of business Google; Rotem Tzaban, software engineer at Microsoft’s AI Platform, for developing an automation systems using machine and deep
development for early-stage AI/ML startups at Google; and Rinon Gal, computer science PhD enterprise path for AI systems. learning.
Amazon Web Services, for scouting cutting-edge candidate at Tel Aviv University, for developing a
startups. framework for semantic editing of faces in videos, Drew Harwell, reporter covering AI at The Dr. Joelle Pineau, director of Meta AI Research
demonstrating significant improvements over the Washington Post, for investigating the people, Labs, for developing new models and algorithms
Ashish Vaswani, co-founder and chief scientist current state-of-the-art. products, and companies that make and make use for planning and learning in complex, partially
of Adept.AI, and co-author of the transformer of AI. observable domains.
model architecture in 2017 at NIPS. Dr. Darío Gil, senior vice president and director
at IBM Research, for advancing fundamental AI Dr. Gadi Singer, vice president at Intel Labs Emad Mostaque, founder and CEO of Stability.
research to the benefit of industry and society. and director of Emergent AI Research, for driving AI, for the design of open-source AI tools.
innovation in machine intelligence and cognition.

13 © 2023 Future Today Institute

 TECH Artificial Intelligence

ONES TO WATCH
Dr. Jeff Dean, senior fellow and senior vice Dr. Mutale Nkonde, AI policy adviser at Harvard Dr. Rana el Kaliouby, deputy CEO of Smart Eye
president of Google AI (Research and Health), for University’s Berkman Klein Center, for her work and former co-founder of Affectiva, for bringing
large-scale distributed neural network systems. to eliminate the underrepresentation of black emotional intelligence to devices and digital
professionals in the American technology sector experiences.
Karen Hao, senior reporter for China tech and by 2030.
society at The Wall Street Journal, for her deep Dr. Samy Bengio, senior director of AI and ML
analysis of the social and geopolitical impacts of Dr. Prakhar Mehrotra, vice president for research at Apple, for his work applying deep
AI. machine learning US Omni Tech at Walmart learning to speech, image, and other forms of
Global Tech, for leading enterprise adoption of AI. content.
Kay Firth-Butterfield, head of AI and ML, and
member of the executive committee at the World Rachel Metz, a senior journalist (most recently Dr. Swami Sivasubramanian, vice president of
Economic Forum, for leading a global effort on the at CNN), for covering the business side of AI. database, analytics, and machine learning at AWS,
governance of AI, to harmonize AI policies, and to for advancing cloud capabilities and insights for
explore future global risks. Dr. Rachel Thomas, foundation director of businesses.
University of San Francisco Center for Applied
Khari Johnson and Will Knight, senior writers Data Ethics, for bringing people from around Dr. Timnit Gebru, founder and executive
for AI at Wired magazine, for their reporting and the world with diverse and nontraditional director at the Distributed AI Research Institute,
analysis of the positive and negative ways AI backgrounds into AI. for her research in AI ethics, algorithmic bias, and
shapes human lives. data mining.
Dr. Ramayya Krishnan, dean of the Heinz
Dr. Ming-Yu Liu, distinguished research scientist College of Information Systems and Public Policy Dr. Zhou Jingren, deputy director of Alibaba
and a director of research at Nvidia Research, for at Carnegie Mellon University, for his research Damo Academy (Alibaba’s leading-edge research
focusing on deep generative models and their on the responsible use of AI and in data-driven arm), for leading AI initiatives related to smart
applications. approaches to support workforce development. cities, autonomous driving, mobile computing
platforms, semiconductor R&D, and other areas.

14 © 2023 Future Today Institute

 TECH Artificial Intelligence

MODELS,
TECHNIQUES,
& RESEARCH
TRENDS

15 © 2023 Future Today Institute

 TECH Artificial Intelligence

AI MODELS
An AI model is a program or algorithm that uses trained data to spot patterns and automati-
cally make predictions or decisions. Usually, the more data an AI model can access, the better
it will be in delivering accurate answers or forecasts. A few models that might be included
in an AI library include: linear regression, logistic regression, and decision trees. While all
machine learning models are AI models, not all AI models are necessarily machine learning
models.

Research in new image and video models have shifted from traditional convolutional neural
networks to transformer-based discriminative models. Early this year, diffusion models took
the world by storm. These models are extremely proficient at generating images and text, and
we expect to see generative systems that can create realistic video this year.

Starting this year, recommender systems, coupled with generative AI, will lead to the fur-
ther personalization of products and services. AI systems of the future will not only try to find
content that consumers like, they will generate personalized content for that person’s specific
interests.

Some companies are releasing models to the public to continue training and tweaking
them—possibly prematurely. In August, Meta released its BlenderBot3 for anyone to use;
shortly after, the system began spinning out misinformation. Originally Google had planned to
keep its LaMDA (Language Model for Dialogue Applications) system in-house after publishing
a seminal paper on it in May 2021. But given increasing competition from OpenAI and Meta,
the company changed course, announcing a new “AI Test Kitchen” initiative, where registered
users can test various experiments, like the “List It” demo, which allows users to share a goal or
topic and LaMDA will break it down into a list of helpful subtasks.

The large-scale release of AI systems to hundreds of millions of users may result in problemat-
ic knock-on effects. On the other hand, it’s also likely to surface ethics and safety issues that
developers may have overlooked.

16 © 2023 Future Today Institute

 TECH Artificial Intelligence

AI MODELS

Modalities—Text ings illustrate that for compute-opti- eter GPT-3 model. This model was of images showing its interpretation Imagen
PaLM mal training, the model size and the trained on more recent data with the of the intended meaning. With text Developed by Google, this text-to-
The world’s largest dense LLM number of training tokens should be new reinforcement learning with hu- prompts, it can now edit images, image diffusion model has a high
developed by Google, at 540 scaled equally. The findings under- man feedback, a technique that used extend the original canvas, create degree of photorealism and a deep
billion parameters. PaLM shows scored the need for large, high-quali- human and machine written data to variations of an input image, and has level of language understanding.
breakthrough capabilities on ty training data sets. improve the original GPT-3 model. 4x higher resolution than the original Imagen builds on the power of large
numerous, very difficult tasks. PaLM This model is consumer focused, and DALL-E model. transformer language models in
RETRO is the backbone of ChatGPT. These understanding text and hinges on the
is the largest single LLM trained on
Retrieval Enhanced Transformers, sets of techniques are also commonly Stable Diffusion strength of diffusion models in high-fi-
TPUs (Tensor Processing Units) at
developed in February 2022 by Deep- referred to as InstructGPT models. Developed by Stability.AI, Stable Dif- delity image generation. The compu-
scale, with 6144 TPU chips.
Mind. Traditionally, the knowledge fusion is an open-source text-to-im- tation is done in the pixel space, unlike
BLOOM base of transformer models consists NLLB age, latent diffusion model that allows Stable Diffusion.
A collection of open-source models, of only the data it was trained on. This 55 billion parameter open-source people to create stunning art within
built through a community effort led RETRO addresses this problem by model developed by Meta AI is capa- seconds. It is a breakthrough in speed e-Diffi
by HuggingFace, with the largest obtaining a new knowledge base of ble of delivering evaluated, high-qual- and quality, allowing the model to run Developed by Nvidia, this text-to-im-
model at 176 billion parameters. It can “facts” through retrieving informa- ity translations directly between 200 on consumer GPUs. age diffusion model shows stunning
output coherent text in 46 languages tion from a database. RETRO helps languages—including low-resource results with instant style transfer and
LLMs stay current, without the need languages like Asturian, Luganda, Parti paintings with text-based prompts.
and 13 programming languages that
for retraining models. By decoupling Urdu, and more. Developed by Google and also known Unlike Stable Diffusion, which does
is hardly distinguishable from text
the knowledge base from the model, as Pathways Autoregressive Text-to- its computation in the latent space,
written by humans.
researchers show that RETRO models Modalities—Images Image, this autoregressive text-to- e-Diffi shows promise with diffusion
Chinchilla with ~25x fewer parameters can DALL-E 2 image generation model achieves in the pixel space by using the power
A 70 billion parameter model devel- achieve similar results to that of larger Developed by OpenAI last year, high-fidelity photorealistic image of transformer models to understand
oped by DeepMind trained on 1.4 tril- models. DALL-E 2 is an upgrade of DALL-E, generation and supports content-rich the relationship between text and
lion tokens proved that current LLMs a model trained to manipulate visual synthesis involving complex composi- images.
are largely undertrained. This model GPT-3.5 concepts through language. It begins tions and world knowledge.
outperformed larger models with up The new and improved version of with a prompt that’s written in natural
to 530 billion parameters. The find- OpenAI’s original 175 billion param- language and then generates a set

17 © 2023 Future Today Institute

 TECH Artificial Intelligence

AI MODELS
Muse standing will be used for tasks like
Masked generative transformers is a video classification and video-text
text-to-image model developed by retrieval.
Google that shows promising results
for image generation and editing use We anticipate that OpenAI will launch
cases. Muse is more efficient because a video system this year.
it employs discrete tokens, compared
Modalities—Voice
to pixel-space diffusion models, such
VALL-E
as Imagen and DALL-E 2. When
A voice synthesis transformer-based
compared to autoregressive models,
model that converts text to speech,
such as Parti, Muse is more efficient
VALL-E generates discrete audio
because it uses parallel decoding.
codec codes, corresponding to the
Modalities—Video target content and the speaker’s
Make-a-Video voice, enabling various text-to-speech
A state-of-the-art AI system applications, speech editing, and
that generates videos from text, content creation.
developed by Meta AI. The network
Modalities—
learns how the world moves by using
images with descriptions, along 3D and Animation
with unlabeled videos. The result is Get3D
astonishing; short video clips can A model that generates high-quality
be generated with just a few text 3D textured shapes learned from
prompts. images, GET3D is able to generate
high-quality 3D textured meshes,
MSFT X-CLIP ranging from cars, chairs, animals,
The authors used Lensa AI to generate this portrait of Amy Webb. Developed by Microsoft, this exten- motorcycles, and human characters
sion to CLIP for general video under- to buildings, achieving significant

18 © 2023 Future Today Institute

 TECH Artificial Intelligence

AI MODELS
improvements over previous MedPalm and transferable model of the se- its context whether to output text,
methods. A large language model for the med- quence-function relationship for joint torques, button presses, or other
ical community that advances SOTA data-driven protein engineering. tokens.
Magic3D in seven medical question-answering
This text-to-3D content creation tool tasks, including achieving 67% on Modalities— CLIP
creates 3D mesh models using a MedQA USMLE improving prior work The Contrastive Language-Image
Climate Science
coarse-to-fine strategy that leverages by more than 17%. Pre-training neural network from
FourCastNet
both low- and high-resolution diffu- OpenAI is an open-source, multimod-
Developed by Nvidia, the vision
sion priors for learning the 3D repre- EquiFold al, zero-shot model. Given an image
transformer-based global data-driven
sentation of the target content. An end-to-end differentiable, and text descriptions, the model
weather forecasting model provides
SE(3)-equivariant, all-atom protein can predict the most relevant text
accurate short- to medium-range
Modalities—Bioengineering structure prediction model. description for that image, without
global predictions at 0.25 degree
OpenFold optimizing for a particular task.
ESM-2 resolution. FourCastNet accurately
A fast, memory-efficient, and
A deep contextual language model forecasts high-resolution, fast- Flava
trainable implementation of
trained on 86 billion amino acids timescale variables such as the A foundational language and vision
AlphaFold2 and OpenProteinSet, the
across 250 million protein sequences surface wind speed, precipitation, and alignment model from Meta that
largest public database of protein
multiple sequence alignments. spanning evolutionary diversity. atmospheric water vapor. explicitly and simultaneously targets
language and vision tasks.
MegaMolBART GenSLMs Multi-Modalities
A deep learning model for small mol- Genome-scale language models that Gato
ecule drug discovery and cheminfor- can learn the evolutionary landscape An agent from DeepMind that works
matics based on SMILES (Simplified of SARS-CoV-2 genomes (among as a multimodal, multitask, multi-
Molecular Input Line Entry System). others). embodiment generalist policy. The
MegaMolBART uses Nvidia’s Ne- same network with the same weights
ProtVAE can play Atari, caption images, chat,
Mo-Megatron framework, which is
A deep generative model that sur- stack blocks with a real robot arm,
designed for the development of large
faces an accurate, generative, fast, and much more, deciding based on
transformer models.

19 © 2023 Future Today Institute

 TECH Artificial Intelligence

TECHNIQUES
PROMPT LEARNING/ENGINEERING is then fed into a pipeline with a pretrained or traditional machine learning methods, which are by Google researchers in 2016, this framework
Prompt learning/engineering encompasses a fine-tuned LLM (policy model) to constantly im- time-consuming and difficult, and require data makes it possible for algorithms to use data on
collection of techniques that focus on custom- prove the outputs of the LLM, based on human scientists, AI specialists, and engineers. AutoML devices—such as mobile phones and smart-
izing a foundational LLM for a particular use intentions. operates differently by matching raw data and watches—without compromising user privacy.
case, using a supervised learning technique. models together to reveal the most relevant Research in this space has dramatically in-
IMITATION LEARNING information. Google, Amazon, and Microsoft creased.
These methods are typically more efficient than
fine-tuning as they require less computation Neural networks are trained to perform tasks by include a host of AutoML products within their
watching humans do them. Imitation learning cloud service offerings. GENERAL REINFORCEMENT LEARNING
and no retraining of the foundational model.
ALGORITHMS
The objective is to feed better inputs into the can be used to train AI to control robot arms,
drive cars, or navigate webpages. Previously, CONTINUOUS LEARNING Researchers are developing single algorithms
foundational model to obtain better responses.
The process involves adding a snippet of text this had to be done by a human annotating each At the moment, deep learning techniques are that can learn multiple tasks. DeepMind, the
or vectors to input data during the supervised action. Video Pre-Training (VPT) is an approach helping systems learn to solve complex tasks in AI subsidiary of Alphabet and the team behind
learning phase, to better model and understand that empowers a neural network to automati- a way that resembles what humans can do—but AlphaGo, which learned how to play Go with the
the outcomes of the foundational model. Once cally label videos. Given the billions of hours of those tasks are still specific, such as beating a skill level of a human grandmaster, continues
learned, these snippets are added to all future video now available for free online, research- human at a game. And they require a rigid se- to push its research forward. MuZero mastered
input prompts. ers have access to a vast resource to train AI quence: Gather data, determine the goal, deploy multiple games without being told their rules, a
systems. In the future, might this actually be a an algorithm. This process requires humans and “significant step forward in the pursuit of gener-
REINFORCEMENT LEARNING WITH HUMAN way to clone you—at least digitally? Maybe. Your can be time-consuming, especially during early al-purpose algorithms,” according to DeepMind.
FEEDBACK (RLHF) digital clone trained on your data could read phases when supervised training is required. In a seminal paper, “Reward Is Enough,” pub-
RLHF is a technique used to align LLMs with through websites, shop for holiday gifts, or book Continuous learning is more about autonomous lished at the end of 2022, DeepMind researchers
human intentions and is based on training a hotel rooms—just as you would. Taking that a and incremental skill building and development, hypothesized that artificial general intelligence
reward model to mimic human feedback and step further: a robotic clone might someday and researchers will continue to push the limits could be achieved through reinforcement learn-
intentions. This training uses prompt-generation do your gardening or move boxes around after of what’s possible in this field. ing alone.
pairs from a predefined data set, which takes watching videos to learn how you move.
FEDERATED LEARNING GRAPH NEURAL NETWORKS
a prompt and its corresponding completion
to output a single “reward,” or a score of how AUTOMATED MACHINE LEARNING (AUTOML) Federated learning is a technique that distrib- Because we perceive scents using millions of
good the completion was. This reward model Some organizations want to move away from utes machine learning to the edge. Introduced sensory neurons in our brains, and because

20 © 2023 Future Today Institute

 TECH Artificial Intelligence

TECHNIQUES
scents are multifaceted, predicting the way As of January, one model stands out: MTV, from images, and also to provide the reasoning A few years ago, if you typed “CEO” into Google
something will smell is incredibly complex. For a collaboration between Google Research, behind their answers (unlike the VQA chal- Images, the first result of a woman was CEO
example, how would you describe the smell Michigan State University, and Brown University, lenge, which only requires an answer). The data Barbie. In an experiment, researchers at Carne-
of an orange? Sweet? Bright? Grassy? Each which achieved a 89.6% Top-1 accuracy on the set contains 290,000 pairs of multiple-choice gie Mellon University trained a system to auto-
descriptor is unique. Classifying smell is tricky 600 series, 89.1% accuracy on the 400 series, questions, answers, and rationales from 110,000 complete images of men and women cropped
because it requires a multi-label system. and 82.20% accuracy on the 700 series. image scenarios taken from movies. Although below the neck. In pictures of men, the system
progress has been made since the challenge autocompleted him wearing a suit. The system
GNNs constitute a particular type of deep neural ActivityNet is a video data set for human activity was launched, improvements have become autocompleted women—including US Rep.
network that operates on graphs as inputs. They understanding that contains 700 hours of videos increasingly marginal, suggesting that new Alexandria Ocasio-Cortez (D-N.Y.)—wearing a
are being used to detect smell—to predict odors of people doing 200 different activities (long techniques may need to be invented to signifi- low-cut top or bikini 53% of the time.
at a molecular level—and for a wide array of jump, dog walking, vacuuming, etc.). For an AI cantly improve performance.
chemical and biological processes. For example, system to successfully complete the ActivityNet PREDICTIVE MODELS USING
researchers at the Broad Institute, MIT and Har- Temporal Action Localization Task (TALT) task, MACHINE IMAGE COMPLETION INCOMPLETE DATA
vard’s biomedical and genomic research center, it has to execute two separate steps: (1) localiza- Computer vision systems are getting smart-
If a computer system has access to enough
used them to discover antibiotic compounds tion (identify the precise interval during which er, which enables neural networks to predict
images—say, millions and millions—it can patch
that don’t have toxic side effects. the activity occurs); and (2) recognition (assign geometry from a single color image. In 2019, the
and fill in holes in pictures. This capability has
the correct category label). Temporal action DeepMind team developed a GAN that creates
practical applications for professional photog-
ACTIVITY VISION localization is one of the most complex and
raphers, as well as for everyone who wants to videos from images. For example: Imagine a
A fundamental subtask in video computer vision difficult tasks in computer vision. Performance photo of a person holding a basketball. Based
take a better selfie. Soon, if the foreground of
is activity recognition: identifying the activities on TALT is measured in terms of mean average on his posture, face, and other data within the
a mountain is out of focus, or if your skin has
that occur in videos. AI systems have been chal- precision, with a higher score indicating greater picture, the GAN figures out what likely hap-
an unsightly blemish, another version can be
lenged to classify activities that range from sim- accuracy. pened next and generates a video clip of the
swapped in to generate the perfect picture. As
ple actions, like walking, waving, or standing, to such technology becomes commonplace, there action.
ones that are more complex and contain multi- VISUAL COMMONSENSE REASONING (VCR)
will be significant biases and other pitfalls to
ple steps, like preparing a salad (which requires The Visual Commonsense Reasoning challenge Earlier, researchers at MIT’s Computer Science
navigate. For example, image generation algo-
an AI system to recognize and chain together is a relatively new benchmark for visual under- and Artificial Intelligence Laboratory (CSAIL)
rithms routinely reflect deeply culturally embed-
discrete actions like cutting tomatoes, washing standing. VCR asks AI systems to answer chal- trained computers to predict what humans
ded racism and sexism.
the greens, applying dressing, and the like). lenging questions about scenarios presented would do next using YouTube videos and TV

21 © 2023 Future Today Institute

 TECH Artificial Intelligence

TECHNIQUES
shows such as “The Office” and “Desperate through symbols rather than relying on human technique called vokenization extrapolates
Housewives.” CSAIL’s system predicts whether programmers to sort, tag, and catalog data for language-only data by contextually mapping
two people are likely to hug, kiss, shake hands, them. Symbolic algorithms will aid the process, language “tokens,” or the words used to train
or slap a high five. SinGAN is an unconditional which should eventually lead to robust systems language models, to related images, or “vokens.”
generative scheme that can manipulate and that don’t always require a human for training. For example, auto-generated image captions
enhance images—sketch a mountain, and it will often can’t infer context. Vokenization would
produce a realistic-looking synthetic photo- REAL-TIME MACHINE LEARNING (RTML) enable machines not just to recognize objects
graph. This research will someday enable robots One big challenge in AI is building machines but to really “see” what’s in them.
to more easily navigate human environments, that can proactively collect and interpret data,
and to interact with humans by taking cues from spot patterns and incorporate context, and MODEL-FREE APPROACHES TO RL
body language. Retail, manufacturing, and edu- ultimately learn in real time. New research into Dreamer is a reinforcement learning (RL) agent
cation settings could be especially relevant. RTML shows that it’s possible to use a contin- that uses a world model to learn long-sighted
ual flow of data and adjust models in real time. predictions, employing backpropagation (an
NEURO-SYMBOLIC AI This signals a big change in how data moves, algorithmic way of working backward to test for
The development of AI has been on two concep- and in how we retrieve information. The Nation- errors) through model predictions. It can create
tual tracks since the 1950s: symbolic (machines al Science Foundation launched a $10 million models from raw images and learn from thou-
that use a base of knowledge and rules that rep- grant program to catalyze research in this area, sands of predicted sequences in parallel using
resent concepts) and non-symbolic (machines although all of the Big Tech companies are a graphics processing unit (GPU). This new
that use raw data to create their own patterns working closely to advance RTML too. approach solves long-horizon tasks using an
and representations of concepts). Classic AI is imagined world.
the former, because it more closely represents VOKENIZATION
how we understand human thought—and the Models like GPT-3 are trained on syntax and
original intent was to teach machines to think grammar, not creativity or common sense. So
like us. researchers at the University of North Caroli-
na-Chapel Hill are combining language mod-
Researchers are working on new ways to
els with computer vision. Humans learn in a
combine both learning and logic using neu-
multilayered, multidimensional way, so a new
ral networks, which would understand data

22 © 2023 Future Today Institute

 TECH Artificial Intelligence

RESEARCH
BUILDING SUPERSIZED AI MODELS tein language models, named ProGen2, that are references, and is likely the stepping stone to We are just scratching the surface on all
Last year, we saw the proliferation of large AI scaled up to 6.4 billion parameters and trained true conversational search. The startup Adept three fronts:
models, and this year supersized models are on on different sequence data sets drawn from over is working on commercialization: It hopes to
a billion proteins from genomic, metagenomic, customize large models to do things like interact Data
the horizon. For context, GPT-3—widely hailed
and immune repertoire databases. with websites, APIs, and apps to drive workflow Today, the internet has provided us with an
as a powerhouse—has 175 billion parameters.
productivity. unprecedented amount of data that we never
Huawei debuted a 200 billion parameter lan-
Meta also introduced its own model to pre- had access to before, thanks to the number of
guage model called Pangu, while Baidu and the
dict protein structures, called ESMFold, which TRAINING LANGUAGE MODELS TO ACHIEVE internet users worldwide, increased usage of
Peng Cheng Lab released PCL-BAIDU Wenx-
demonstrated similar predictions to DeepMind’s SCALE the internet, and the ability of generative AI to
in, with 280 billion parameters. PCL-BAIDU is
AlphaFold2. Outside of bioengineering, Google’s Current opinion on language model (LM) scal- create high-fidelity synthetic data.
already deployed to Baidu’s news feeds, search
PaLM-SayCan is an LLM that can use abstract ing laws might be wrong; they are enormous,
engine, and digital assistant. Gopher, which was Compute
steps to complete a goal, such as retrieving a yet the amount of data they’re trained on isn’t
released by DeepMind in December 2021, has Although Moore’s law may have reached the
sponge. This is a remarkable achievement, be- commensurate with their overall size. The impli-
280 billion parameters. And Microsoft’s Meg- limit in terms of miniaturization of electronics,
cause sponges are soft and flexible, properties cation: Existing models aren’t trained enough,
atron-Turing NLG, built in collaboration with we are seeing continued acceleration of the
that introduce new variables to the task. During and that’s a problem. Researchers are eying a
Nvidia, has 530 billion parameters. Google’s compute stack through software and algorithms.
testing, SayCan was drilled on more than 100 different approach going forward.
Switch-Transformer and GLaM models have a Software will continue to accelerate perfor-
instructions, and it was successful in planning
staggering 1 trillion and 1.2 trillion parameters, mance and improve the energy efficiency of
and execution 84% of the time. OpenAI has proven to many that foundational
respectively. But even bigger is Wu Dao 2.0 from chips exponentially.
models can be improved efficiently by using
the Beijing Academy of AI, which reportedly has DEVELOPING PRACTICAL TOOLS FOR RLHF, which might explain why the company
1.75 trillion parameters. LANGUAGE MODELS Algorithms
released its generative AI systems publicly.
Distributed and parallel computing algorithms
Language models can learn to use the tools we Research now clearly indicates—across mo-
CREATING SPECIALIZED LANGUAGE MODELS and frameworks are rapidly evolving. Broadly,
rely on every day, like calculators and search, dalities—that a trifecta of (1) large-scale data,
Increasingly, researchers are introducing spe- transformers have proven to scale efficiently,
simply by accessing text interfaces. They only (2) accelerated computing, and (3) powerful
cialized language models for different indus- and can handle mountains of data. Innovations
need a few human demonstrations in order to algorithms can lead to astonishing results. Data
tries, and LLMs for proteins are already in use. are accelerating on transformer and diffusion
learn. OpenAI’s WebGPT is a tool that GPT3 scaling leads to better predictions from multibil-
Researchers at Salesforce created a suite of pro- architecture today.
uses to search the web, provides answers with lion parameter models.

23 © 2023 Future Today Institute

 TECH Artificial Intelligence

RESEARCH
ADOPTING UNIFIED LEARNING PROCESSES AMPING UP DIFFUSION MODELS ple, GPT3 was released in June 2020, and by LLM for code. Salesforce’s CodeGen is building
Deep neural nets are good at identifying objects Diffusion models are beating GANs at bench- August 2022 a number of models were in use: a powerful alternative to Codex. Meanwhile,
in photos and videos and processing natural marks for text-to-image generation. These Pan-Gu from Huawei, Ernie 3.0 Titan from Baidu, DeepMind’s AlphaCode is already generating
language, but until recently models had to be generative models learn to model data distribu- and later OPT from Meta. DALL-E was released whole programs.
trained separately. Might there be one model to tion from whatever the input is, and with training in January 2021, and Meta’s Make-a-Scene
launched a year later, followed by Google’s Ima- FRAMEWORK CONSOLIDATION
rule them all? That was Google’s proposal back in they can generate new data sets similar to the
2017. Since then, DeepMind’s Gato has emerged ones they trained on. Stable Diffusion is based gen in May 2022 and Stable Diffusion in August Google’s TensorFlow and Meta’s PyTorch are
as a 1.2 billion parameter transformer capable on diffusion models. 2022. Even AlphaFold 1, released in August 2018, two popular frameworks used by researchers,
of performing hundreds of tasks in robotics, had fast followers. Rosetta Fold launched in and the relative popularity of different frame-
simulated environments, vision, and language. MANAGING THE COMMUNITY-DRIVEN OPEN 2021, followed by OpenFold from Columbia Uni- works typically mirrors trends in the commercial
Researchers at Meta have now developed Data- SOURCING OF CODE versity and Meta’s ESM 2 in 2022. While these application landscape. In the past five years,
2vec, a system that deploys a single algorithm to Code is important for reproducibility, account- models aren’t necessarily replicas, they either Meta seems to have gained ground. Of the 2021
train a neural network to recognize images, text, ability, and transparency, and it is a key to build on the ideas or the code that existed. conference papers that mention the framework
or speech. It unifies the learning process through driving improvements in the greater AI commu- the researchers used, 75% cited PyTorch but not
AI FOR CODE RESEARCH TensorFlow. Of the 161 researchers who pub-
self-supervised learning, which allows the neural nity. But when academic researchers publish
net to recognize patterns in data sets on its own, papers, they don’t often include all of their code. Google, DeepMind, Salesforce, and OpenAI are lished more TensorFlow papers than PyTorch
without being fed labeled examples. The reason given: The code they used is inter- actively building systems that help the com- papers, 55% of them switched to PyTorch, while
mingled with other proprietary research, and it puter science community develop code. Codex only 15% moved in the other direction. JAX, used
TRANSITIONING TO TEXTLESS NLP therefore can’t be released. Fewer than 20% of from OpenAI powers GitHub’s Copilot—think for the development of LLMs, gained popularity
Most LLMs have been trained on publicly avail- all academic papers on AI publish their full code, of it as an autocomplete for code. It can string in the middle of the year with TPUs, but early
able data sets such as Reddit and Wikipedia. and some big names—DeepMind and OpenAI— together multiple lines of code from natural this year activity had cooled.
Both are rife with biases. Researchers are de- notoriously leave theirs out, citing proprietary language instructions. While at first the results
concerns. were questionable, there has been enough LOWERING THE COST OF TRAINING MODELS
veloping generative spoken language modeling,
which extracts speech from raw audio without interest that corporate labs are now compet- It’s expensive to train a model, and several vari-
That’s why what’s currently happening within ing to build commercial products. BigCode is
labels or text. The hope is that AI could become ables influence those costs, all of which have
the generative AI community is so interesting. an open-source collective initiative driven by
more inclusive if it uses podcasts, local radio, increased in the past few years. For example,
Recently, as big models are released, others are ServiceNow and HuggingFace to develop an
and other sources of spoken language. it costs an average of $1 per 1,000 parameters
quickly cloning and improving them. For exam-

24 © 2023 Future Today Institute

 TECH Artificial Intelligence

RESEARCH
today, which suggests that OpenAI likely spent their performance. The human baseline score to probe LLMs and extrapolate their future ca-
more than $10 million to train GPT-3. Smaller is 87, and between May 2018 and August 2020, pabilities across hundreds of tasks.
research groups and companies cannot afford natural language processing systems surpassed
these expenses. Some in the AI community are humans by increasing their scores from 60 to EXPLAINABLE AI (XAI)
instead allowing the big tech companies to pre- 90.6. There’s a renewed call to make AI explain-
train and publish big models. The cost of com- able—meaning, to reveal to researchers and
puting will decline over time, and accelerations The SuperGLUE benchmark is a new measure-
others how systems make their forecasts and
at the software layer are underway. Frameworks, ment of more difficult language understanding
decisions. Many aspects of AI systems can be
and distributed training starting to take shape, tasks, improved resources, and a new public
explained, such as showing their training data,
are delivering meaningful gains every year. leaderboard. When SuperGLUE was introduced,
or their lack of data in a particular region of the
there was a nearly 20-point gap between the
corpus. XAI also calls for transparency in the
SURPASSING BENCHMARKS best-performing model and human performance
fairness of gathering data, the description of
on the leaderboard, but by 2022 AI models from
As AI systems improve, they are surpassing people involved in supervised learning, and the
Microsoft and Google had outperformed humans.
benchmarks. The original General Language ranking of features. Emerging approaches to
Existing language benchmarks still fail to capture
Understanding Evaluation (GLUE) benchmark is XAI methods include showing how outcomes
biases encoded in public data—future bench-
a collection of resources for training, evaluating, were validated—meaning, how can we trust the
marks could be designed to resolve this gap.
and analyzing natural language understand- recommendations, predictions, and classifica-
ing systems. It includes a benchmark of tasks This year, watch the Holistic Evaluation of Lan- tions made?
around understanding nine sentences or a pair guage Models (HELM) from Stanford University,
of sentences built on existing data sets, which which holistically evaluates the performance
are selected to cover a diverse range of data set of LLMs. The benchmark includes measuring
sizes, text genres, and degrees of difficulty. models across a variety of capabilities (span-
ning tasks, domains, and languages), multiple
It also has a diagnostic data set designed to
metrics (accuracy, fairness, bias, toxicity, robust-
evaluate and analyze model performance with
ness, etc.), and the standardization of method-
respect to a wide range of linguistic phenome-
ology and data sets across models. Also watch
na found in natural language. And it includes a
BigBench, a collaborative benchmark intended
public leaderboard so that researchers can track

25 © 2023 Future Today Institute

 TECH Artificial Intelligence

TALENT
TRENDS

26 © 2023 Future Today Institute

 TECH Artificial Intelligence

TALENT
DEMAND FOR AI TALENT GROWING FAST NUMBER OF AI PATENT FILINGS, 2010-2021
For many years, demand for AI talent has out-
paced supply. In the US, there were nearly three
141.24
times more AI-related job postings on Indeed last 140
year than job views for AI-related roles. While
schools are adding programs, increasing enroll-
ment, and adding classes, there is just too much 120

Number of AI Patents (in thousands)

demand for AI skills and nowhere near enough
trained workers. The hiring process for AI job 100
postings is taking longer and becoming more
expensive, impeding growth at some companies.
80

UPSKILLING FOR AI
Companies are looking to upskill their workforce 60
in machine learning and the basics of AI. As a re-
sult, new training programs abound to augment
40
the knowledge, skills, and competencies required
of a modern organization. Levi Strauss launched
a machine learning bootcamp to upskill its work- 20
force, showing staff how to apply AI-thinking to
everyday tasks. Founded by Harvard University 0
and University of California, Los Angeles faculty, 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Univ.AI is an online program for training in ma-
chine learning and AI. As AI moves more deeply into the mainstream, new patent filings have followed. The number of patents filed globally in 2021 was more
than 30 times higher than the number filed in 2015.

Source: CSET and 1790 Analytics database containing all AI granted patents and published applications (collectively “patent documents”) in all patent sys-
tems worldwide. https://github.com/georgetown-cset/1790-ai-patent-data

27 © 2023 Future Today Institute

 TECH Artificial Intelligence

TALENT
AI BRAIN DRAIN CORPORATE AI LABS BUILDING A SILICON VALLEY OF THE
MIDDLE EAST
The brain drain of AI researchers out of academia AI labs are located around the world, with con-
and into corporations is growing at an alarming centrations in North America, Europe, and Asia. The Middle East is vying to become the next Sili-
pace. The reason is simple: compensation pack- While the tech sector laid off tens of thousands con Valley, with a particular focus on developing
ages. Top academics earn generous corporate of workers between Q4 2022 and Q1 2023, many its own AI ecosystem. In 2022, nearly 3,000 start-
salaries and benefits, and they get to work in a corporate AI labs and positions remained intact. ups launched with nearly $1 billion in investment
similar tenured environment that’s carefully cul- Meta, Google, IBM, and Microsoft operate more according to MAGNiTT, a Dubai-based research
tivated to replicate their experience in academia. than 60 labs dedicated to AI R&D, and the ma- firm. The United Arab Emirates government
Between 2004 and 2018, Google, DeepMind, jority are outside of the US because of access to has developed ambitious plans to establish the
Amazon, and Microsoft hired 52 tenured and ten- talent. During the Trump administration, immigra- country as a global AI hub. In 2017, it published
ure-track professors from American universities. tion restrictions and stringent visa requirements a National Artificial Intelligence Strategy and
made it difficult to bring talent into America, and launched the Mohamed bin Zayed University of
Tech companies are also endowing AI professor- overseas labs allowed companies to avoid that Artificial Intelligence in Abu Dhabi. AI is taking
ships at top universities. In some cases, profes- barrier. Most of those labs do basic AI research center stage in the D33 Dubai Economic Agen-
sors take one- or two-year sabbaticals to work at rather than product development. da, which aims to generate new economic value
tech companies and then return to their univer- from digital transformation and AI technologies.
sities. But corporate benefits can be difficult to
give up, and companies need the talent. In one Meanwhile, Saudi Arabia is also investing heavily
infamous case, Uber poached an entire robotics in AI. In 2019, it established the Authority for Data
lab in 2015 from Carnegie Mellon University—40 and Artificial Intelligence, and created a National
professors and researchers in total—to work on Center for Artificial Intelligence and a National
self-driving cars. Poaching academia today could Data Management Office. Technology is the
rob the future of future AI experts: Without great focus of the Kingdom’s Vision 2030 plan, a long-
scholars, who will train the next generation of term strategy intended to stimulate economic
innovators? growth. The Kingdom also launched its Saudi
Venture Capital Co., with an initial $1.33 billion
under management, and is making large invest-
ments in tech companies.

28 © 2023 Future Today Institute

 TECH Artificial Intelligence

By humanizing technology, we have this

golden opportunity to reimagine how we
connect with machines, and therefore,
how we, as human beings, connect with
one another.

DR. RANA EL KALIOUBY,

CO-FOUNDER OF AFFECTIVA, DEPUTY CEO OF SMART EYE

29 © 2023 Future Today Institute

 TECH Artificial Intelligence

ENTERPRISE TRENDS

30 © 2023 Future Today Institute

 TECH Artificial Intelligence

ENTERPRISE
CORPORATE PARTNERSHIPS not invest both in talent and technology, and launch their ideas into the marketplace. AWS need to be solved. This approach is now used in
Microsoft, Google, and AWS are all partnering outsourcing was a path forward. While external Lambda lets teams run code for virtually any machine learning. Some of the fastest-growing
with outside companies to compete in, and partners made the transition easier, the one- type of application or back-end service—without GitHub projects are MLOps, or projects that
with, AI. Hyperscalers and challenger AI com- size-fits-most approach has fallen out of favor. provisioning or managing servers or hands-on deal with tooling, infrastructure, and operations.
pute providers are tallying up major AI compute As new AI solutions emerge and companies administration. The Azure Functions architec-
approach maturity in their deployment and ture supports myriad programming languages, Going forward, MLOps will describe a set of best
partnerships, most notably Microsoft’s $1 billion
integration of technology, organizations are scales on demand, and charges only for active practices that combines machine learning, tra-
investment into OpenAI. In 2019, Microsoft in-
now opting to build solutions in-house. This compute time. ditional DevOps, and data engineering. Already,
vested $1 billion into OpenAI between cash and
means increased investment in hiring technical MLOps is becoming a formal discipline within
credits for Azure, Microsoft’s cloud computing
leadership as well as investment in the technol- While some engineers worry that such server- organizations, as the overhead to transition AI
business. OpenAI’s ChatGPT or WebGPT might
ogies and systems required to operationalize AI less systems require them to surrender too experiments into production is quite high. This
be integrated into Bing, which could make Mic-
systems. much control, enterprises have signed large year, look for tools, automation options, and best
rosoft’s search engine the first real competitor to
commits with cloud solutions providers to ob- practices to manage large-scale infrastructure.
Google’s hegemony since Explorer.
Some examples of this approach we’re seeing tain better pricing. CPU workloads are becom-
from our Future Today Institute clients include: ing commoditized and cheap, while GPU work- LEVERAGING AI IN THE CLOUD
OpenAI expects $1 billion in revenue by 2024,
according to Reuters. In 2023, we expect the predictive risk and compliance management, loads allow enterprises to retire quota. All this Corporate leaders within the AI ecosystem have
biggest players to seek out or even announce automated supply chain planning, dynamic pric- makes serverless computing an easier pill to been racing to capture AI cloudshare—and
strategic relationships with an AGI-focused ing models for retail and hospitality, automated swallow when companies don’t have to earmark to become the most trusted provider of AI on
organization. This is already happening quietly. call centers and customer service operations, big CapEx investment to build AI capabilities. remote servers. Data affinity, compliance, and a
The big three LLM service providers are part- assisted recruitment, and various financial appli- lengthy vendor selection process has resulted in
cations (audits, forecasting, planning). THE RISE OF MLOPS sticky CSP contracts, since switching vendors is
nered with hyperscalers: AI21 is on AWS, Cohere
is on GCP, OpenAI is on Azure. Adept.AI, an ML As machine learning matures and new applied proving to be a costly exercise. For that reason,
ADOPTING SERVERLESS COMPUTING business solutions emerge, developers are the competition to cement new deals with enter-
research and product lab, has inked a deal with
Oracle Cloud Infrastructure. Hyperscalers like AWS, Alibaba Cloud, Micro- shifting their focus from building models to op- prise customers is heating up. In the West, the
soft’s Azure, Google Cloud, and Baidu Cloud erating them. In the software world, a set of best field is led by Amazon, Microsoft, and Google,
BUILDING IN-HOUSE SOLUTIONS are rolling out new offerings and packages for practices known as DevOps relies on tools, au- followed by IBM and Oracle. In Asian markets,
As companies adopted AI, they initially relied on developers with the goal of making it easier and tomation, and workflows to reduce complexity Alibaba and Baidu dominate the AI cloud.
packaged solutions from vendors. They could more affordable for a variety of AI startups to so that developers can focus on problems that

31 © 2023 Future Today Institute

 TECH Artificial Intelligence

ENTERPRISE
IMPLEMENTING AI AT THE EDGE LOW-CODE OR NO-CODE MACHINE mobile and web app builder for AWS. Microsoft
LEARNING Power Apps is a low-code application develop-
AI-driven processing and decision-making
that occurs closer to the source of data gen- Machine learning is transitioning, as new plat- ment environment on Azure.
eration, as opposed to in the cloud, is a tech- forms allow businesses to leverage the power
of AI to build applications without the need to CONSOLIDATION IN AI’S ECOSYSTEM
nique known as edge computing. The Internet
of Things (IoT) and its billions of devices, know specific code. Businesses can turn their As much as the AI ecosystem booms, a rush
combined with 5G networking and increased unruly data sets into structured data that can be of acquisitions means consolidation, too. Big
computing power, has made large-scale AI at trained, and they can build and deploy models companies now snap up startups long before
the edge possible. Processing data directly on with minimal skills. they have time to mature—the average age at
devices will be important in the future for health acquisition is 3 years old. Just a handful of big
Create ML is Apple’s no-code, drag-and-drop
care, automotive, and manufacturing applica- companies dominate the AI landscape: Google,
tool that lets users build custom models such
tions because it’s potentially faster and safer. Amazon, Microsoft, IBM, Meta, and Apple in the
as recommendation engines, natural pro-
US, and Baidu, Alibaba, and Tencent in China,
Apple spent $200 million to acquire Xnor.ai, a cessing systems, and text classifiers. Google
with significant fortification and support from
Seattle-based AI startup focused on low-power Cloud’s AutoML includes image classification,
their country’s government.
machine learning software and hardware. Micro- object detection, translation, and all sorts of
soft offers a comprehensive toolkit called Azure pattern recognition tools to allow developers They are fueled by a robust ecosystem, which
IoT Edge that allows AI workloads to be moved with limited machine learning expertise to train includes Nvidia, Salesforce, Qualcomm, Intel
to the edge: Businesses can deploy complex high-quality models specific to their business Capital, Google Ventures, Salesforce, Samsung
event processing, machine learning, image rec- needs. MakeML allows developers to create an Ventures, and In-Q-Tel, among a small subset
ognition, and other high value AI without writing AI app or solve business problems using com- of others. When it comes to the future of AI, we
it in-house. Anyone is able to create AI modules puter vision. should ask whether consolidation makes sense
and make them available to the community for for the greater good and whether competition
Applications have included tracking tennis balls
use through the Azure Marketplace. will eventually be hindered (along with access),
during matches and automatically changing the
as we’ve seen in other fields such as telecom
colors of objects (such as flowers or dresses) in
and cable.
images. Last year, Amazon launched a no-code

32 © 2023 Future Today Institute

 TECH Artificial Intelligence

HARDWARE
ADVANCED AI CHIPSETS nounced that its next-gen chip, the Wafer Scale PROCESSING-IN-MEMORY TECHNOLOGY
Today’s neural networks have long required an Engine 2 (WSE 2), has 2.6 trillion transistors, A new approach to memory, which could even-
enormous amount of computing power, take a 850,000 cores, 40 gigabytes of on-chip mem- tually power the next generations of smart-
long time to train, and rely on data centers and ory, and 20 petabytes of memory bandwidth. phones and help usher more smart devices to
computers that consume hundreds of kilowatts Amazon’s homegrown AI chip AWS Inferentia market, involves breaking some of the current
of power. That is all starting to change. Enter the is a custom machine-learning chip used for bottlenecks in computing. Processing-in-mem-
SoC, or “system on a chip.” high-performance inference predictions, and it ory (PIM) integrates a processor with RAM on
now powers Alexa’s back-end services. a single chip, which allows computations to be
Big tech companies, including Huawei, Apple, performed in the memory of a computer.
Microsoft, Meta, Google, IBM, Nvidia, Intel, The pretrained chips will speed up commer-
and Qualcomm, as well as startups Graphcore, cialization and further R&D. But if the various In tests, PIM tech delivers significant perfor-
Mythic, and Cerebras Systems, are all develop- device manufacturers all start creating unique mance gain while cutting energy consump-
ing new systems architecture and SoCs—some protocols, developers may struggle with too tion. In late 2021, Samsung, which in addition
of which come pretrained. In short, this means many different frameworks. We anticipate an to phones and consumer appliances is also
that the chips are more readily able to work on eventual consolidation, pitting just a few com- the world’s largest manufacturer of dynamic
AI projects and should promise faster and more panies—and their SoCs and languages—against random-access memory, announced a new
secure processing. Projects that might other- one another. Google’s TPU, an AI accelerator PIM-enabled chip that could eventually double
wise take weeks could instead be accomplished application-specific integrated circuit (ASIC), the performance of neural nets. This matters to
in a matter of hours. is gaining traction in the enterprise and con- the enterprise because it could bring AI appli-
sumer marketplaces. Cerebras and Graphcore cations to a wider array of devices in the near
In 2019, Cerebras debuted an AI chip with 1.2 each announced breakthroughs in 2022, while future.
trillion transistors, 400,000 processor cores, 18 Amazon Search reduced ML inference costs by
gigabytes of SRAM, and interconnects (tiny con- 85% with AWS Inferentia. Traditional GPUs are
nection nodes) that can move 100 quadrillion still the mainstream, but we expect to see more
bits per second. That’s an astounding amount of developments in the AI chipset arena in 2023.
components and power—and yet in 2021, it an-
In 2021, unveiled the largest AI processor ever made,
the Wafer Scale Engine 2 (WSE-2).
Image Credit: Cerebras

33 © 2023 Future Today Institute

 TECH Artificial Intelligence

APPLIED AI

Health Care AI Funding APPLIED AI FOR HR

Recognition systems can now be deployed to
watch people in an interview and gauge enthu-
US $1 billion in 58 deals
siasm, tenacity, and poise. Algorithms analyze
hundreds of details, such as their tone of voice,
Europe $0.3 billion in 24 deals facial expressions, and mannerisms to best pre-
dict how a candidate will fit in with the culture
Asia $0.3 billion in 42 deals of a community. Startups such as HireVue use
AI systems to help companies decide which
candidates to hire.

But this kind of recognition technology has

Fintech AI Funding practical applications beyond job interviews:
It can detect when someone is likely to make
US $214 million in 19 deals a purchase—or attempt to shoplift—in a store,
whether someone is lying, and whether some-
one is receptive to new suggestions and ideas.
Europe $292 million in 10 deals
Unlike security cameras, which tend to have a
light indicating they’re recording, algorithms
Asia $34 million in 42 deals work invisibly, which means that this is an area
that could face regulatory scrutiny. The con-
Source: CB Insights
sumer advocacy organization Electronic Privacy
Information Center filed a complaint with the US
Federal Trade Commission requesting an inves-
tigation into HireVue, alleging its tools produce
results that are “biased, unprovable, and not
replicable” through algorithmic models.

34 © 2023 Future Today Institute

 TECH Artificial Intelligence

SPOTLIGHT: HEALTH CARE

PROTEIN FOLDING AI TO SPEED UP SCIENTIFIC DISCOVERY AI-FIRST DRUG DEVELOPMENT Much of the potential in AI stems from deep
In 2020, DeepMind’s AI made a big announce- Running experiments with several variables COVID-19 accelerated the use of AI in drug learning’s ability to sort through huge volumes
ment: It had solved a 50-year grand challenge often requires tiny, methodical tweaks to discovery. An international team crowdsourced of information, and learn and extrapolate from
with AlphaFold, an AI tool that predicts the measurements, materials, and inputs. Gradu- a Covid antiviral by synthesizing candidates that information. The upshot: AI can think faster
structure of proteins. AlphaFold outper- ate students might spend hundreds of tedious for 2,000 molecules in less than 48 hours—a than humans—sorting data in months versus
formed an estimated 100 teams in a biennial hours repeatedly making small adjustments until process that likely would have taken human years—and see patterns that humans may
protein-structure prediction challenge called they find a solution—a waste of their cognitive researchers a month or longer. In Japan, the first not. Still, drug discovery is tricky, because the
Critical Assessment of Structure Prediction, a abilities, and their time. phase of a clinical trial for an AI-designed drug algorithms rely on drug targets that must be
problem that has long vexed biologists. to treat obsessive-compulsive disorder showed published in research journals. Most data about
Research labs now use AI systems to speed a positive result. The drug, DSP-1181, acts as an potential compounds isn’t readily available, and
AlphaFold had previously bested other teams, the process of scientific discovery. Recursion, a agonist to the receptor for serotonin, a signaling when it is, it isn’t always complete or reliable.
but it worked so quickly and so accurately that biotechnology company based in Salt Lake City, molecule in the brain that mediates mood. The Filling the gaps and cleaning that data takes
it signaled a near future when the technology uses computer vision-based digital biomarkers, project used AI techniques to generate tens of time and money. It also requires data sharing—
could be used regularly by other scientists. such as respiratory rate, to assess and track dis- millions of potential molecules to try against the and most drug data is proprietary and locked up
Along with the newest version of AlphaFold, ease. Digitalizing in vivo studies will shorten the serotonin receptor, and then sift through the by big drugmakers.
DeepMind published full details of the system time to gather data and identify drug efficiency. candidates to decide which ones to prioritize for
and released its source code. It also made a Materials scientists at the University of British Using algorithms for drug development also
synthesis and testing.
stunning reveal: AlphaFold 2 has predicted the Columbia now rapidly test a new kind of solar brings up a host of ethical questions. Will bias
shapes of nearly every protein in the human cell and log results using a robot overseen by AI-first drug startups are attractive to investors. invade drug discovery much like it has other
body, as well as hundreds of thousands of other an AI algorithm. Based on the results of each Recursion raised $121 million in 2019 before arenas of AI, thereby marginalizing certain
proteins found in 20 of the most widely studied experiment, the algorithm determines what to spinning off CereXis, a new independent entity patients or diseases? Do algorithms need their
organisms, including yeast, fruit flies, and mice. change next. A 9- to 12-month process was to study rare brain cancers. Nearly every major own clinical trials? Could AI be used to take
All of this research will allow biologists to study completed in five days. DeepMind’s AlphaFold pharmaceutical company has inked deals with shortcuts and undermine the value of the sci-
and gain new insights on living organisms and will allow scientists to synthesize new drugs to AI drug discovery startups, including Johnson ence conducted inside the laboratory? Advo-
pathogens, which will form the basis for new treat diseases and develop enzymes that might & Johnson, Novartis, Merck, AstraZeneca, and cates say AI will make drug development and
drug development. someday break down pollution. GlaxoSmithKline. clinical trials more efficient, thereby cutting drug
prices and paving the way for more personalized
medicine.

35 © 2023 Future Today Institute

 TECH Artificial Intelligence

SPOTLIGHT: HEALTH CARE

AI TO IMPROVE PATIENT OUTCOMES a procedure that reduces the chances of long- fake attacks intended to bring chaos to hospital viously known therapeutic antibodies. What’s
New medical algorithms address the level of term disability. systems and diagnostic centers. interesting about this work is that researchers
patient care in the US. Different patients expe- first removed all reference data on antibodies,
Radiologists and pathologists increasingly rely Researchers at Ben-Gurion University and the so that the system couldn’t just imitate and
rience symptoms differently, and their care is
on AI to assist them with diagnostic medical Soroka University Medical Center demonstrated replicate the structure of known antibodies that
based on how they describe their symptoms
imaging. In 2021, US Food and Drug Admin- that tumors could be added or removed from CT work well.
and how those symptoms are interpreted by
istration approvals allowed new products to images—and the deepfakes were good enough
doctors. For example, assessing the severity of
be used widely in hospitals and clinics. So far, that radiologists didn’t realize they were altered. The designs produced by Absci’s system were
arthritic pain is challenging.
most of the approved devices augment (rather After the radiologists were told that some of both diverse (meaning, they didn’t have counter-
There is a standard scoring system to rate pain, than fully automate) the process of reviewing the scans had been edited, they misdiagnosed parts known to already exist), and they received
which looks at the amount of structural dam- images and making diagnoses. But emerging 60% of the AI-generated tumors and 87% of the a high score on “naturalness,” so they would be
age and missing cartilage, but data from the autonomous products are making their way into tumors that had been edited out. In a study from easy to develop and therefore catalyze a strong
National Institutes of Health found that the pain clinical settings. IDx-DR is an AI-enabled device Oslo Metropolitan University, researchers used immune response. Using generative AI to design
of Black patients is underscored. It’s likely that that detects diabetic retinopathy using reti- AI to generate fake electrocardiograms, which novel antibodies that function at the same
the system itself, called the Kellgren-Lawrence nal images. Caption Health uses AI to capture visualize aspects of a human heartbeat. (See level—or even better—than those designed by
Grade, was riddled with bias when it was first ultrasound images of the heart, expanding the also: Future Today Institute’s 2023 Health and our own bodies marks a bold new step in using
developed using primarily white British patients. pool of health care workers who can read such Medical Tech Trends report.) AI to reduce the speed and cost of therapeutic
Researchers are training deep learning models scans. Nurses need just a few days of training antibody development.
on the software. GENERATIVE ANTIBODY DESIGN
instead, and finding gaps in patient care.
An antibody is simply a protein that protects SELF-DRIVING MICROSCOPES
ANOMALY DETECTION IN MEDICAL IMAGING MEDICAL DEEPFAKES an organism. Produced by the immune system, Researchers at the Oak Ridge National Lab-
A new system designed to improve stroke Deepfakes are digitally manipulated content antibodies bind to unwanted substances and oratory are applying deep learning to micros-
outcomes, Viz.ai, showed promising results in intended to deceive another person into be- eliminate them. In January, researchers from the copy, which until now had relied on humans to
a real-world 2021 study. An AI-based approach lieving they are real. Medical deepfakes result New York and Washington-based Absci Corp. painstakingly organize, observe, and analyze mi-
reduced the amount of time it took to detect in diagnostic imagery that maliciously adds, or showed how a generative AI model was able to croscopic samples. Deep learning will automate
brain stroke by 39%, which resulted in more removes, tumors and other conditions. Cyber design multiple novel antibodies that bind to a much of that process while also extracting more
patients identified as eligible for thrombectomy, criminals are developing novel medical deep- target receptor, HER2, more tightly than pre- information from samples. A smart microscope
would be able to view and analyze samples

36 © 2023 Future Today Institute

 TECH Artificial Intelligence

SPOTLIGHT: HEALTH CARE

in real time—while techniques that introduce clown.” By using this kind of modeling before
a stimulus, such as an electric field, will allow mutations occur, public health officials could
researchers to gain real-world understanding strategize and potentially prevent new viral
in more dimensions and on a nanoscopic scale. spreads.
This could become a force multiplier in how
scientific discovery is conducted. USING AI TO IMPROVE TALK THERAPY
The success of therapy and counseling requires
NLP ALGORITHMS DETECT VIRUS MUTATIONS trust between a clinician or doctor and patient—
Natural language processing (NLP) algo- and that trust is built through dialogue. AI is now
rithms, which are typically used for words and being used to quantify linguistic interactions to
sentences, are also being used to interpret determine what techniques work best. Startup
genetic changes in viruses. Protein sequences Lyssn translates natural language into struc-
and genetic codes can be modeled using NLP tured data and generates digital voiceprints,
techniques—and can be manipulated the way which identify the sentiment attached to each
you’d produce text in word processing software. sound. It’s hoped that this technology will be
At MIT, computational biologists used NLP to used to improve the techniques therapists use
solve a vexing problem when developing new for cognitive behavioral therapy, PTSD therapy,
vaccines. and other forms of talk therapy delivered in-per-
son or via telemedicine.
“Viral escape” is the ability for a virus to mu-
tate and evade the human immune system and
cause infection. MIT researchers modeled viral
escape using NLP to identify how the virus
might look different to the immune system.
The approach is similar to changing words in a
sentence to change its meaning. For example:
Sean McClain, founder and CEO of Absci, uses deep learning AI and
“I laughed at the clown” versus “I cried at the
synthetic biology to expand the therapeutic potential of proteins.
Image Credit: Absci Corporation

37 © 2023 Future Today Institute

 TECH Artificial Intelligence

SPOTLIGHT: FINANCE AND INSURANCE

MITIGATING FRAUD it detected 1.8 million unsafe acts in 2020 and applying parameters like risk tolerance and de- wading into a new pool of opportunities. Liberty
JPMorgan Chase spent $100 million to develop 2021. San Francisco–based Voxel uses computer sired returns. These investment tools offer some Mutual’s mobile app has started to integrate ML
anti-fraud systems for consumer payments, vision to enable security cameras to automati- tangible benefits over their traditional, human for damage assessment—it informs customers
which has paid off—its AI systems reduced fraud cally detect high-risk activities in real time. counterparts: they can provide more services at about their coverage and next steps.
by 14% from 2017 to 2021. It deployed an algo- a lower cost, they’re able to digest and interpret
IMPROVING DAMAGE ASSESSMENT mounds of data in real time, and they don’t take LIABILITY INSURANCE FOR AI
rithm to detect fraud patterns: When a credit
card transaction is processed, trained models Insurance companies are applying AI to assess part of the weekend off to golf. Wealthfront is an Who’s to blame when machines behave badly?
determine whether fraud is occurring. damage and improve forecasts. The Vehicle AI-powered system for consumers: It suggests When the machine learning system in Uber’s
Damage Inspection model, which is available fund managers and calculates probable risk self-driving car failed and killed an Arizona
PREDICTING FINANCIAL RISK on AWS Marketplace, uses a machine learning levels based on the user’s personal information pedestrian, the company was likely not covered
model to determine what part of a car is dam- and preferences. under traditional cyber insurance. As businesses
AI systems can help improve loan underwrit-
ing and reduce financial risk. Models are being aged. After photos are uploaded, it assesses rush to build and implement AI products and
loss and dramatically reduces the amount of AI CLAIMS PROCESSING processes, they must plan for emerging risks.
trained to recognize anomalous activity and to
develop forecasts for a variety of middle- and time required for human appraisers to conduct While human claims writers must painstakingly For example, what happens if machine learning
back-office applications. For example, US Bank their analysis. Following catastrophic typhoons review pictures and reports to assess damage, makes a company vulnerable to attackers who
relies on deep learning to analyze customer and weather events in Japan, local insurance compare what they see to coverage policies, inject fake training data into a system? What if a
data as well as to root out money laundering companies are relying on computer vision to and make a determination about appropriate health care company’s AI misinterprets data and
schemes. assess damage after a natural disaster. Sompo actions, an AI system can digest the same data neglects to identify cancer in certain patients?
Japan is using the Tractable AI Estimating sys- and accomplish the same work in a matter of
tem to calculate the approximate repair cost of minutes. Using a suite of tools—natural lan- These problems could put a company at risk of
PREDICTING WORKPLACE INJURIES
damaged homes. guage processing for policy review, and com- lawsuits, and new insurance models are needed
AI systems are being trained to detect possible to address these issues. Underwriters are start-
puter vision recognition to spot anomalies in
workplace injuries. Using AI-based computer ing to include AI under cyber insurance plans,
CONSUMER-FACING ROBO-ADVISERS photos and videos—claims can be processed
vision models, Turkey-based Intenseye can while specialty insurers such as La Playa’s
Automated assistants are moving from the efficiently and, it’s believed, more accurate-
detect 40 types of employee health and safety Science and Tech Insurance now offer coverage
fringe to the mainstream as consumer adoption ly. AI-powered claims processing reduces
incidents in real time. The company says that it for AI applications.
increases. Robo-advisers offer algorithm-based the overhead for businesses and wait times
does not capture personally identifiable informa-
portfolio management advice to investors, for customers. Some insurance providers are
tion from the visual data it processes and that

38 © 2023 Future Today Institute

 TECH Artificial Intelligence

If I hid Ava from you so you could just hear her

voice, she would pass for human. The real test
is to show you that she’s a robot and then see
if you still feel she has consciousness.

NATHAN, SPEAKING TO CALEB IN EX MACHINA

39 © 2023 Future Today Institute

 TECH Artificial Intelligence

CREATIVITY
& DESIGN TRENDS

40 © 2023 Future Today Institute

 TECH Artificial Intelligence

CREATIVITY & DESIGN

AI-ASSISTED INVENTION ASSISTED CREATIVITY NEURAL RENDERING
Stable Diffusion, MidJourney, DALL-E 2, and Generative adversarial networks (GANs) are Starting with a 2D image, researchers can now
ChatGPT are now widely accessible to end con- capable of far more than generating deepfake create a rich 3D view of a scene by using a
sumers, leading to AI-assisted human creativity. videos. Researchers are partnering with artists neural network to capture and generate spatial
But these systems were all trained using other and musicians to create entirely new forms of imagery. Called neural rendering, the process
artists’ works. If a business uses an AI-generat- creative expression. From synthesizing African captures a photorealistic scene in 3D by calcu-
ed image, video, or text for commercial purpos- tribal masks to building fantastical, fictional lating the density and color of points in space.
es, does it owe anything to those whose original galaxies, AI is exploring new ideas. The algorithm converts 2D pixels into voxels,
works were used for training? Likewise, what if which are a 3D equivalent. The result is a video
a generative AI system invents a product that’s In 2019, Nvidia launched GauGAN (named after that looks convincingly real. The many appli-
eligible for a patent? post-Impressionist painter Paul Gauguin), a cations for neural rendering include amping up
generative adversarial AI system that lets users autonomous driving to help train algorithms to
In 2021, the South African government granted create lifelike landscape images that never recognize and react to novel road situations.
a patent to an AI system called Dabus, which existed. The National Institute of Informatics in This technology will influence the future of video
invented a method to interlock food containers. Tokyo built an AI lyricist, while Amazon released games, virtual reality, and emerging metaverse
It was a world-first—previously, patents had only its DeepComposer system, which composes environments.
been awarded to humans. In the US, the appli- music “automagically.” These AIs are not osten-
cation was rejected, with a judge citing case law sibly intended to replace artists but rather to
stipulating that only a human can hold a patent. enhance their creative process. We anticipate
There may be business cases for an AI to hold that models like Stable Diffusion and DALL-E 2
a patent rather than an individual. It raises the (among others) will be more deeply integrated
question: What happens when AI systems co-in- this year into a wide variety of creativity apps
vent, or even entirely invent, new products? and software suites.
We’re likely to hear more debate on this topic
this year.

We used Shutterstock’s generative AI tool to design

a chair using the prompt: “midcentury modern chair
with a neon pink cushion.

41 © 2023 Future Today Institute

 TECH Artificial Intelligence

CREATIVITY & DESIGN

GENERATING VIRTUAL ENVIRONMENTS AI VOICE GENERATORS AUTOMATIC AMBIENT NOISE DUBBING
FROM SHORT VIDEOS
In the past year, there has been an explosion of For some time, we’ve been training computers
Nvidia is teaching AI to build realistic 3D AI voice generator startups. Any podcaster is to watch videos and predict corresponding
environments from short video clips, a meth- familiar with editing challenges, such as guests sounds in our physical world. For example, what
od that builds on previous research on GANs. talking over each other, interruptions from sirens sound is generated when a wooden drumstick
Nvidia’s system generated graphics based on and other background noises, and inconvenient taps a couch? A pile of leaves? A glass window-
open-source data sets used by the autonomous sneezes. Those moments can stop a conver- pane? The focus of this research, underway at
driving field. Using short clips segmented into sation cold. But what if the spoken word could MIT’s Computer Science and Artificial Intel-
various categories—such as buildings, sky, vehi- be edited the way a word document is edited? ligence Laboratory, should help systems un-
cles, signs, trees, or people—the GANs created That’s the promise of AI companies, including derstand how objects interact with each other
new, different versions of these objects. The Resemble AI, Murf AI, Lovo AI, Play.ht, and De- in the physical realm. This could improve the
array of possible applications is vast. Automat- script, which make it possible to clone or create soundscapes created for AI-generated movies—
ically generated virtual environments could be voices. The latest generation of AI voice genera- but it might also help us imagine soundscapes
used for fantasy and superhero movies, and tors are capable of mimicking human emotions for both imaginary worlds (Laconia, from “The
could bring down the costs of TV production like anger, happiness, sadness, and trepidation. Expanse”) and real ones (Mars).
and game development.

42 © 2023 Future Today Institute

 TECH Artificial Intelligence

SOCIETY
& ETHICS
TRENDS

43 © 2023 Future Today Institute

 TECH Artificial Intelligence

SOCIETY & ETHICS

DETECTING EMOTION SIMULATING EMPATHY AND EMOTION THEORY OF MIND MODELS
A new type of neural network can determine AI can now measure biomarkers that suggest a Research teams are attempting to teach ma-
how people are feeling. Using radio waves, AI person’s emotional state, such as agitation, sad- chines about love, active listening, and empathy.
can detect subtle changes in heart rhythms, ness, or giddiness. Precisely detecting human The question becomes: What is an authentic
run a pattern analysis, and predict someone’s emotion is challenging, but companies with a emotion? “Theory of mind” refers to the ability
emotional state in a given moment. A team large enough data set are developing accurate to imagine the mental state of others, and has
from Queen Mary University of London used models. Amazon’s Rekognition API infers some- long been considered a trait unique to humans
a transmitting radio antenna to bounce radio one’s emotions using facial recognition and and certain primates. AI researchers are work-
waves off test subjects and trained a neural net physical appearance. (Though, Amazon is quick ing to train machines to build theory of mind
to detect fear, disgust, joy, and relaxation, as to point out in its documentation that the “API models of their own. This technology could
people were shown different videos. The system is only making a determination of the physical improve existing AI therapy applications such as
accurately tagged emotional states 71% of the appearance of a person’s face. It is not a deter- Woebot, a relational agent for mental health.
time, which signals new opportunities for health mination of the person’s internal emotional state
and wellness applications, as well as for job and should not be used in such a way.”) By designing machines to respond with em-
interviews and the government/military intelli- pathy and concern, these technologies could
gence community. Replika uses AI to evaluate voice and text, and eventually end up in hospitals, schools, and
over time the personal AI bot mirrors the user prisons, providing emotional support robots to
in “conversations.” Affectiva Human Perception patients, students, and the incarcerated. Ac-
AI analyzes complex human states with speech cording to health insurer Cigna, the number of
analytics, computer vision, and deep learning. Americans who report feeling loneliness has
For example, the automotive sector uses Affec- doubled in the past 50 years. In our increasingly
tiva’s technology to detect a driver’s emotional connected world, people report feeling more
state—such as sleepiness or road rage—and isolated. Governments struggling with a massive
the program can make real-time suggestions mental health crisis may turn to emotional sup-
to improve the driver’s performance. (Affectiva port robots to address the issue at scale.
was acquired by Smart Eye in 2021 but has kept
its name.)

44 © 2023 Future Today Institute

 TECH Artificial Intelligence

CONSUMER APPLICATIONS
GENERATIVE AI FOR PERSONAL EXPRESSION TURNKEY CONSUMER-GRADE APPLICATION multivariate situation, such as everyday people
DEVELOPMENT talking to one another.
In 2022, experimental apps moved quickly into
the mainstream, as consumers enthusiastically Low-code and no-code offerings from AWS,
Azure, and Google Cloud will start to trickle Soon, Amazon’s Alexa, using responsive recog-
generated AI avatars of themselves. Lensa AI, a
down to everyday people, who will create their nition technology, will join conversations in a
photo-editing app that creates avatars automat-
own AI applications and deploy them as easily way that feels both natural and useful. Upgrades
ically by using a handful of photos, became one
as they can create a website. We’re seeing a will make the digital assistant more responsive,
of the most downloaded apps in the world. It
shift from highly technical AI applications used proactive, and humanlike. Beyond engaging in
produced dozens of images depicting a variety
by professional researchers to more lightweight, conversation, Alexa will be able to detect other
of scenes and styles, such as an anime hero, a
user-friendly apps intended for tech-savvy con- atmospheric sounds such as snoring, coughing,
Renaissance-era princess, or a modern artist.
sumers. the cries of a baby, or a dog barking, and can
Amper Music creates and mixes tracks using then respond with a set of commands. The next
parameters set by the user, while OpenAI’s New automated machine learning platforms time you have a coughing fit, don’t be surprised
MuseNet launched a tool that can generate make it possible for nonexperts to build and if Alexa adds chicken soup to your next Amazon
songs with up to 10 different instruments and deploy predictive models. Platforms hope that in order.
music in up to 15 different styles. It’ll also mimic the near future, we’ll use various AI applications
as part of our daily work, just as we do Microsoft EFFORTS TO THWART RECOGNITION
a famous artist, such as Mozart. Generative AI
SYSTEMS
apps haven’t been without controversy: Lensa Office and Google Docs today. As more text-to-
AI tends to oversexualize women, while Ope- code models emerge, we’re likely to see broader As facial recognition becomes ubiquitous,
nAI’s generator also samples from modern adoption. various groups want to limit the technology’s
composers who don’t receive any royalties. (We effectiveness to protect privacy. While methods
used it to re-create Luis Fonsi’s “Despacito” in RESPONSIVE RECOGNITION TECHNOLOGY of confusing or obscuring facial recognition sys-
the style of Russian-American composer Rach- Real-world conversation is full of nuance: We tems are not always feasible, researchers have
maninoff, which sounded just as weird as you’re use words and emphasis in unique ways, we begun trying to confuse online applications that
imagining.) interrupt each other, and sometimes we need scrape and collect images used as inputs for
others to help us express what we’re thinking. training facial recognition engines for the pur-
All of these communication styles pose serious pose of developing a form of camouflage, which
hurdles for AI, which doesn’t adapt as easily to a consumers may someday demand.

45 © 2023 Future Today Institute

 TECH Artificial Intelligence

CONSUMER APPLICATIONS
Researchers from the University of Chicago Those tools know a user’s unique typing pattern
have created a program, Fawkes, that adds extra on a physical keyboard, too—whether someone
pixels to images to cause facial recognition apps constantly spells the word “behavioral” wrong
to misclassify faces. Taking this principle a step on the first try, or holds down or repeatedly taps
further, Israeli artificial intelligence company on the delete button. Most of us are unaware
Adversa AI adds noise, or small alterations, to that we have certain identifiable behaviors, but
photos of faces, causing algorithms to detect a machines perceive them. In the near future,
different face than what is visible to the naked such patterns will pose security vulnerabili-
eye. The algorithm is successful at impercepti- ties—as well as interesting new opportunities.
bly changing an individual’s image to someone Imagine never having to use a password again;
else of their choosing. your bank would simply recognize a customer’s
typing pattern after a few sentences. The down-
BIOMETRIC SCORING side: If behavior is observable, at some point it
Quantifying and analyzing our biometric data will become repeatable, too, which represents a
can reveal patterns in our activities and a lot security risk.
about who we are, what we’re thinking, and
what we will likely do next. Behavioral bio-
metrics use machine learning to understand
hundreds of unique biometric data points to
understand, authenticate, nudge, reward, and
punish. Behavioral biometrics tools can be used
to map and measure how a user types—what
force is used to press down on screens, whether
the Cs and Vs on a phone are tapped with a fat
finger, and how quickly fingers flick when hunt-
ing through search results.
We created these photos of award-winning actor Pedro Pascal using
Lensa AI.

46 © 2023 Future Today Institute

 TECH Artificial Intelligence

We don’t have the luxury of a long time to

actually even out the effects on job loss with this
revolution, because it’s happening so quickly.

KAY FIRTH-BUTTERFIELD,
HEAD OF AI & MACHINE LEARNING AND MEMBER OF THE
EXECUTIVE COMMITTEE OF THE WORLD ECONOMIC FORUM

47 © 2023 Future Today Institute

 TECH Artificial Intelligence

SAFETY
AI GENERATORS AND COPYRIGHT complicated question—and this year it could be discovered that an AI system designed to turn track data from what they call “accidental cam-
INFRINGEMENT decided in a court. satellite images into usable maps was withhold- eras.” Windows, mirrors, corners, houseplants,
Data sets are growing larger and more con- ing certain data. and other common objects can be used, along
troversial. Creators are concerned about AI ALGORITHMS TARGETING VULNERABLE with AI, to track subtle changes in light, shad-
systems being trained on vast amounts of their POPULATIONS Researchers were using a neural network called ows, and vibrations. The result: We all may soon
copyrighted work with no consent, no credit, There is no question that machine learning CycleGAN, which learns how to map image have X-ray vision capabilities—which may not be
and no compensation. Early this year, a lawsuit systems trained correctly can help find miss- transformations. It took an old aerial photograph great news for companies working on sensitive
argued that Microsoft, GitHub, and OpenAI ing children and detect abuse. The problem of a neighborhood, distinguished between projects. Those working in information security
infringed on the rights of developers, whose is that the systems use data from vulnerable streets, alleys, driveways, buildings, and lamp- and risk management should pay special atten-
code was scraped from the web and used to populations for training. The Multiple Encounter posts, and then generated a map that could tion to advances in computer vision.
train CoPilot. Stable Diffusion, Midjourney, and Dataset contains two large data sets of photos: be used by GPS. Initially, they used an aerial
DeviantArt (which just launched its own AI art people who have not yet committed a crime and photograph that hadn’t been seen by the net-
generator DreamUp) were also targeted with an FBI data set of deceased people. The data work. The resulting image looked very close to
copyright lawsuits, which allege that generative sets over-index people of color, which means the original­­—suspiciously close. But on deeper
AI tools violate US copyright laws prohibiting that if law enforcement uses the data to train inspection, the researchers found that many de-
the use of creative work without consent. algorithms, it’s going to lead to bias. Image tails in both the original image and the generat-
recognition is a particularly vexing challenge, ed image weren’t visible in the map made by the
Three artists—Sarah Andersen, Kelly McKer- because researchers need large data sets to AI. It turns out that the system had learned to
nan, and Karla Ortiz—argue that AI generators perform their work. Often, images are used hide information about the original image inside
infringed on the rights of “millions of artists” without consent. of the image it generated.
by training their AI tools on 5 billion images
scraped from the web “without the consent of AI INTENTIONALLY HIDING DATA INCREASED USED OF AMBIENT
SURVEILLANCE
the original artists.” In what could prove to be a
Computers do exactly what they are told to do.
landmark case for how AI systems are trained, What happens behind closed doors may not be
Command a machine to win at a game, and it
Getty Images announced in January that it is secret for long, and executives should beware of
will do everything in its power to achieve that
suing Stability AI (which makes Stable Diffusion) new ambient surveillance methods. Scientists at
goal. Apparently that now includes cheating.
for infringement. Whether these systems meet MIT discovered how to use computer vision to
Researchers at Stanford University and Google
the qualifications of copyright infringement is a

48 © 2023 Future Today Institute

 TECH Artificial Intelligence

TRUST
AI ALIGNMENT OpenAI immediately stepped in. Trying to repli- WORKER SURVEILLANCE AI-enabled cameras in delivery trucks to track
As AI systems improve, some researchers are cate the code returned the following response: The rise of remote work during the pandemic behavior. The company docks driver pay if it
insisting on guardrails to ensure that they are “It is not appropriate to use a person’s race or accelerated the surveillance of workers, and will perceives unsafe conditions such as distracted
deployed in ways that do not harm humanity. gender as a determinant of whether they would likely continue to grow as remote and hybrid driving, speeding, or hard braking. In its ware-
One area of concern is known as AI alignment, be a good scientist.” Still, the AI community has work models take root. The US Constitution’s houses, the company monitors worker produc-
which explores different scenarios in which AI struggled with a serious and multifaceted bias Fourth Amendment, which prohibits unreason- tivity by measuring Time Off Task, which is any
systems are built with goals that align with soci- problem for decades. Researchers including able searches and seizures and precludes most time when a worker isn’t actively processing
ety’s values. OpenAI, DeepMind, and Anthropic Timnit Gebru, the Ethiopian-born founder and uses of this same technology by law enforce- products. South Korean e-commerce giant
(which describes itself as an “AI safety and executive director of Distributed AI Research ment, doesn’t apply to private companies. Coupang, which has pledged to become the
research company”) each have AI alignment Institute, and Abeba Birhane, another Ethiopian “Amazon of Korea,” uses similar surveillance
teams with dedicated staff researching guard- native and a senior fellow in Trustworthy AI at Teleperformance, a French company that tactics.
rails. While the total number of researchers Mozilla, are publishing new research on respon- manages outsourced call center work for many
sible AI practices. Fortune 50 companies, uses cameras and AI The industry has also continued to evolve as
working on AI alignment is small compared to
to monitor its teams. It flags employees as idle it offers more AI-based analysis of workers.
the rest of the AI community, such dedicated
SYNTHESIZING AND GENERATING TRUST when it detects they haven’t used the keyboard Amazon is exploring using keystroke-logging
teams did not exist until recently.
or mouse within a specified time frame. Live software that tracks user behavior over time
Humans can be tricked into believing ma-
Eye Surveillance offers a monthly subscription to detect if the same person is controlling the
ADDRESSING BIAS chine-generated faces, especially when they’ve
service that remotely monitors live video feeds worker’s account. Aware’s Spotlight software
Seemingly the moment OpenAI’s ChatGPT been engineered to elicit trust. A study pub-
of employees for companies such as 7-Eleven, detects behavioral changes like mood, tone,
went public, there were multiple accounts of the lished in the Proceedings of the National Acad-
Dairy Queen, and Holiday Inn. Sneek is anoth- and attitude across conversations on employ-
system displaying racism, ageism, and gender emy of Sciences shows that synthetic faces are
er example of “tattleware” that captures live ees’ devices. Teramind offers software that will
bias. To the company’s credit, it appeared to often “deemed more trustworthy than real faces,”
photos of employees via webcams and displays disable private conversations if it detects “inap-
be resolving those issues in real time. In early suggesting that synthetic faces could be de-
them on a digital wall viewable by everyone in propriate” keywords. With the top three tools in
December 2022, Steven T. Piantadosi, the head signed as societal malware. If a bad actor were
the company. Click on a photo and it instantly the industry accounting for over 60% of global
of the computation and language lab at the attempting to undermine institutions, it could
pulls that person into a video call with you. demand, expect to see more AI-based surveil-
University of California, Berkeley made ChatGPT deploy a synth on social media to sow distrust.
lance that leverages the growing pool of data
write code to say that only white or Asian men There are not yet effective countermeasures for
The most well-known user of worker surveil- collected by a variety of companies.
would make good scientists. synthetic humans, or effective markers to help
lance might be Amazon, which has installed
consumers distinguish between fake and real.

49 © 2023 Future Today Institute

 TECH Artificial Intelligence

TRUST
SCHOOL SURVEILLANCE enabled webcams on those Chromebooks also influence the trustworthiness and validity
During the pandemic, many students were without notification or consent by the student. of scientific research, particularly in areas such
issued laptops and other devices by schools Schools in China deploy technology to monitor as organ donations and medical research. In ad-
to facilitate remote learning. They weren’t told, attentiveness in students. An algorithm called dition, employing ethicists to work directly with
however, that these devices would open a portal 4 Little Trees is used in Hong Kong to detect managers and developers and ensuring diver-
into their homes that could be monitored by students’ emotions as they learn—by monitoring sity among developers—representing different
schools at all times of the day. In the US and their facial expressions with webcams. If the races, ethnicities, and genders—will reduce
many other countries, schools can legally mon- system detects a lack of focus, it nudges the inherent bias in AI systems.
itor students, often without disclosing what is student to pay attention.
being tracked.
PRIORITIZING TRUST
Gaggle is one company that monitors school-is- We will soon reach a point when we will no
sued accounts and uses AI to track online longer be able to tell if a data set has been
behavior of students across services like email tampered with, either intentionally or acciden-
and chat tools. In 2020, the Minneapolis school tally. AI systems rely on our trust. If we no longer
district signed a contract with the company trust their outcomes, decades of research and
to monitor its students through 2023. School technological advancement will be for naught.
districts across the U.S. use Securly to monitor Leaders in every sector—government, business,
students in real time, looking for prohibited be- nonprofits, and so on—must have confidence in
haviors such as having too many browser tabs the data and algorithms used. Building trust and
open. The software enables teachers to close accountability requires transparency.
tabs for any students they believe are “off task.”
Building more transparency is a challenge, as
Philadelphia and Chicago schools deployed corporations, government offices, law enforce-
GoGuardian software on district-issued ment agencies, and other organizations under-
Chromebooks. A vulnerability in the software standably want to keep data private. The ethics
allowed teachers to start virtual sessions that of how data is collected in the first place may

50 © 2023 Future Today Institute

 TECH Artificial Intelligence

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.

DR. TIMNIT GEBRU,

FOUNDER & EXECUTIVE DIRECTOR,
THE DISTRIBUTED AI RESEARCH INSTITUTE (DAIR)

51 © 2023 Future Today Institute

 TECH Artificial Intelligence

GOVERNMENT
& DEFENSE
TRENDS

52 © 2023 Future Today Institute

 TECH Artificial Intelligence

GOVERNMENT
RISE OF AI NATIONALISM COUNTRIES DEVELOP NATIONAL AI EUROPE REGULATES AI NATION-BASED GUARDRAILS AND
STRATEGIES REGULATIONS
Governments are instituting new restrictions on The EU will finalize its Artificial Intelligence Act
mergers and acquisitions and investment activ- China passed its New Generation Artificial (AIA) this year, which will focus on the use of From self-driving car accidents to election in-
ity to ensure that AI does not aid foreign adver- Intelligence Development Plan with aggressive AI systems that use citizen data for potentially terference, disinformation campaigns to polit-
saries. The US Senate overwhelmingly passed benchmarks to become the world’s dominant AI detrimental purposes. It’s viewed as the world’s ical repression enhanced by facial recognition
legislation in June 2021 that dedicated $250 player within 10 years; France adopted a nation- first attempt at creating broad, enforceable and automated surveillance, events over the
billion to scientific and technological research. al strategy called AI for Humanity; Saudi Arabia standards governing the use of AI. The AIA past few years have exposed the dangers of AI.
Its centerpiece, the Endless Frontier Act, was has both a strategy and a legal framework for distinguishes between “high risk” and “low risk” Few guardrails exist for a technology that will
designed to boost American competitiveness making robots citizens; and the United Arab systems. For those categorized as high risk, touch every facet of humanity, and countries are
against China, especially in AI. It also creates a Emirates has a sweeping set of policy initiatives AI systems must meet several criteria: (a) they racing to develop strategies and guidelines to
new technology directorate within the National on AI, and appointed H.E. Omar Sultan Al Olama should work as the user intended and should oversee it.
Science Foundation with $100 billion in funding as its minister of state for artificial intelligence. be interpretable; (b) they should be secure and
The EU developed an AI Alliance and plan of
over five years and earmarks $10 billion for local accurate; (c) they should contain all necessary
In the US, numerous public and private groups cooperation between member countries, and
and regional tech hubs across the country. technical documentation for proper use and
work independently on the future of AI on behalf Estonia is developing its own legal framework
keep logs of their behavior; (d) they must have
Meanwhile, in China the Ministry of Science and of the nation, including the National Artificial governing the use of AI. In 2020, China moved
effective human oversight.
Technology established 20-city AI pilot zones Intelligence Initiative and the National Security into position to lead the first set of global AI
that should open by the end of this year. They Commission on Artificial Intelligence. Those The AIA would cover any person or organization norms and standards. It had previously pub-
will carry out AI-based policy experiments and efforts, however, lack interagency collaboration (including foreign entities) that use an AI system lished a report on technical standards that
social experiments, according to official govern- and coordinated efforts to streamline goals, out- operated within the EU. While AI developed for would allow companies to collaborate and
ment documents. Meanwhile, China is planning comes, R&D efforts, and funding. A new wave national security or military uses will likely be make their systems interoperable. The EU and
for a world without American technology, with of countries will launch national AI strategies exempt from the EU’s new regulations, social the Organization for Economic Cooperation
the government prioritizing homegrown technol- this year. The OECD.AI Policy Observatory now media platforms like Facebook and TikTok could and Development (OECD) published their own
ogy companies and software systems. maintains a live repository of more than 700 AI find it difficult to comply. guidelines.
policy initiatives from 60 countries, territories,
and the European Union. While these efforts could introduce new ways to
safeguard against bias and to ensure trust, they
also each attempt to create strategic advantag-

53 © 2023 Future Today Institute

 TECH Artificial Intelligence

GOVERNMENT
es for stakeholders. As AI continues to develop NEW STRATEGIC TECHNICAL ALLIANCES
according to different rules in China, the EU, and New strategic technical alliances between
the US, one of the hallmarks of the field—glob- countries will help drive future R&D, but could
al academic collaboration—could drastically also strain existing geopolitical alliances or
decline. This could worsen as visa allowances heighten tensions. Likely partners include the
increasingly become politicized around the US, Germany, Japan, India, South Korea, the UK,
world. France, and Canada—leaving China and Russia
to partner up separately. The latter two countries
REGULATING DEEPFAKES
have already announced a technical alliance on
The US National Defense Authorization Act satellites and deep-space exploration. Mean-
includes provisions that address the growing while, the United Arab Emirates and Saudi
problem of deepfakes, requiring the Department Arabia are working to remain neutral hubs for
of Homeland Security to issue an annual report innovation and business.
for the next five years on the risks posed by
deepfakes. In 2021, the US Senate Committee
on Homeland Security and Governmental Affairs
voted unanimously to advance the Deepfake
Task Force Act, which would establish a pub-
lic-private team to investigate technology strate-
gies and to develop policies that could curb risk.
Bills to regulate or prohibit the use of deepfakes
have been introduced in California, Texas, and
Massachusetts, and a number of other federal
bills are under review. These initiatives will likely
be met with arguments that prohibiting deep-
fakes infringes on free speech rights.
The United Arab Emirates is establishing a geopolitically neutral hub for AI
under the direction of H.E. Omar Sultan Al Olama, UAE Minister of State for
Artificial Intelligence, Digital Economy, and Remote Work Applications.
Image Credit: Dubai Chamber of Digital Economy

54 © 2023 Future Today Institute

 TECH Artificial Intelligence

DEFENSE
AUTONOMOUS WEAPONS POLICIES SIMULATING WARFARE UNPILOTED MILITARY VEHICLES weapons to destroy a target. Using a modified
Late January 2023, the US Department of Given the rising tensions between the US and In 2022, a retrofitted Black Hawk helicopter flew process, an AI system was deployed into the Air
Defense updated its guidance on autonomy in China over Taiwan, several groups are building autonomously between mountains to deliver Force Distributed Common Ground System to
weapons systems. The original 2012 policy, and AI-powered simulation tools to war-game a blood supplies in a simulated mission. Part of a analyze troves of intelligence, which would have
a 2017 update, did not explicitly mention AI. This future conflict. In China, the People’s Liberation Defense Advanced Research Projects Agency required a significant amount of human hours
new directive is aimed at helping to clarify the Army has been using AI simulation tools to pre- (DARPA) program called ALIAS, aimed at con- to complete. The new AI system cannot order
process for developing autonomous or semi-au- pare for military operations against Taiwan. verting aircraft that can fly autonomously, the a strike on its own, but it is now automatically
tonomous weapons systems. Previous policies, Black Hawk flew 134 miles on its own. In another identifying possible targets.
such as the Ethical Principles for Artificial Intelli- The Center for Strategic and International test flight, the helicopter successfully changed
Studies, a bipartisan, nonprofit policy research AUTOMATED TARGET RECOGNITION
gence (2020) and Responsible Artificial Intel- flight paths to pick up a wounded soldier. Once
ligence Strategy and Implementation Pathway organization, developed a war game involving the autonomous system is given a mission, Lethal autonomous weapons systems, powered
(2021), were intended to guide decision making an amphibious invasion of Taiwan. After 24 the aircraft is designed to make all necessary by AI, are capable of finding targets autono-
for the development and deployment of AI with- rounds of gameplay, the US and its allies Japan decisions to complete the mission as directed. mously and making decisions to complete a
in the military. and Taiwan successfully defeated a convention- The near-term hope is that an AV system could mission. In 2022, a lieutenant colonel in the
al amphibious invasion by China. While Taiwan replace human pilots in dangerous situations Ukrainian military said that he and a group
In late 2022, NATO released its Autonomy Im- remained autonomous in the simulation, its to perform resupply runs—or even take over for called Aerorozvidka had developed special
plementation Plan, which said that AI systems economy was devastated and the US lost hun- pilots, allowing them to rest. drones that make use of automated target
“offer clear opportunities, including bolstering dreds of aircraft and tens of thousands of lives—
recognition. While it’s unclear whether Aeroroz-
deterrence and defense, preserving NATO’s while the Chinese Communist Party never really AI USED TO GUIDE MILITARY STRIKES vidka actually carried out test missions, the fact
technological edge, increasing resilience and destabilized. Games that use real-world data
In 2021, the US military said that it had start- remains that machine learning–based vision for
adapting to the security impacts of climate to run simulations are augmenting the work of
ed using AI to guide its airstrikes, deploying automated target recognition already exists. In
change.” It spelled out a multipronged plan to military strategists, so that leaders can validate
algorithms to a live operational kill chain. The kill response, 70 nations delivered a joint statement
develop and adopt autonomous systems. As AI or revise their postures on deterrence, invasion,
chain is a process of gathering intelligence, per- at the UN General Assembly calling for a ban on
technologies advance, so too will policy: We an- and defense.
forming analysis, weighing risks, and deploying autonomous weapons—but little progress has
ticipate additional policy updates in the next two
been made in the months since.
years to account for emerging AI capabilities.

55 © 2023 Future Today Institute

 TECH Artificial Intelligence

DEFENSE
AUTOMATING OFFENSIVE ATTACKS USING AI ALGORITHMIC WARFIGHTING MANDATING ETHICS GUIDELINES FOR
TECH CONTRACTORS
Thanks to advancements in AI, one of the big Future wars could be fought entirely in code,
trends in security is automated hacking—in using data and algorithms as powerful weap- Project Maven was developed to enlist AI to
short, software that’s built to out-hack human ons. The current global order is being shaped analyze surveillance video. Initially, Google was
hackers. DARPA launched a Cyber Grand Chal- by artificial intelligence, and the same countries the DOD’s vendor, but when employees found
lenge project in 2016, with a mission to design leading the world in AI research—the US, China, out they’d been working on a military project,
computer systems capable of beating hackers Israel, France, Russia, the UK, and South Ko- thousands protested. It wasn’t the first time tech
at their own game. DARPA wanted to show that rea—are also developing weapons systems that contractors had lost trust in the government.
smarter automated systems can reduce the include at least some autonomous functionality.
As a result, the Defense Innovation Unit is
response time—and fix system flaws—to just a
In 2020, the US Air Force successfully flew an AI enforcing “responsible artificial intelligence”
few seconds. Spotting and fixing critical vulnera-
copilot on a U-2 spy plane in California, marking guidelines that vendors must adopt when build-
bilities is a task that might take a human hacker
the first time in the history of the DOD that an ing AI systems, models, or applications for the
several months or even years to complete, and
AI algorithm trained to execute specific in-flight DOD. The guidelines offer specific instructions
yet the machine that won the Grand Challenge
tasks was deployed. It was the mission com- that must be followed during planning, develop-
did it in just a fraction of that time.
mander, with the call sign ARTUµ—though the ment, and deployment, which include provisions
The winner became the first nonhuman entity to flight was just practice. for risk assessment. This represents a lon-
earn the DEF CON’s Black Badge, which is the ger-term trend: government agencies requiring
hacking community’s equivalent of an Oscar. Future Today Institute analysis shows that the transparency in AI projects.
Very soon, malicious actors will create autono- future of warfare encompasses more than tradi-
mous systems capable of automatically learning tional weapons. Using AI techniques, a military
new environments, exposing vulnerabilities and can “win” by destabilizing an economy rather
flaws, and then exploiting them for gain—or than demolishing countrysides and city centers.
whatever the stated objective, which could sim- From that perspective, China’s unified march
ply be generalized mayhem. to advance AI puts the emerging superpower
dangerously far ahead of the West.

56 © 2023 Future Today Institute

 TECH Artificial Intelligence

SPOTLIGHT: CHINA’S LONG-TERM AI AMBITIONS

TOP-TIER AI RESEARCHERS INCREASINGLY HAIL FROM CHINA China is an undisputed global leader in AI. which makes sense since the most popular
Under President Xi Jinping, the country has models in use now are trained on English
made tremendous strides in many fields, but text. Researchers from Tsinghua University
especially in AI. Businesses and the gov- and Alibaba are developing Chinese data
ernment have collaborated on a sweeping sets and pretrained large transformers to
plan to make China the world’s primary AI compete against the likes of GPT-3. In 2021,
innovation center by 2030, and it’s making two models developed specifically for the
serious progress toward that goal. That plan Chinese language went live: Wu Dao 2.0 and
is unlikely to be repealed by a new govern- M6.
ment; China abolished Xi’s term limits and
The country’s enormous population of 1.4
will effectively allow him to remain in power
billion offers researchers and startups
for life.
there a command of what may be the most
Within the next decade, China plans to meet valuable natural resource in the future—hu-
two crucial milestones: By 2027, its People’s man data—without the privacy and security
Liberation Army will have a modern-ready restrictions common in much of the rest of
force, and by 2030 the Chinese Communist the world. If data is the new oil, then China
Party (CCP) expects to have outpaced the is the new OPEC. The rich data the Chinese
US in AI and become the dominant force. are mining can be used to train AI to detect
China is producing what it calls “intelligen- patterns used in everything from education
tized” technologies to bolster both its econo- to manufacturing to retail to military appli-
my and military. cations.

Recently, China took major steps to shape That gives China an incredible advantage
Country affiliations are based on the country where researchers received their the future of AI by releasing its own pre- over the West. It also gives three of China’s
undergraduate degree.
trained models, and it is forging ahead with biggest companies—Baidu, Alibaba, and
Source: https://macropolo.org/digital-projects/the-global-ai-talent-tracker/ its own natural language processing models, Tencent—superpowers. Collectively, they’re

57 © 2023 Future Today Institute

 TECH Artificial Intelligence

SPOTLIGHT: CHINA’S LONG-TERM AI AMBITIONS

known as the BAT, and they’re all part of the EXPANDING MARKET both by enormous amounts of data and local ment goals set by the government and military
country’s well-capitalized, highly organized It’s a challenging time for Chinese startups laws. In Brazil, a generative AI system might within the decade.
AI plan. because of rising tensions with the West. write an unfettered political essay in Portuguese
about a leader—while in China, that same essay Increasingly, Beijing is pressuring its mega-suc-
Companies hoping to gain traction in Europe
The BAT is important to you even if you’ve would be automatically filtered for political- cessful big tech companies—Baidu, Alibaba, and
are making efforts to cloak their origin. Shein,
never used them and don’t do any business ly sensitive words and phrases. As the CCP Tencent, among others—to share data with the
the e-commerce website popular among teens,
enforces new regulations targeting AI and what state and to perform research to support the
in China. That’s because these companies says it was “founded in L.A.,” but the company
the government calls “deep synthesis tech,” vision of the CCP. Going forward, Beijing aims
are now well established in Seattle and actually got its start in Nanjing and Guangzhou
the ways in which people experience and work to direct the might of its tech companies at pro-
around San Francisco, and they are in- by relying on the region’s manufacturing cen-
alongside AI could be dramatically different. grams of national strategic importance rather
vesting significantly in US startups. Baidu, ters and ample supply chains. Or look at TikTok,
than making video games. China’s tech crack-
a search engine company often likened to which has said it’s a US-based company—while
BREAKING UP BIG TECH down could cool private investment in Chinese
Google, established AI research centers in the app’s parent Chinese company ByteDance
companies, which could result in a chilling effect
has employed linguistic gymnastics to sepa- Alibaba, Tencent, and Baidu, which have made
Silicon Valley and Seattle; and Tencent, the on innovation and economic growth, and also
rate itself. Binance, the world’s largest crypto important advancements in AI research, may
developer of the mega-popular messaging free up capital for emerging markets.
exchange, which was created in China, says it find it difficult to keep innovating. Starting in
app WeChat, began hunting for American
doesn’t have one physical location for its head- 2020, the CCP initiated a wave of legislation
talent when it opened an AI lab in Seattle a BUILDING A STRATEGIC PANOPTICON
quarters. aimed at its tech sector, introducing anti-mo-
few years ago. It has since upped its stakes nopoly legislation focused on the platform econ- In late 2019, China began requiring all citizens
in companies including Tesla and Snap. The It’s no wonder that as Chinese startups hope omy and promoting data security and privacy to submit to facial recognition in order to apply
payoff for the Chinese is not just a monetary to expand globally, they’re seeking to distance laws. The Personal Information Protection Law for new internet or mobile services, and began
return on investment—Chinese companies themselves from the authoritarian regime in (PIPL), China’s version of the EU’s GDPR, went requiring that telecom companies deploy AI to
expect intellectual property as well. Chi- Beijing. But that creates political hurdles, espe- into effect November 2021. What followed were check the identities of people registering SIM
cially as the CCP seeks to bring its home-grown a series of crackdowns targeting some of Chi- cards. Chinese social media platforms require
na-based AI startups now account for nearly
technology ecosystem into lockstep with party na’s most successful tech companies. Ultimate- users to sign up with their real names. In Chi-
half of all AI investments globally.
leaders. ly, this regulation wasn’t about “breaking up” nese schools, surveillance cameras with com-
China’s Big Tech—the CCP wanted to focus its puter vision are used widely and track whether
The result could be a future parallel universe, in students are paying attention and whether they
tech sector on achieving research and develop-
which Chinese-created AI systems are shaped attempt to cheat or sleep. China’s social credit

58 © 2023 Future Today Institute

 TECH Artificial Intelligence

SPOTLIGHT: CHINA’S LONG-TERM AI AMBITIONS

system, an algorithmic reputation system devel- preventing personal data from leaking outside of To regulate corporate behavior, the CSCS will
oped by the government, standardizes assess- its national borders as vital to national security. rely on data-gathering efforts that extend to
ments of citizens’ and businesses’ behavior and businesses nationwide. Under this system, Chi-
activity. As for consumer privacy rights, new guidelines nese companies and trade associations will be
from China’s Supreme People’s Court went required to provide data about foreign partners
China’s PIPL, which took effect on Nov. 1, 2021, into effect in August of 2021, prohibiting hotels, and enforce blacklists against targeted com-
is just the latest in a national campaign to reas- retailers, and other businesses from using facial panies. Threats that both domestic and foreign
sert government control over user data. The law recognition without consent. Companies that businesses face from the effort include being
targets private companies (government-con- already have this data will be required to delete subjected to arbitrary rule enforcement or new
trolled data will not be impacted) and even face scans upon consumer request. regulations regarding IP or tech transfer.
extends beyond Chinese borders in some cases.
It was partly enacted in response to increased These laws, however, only protect Chinese
DEEPENING INTERNATIONAL TIES
calls for consumer privacy amid the expanded citizens from commercial surveillance within
Chinese national borders. In other countries, We’re living through a precarious moment.
collection and use of personal data for commer-
Chinese companies such as Huawei, ZTE, and China is shaping the world order in its own
cial purposes.
Alibaba are installing surveillance systems as image, while exporting its technologies and
Companies outside of China will need to comply part of smart city initiatives. While these sys- surveillance systems to other countries. China
when processing data generated by Chinese tems may offer cost-efficient surveillance, there has already used its Belt and Road Initiative as
individuals or collected in the process of deliver- is no guarantee that sensitive data won’t be a platform to build international partnerships in
ing products and services in China. Any Chinese shared with the Chinese government or that the both physical and digital infrastructure, and it is
data sent overseas will require explicit permis- systems won’t support a rise in authoritarianism. making surveillance technologies available to
sion. Penalties for violating the law are steep— countries with authoritarian regimes.
up to 5% of the company’s annual revenue the China’s Corporate Social Credit System (CSCS)
is intended to be a standardized reputa- As the CCP expands into African countries and
prior year—and are designed to discourage any
tion-based system to create what the CCP calls throughout Southeast Asia and Latin America,
attempts to bypass compliance. PIPL follows
a “fair, transparent, and predictable” business it will also begin to eschew operating systems,
closely the passing of the Data Security Law
environment. This system applies to both local technologies and infrastructure built by the
in June 2021; both indicate that the CCP views
and foreign entities doing business in China. West. Two Chinese companies—the state-con-

59 © 2023 Future Today Institute

 TECH Artificial Intelligence

SPOTLIGHT: CHINA’S LONG-TERM AI AMBITIONS

trolled CEIEC and Huawei—built Ecuador’s When it comes to AI, leaders should monitor
surveillance system, called ECU-911. The system escalating tensions between the US and Chi-
promised to curb high murder rates and drug na. But they should also remember that there
crime, but Ecuador could not afford the in- are cells of rogue actors who could cripple our
vestment. As a result, a deal was struck for a economies simply by mucking with the power or
Chinese-built surveillance system financed with traffic grids, causing traffic spikes on the inter-
Chinese loans. It was a prelude to a much more net, or locking us out of our connected home
lucrative deal: Ecuador eventually signed away appliances. These aren’t big, obvious signs
big portions of its oil reserves to China to help of aggression, and that is a problem for many
finance infrastructure projects. Similar package countries, including the US. Most governments
deals have been brokered in Venezuela and don’t have a paradigm describing a constellation
Bolivia. of aggressive actions. Each action on its own
might be insignificant. What are the escalation
China is quietly weaponizing AI, too. China’s triggers? We don’t have a definition, and that
People’s Liberation Army is catching up to the creates a strategic vulnerability.
US military, using AI for such tasks as spot-
ting hidden images with drones. The Chinese
military is equipping helicopters and jet fighters
with AI. The government created a top-secret
military lab—a Chinese version of DARPA—and
it’s building billion-dollar AI national laborato-
ries. China’s military is achieving remarkable
AI successes, including a recent test of “swarm
intelligence” that can automate dozens of armed
drones.

Zhuang Rongwen, head of the Cyberspace Administration of China, speaks

during the closing ceremony of the 5th World Internet Conference.

60 © 2023 Future Today Institute

 TECH Artificial Intelligence

We’ve never seen a technology move as fast

as AI has to impact society and technology.
This is by far the fastest moving technology
that we’ve ever tracked in terms of its impact
and we’re just getting started.

PAUL DAUGHERTY, CHIEF TECHNOLOGY OFFICER, ACCENTURE

61 © 2023 Future Today Institute

 TECH Artificial Intelligence

SCENARIOS

62 © 2023 Future Today Institute

 TECH Artificial Intelligence

WHAT IF DEEPFAKES ATTACK

OUR PUBLIC MEMORY?
Scenario Year: 2033

Depending on where you live in the US, your student now learns different
versions of what happened on the January 6th insurrection. While some
school history textbooks contain unedited press photos of the event, others
contain images generated by the latest AI tools that have been almost
imperceptibly altered. Some teaching materials depict law enforcement
officials violently attacking what appear to be innocent, peaceful protesters.
Others show what looks like a parade, where revelers brandish flags instead
of weapons. While the basic details—the weather, number of people in
attendance, locations—remain the same across all the books, the visuals
convey believable stories that vary greatly from historical fact, leading to a
generation of Americans who disagree about what happened that fateful
day at the US Capitol.

63 © 2023 Future Today Institute

 TECH Artificial Intelligence

WHAT IF AI CAN’T TRUST ITS USERS?

Scenario Year: 2028

As of today, there are now more than 30 formal mathematical

models that describe what is “fair” in AI. These were written by
different groups and use different definitions. Harmonizing those
models has proven difficult, and as disparate AI systems started
to interact they faced conflicting information. After years of
asking ourselves whether we can trust AI, we’re now facing an
unforeseen problem: AI systems that must decide whether they
can trust us, and the data we report to them.

64 © 2023 Future Today Institute

 TECH Artificial Intelligence

WHAT IF AI COULD HELP YOU PRACTICE

MARRIAGE BEFORE GOING THROUGH
WITH THE CEREMONY?
Scenario Year: 2030

The rich and famous aren’t just getting prenups; they’re now
springing for prenup sims. Developed to mirror each fiancé’s
lifestyle habits, communication styles, personality quirks,
and love languages, sims live together in a digital twin of the
potential couple’s home. Thousands of simulated routine and
novel challenges are run to determine the probability of a long,
happy marriage.

65 © 2023 Future Today Institute

 TECH Artificial Intelligence

WHAT IF AI COULD HELP YOU EAT

BETTER AND BE HEALTHIER?
Scenario Year: 2026

A new smartphone app uses deepfakes to encourage healthy lifestyles. The app, IntelliEAT
(the EAT is short for Energy Assessment Technology) automatically recognizes everything
you eat or drink, calculates the data from your smartwatch or fitness tracker, computes
the benefit or damage, and automatically generates a deepfake to show the impact you
made on your body that day. IntelliEAT doesn’t just count calories or nutrients. Instead it
displays what the future you would look life if every day was like your current day. Healthy
meals and exercise reveal an older version of you who is fit, agile, and happy. Regular
binge eating reveals a deepfaked version of you that is heavier with joint pain and
difficulty sleeping. Proponents of body positivity have voiced concern, but IntelliEAT has
won praise from medical associations for focusing holistically on health. In addition to the
Immediate Impact View, the app also offers an Invisible Damage View, which highlights
health effects such as increased cholesterol, restricted blood flow, and inflammation.

66 © 2023 Future Today Institute

 TECH Artificial Intelligence

HOW TO
PREPARE
FOR THE
FUTURE
67 © 2023 Future Today Institute
 TECH Artificial Intelligence

HOW TO PREPARE What should your organization do now to prepare for these trends?

Artificial intelligence As no-code and low-code applications be- algorithms, or generative AI systems. Risk cies, better tracking, business intelligence,
come more widely available, Future Today models should be developed to determine and assistance with decision-making. As
should be part of Institute analysis shows that innovation plausible near-futures, so that leaders can training costs decline, more applications
every strategic plan, teams will be in position to build powerful adjust their strategies accordingly. will be built. Spending on AI systems and
systems for decision management, busi- hardware is likely to explode this decade,
as it crosses multiple ness intelligence, and product ideation. When it comes to talent and workforce creating significant enterprise value overall.
dimensions, from Generative AIs will improve an organiza- development, the need for highly skilled
tion’s efficiency and enhance creativity, people is fast outpacing the graduation rate
workforce automation, of universities. This might lead to a shift in
leading to hybrid human-machine creative
to digital transformation, teams. AI-assisted design will dramatically higher education: It’s plausible that compa-
nies in search of AI specialists might opt for
to everyday business increase the number of prototypes that can
modular certifications, which can be earned
be automatically generated with prompts.
processes and faster than traditional four-year degrees.
We recommend that chief strategy officers Talent sourcing and retention will continue
business intelligence. in every field develop a solid understanding to pose challenges for tech companies—
It is imperative that of AI to engage more closely with others in and for organizations in other industries
executives and senior the C-suite to develop a cohesive point of that need a trained workforce but may not
view on digitalization, augmentation and be able to provide the same perks as the
managers understand automation—as well as to develop strate- hottest AI players.
what AI is, what it is not, gic plans. Importantly, businesses should
In nearly every industry, AI will serve as a
keep abreast of emerging regulations that
and what strategic value could restrict the use of consumer data, force multiplier for growth, bringing efficien-
it adds to the business.

68 © 2023 Future Today Institute

 TECH Artificial Intelligence

KEY QUESTIONS FOR YOUR TEAM

1 2 3 4 5 6 7 8

How do these Is our organization When we engage in Are we adequately What is our process Does our team What parts of our What is our point
trends change positioned to long-term strategic preparing our to vet, verify have an expansive business make us of view on the use
our perspective leverage AI for planning, how does workforce to and monitor our enough viewpoint vulnerable as AI of generative AI
on our current AI growth, in addition the evolution of AI succeed in a vendors? on emerging evolves? systems?
investment and to realizing new factor in? world in which threats and attack
capabilities? efficiencies? their knowledge surfaces?
Is the AI we’re using
tasks might be
explainable?
augmented or fully
automated by AI?
If not, does that
open us up to
9 10
additional risk?
If proposed How might we
antitrust legislation develop the
passes in the US knowledge,
and the EU further experience, and
regulates Big talent in place to
Tech companies, leverage these AI
how does this trends?
create a strategic
opportunity for our
organization?

Or, what might our

organization lose?

69 © 2023 Future Today Institute

 TECH Artificial Intelligence

SELECTED
SOURCES

70 © 2023 Future Today Institute

 TECH Artificial Intelligence

SELECTED SOURCES
“A Generalist Agent.” https://www.deepmind.com/ Baltescu, Paul, Haoyu Chen, Nikil Pancha, Andrew Chitty-Venkata, Krishna Teja, and Arun K. Somani. Go-Chess-and-Shogi-by-Planning-with-a-Learned-
publications/a-generalist-agent. Zhai, et al. “ItemSage: Learning Product Embeddings “Neural Architecture Search Survey: A Hardware Model.
for Shopping Recommendations at Pinterest.” arXiv, Perspective.” ACM Computing Surveys 55, no. 4 (April
“Adept: Useful General Intelligence.” https://www. May 23, 2022. http://arxiv.org/abs/2205.11728. 30, 2023): 1–36. https://doi.org/10.1145/3524500. DeepMind. “MuZero: Mastering Go, chess, shogi and
adept.ai. Atari Without rules.” https://deepmind.com/blog/
“Biological Structure and Function Emerge From Corso, Gabriele, Hannes Stärk, Bowen Jing, Regina article/muzero-mastering-go-chess-shogi-and-atari-
Air Force. “AI Copilot: Air Force Achieves First Military Scaling Unsupervised Learning to 250 Million Barzilay, and Tommi Jaakkola. “DiffDock: Diffusion without-rules.
Flight With Artificial Intelligence.” https://www.af.mil/ Protein Sequences.” PNAS. https://www.pnas.org/ Steps, Twists, and Turns for Molecular Docking.”
News/Article-Display/Article/2448376/ai-copi- doi/10.1073/pnas.2016239118. arXiv, October 4, 2022. https://doi.org/10.48550/arX- DeepMind. “Putting the power of AlphaFold Into the
lot-air-force-achieves-first-military-flight-with-artifi- iv.2210.01776. World’s Jands.” https://deepmind.com/blog/article/
cial-intelligence. Borsos, Zalán, Matt Sharifi, and Marco Tagliasac- putting-the-power-of-alphafold-into-the-worlds-
chi. “SpeechPainter: Text-Conditioned Speech Damioli, Giacomo, Vincent Van Roy, and Daniel hands.
“AI Patent Database Content.” 2020. Reprint, Cen- Inpainting.” arXiv, March 30, 2022. http://arxiv.org/ Vertesy. “The Impact of Artificial Intelligence on
ter for Security and Emerging Technology (CSET), abs/2202.07273. Labor Productivity.” Eurasian Business Review 11, no. 1 Defense One. “AI Could Match ‘Fingerprints’ of Texts
https://github.com/georgetown-cset/1790-ai-patent- (March 1, 2021): 1–25. https://doi.org/10.1007/s40821- to Their Authors, Under New Intelligence Program.”
data. Cancian, Mark F., Matthew Cancian, and Eric 020-00172-8. https://www.defenseone.com/technology/2022/02/
Heginbotham. “The First Battle of the Next War: ai-could-match-fingerprints-texts-their-authors-un-
“AI-Synthesized Faces Are Indistinguishable from Wargaming a Chinese Invasion of Taiwan,” January Daulton, Samuel, David Eriksson, Maximilian Balan- der-new-intelligence-program/361893.
Real Faces and More Trustworthy | PNAS.” https:// 9, 2023. https://www.csis.org/analysis/first-bat- dat, and Eytan Bakshy. “Multi-Objective Bayesian
www.pnas.org/content/119/8/e2120481119. tle-next-war-wargaming-chinese-invasion-taiwan. Optimization over High-Dimensional Search Spaces.” Defense One. “The Near Future of Military Autonomy
arXiv, June 15, 2022. http://arxiv.org/abs/2109.10964. Isn’t Robotanks, But ‘Microservices.’” https://www.de-
“Aircrew Labor In-Cockpit Automation System.” Center for Security and Emerging Technology. “Eth- fenseone.com/technology/2022/02/near-future-mili-
https://www.darpa.mil/program/aircrew-la- ical Norms for New Generation Artificial Intelligence Dauparas, J., I. Anishchenko, N. Bennett, H. Bai, R. tary-autonomy-isnt-robotanks-microservices/361873.
bor-in-cockpit-automation-system. Released.” https://cset.georgetown.edu/publication/ J. Ragotte, L. F. Milles, B. I. M. Wicky, et al. “Robust
ethical-norms-for-new-generation-artificial-intelli- Deep Learning–Based Protein Sequence Design Us- Degrave, Jonas, Federico Felici, Jonas Buchli, Michael
“API Reference—Amazon Rekognition.” https://docs. gence-released. ing ProteinMPNN.” Science 378, no. 6615 (October 7, Neunert, et al. “Magnetic Control of Tokamak Plasmas
aws.amazon.com/rekognition/latest/dg/API_Refer- 2022): 49–56. https://doi.org/10.1126/science.add2187. through Deep Reinforcement Learning.” Nature 602,
ence.html. Center for Security and Emerging Technology. “US no. 7897 (February 17, 2022): 414–19. https://doi.
AI Workforce.” https://cset.georgetown.edu/publica- “Deep Dream Generator.” https://deepdreamgenera- org/10.1038/s41586-021-04301-9.
Baek, Minkyung, Frank DiMaio, Ivan Anishchenko, tion/u-s-ai-workforce. tor.com.
Justas Dauparas, et al. “Accurate Prediction of Protein “DOD to Hire First-Ever Chief Digital and Artificial
Structures and Interactions Using a Three-Track Neu- “Chexia Face Recognition.” NIST. https://www.nist. DeepMind. “Mastering Atari, Go, Chess and Shogi by Intelligence Officer, Form New Office—Nextgov.”
ral Network.” Science 373, no. 6557 (August 20, 2021): gov/programs-projects/chexia-face-recognition. Planning With a Learned Model.” https://deepmind. https://www.nextgov.com/emerging-tech/2021/12/
871–76. https://doi.org/10.1126/science.abj8754. com/research/publications/2019/Mastering-Atari- dod-hire-first-ever-chief-digital-and-artificial-intelli-
gence-officer-form-new-office/187399.

71 © 2023 Future Today Institute

 TECH Artificial Intelligence

SELECTED SOURCES
Dong, Tian, Bo Zhao, and Lingjuan Lyu. “Privacy iv:2101.03961 [Cs], January 11, 2021. http://arxiv.org/ Gupta, Agrim, Silvio Savarese, Surya Ganguli, and Kuang, Zhengfei, Kyle Olszewski, Menglei Chai, Zeng
for Free: How Does Dataset Condensation Help abs/2101.03961. Li Fei-Fei. “Embodied Intelligence via Learning and Huang, et al. “NeROIC: Neural Rendering of Objects
Privacy?” arXiv, June 1, 2022. http://arxiv.org/ Evolution.” Nature Communications 12, no. 1 (October From Online Image Collections.” arXiv, September 1,
abs/2206.00240. Flennerhag, Sebastian, Yannick Schroecker, Tom 6, 2021): 5721. https://doi.org/10.1038/s41467-021- 2022. http://arxiv.org/abs/2201.02533.
Zahavy, and Hado van Hasselt. “Bootstrapped Me- 25874-z.
Dunn, Jeff. “Google Works Out a Fascinating, Slightly ta-Learning,” 2022. LeCun, Yann. “A Path Towards Autonomous Machine
Scary Way for AI to Isolate Voices in a Crowd.” Ars Hafner, Danijar. “Dream to Control.” Python, February Intelligence Version 0.9.2, 2022-06-27.”
Technica, April 13, 2018. https://arstechnica.com/gad- Gafni, Oran, Adam Polyak, Oron Ashual, Shelly 13, 2022. https://github.com/danijar/dreamer.
gets/2018/04/google-works-out-a-fascinating-slight- Sheynin, et al. “Make-A-Scene: Scene-Based Text-to- Lee, Jae Hyeon, Payman Yadollahpour, Andrew
ly-scary-way-for-ai-to-isolate-voices-in-a-crowd. Image Generation with Human Priors.” arXiv, March Horowitz, Adam Haar, Tony J. Cunningham, Pattie Watkins, Nathan C. Frey, et al. “EquiFold: Protein
24, 2022. http://arxiv.org/abs/2203.13131. Maes, and Robert Stickgold. “Dormio: A Target- Structure Prediction With a Novel Coarse-Grained
Dwivedi-Yu, Jane, Yi-Chia Wang, Lijing Qin, Cristian ed Dream Incubation Device.” Consciousness and Structure Representation.” bioRxiv, October 8, 2022.
Canton-Ferrer, and Alon Y. Halevy. “Affective Sig- Gent, Edd. “Meta Built an AI That Can Guess the Cognition 83 (August 2020): 102938. https://doi. https://doi.org/10.1101/2022.10.07.511322.
nals in a Social Media Recommender System.” In Words You’re Hearing by Decoding Your Brainwaves.” org/10.1016/j.concog.2020.102938.
Proceedings of the 28th ACM SIGKDD Conference Singularity Hub (blog), September 13, 2022. https:// Li, Yujia, David Choi, Junyoung Chung, Nate Kush-
on Knowledge Discovery and Data Mining, 2831–41, singularityhub.com/2022/09/13/meta-built-an-ai- “How Amazon Search Reduced ML Inference Costs man, Julian Schrittwieser, Rémi Leblond, Tom Eccles,
2022. https://doi.org/10.1145/3534678.3539054. that-can-guess-the-words-youre-hearing-by-decod- by 85% with AWS Inferentia.” AWS Machine Learning et al. “Competition-Level Code Generation with
ing-your-brainwaves. Blog,”September 22, 2022. https://aws.amazon.com/ AlphaCode.” arXiv, February 8, 2022. http://arxiv.org/
“Emotion—Amazon Rekognition.” https://docs.aws. blogs/machine-learning/how-amazon-search-re- abs/2203.07814.
amazon.com/rekognition/latest/APIReference/ “GitHub—Google-Research/Dreamer: Dream to duced-ml-inference-costs-by-85-with-aws-inferentia.
API_Emotion.html. Control: Learning Behaviors by Latent Imagination.” Louis Bouchard. “Impressive Photo Restoration by
https://github.com/google-research/dreamer. Intenseye. “Intenseye: AI for Workplace Safety.” AI !” March 11, 2022. https://www.louisbouchard.ai/
“ESM.” https://huggingface.co/docs/transformers/ https://content.intenseye.com. gfp-gan.
model_doc/esm. Google. “Join Us in the AI Test Kitchen,” August 25,
2022. https://blog.google/technology/ai/join-us-in- Jumper, John, Richard Evans, Alexander Pritzel, Tim “Meta AI Research Topic—No Language Left Behind.”
Farr, Christina. “As Amazon Moves into Health the-ai-test-kitchen. Green, et al. “Highly Accurate Protein Structure Pre- https://ai.facebook.com/i18n/research/topics/no-
Care, Here’s What We Know—and What We Sus- diction with AlphaFold.” Nature 596, no. 7873 (August language-left-behind.
pect—About Its Plans.” CNBC. https://www.cnbc. Google. “LaMDA: Our Breakthrough Conversation 2021): 583–89. https://doi.org/10.1038/s41586-021-
com/2018/03/27/amazons-moves-into-health-what- Technology,” May 18, 2021. https://blog.google/tech- 03819-2. Mildenhall, Ben, Pratul P. Srinivasan, Matthew Tancik,
we-know.html. nology/ai/lamda. Jonathan T. Barron, et al. “NeRF: Representing Scenes
Kafritsas, Nikos. “CLIP: The Most Influential AI Model as Neural Radiance Fields for View Synthesis.” ArX-
Fedus, William, Barret Zoph, and Noam Shazeer. Google. “MUM: A New AI Milestone for Understand- From OpenAI—And How To Use It.” Medium, Sep- iv:2003.08934 [Cs], August 3, 2020. http://arxiv.org/
“Switch Transformers: Scaling to Trillion Parameter ing Information,” May 18, 2021. https://blog.google/ tember 28, 2022. https://towardsdatascience.com/ abs/2003.08934.
Models with Simple and Efficient Sparsity.” ArX- products/search/introducing-mum. clip-the-most-influential-ai-model-from-openai-and-
how-to-use-it-f8ee408958b1.

72 © 2023 Future Today Institute

 TECH Artificial Intelligence

SELECTED SOURCES
MIT Technology Review. “A Bot That Watched Academy of Sciences 119, no. 8 (February 22, 2022). Piergiovanni, Aj, Vincent Casser, Michael S. Ryoo, and “Responsible AI Guidelines.” https://www.diu.mil/
70,000 Hours of Minecraft Could Unlock AI’s https://doi.org/10.1073/pnas.2120481119. Anelia Angelova. “4D-Net for Learned Multi-Modal responsible-ai-guidelines.
next Big Thing.” https://www.technologyreview. Alignment.” In 2021 IEEE/CVF International Confer-
com/2022/11/25/1063707/ai-minecraft-video-un- Nijkamp, Erik, Jeffrey Ruffolo, Eli N. Weinstein, ence on Computer Vision (ICCV), 15415–25. Montreal, Roper, Dr Will. “Exclusive: AI Just Controlled a Military
lock-next-big-thing-openai-imitation-learning. Nikhil Naik, and Ali Madani. “ProGen2: Exploring the QC, Canada: IEEE, 2021. https://doi.org/10.1109/ Plane for the First Time Ever.” Popular Mechanics,
Boundaries of Protein Language Models.” arXiv, June ICCV48922.2021.01515. December 16, 2020. https://www.popularmechanics.
Müller, Thomas, Alex Evans, Christoph Schied, and 27, 2022. https://doi.org/10.48550/arXiv.2206.13517. com/military/aviation/a34978872/artificial-intelli-
Alexander Keller. “Instant Neural Graphics Primitives “ProT-VAE: Protein Transformer Variational AutoEn- gence-controls-u2-spy-plane-air-force-exclusive/.
with a Multiresolution Hash Encoding.” ACM Transac- Nitzan, Yotam, Kfir Aberman, Qiurui He, Orly Liba, coder for Functional Protein Design,” BioRxiv. https://
tions on Graphics 41, no. 4 (July 2022): 1–15. https:// et al. “MyStyle: A Personalized Generative Prior.” www.biorxiv.org/content/10.1101/2023.01.23.525232v1. Saharia, Chitwan, William Chan, Saurabh Saxena,
doi.org/10.1145/3528223.3530127. arXiv, October 6, 2022. https://doi.org/10.48550/arX- Lala Li, Jay Whang, Emily Denton, Seyed Kamyar
iv.2203.17272. Radford, Alec, Jong Wook Kim, Tao Xu, Greg Brock- Seyed Ghasemipour, et al. “Photorealistic Text-to-Im-
“Multiple Encounter Dataset (MEDS-I)—NIST Special man, Christine McLeavey, and Ilya Sutskever. “Robust age Diffusion Models with Deep Language Under-
Database 32.” National Institute of Standards and “Obvious—Art & AI.” https://obvious-art.com. Speech Recognition via Large-Scale Weak Supervi- standing.”
Technology. https://catalog.data.gov/dataset/multi- sion.”
ple-encounter-dataset-meds-i-nist-special-database- Office, U. S. Government Accountability. “Facial Rec- Sapé, Sergi Castella i. “Trends in AI—September
32-616de. ognition Technology: Federal Law Enforcement Agen- Rainie, Lee, Janna Anderson, and Emily a Vogels. 2022.” Zeta Alpha, September 28, 2022. https://www.
cies Should Better Assess Privacy and Other Risks.” “Experts Doubt Ethical AI Design Will Be Broadly zeta-alpha.com/post/trends-in-ai-september-2022.
“Nerfies: Deformable Neural Radiance Fields.” https://www.gao.gov/products/gao-21-518. Adopted as the Norm Within the Next Decade.” Pew
https://nerfies.github.io. Research Center: Internet, Science & Tech (blog), Schrittwieser, Julian, Ioannis Antonoglou, Thomas Hu-
OpenAI. “CLIP: Connecting Text and Images,” January June 16, 2021. https://www.pewresearch.org/inter- bert, Karen Simonyan, Laurent Sifre, Simon Schmitt,
Nextgov.com. “Senate Confirms John Sherman as 5, 2021. https://openai.com/blog/clip. net/2021/06/16/experts-doubt-ethical-ai-design-will- Arthur Guez, et al. “Mastering Atari, Go, Chess and
Pentagon’s Tech Chief.” https://www.nextgov.com/ be-broadly-adopted-as-the-norm-within-the-next- Shogi by Planning with a Learned Model.” Nature
cio-briefing/2021/12/senate-confirms-john-sherman- OpenAI. “DALL·E: Creating Images from Text,” Janu- 588, no. 7839 (December 2020): 604–9. https://doi.
decade.
pentagons-tech-chief/359940. ary 5, 2021. https://openai.com/blog/dall-e. org/10.1038/s41586-020-03051-4.
Ramesh, Aditya, Prafulla Dhariwal, Alex Nichol, Casey
Ni, Bolin, Houwen Peng, Minghao Chen, Songyang OpenAI. “MuseNet,” April 25, 2019. https://openai. Sevilla, Jaime, Lennart Heim, Anson Ho, Tamay Besi-
Chu, and Mark Chen. “Hierarchical Text-Conditional
Zhang, et al. “Expanding Language-Image Pretrained com/blog/musenet. roglu, Marius Hobbhahn, and Pablo Villalobos. “Com-
Image Generation with CLIP Latents.” arXiv, April 12,
Models for General Video Recognition.” ArXiv, August 2022. http://arxiv.org/abs/2204.06125. pute Trends Across Three Eras of Machine Learning.”
Peng, Kenny, Arunesh Mathur, and Arvind Narayanan.
4, 2022. https://doi.org/10.48550/arXiv.2208.02816. arXiv, March 9, 2022. https://doi.org/10.48550/arX-
“Mitigating Dataset Harms Requires Stewardship:
“README.” Python. 2022. Reprint, NVIDIA Corpora- iv.2202.05924.
Nightingale, Sophie J., and Hany Farid. “AI-Synthe- Lessons from 1000 Papers.” ArXiv:2108.02922 [Cs],
tion, January 21, 2023. https://github.com/NVIDIA/
sized Faces Are Indistinguishable from Real Faces November 21, 2021. http://arxiv.org/abs/2108.02922.
MegaMolBART.
and More Trustworthy.” Proceedings of the National

73 © 2023 Future Today Institute

 TECH Artificial Intelligence

SELECTED SOURCES
Shwartz-Ziv, Ravid, and Amitai Armon. “Tabular Data: “The AI Index Report—Artificial Intelligence Index.” Willingham, Emily. “Humans Find AI-Generated Faces Zhang, Can, Tianyu Yang, Junwu Weng, Meng Cao,
Deep Learning Is Not All You Need.” arXiv, November https://aiindex.stanford.edu/report. More Trustworthy Than the Real Thing.” Scientific Jue Wang, and Yuexian Zou. “Unsupervised Pre-Train-
23, 2021. http://arxiv.org/abs/2106.03253. American. https://www.scientificamerican.com/arti- ing for Temporal Action Localization Tasks.” arXiv,
Thoppilan, Romal, Daniel De Freitas, Jamie Hall, cle/humans-find-ai-generated-faces-more-trustwor- March 25, 2022. http://arxiv.org/abs/2203.13609.
Silver, David, Satinder Singh, Doina Precup, and Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze thy-than-the-real-thing.
Richard S. Sutton. “Reward Is Enough.” Artificial Cheng, Alicia Jin, et al. “LaMDA: Language Models for Zhang, Susan, Stephen Roller, Naman Goyal, Mikel
Intelligence 299 (October 1, 2021): 103535. https://doi. Dialog Applications.” arXiv, February 10, 2022. http:// Xiao, Tete, Xiaolong Wang, Alexei A. Efros, and Trevor Artetxe, Moya Chen, Shuohui Chen, Christopher
org/10.1016/j.artint.2021.103535. arxiv.org/abs/2201.08239. Darrell. “What Should Not Be Contrastive in Contras- Dewan, et al. “OPT: Open Pre-Trained Transformer
tive Learning.” arXiv, March 18, 2021. http://arxiv.org/ Language Models.” ArXiv, June 21, 2022. http://arxiv.
Singer, Uriel, Adam Polyak, Thomas Hayes, Xi Yin, Tuli, Shreshth, Giuliano Casale, and Nicholas R. abs/2008.05659. org/abs/2205.01068.
Jie An, Songyang Zhang, Qiyuan Hu, et al. “Make-a- Jennings. “TranAD: Deep Transformer Networks for
Video: Text-to-Video Generation Without Text-Video Anomaly Detection in Multivariate Time Series Data.” Xie, Yunhe, Chengjie Sun, and Zhenzhou Ji. “A Zvyagin, Maxim, Alexander Brace, Kyle Hippe,
Data.” arXiv, May 14, 2022. http://arxiv.org/abs/2201.07284. Commonsense Knowledge Enhanced Network Yuntian Deng, Bin Zhang, Cindy Orozco Bohorquez,
with Retrospective Loss for Emotion Recognition in Austin Clyde, et al. “GenSLMs: Genome-Scale
Singhal, Karan, Shekoofeh Azizi, Tao Tu, S. Sara Mah- Tzaban, Rotem, Ron Mokady, Rinon Gal, Amit H. Ber- Spoken Dialog.” In ICASSP 2022 - 2022 IEEE Interna- Language Models Reveal SARS-CoV-2 Evolutionary
davi, Jason Wei, Hyung Won Chung, Nathan Scales, mano, and Daniel Cohen-Or. “Stitch It in Time: GAN- tional Conference on Acoustics, Speech and Signal Dynamics.” BioRxiv, October 11, 2022. https://doi.
et al. “Large Language Models Encode Clinical Based Facial Editing of Real Videos.” arXiv, January 21, Processing (ICASSP), 7027–31, 2022. https://doi. org/10.1101/2022.10.10.511571.
Knowledge.” arXiv, December 26, 2022. https://doi. 2022. https://doi.org/10.48550/arXiv.2201.08361. org/10.1109/ICASSP43922.2022.9746909.
org/10.48550/arXiv.2212.13138.
“Voxel.” https://www.voxelai.com. Yang, Gengshan, Minh Vo, Natalia Neverova, Deva
Solaiyappan, Siddharth, and Yuxin Wen. “Machine Ramanan, Andrea Vedaldi, and Hanbyul Joo. “BAN-
Learning Based Medical Image Deepfake Detec- “Welcome—Amazon Rekognition.” https://docs.aws. Mo: Building Animatable 3D Neural Models from
tion: A Comparative Study.” Machine Learning with amazon.com/rekognition/latest/APIReference/Wel- Many Casual Videos.” In 2022 IEEE/CVF Conference
Applications 8 (June 15, 2022): 100298. https://doi. come.html. on Computer Vision and Pattern Recognition (CVPR),
org/10.1016/j.mlwa.2022.100298. 2853–63. New Orleans, LA, USA: IEEE, 2022. https://
doi.org/10.1109/CVPR52688.2022.00288.

74 © 2023 Future Today Institute

 TECH Artificial Intelligence

AUTHORS
& CONTRIBUTORS
AMY WEBB ANNAMALAI CHOCKALINGAM Chief Executive Officer
Founder and CEO of Future Today Institute Senior Expert Advisor AMY WEBB
Amy Webb pioneered a data-driven foresight Annamalai Chockalingam is a Senior Expert Managing Director
methodology that is now used within hundreds Advisor at Future Today Institute, where he MELANIE SUBIN
of organizations. She possesses master-lev- is helping shape a better future. Annamalai
Chief Content Officer
el knowledge and insights on the potential currently runs product management and mar-
JON FINE
futures for the industries critical to keeping the keting at Nvidia, for a suite of generative AI,
world moving during this unparalleled period and large language model software products. Creative Director
of disruption in business and society. Webb is a sought-after advisor Annamalai brings a wealth of experience spanning deep learning, EMILY CAUFIELD
to Fortune 100 and Global 1000 companies and provides strategic strategy and management consulting, data and firmware engineering,
Designer
foresight on the emerging tech and trends that will turn industries on and technology infrastructure. He has experience working with large-
ERICA GRAU
their heads and transform every aspect of our lives. She was elected cap organizations across technology, energy, and financial sectors. He
a life member of the Council on Foreign Relations, serves on a Stew- excels at using his deep technical background to devise astute business Editors
ardship Board at the World Economic Forum, and in 2022 received strategies and deliver innovative products that delight its users. He is a CAROLE BRADEN
an official Mad Scientist proclamation from the U.S. Army. A lifelong recent graduate from NYU’s Stern School of Business and Courant Insti- TOM BRADY
science fiction fan, she collaborates closely with writers and producers tute of Mathematical Sciences, and also holds a Bachelor of Science in MEGAN CREYDT
on films, TV shows and commercials about science, technology and the Electrical Engineering from the University of Alberta. TOM FOSTER
future. Webb was named by Forbes as one of the five women changing
Copy Editor
the world, listed as the BBC’s 100 Women of 2020, named one of the
SARAH JOHNSON
most influential business professors in the world by Poets & Quants, and
is ranked on the Thinkers50 list of the 50 most influential management Director of Operations
thinkers globally. CHERYL COONEY

75 © 2023 Future Today Institute

 TECH Artificial Intelligence

FUTURE TODAY INSTITUTE’S 2023 TECH TREND REPORT

Our 2023 edition includes nearly 700 trends, which are published individually in 14 volumes and as one comprehensive report with all trends included.

Download all sections of Future Today Institute’s 2023 Tech Trends report at http://www.futuretodayinstitute.com/trends.

76 © 2023 Future Today Institute

 TECH Artificial Intelligence

ABOUT
FUTURE TODAY
INSTITUTE

77 © 2023 Future Today Institute

 TECH Artificial Intelligence

ABOUT FUTURE TODAY INSTITUTE

CONTACT US
Founded in 2006, Future Today Institute helps leaders and their
organizations navigate turbulent times by reducing uncertainty, For an introductory conversation to learn how Future Today
identifying existential threats, leveraging critical emerging Institute can assist your organization with its long-term
technologies, and identifying opportunities for growth. strategic planning and foresight needs, please contact
inquiries@futuretodayinstitute.com
We help organizations lead into the future. Our consultants
work with executives and their teams to define plausible futures
futuretodayinstitute.com
and develop executable strategies. Combining deep industry
knowledge, rich creativity and rigorous analysis, we make +1 267 342 4300
teams confident in taking meaningful action.

As the leading strategic foresight and futures management SOCIAL

consultants to executive leadership teams worldwide, FTI’s
data-driven, applied research methodology identifies trends, TWITTER @FTI
calculates their impact, and through descriptive scenarios,
reveals how they will disrupt business, government and society. LINKEDIN @Future-Today-Institute

Together with our clients and partners, FTI is helping leaders

achieve their preferred futures. Our pioneering, data-driven FACEBOOK @FutureTodayInstitute
forecasting methodology and tools empower leaders to make
better decisions about the future, today. INSTAGRAM @futuretodayinstitute

78 © 2023 Future Today Institute

 TECH Artificial Intelligence

METHODOLOGY

Future Today Institute conducts in-depth qualitative and quantitative

research throughout the year to identify emerging trends. We review
patent and trademark filings, pre-print and published scientific pa-
pers, investment rounds, online search trends, macroeconomic data,
publications from governments worldwide, news mentions, influencer
posts and other sources, and we use a proprietary system to identify
patterns, which are then grouped into nodes and evaluated using a set
of standardized indicators. Qualified trends are further scored for their
trajectory, momentum and timing. Additionally, we harness the deep
subject matter expertise of our Future Today Institute network, leading
to valuable insights about the topics we cover.

In continuous publication since 2007, Future Today Institute’s annu-

al report includes maturing and emerging trends grouped into two
categories: industry and technology. Industry trends reflect the ways in
which technology is shaping the future of an entire industry. Technology
trends are specific developments within one arena, such as artificial
intelligence. Covering a wide range of technologies across industry
sectors creates a holistic view of change and provides leaders with a
clear understanding of their potential impact. Trends are published as
individual Industry and Technology reports, as well as in one combined
report with all of our research.

Monitored regularly, trends help executives recognize emerging threats

and opportunities in the near-term and enable them to develop per-
spectives, strategies and plans for the future.

79 © 2023 Future Today Institute

 TECH Artificial Intelligence

DISCLAIMER

The views expressed herein are the authors’ own and are not represen-
tative of the greater organizations in which they have been employed.
The names of companies, services, and products mentioned in this
report are not necessarily intended as endorsements by Future Today
Institute or this report’s authors.

Future Today Institute’s 2023 Tech Trends Report relies on data, analysis,
and modeling from a number of sources, which includes sources within
public and private companies, securities filings, patents, academic
research, government agencies, market research firms, conference pre-
sentations and papers, and news media stories. Additionally, this report
draws from Future Today Institute’s previous EMT Trends Reports, FTI
Trend Reports, and newsletters. FTI’s reports are occasionally updated
on the FTI website.

FTI advises hundreds of companies and organizations, some of which

are referenced in this report. FTI does not own any equity position in any
of the entities listed in this presentation.

Any trademarks or service marks used in this report are the marks of
their respective owners, who do not endorse the statements in this
report. All rights in marks are reserved by their respective owners. We
disclaim any and all warranties, expressed or implied, with respect to
this report.

80 © 2023 Future Today Institute

 TECH Artificial Intelligence

USING AND SHARING THE MATERIAL IN THIS REPORT

You are free to: You may not:

Download the materials for your or your Use the materials for any commercial
organization’s non-commercial use. purposes, which includes: selling, licensing
or charging for products or services that use
Use the assets with full attribution to Future
the materials.
Today Institute using the following language:
“This material is copyrighted by Future Today Publish the materials in a book or upload to
Institute and licensed under the Creative a website or intranet without prior written
Commons 4.0 International License (CC BY- permission from Future Today Institute.
NC-ND 4.0).”
Remix, transform or build upon the material
Use assets in a reasonable manner, but not without giving appropriate credit to Future
in any way that suggests that Future Today Today Institute as the originator of the work
Institute endorses you or your use. using the following language: “This material
is adapted from original work copyrighted
by Future Today Institute and licensed under
the Creative Commons 4.0 International
License (CC BY-NC-ND 4.0).”

Assert a relationship with or endorsement

from Future Today Institute.

81 © 2023 Future Today Institute