Deutsche Bank

Research

Generative AI and ChatGPT 101

#PositiveImpact

May 2023

Adrian Cox
Thematic Research
Distributed on: 23/05/2023 04:31:08 GMT

(+44)-20-7541-7775
adrian.cox@db.com

Jim Reid
(+44)-20-754-72943
jim.reid@db.com

Galina Pozdnyakova
What is How does it How can you use What does generative (+44)-20-754-74994
generative AI? work? chatbots? AI mean for the future? galina.pozdnyakova@db.com

Deutsche Bank
IMPORTANT Research
RESEARCH | AdrianAND
DISCLOSURES CoxANALYST
(+44) 20CERTIFICATIONS
7541 7775 | May 2023
ARE LOCATED IN APPENDIX 1. MCI (P) 097/10/2022. UNTIL 19th MARCH 2021 INCOMPLETE DISCLOSURE 0
0
INFORMATION MAY HAVE BEEN DISPLAYED, PLEASE SEE APPENDIX 1 FOR FURTHER DETAILS.
7T2se3r0Ot6kwoPa Image by DALL-E
 Six months of hype, chat and fears

Next week will mark six months since the launch of 1950: Alan Turing proposes
ChatGPT, the chatbot based on OpenAI’s GPT-3.5 large a learning machine
language model. 1966 First chatbot ELIZA
developed
There has been an extraordinary flurry of launches, hype
and activity since then. Even in the past week, OpenAI, 1980s Early development of
Google and Meta announced fresh products, OpenAI’s neural networks
CEO testified in Congress, and the G7 called for
“guardrails” on AI, among many other developments. 2011 Apple releases first
iPhone with AI assistant Siri
Whether you agree with the researchers who say the
underlying technology has “sparks of Artificial General
2012 AlexNet takes leap
Intelligence” or you think it is just a “stochastic parrot”, it
forward in image recognition
is vital to understand how it works and what it could mean.
2017 “Attention is all you need”
This is the latest in our 101 series for generalists, aiming to paper announces transformers
give a clear overview of generative AI, addressing what it
is, how it works, how to use it, and what it means for jobs, 2018 Google introduces
the economy and society. It is also part of an ongoing BERT large language model
series starting this week looking more deeply into the topic.
2019 OpenAI creates GPT-2
In the meantime, one thing seems certain. Amara’s law, LLM with 1.2bn parameters
that “we tend to overestimate the effect of a technology in
the short run and underestimate the effect in the long run,” Nov 30, 2022: OpenAI
could have been written for (or indeed by) ChatGPT. launches ChatGPT chatbot

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 1
1
 What is special about generative AI?
Almost anyone with internet access can right now operate a chatbot based on a large language model (LLM) at little
cost and with little or no training

1. General (multi-purpose) tech 2. Ease of use 3. Rapid adoption

Months
Search
Coding Tell me how to access and use Telephone

Key technologies
ChatGPT in a six-line rhyming

(adoption)
TV
poem
WWW
Translation
Sentiment Smartphone
analysis To access ChatGPT, heed this
rhyme, Uber
Visit websites where it's prime. Spotify

Apps hitting 100m users

Summary
In the chatbox, type your quest,
Hit send, put it to the test. Pinterest
Human-
like Q&A Read responses as they show, Twitter
Image Keep chatting, watch the
knowledge flow. Facebook
generation
Instagram

Entity ChatGPT Two months

recognition
0 100 200 300

Source: Deutsche Bank, company reports, media headlines; ChatGPT content written with ChatGPT

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 2
2
What is a large language model (LLM) and where does it fit into AI?
GPT-4 is the underlying technology for ChatGPT and is currently the most high-profile LLM, an AI model trained on a
vast amount of data, capable of understanding and generating human-like text

Artificial intelligence Machines/software capable of intelligent behaviour:

three levels
1. Narrow AI: performs limited tasks
2. Artificial General Intelligence (AGI)
3. Superintelligence

Machine learning Learning and acting on data without explicit

instructions
Neural networks
Models inspired by the human brain

Deep learning Uses multiple hidden layers of artificial neural

networks to enhance performance

Foundation
model Understands and generates text and/or images, code
etc; need a lot of data for training

LLM
Focused on language tasks

Source: Deutsche Bank.

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 3
3
What is the underlying deep learning technology for LLMs?
An image classification example

Input Feature extraction Classification Output

Traditional machine
CAT
learning:
Experts input
NOT CAT
structured data

Input Feature extraction + Classification Output

Hidden layers → deep

Deep learning with
neural networks:
Can learn far more
CAT
complex structures
and relationships (eg
NOT CAT
pixels) without needing
human help to identify Input Output
relevant features “Neuron/node” “Weight”

Source: Deutsche Bank.

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 4
4
 What are the key neural networks and what are they made of?
Neural networks have become increasingly sophisticated over the past decade, particularly with the breakthrough of
Convolutional Neural Networks in 2012 and Transformers in 2017

Key neural network models: Common components:

(Approximately ranked from simple to complex)
• Input layer – takes in the data, e.g. text, and feeds it into
1. Feedforward: one-directional, no looping/feedback the next layer

2. Recurrent (includes Long Short-Term Memory, LSTMs): • Hidden layers – in between input and output layers,
good for sequential data like text; have a feedback responsible for data computation and transformation
connection, enabling them to “remember” previous inputs
• Output layer – the final result, the form depends on the
3. Convolutional: layers “convolve” the input, making them initial task – e.g. classification, next word prediction etc
particularly good at handling images
• Neurons – the nodes within layers where weights,
4. Generative Adversarial Networks: consist of two networks activation functions and biases are applied to inputs
that compete and can generate new data
• Parameters (weights and biases) are learned
5. Autoencoders: unsupervised, used for data compression during training

6. Transformers: best for natural language processing (NLP): • Activation functions decide whether a neuron
• Process all pair-relationships between input tokens should be activated or not, adding non-linearity
• Use “self attention” to weigh the relevance of, for
example, different words in a sentence and capture long- • Tokens – unit of text, such as a word or subword, that is
range dependencies in data an input to a language model
• Form basis for models like OpenAI’s GPT, Google’s BERT
and PaLM, Meta’s LLaMA and DeepMind’s Chinchilla
Source: Deutsche Bank.

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 5
5
What was the tipping point for generative AI?
Massive computing power, data and the invention of the transformer model are the oxygen, fuel and heat that have
enabled generative AI to catch fire

Transformer architecture
Transformer model: the Rosetta Stone of SOTA (State of
the Art) large language models

Source: Deutsche Bank. From “Attention Is All You Need”, Vaswani et al, 2017

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 6
6
 How large was the advance in computing power to get us here?
Recent models use several orders of magnitude more processing power, or “compute”, than past models, with more
parameters and more complex model architectures
Computation used to train notable AI systems
Total petaFLOP (10¹⁵ floating-point operations), log scale
GPT-4
100,000,000,000
PaLM (540B) Minerva (540B)
10,000,000,000
GPT-3 175B (davinci) LaMDA
1,000,000,000 AlphaCode Chinchilla
100,000,000 T5-11B
T5-3B
10,000,000
Seq2Seq LSTM GPT-2
1,000,000
HuBERT
100,000 BERT-Large AlexaTM 20B
RNNsearch-50*
10,000

1,000 GPT
Transformer
100 Others

10 Vision
Speech
1
Language
0
Multimodal
0
Jan-10 Feb-11 Mar-12 Apr-13 May-14 Jun-15 Jul-16 Sep-17 Oct-18 Nov-19 Dec-20 Jan-22 Feb-23

Source: Sevilla et al. (2023), Our World in Data, Deutsche Bank. Note: Computation is estimated based on published results in the AI literature and comes with some uncertainty. The
Authors of the study expect the estimates to be correct within a factor of 2.

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 7
7
How many parameters do LLMs have now?
The number of parameters in LLMs, a key measure of size, has surged in the past five years. Even if estimates are true,
GPT-4 will not be the first 1tn parameter model. China’s multimodal Wu Dao 2.0 reportedly has 1.75tn parameters

Number of parameters in B Selected multimodal and language AI systems with more than 50B parameters
2000 Wu Dao 2.0 Multimodal
Language
1800

1600

1400

1200 M6-T [GPT-4?] *

1000

800 PALM
Megatron-Turing
600
Minerva
400
GPT-3 (davinci)
FLAN LaMDA
200 Flamingo LLaMA
NLLB
0
Mar-20 Sep-20 Apr-21 Oct-21 Chinchilla May-22 Dec-22

Source: Sevilla et al. (2023), Our World in Data, Deutsche Bank. *Note GPT-4 parameter count has not been announced; industry estimates that it is around 1 trillion parameters. Parameters are
estimated based on published results in the AI literature and come with some uncertainty. The authors of the study expect the estimates to be correct within a factor of 10.

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 8
8
 Why do more parameters matter?
LLMs show emergent abilities, where capabilities suddenly (and often unexpectedly) appear when the model reaches
a certain size

Capabilities Examples of emergent abilities in the few-shot prompting setting

Parameters

Source: Deutsche Bank, “Emergent Abilities of Large Language Models”, Wei et al. (2022); IPA is International Phonetic Alphabet, NLU is Natural Language Understanding

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 9
9
What are the steps in preparing and using an LLM?
LLMs are pretrained then often fine-tuned on data to acquire a skillset before the user prompts them to perform a task

Huge amounts of
unstructured data
1) Pretraining
(often self-supervised)

LLM
5) Inference Generated
output (text,
images etc)
2) (Typically)
Fine-tuning with
specific labelled 4) Prompt
task-focused data (Context input)
(eg medical, legal)

3) Chatbot overlay 6) Output taken into

sometimes further account as context for
trained with human subsequent prompts in
feedback to improve chat mode
how it behaves

Source: Deutsche Bank.

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 10
10
 Where do LLMs fit in with other technology?
The “tech stack” is built on hardware and cloud computing. LLMs can often support multiple applications which
interact with them via APIs (Application Programming Interfaces). Apps can also be given extra skills with “plug-ins”

Text Image Coding Multimodal

OpenAI/M’soft Google Others

OA DALL-E 2 Github Copilot

Applications ChatGPT Jasper Runway
Bard Stable Diffusion OA Codex
Bing YouChat etc Gen-2 etc
Midjourney etc OA ChatGPT etc

API interface

Meta’s
Foundation BERT Numerous other closed and open
GPT-3 Deep Mind’s LLaMA
model/LLM PaLM source models, some hosted on sites
GPT-4 etc Chinchilla (open
LaMDA like Hugging Face
source)
Google
Cloud Cloud Amazon’s
Azure Alibaba Cloud, Oracle Cloud, IBM Cloud etc
computing Platform AWS
(GCP)

Hardware NVIDIA, AMD, etc GPUs (Graphics Processing Units)

Source: Deutsche Bank.

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 11
11
 How do you use a chatbot?
ChatGPT and its competitors look and feel like search engines but it is more effective to interact with them like
speaking with a human, giving lots of context, and expecting that answers may be more creative than factual

Basic approach Shortened example prompts for various use cases

1. Ask like you’re having a conversation with a person Business

2. Give context: more is better, inc goal, tone of voice etc; • Identify 10 ways a finance company can use AI…
you can paste in relevant text and data (the amount it can • You are a customer service rep. Write an apology for…
handle depends on the model) • Take this data [eg pasted table] and identify themes…
3. Rephrase, correct and clarify if you don’t get the answer • Use this job description and cv to write a cover letter…
you want – you don’t need to restate the whole context • Tell me which angel investors have invested in [name],
give their email addresses and draft emails asking for…
Useful points to include • Write a business plan and slogan for a business that…
Education
“Act as a [ROLE]”: doctor, teacher, legal expert, sales manager… • Summarise War and Peace in 500 words for kids…
“Addressing an [AUDIENCE]”: university student, 7-year-old… • Teach me basic calculus with analogies and a quiz…
“In a [FORMAT]”: 500-word essay, sales pitch, analogy, rhyming Code
poem, social media post, table, code, in the style of Mary Shelley… • Write code to convert a video file into a smaller format…
• Explain the security flaw in this code snippet…
Digging deeper General text and reasoning
• Write a humorous poem welcoming wedding guests…
1. Specify steps: use chained prompting for complex tasks: • I have this [list] in the fridge… Give me a healthy recipe…
ie write an article starting with the outline… [Output]… • Parrot crosses a road. Complete this joke… Why funny?
Now write three key messages for each subheading… etc • Compose a song about a fund manager inc chords…
2. Give sample answers in your question (ie “one-/few-shot • Why might my dog be sick? It has these symptoms…
prompt”) instead of an open-ended “zero-shot prompt” • Write a five-day itinerary in a table format for a 30-year-
3. Ask the chatbot to suggest prompts to achieve your task old wheelchair user visiting Berlin for the first time…

Source: Deutsche Bank.

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 12
12
What are the immediate limitations and concerns?
LLMs are not search engines, though some can be paired with search engines for greater accuracy. Their output
reflects the limitations and biases of their training data, encompassing billions of words and images on the internet

Five limitations and concerns about generative AI and the risks for enterprises

1. Hallucinations Libel risk

2. Bias Reputational risk

3. Data privacy Security, confidentiality risk

4. Unclear sourcing Legal risk, IP/copyright

5. Deliberate misuse Operational risk

Additionally: Limited range of training data, eg GPT-4’s “knowledge cutoff” is Sept 2021

Source: Deutsche Bank, media headlines

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 13
13
What job tasks will be most affected?
Roles will be “unbundled” with a net effect of new work tasks, many higher value; occupations requiring science and
critical skills are less likely to be impacted by LLMs while programming and writing roles are more likely to be impacted

Effect of AI on 13 main tasks performed by

Regression of occupation-level, human-
customer service agents
annotated exposure to GPTs on skill importance
Existing tasks: Post AI:
Programming
Fully automated tasks, New, high-value tasks
no longer exist

Writing

Critical thinking

Exposure *:
Enhanced High
Learning strategies
Medium
Low

Science

Unchanged
-0.4 -0.2 0 0.2 0.4 0.6 0.8

Lower LLM exposure Higher LLM exposure

Source: Customer service example: “Radically Human”, Paul Daugherty, H. James Wilson (2022); “GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language
Models”, Eloundou et al. (2023), Deutsche Bank. Note: Low exposure (alpha in the original study) - lower bound of the share of exposed tasks within a role; medium (beta)– exposure when using
the technology via additional tools and applications needs additional investment; high (zeta) – upper bound of exposure to LLMs. See the paper for in-depth description.

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 14
14
 How are corporates beginning to use generative AI?
Early adopters have often been individuals doing creative tasks and coding; big tech is now integrating AI into existing
products; companies may delay rolling out pure AI solutions until they can ensure quality, reliability and governance

Six busy months of AI announcements and launches Businesses that say they are using chatbots

Nov 2022 Instacart: biggest US online delivery company

30: Open AI launches ChatGPT, built on the GPT-3.5 LLM • Adding ChatGPT to shopping app, so customers can ask
for recommendations for recipes, healthy food etc
Jan 2023 WPP: world’s largest marketing services company
23: Microsoft makes reported $10bn investment in OpenAI • AI is “fundamental” to future business, used to create ad
campaigns and develop personalised advertisements
Feb 2023 Allen & Overy: one of London’s “magic circle” law firms
7: Microsoft launches new GPT-powered Bing search engine • Using Harvey, built using OpenAI’s GPT tech, to help
24: Meta announces launch of lightweight LLaMA LLMs for come up with first drafts of legal documents
researchers; it leaks a few days later and rapidly becomes Zurich Insurance: Swiss insurance company
the basis for numerous open-source uses • Investigating using ChatGPT to extract data from claims
descriptions and write code for statistical models
March 2023 Duolingo: language learning app
14: OpenAI launches GPT-4, said in research to show sparks • Launched Duolingo Max offering tailored language
of Artificial General Intelligence; Anthropic launches Claude lessons and features to roleplay and explain answers
21: Google begins widespread rollout of new Bard chatbot Salesforce: software-as-a-service pioneer
21: Microsoft adds Image Creator to Bing • Introducing ChatGPT to Slack platform to summarise
conversations, answer questions and draft responses
May 2023 Snap: maker of Snapchat app
2: Amazon said to be bringing AI chat capabilities to Alexa • Launched My AI, a customised ChatGPT, for users
10: Google presents AI-powered tools for Workspace inc Mattel: US toymaker, owner of Barbie and Hot Wheels
“Help me write” for emails and image generation for Slides • Using DALL-E 2 image generator to design new toys

Source: Media reports, company announcements, Deutsche Bank.

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 15
15
How fast will things change?
AI text may match the best human writing soonest, followed by code and images, with video games and movies
potentially becoming “personalised dreams” by the end of the decade

Source: Sequoia Capital blog post: “A timeline for how we might expect to see fundamental models progress and the associated applications that become possible. 2025 and beyond is just a guess”

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 16
16
 What are potential bottlenecks to AI development right now?
Some experts say that growth will slow as costs spiral and raw materials such as data and hardware become scarce

1. Access to GPUs 2. Cost of training 3. Environmental costs

There is an acute shortage of GPUs, Estimated training cost of select large CO2 equivalent emissions (tonnes) by
almost all of which are made by Nvidia, for language and multimodal models, $mn * selected machine learnings models and
supercomputers used to train LLMs. Lead real life examples, 2022
times are at least nine months and more BLOOM
usually 18 months for the new H100 chip. GLM-130B
GPT-3 (175B)
The OpenAI supercomputer used to train Minerva (540B)
OPT-175B
GPT-4 reportedly uses 10,000 of the older

2022
Stable Diffusion
Nvidia A100 GPUs. PaLM (540B)
Gopher (280B)
H100s have a retail price of around $10k Chinchilla
but can trade on the grey market for as GPT-NeoX-20B OPT (175B)
much as $40k AlphaCode
Gopher
Car, avg. incl. fuel, 1
Megatron-Turing
lifetime
Nvidia’s stock has almost doubled Codex
since ChatGPT launch ERNIE 3.0
2021

HyperClova BLOOM (176B)

250 Nvidia GPT-J-6B
230 AMD GPT-Neo
Semiconductor index American life, avg., 1 year
210 Wu Dao - Wen Yuan
Nasdaq 100 DALL-E
190
GPT-3 175B Low Human life, avg., 1 year
170
2020

Turing NLG
150 Meena Mid
130 T5-11B Air travel, 1 passenger,
2019

High NY-SF
110 GPT-2
90 0 5 10 15 0 200 400 600
Nov Dec Jan Feb Mar Apr May

Source: Bloomberg Finance LP; 2023 AI Index Report; Luccioni et al., 2022; Strubell et al., 2019; Deutsche Bank. * Note: low refers to potential underestimate, mid refers to what is believed to be
a mid-level estimate and high refers to a potential overestimate.

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 17
17
How do the general public and experts feel about AI?
Concerns outweighed excitement even before the launch of ChatGPT last November

Americans' feelings toward increased use of AI Artificial General Intelligence (AGI) and major
programs in daily life, 2022 risks, according to the natural language
processing (NLP) community, 2022
No answer
80%

% respondents that “agree” or “weakly agree”

70%

More excited 60%

than
concerned 50%

Equally 40%
concerned and
excited 30%

20%
More
concerned 10%
than excited
0%
AI could soon AGI is an Recent AI decisions
lead to important progress is could cause
revolutionary concern moving us nuclear-level
societal otwards AGI catastrophy
change

Source: Pew Research, 2022; Michael et al. 2022; 2023 AI Index Report; Deutsche Bank.

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 18
18
 Are we on the brink of Artificial General Intelligence? And is that good?
Experts differ widely between sceptics, who say AI is still rudimentary, despite appearances, and may never match
human intelligence, and those who say AGI is imminent, for better or worse; there are also many overlaps

“Intelligence, by humans and “A fork is technology. I can use it to

machines, is what fuels progress… eat spaghetti or I can stab you in
I think there’s going to be new the hand… Saying “no more forks”
ideas that are gonna make those means you miss out on all of the
systems much more controllable.” benefits.” Matt Brittin, Google
Yann LeCun, Meta Optimistic
“AI will lead to a low cost and better
quality life for millions of people…
Right now, I don't see a clear path
for AI to surpass human-level
intelligence.” Andrew Ng, AI “AI has much to offer humanity. It
entrepreneur and academic could also wreak terrible harm. It
must be controlled.” Stuart Russell,
“There are lots of dudes in our academic computer scientist
industry who struggle to tell the
difference between science fiction
and real life.” Andrew Feldman, AI
Sceptical Pessimistic
entrepreneur
"Right now, they're not more “If somebody builds a too-powerful
intelligent than us, as far as I can AI, under present conditions, I
tell. But I think they soon may be… expect that every single member of
So we need to worry about that.” the human species and all biological
Geoff Hinton, academic and life on Earth dies shortly thereafter.”
“godfather of AI” Eliezer Yudkowsky, academic

Source: Media reports, Deutsche Bank: NB: Categories are rough and flexible, for indicative purposes only. They reflect emphasis of specific quotes and may not fully reflect views of the speakers.

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 19
19
 What are policymakers doing?
Governments and regulators are playing catchup with fast-evolving AI to set “guardrails” to regulate and oversee it
without discouraging investment and innovation and losing a competitive advantage

Challenges Divergent approaches emerging Notable developments

“Like trying to nail Jell-O to the wall.” US/UK Jan

Bill Clinton, US President, 2000, on • Light touch, risk-based regulation • US Dept of Commerce releases
China’s attempts to crack down on • Applying existing regulations voluntary AI risk framework
the internet • No new regulatory bodies March
• Sector-specific, focusing on use of • More than 1,000 tech leaders inc
1) Technology still emerging AI rather than the technology itself Elon Musk publish letter calling for
“pause” in AI development
• Hard to predict future risk and European Union • Italy temporarily bans ChatGPT
benefits; new laws may be • Multifaceted inc specific legislation • UK publishes AI policy paper calling
obsolete by the time are enacted • Proposed AI Act in legislative for “pro-innovation approach”
process is detailed and prescriptive, April
2) Lack of deep understanding eg classifying AI uses by degrees of • China’s internet regulator proposes
risk; minimum requirements for AI new rules applying to generative AI
• Asymmetry between lawmakers/ systems; new national/EU bodies • Reports that US Senator Schumer
public and industry/academia • Builds on tech rules such as GDPR is developing AI regulation plan
May
3) Cross-sectoral application China • OpenAI CEO Sam Altman appears
• Piecemeal approach at US Senate hearing
• AI affects multiple industries with • Focus on security, eg: mandatory • G7 leaders call for AI “guardrails”
varying risks and impacts security assessment for new Upcoming…
products; must meet China’s values • Late 2023: EU AI Act set to pass

Source: Government and regulatory announcements, media reports, Deutsche Bank.

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 20
20
What happens next?
The rapid pace of eye-catching launches is likely to slow as the focus turns to refining and integrating the new tech into
infrastructure, governance and culture; industry and governments will need to work together to avoid a backlash

Users Regulators and society

• Individuals: rapid adoption for personal/semi-professional use Widening anxiety threatening a
• Companies: FOMO (fear of missing out) balanced by governance concerns; potential backlash with concerns about:
increasing use in customer services, data retrieval, content creation, coding • Jobs: some white-collar jobs will be
• General: increasingly sophisticated use of tech (eg plugins) and techniques enhanced while others may be at risk
(eg chained prompts through eg AutoGPT for complex multi-step tasks) • Education: eg UK teacher coalition
warned in May of “severe” threat
• Privacy: Spain, Canada and others
App developers have considered or started probes
• Explosion in new applications built on top of LLMs • Truth: US presidential election in
• Trying out of open-source models, fine-tuned and stacked for specific tasks 2024 will be a litmus test of AI-
enabled hyper-personalised
campaigning and potential
LLM developers misinformation by “bad actors”
• Race to offer integrated enterprise solutions with reliable, ringfenced data • Ethics: “alignment” to human values
• Further advances in 1-trillion+ parameter multimodal LLMs, notably GPT-5, Pressure to legislate and retrain
rumoured to finish training by year-end and to have video capabilities • Policymakers caught between desire
• Parallel advances in open-source lightweight alternative LLMs that offer for popular precautions and a “race
similar performance to large models, trained more efficiently and cheaply with to the bottom” to encourage home-
pruned datasets, and that users can run on their own hardware grown development and use
• Governments will be urged to
provide financial and educational
Hardware safety nets for those displaced by AI
• Ongoing pressure on supplies of key GPUs and networking equipment to mitigate social upheaval
Source: Deutsche Bank.

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 21
21
Appendix 1
Important Disclosures
*Other information available upon request

*Prices are current as of the end of the previous trading session unless otherwise indicated and are sourced from local exchanges
via Reuters, Bloomberg and other vendors. Other information is sourced from Deutsche Bank, subject companies, and other
sources. For further information regarding disclosures relevant to Deutsche Bank Research, please visit our global disclosure look-
up page on our website at https://research.db.com/Research/Disclosures/FICCDisclosures. Aside from within this report, important
risk and conflict disclosures can also be found at https://research.db.com/Research/Disclosures/Disclaimer. Investors are strongly
encouraged to review this information before investing.

Analyst Certification
The views expressed in this report accurately reflect the personal views of the undersigned lead analyst about the subject issuers and the securities of those
issuers. In addition, the undersigned lead analyst has not and will not receive any compensation for providing a specific recommendation or view in this report.
AdrIan Cox, Jim Reid, Galina Pozdnyakova

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 22
22
Additional Information

The information and opinions in this report were prepared by Deutsche Bank AG or one of its affiliates (collectively "Deutsche Bank"). Though the information herein is believed to be reliable and has been
obtained from public sources believed to be reliable, Deutsche Bank makes no representation as to its accuracy or completeness. Hyperlinks to third-party websites in this report are provided for reader
convenience only. Deutsche Bank neither endorses the content nor is responsible for the accuracy or security controls of those websites.

If you use the services of Deutsche Bank in connection with a purchase or sale of a security that is discussed in this report, or is included or discussed in another communication (oral or written) from a Deutsche
Bank analyst, Deutsche Bank may act as principal for its own account or as agent for another person.

Deutsche Bank may consider this report in deciding to trade as principal. It may also engage in transactions, for its own account or with customers, in a manner inconsistent with the views taken in this research
report. Others within Deutsche Bank, including strategists, sales staff and other analysts, may take views that are inconsistent with those taken in this research report. Deutsche Bank issues a variety of research
products, including fundamental analysis, equity-linked analysis, quantitative analysis and trade ideas. Recommendations contained in one type of communication may differ from recommendations contained in
others, whether as a result of differing time horizons, methodologies, perspectives or otherwise. Deutsche Bank and/or its affiliates may also be holding debt or equity securities of the issuers it writes on.
Analysts are paid in part based on the profitability of Deutsche Bank AG and its affiliates, which includes investment banking, trading and principal trading revenues.

Opinions, estimates and projections constitute the current judgment of the author as of the date of this report. They do not necessarily reflect the opinions of Deutsche Bank and are subject to change without
notice. Deutsche Bank provides liquidity for buyers and sellers of securities issued by the companies it covers. Deutsche Bank research analysts sometimes have shorter-term trade ideas that may be inconsistent
with Deutsche Bank's existing longer-term ratings. Some trade ideas for equities are listed as Catalyst Calls ideas on the Research Website (https://research.db.com/Research/), and can be found on the general
coverage list and also on the covered company’s page. A Catalyst Call idea represents a high-conviction belief by an analyst that a stock will outperform or underperform the market and/or a specified sector over
a time frame of no less than two weeks and no more than three months. In addition to Catalyst Calls ideas, analysts may occasionally discuss with our clients, and with Deutsche Bank salespersons and traders,
trading strategies or ideas that reference catalysts or events that may have a near-term or medium-term impact on the market price of the securities discussed in this report, which impact may be directionally
counter to the analysts' current 12-month view of total return or investment return as described herein. Deutsche Bank has no obligation to update, modify or amend this report or to otherwise notify a recipient
thereof if an opinion, forecast or estimate changes or becomes inaccurate. Coverage and the frequency of changes in market conditions and in both general and company-specific economic prospects make it
difficult to update research at defined intervals. Updates are at the sole discretion of the coverage analyst or of the Research Department Management, and the majority of reports are published at irregular
intervals. This report is provided for informational purposes only and does not take into account the particular investment objectives, financial situations, or needs of individual clients. It is not an offer or a
solicitation of an offer to buy or sell any financial instruments or to participate in any particular trading strategy. Target prices are inherently imprecise and a product of the analyst’s judgment. The financial
instruments discussed in this report may not be suitable for all investors, and investors must make their own informed investment decisions. Prices and availability of financial instruments are subject to change
without notice, and investment transactions can lead to losses as a result of price fluctuations and other factors. If a financial instrument is denominated in a currency other than an investor's currency, a change
in exchange rates may adversely affect the investment. Past performance is not necessarily indicative of future results. Performance calculations exclude transaction costs, unless otherwise indicated. Unless
otherwise indicated, prices are current as of the end of the previous trading session and are sourced from local exchanges via Reuters, Bloomberg and other vendors. Data is also sourced from Deutsche Bank,
subject companies, and other parties.

The Deutsche Bank Research Department is independent of other business divisions of the Bank. Details regarding our organizational arrangements and information barriers we have to prevent and avoid
conflicts of interest with respect to our research are available on our website (https://research.db.com/Research/) under Disclaimer.

Macroeconomic fluctuations often account for most of the risks associated with exposures to instruments that promise to pay fixed or variable interest rates. For an investor who is long fixed-rate instruments
(thus receiving these cash flows), increases in interest rates naturally lift the discount factors applied to the expected cash flows and thus cause a loss. The longer the maturity of a certain cash flow and the
higher the move in the discount factor, the higher will be the loss. Upside surprises in inflation, fiscal funding needs, and FX depreciation rates are among the most common adverse macroeconomic shocks to
receivers. But counterparty exposure, issuer creditworthiness, client segmentation, regulation (including changes in assets holding limits for different types of investors), changes in tax policies, currency
convertibility (which may constrain currency conversion, repatriation of profits and/or liquidation of positions), and settlement issues related to local clearing houses are also important risk factors. The sensitivity
of fixed-income instruments to macroeconomic shocks may be mitigated by indexing the contracted cash flows to inflation, to FX depreciation, or to specified interest rates – these are common in emerging
markets. The index fixings may – by construction – lag or mis-measure the actual move in the underlying variables they are intended to track. The choice of the proper fixing (or metric) is particularly important in
swaps markets, where floating coupon rates (i.e., coupons indexed to a typically short-dated interest rate reference index) are exchanged for fixed coupons. Funding in a currency that differs from the currency in
which coupons are denominated carries FX risk. Options on swaps (swaptions) the risks typical to options in addition to the risks related to rates movements.

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 23
23
 Additional Information

Derivative transactions involve numerous risks including market, counterparty default and illiquidity risk. The appropriateness of these products for use by investors depends on the investors' own circumstances,
including their tax position, their regulatory environment and the nature of their other assets and liabilities; as such, investors should take expert legal and financial advice before entering into any transaction
similar to or inspired by the contents of this publication. The risk of loss in futures trading and options, foreign or domestic, can be substantial. As a result of the high degree of leverage obtainable in futures and
options trading, losses may be incurred that are greater than the amount of funds initially deposited – up to theoretically unlimited losses. Trading in options involves risk and is not suitable for all investors. Prior
to buying or selling an option, investors must review the "Characteristics and Risks of Standardized Options”, at http://www.optionsclearing.com/about/publications/character-risks.jsp. If you are unable to
access the website, please contact your Deutsche Bank representative for a copy of this important document.

Participants in foreign exchange transactions may incur risks arising from several factors, including the following: (i) exchange rates can be volatile and are subject to large fluctuations; (ii) the value of currencies
may be affected by numerous market factors, including world and national economic, political and regulatory events, events in equity and debt markets and changes in interest rates; and (iii) currencies may be
subject to devaluation or government-imposed exchange controls, which could affect the value of the currency. Investors in securities such as ADRs, whose values are affected by the currency of an underlying
security, effectively assume currency risk.

Unless governing law provides otherwise, all transactions should be executed through the Deutsche Bank entity in the investor's home jurisdiction. Aside from within this report, important conflict disclosures
can also be found at https://research.db.com/Research/ on each company’s research page. Investors are strongly encouraged to review this information before investing.

Deutsche Bank (which includes Deutsche Bank AG, its branches and affiliated companies) is not acting as a financial adviser, consultant or fiduciary to you or any of your agents (collectively, “You” or “Your”)
with respect to any information provided in this report. Deutsche Bank does not provide investment, legal, tax or accounting advice, Deutsche Bank is not acting as your impartial adviser, and does not express
any opinion or recommendation whatsoever as to any strategies, products or any other information presented in the materials. Information contained herein is being provided solely on the basis that the recipient
will make an independent assessment of the merits of any investment decision, and it does not constitute a recommendation of, or express an opinion on, any product or service or any trading strategy.

The information presented is general in nature and is not directed to retirement accounts or any specific person or account type, and is therefore provided to You on the express basis that it is not advice, and You
may not rely upon it in making Your decision. The information we provide is being directed only to persons we believe to be financially sophisticated, who are capable of evaluating investment risks
independently, both in general and with regard to particular transactions and investment strategies, and who understand that Deutsche Bank has financial interests in the offering of its products and services. If
this is not the case, or if You are an IRA or other retail investor receiving this directly from us, we ask that you inform us immediately.

In July 2018, Deutsche Bank revised its rating system for short term ideas whereby the branding has been changed to Catalyst Calls (“CC”) from SOLAR ideas; the rating categories for Catalyst Calls originated in
the Americas region have been made consistent with the categories used by Analysts globally; and the effective time period for CCs has been reduced from a maximum of 180 days to 90 days.

United States: Approved and/or distributed by Deutsche Bank Securities Incorporated, a member of FINRA, NFA and SIPC. Analysts located outside of the United States are employed by non-US affiliates that
are not subject to FINRA regulations.

European Economic Area (exc. United Kingdom): Approved and/or distributed by Deutsche Bank AG, a joint stock corporation with limited liability incorporated in the Federal Republic of Germany with its
principal office in Frankfurt am Main. Deutsche Bank AG is authorized under German Banking Law and is subject to supervision by the European Central Bank and by BaFin, Germany’s Federal Financial
Supervisory Authority.

United Kingdom: Approved and/or distributed by Deutsche Bank AG acting through its London Branch at Winchester House, 1 Great Winchester Street, London EC2N 2DB. Deutsche Bank AG in the United
Kingdom is authorised by the Prudential Regulation Authority and is subject to limited regulation by the Prudential Regulation Authority and Financial Conduct Authority. Details about the extent of our
authorisation and regulation are available on request.

Hong Kong SAR: Distributed by Deutsche Bank AG, Hong Kong Branch except for any research content relating to futures contracts within the meaning of the Hong Kong Securities and Futures Ordinance Cap.
571. Research reports on such futures contracts are not intended for access by persons who are located, incorporated, constituted or resident in Hong Kong. The author(s) of a research report may not be
licensed to carry on regulated activities in Hong Kong and, if not licensed, do not hold themselves out as being able to do so. The provisions set out above in the “Additional Information” section shall apply to the
fullest extent permissible by local laws and regulations, including without limitation the Code of Conduct for Persons Licensed or Registered with the Securities and Futures Commission. This report is intended
for distribution only to “professional investors” as defined in Part 1 of Schedule 1 of the SFO. This document must not be acted or relied on by persons who are not professional investors. Any investment or
investment activity to which this document relates is only available to professional investors and will be engaged only with professional investors.

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 24
24
Additional Information

India: Prepared by Deutsche Equities India Private Limited (DEIPL) having CIN: U65990MH2002PTC137431 and registered office at 14th Floor, The Capital, C-70, G Block, Bandra Kurla Complex Mumbai (India)
400051. Tel: + 91 22 7180 4444. It is registered by the Securities and Exchange Board of India (SEBI) as a Stock broker bearing registration no.: INZ000252437; Merchant Banker bearing SEBI Registration no.:
INM000010833 and Research Analyst bearing SEBI Registration no.: INH000001741. DEIPL may have received administrative warnings from the SEBI for breaches of Indian regulations. Deutsche Bank and/or its
affiliate(s) may have debt holdings or positions in the subject company. With regard to information on associates, please refer to the “Shareholdings” section in the Annual Report at:
https://www.db.com/ir/en/annual-reports.htm

Japan: Approved and/or distributed by Deutsche Securities Inc.(DSI). Registration number - Registered as a financial instruments dealer by the Head of the Kanto Local Finance Bureau (Kinsho) No. 117. Member
of associations: JSDA, Type II Financial Instruments Firms Association and The Financial Futures Association of Japan. Commissions and risks involved in stock transactions - for stock transactions, we charge
stock commissions and consumption tax by multiplying the transaction amount by the commission rate agreed with each customer. Stock transactions can lead to losses as a result of share price fluctuations
and other factors. Transactions in foreign stocks can lead to additional losses stemming from foreign exchange fluctuations. We may also charge commissions and fees for certain categories of investment
advice, products and services. Recommended investment strategies, products and services carry the risk of losses to principal and other losses as a result of changes in market and/or economic trends, and/or
fluctuations in market value. Before deciding on the purchase of financial products and/or services, customers should carefully read the relevant disclosures, prospectuses and other documentation. "Moody's",
"Standard & Poor's", and "Fitch" mentioned in this report are not registered credit rating agencies in Japan unless Japan or "Nippon" is specifically designated in the name of the entity. Reports on Japanese
listed companies not written by analysts of DSI are written by Deutsche Bank Group's analysts with the coverage companies specified by DSI. Some of the foreign securities stated on this report are not
disclosed according to the Financial Instruments and Exchange Law of Japan. Target prices set by Deutsche Bank's equity analysts are based on a 12-month forecast period.

Korea: Distributed by Deutsche Securities Korea Co.

South Africa: Deutsche Bank AG Johannesburg is incorporated in the Federal Republic of Germany (Branch Register Number in South Africa: 1998/003298/10).

Singapore: This report is issued by Deutsche Bank AG, Singapore Branch (One Raffles Quay #18-00 South Tower Singapore 048583, +65 6423 8001), which may be contacted in respect of any matters arising
from, or in connection with, this report. Where this report is issued or promulgated by Deutsche Bank in Singapore to a person who is not an accredited investor, expert investor or institutional investor (as
defined in the applicable Singapore laws and regulations), they accept legal responsibility to such person for its contents.

Taiwan: Information on securities/investments that trade in Taiwan is for your reference only. Readers should independently evaluate investment risks and are solely responsible for their investment decisions.
Deutsche Bank research may not be distributed to the Taiwan public media or quoted or used by the Taiwan public media without written consent. Information on securities/instruments that do not trade in
Taiwan is for informational purposes only and is not to be construed as a recommendation to trade in such securities/instruments. Deutsche Securities Asia Limited, Taipei Branch may not execute transactions
for clients in these securities/instruments.

Qatar: Deutsche Bank AG in the Qatar Financial Centre (registered no. 00032) is regulated by the Qatar Financial Centre Regulatory Authority. Deutsche Bank AG - QFC Branch may undertake only the financial
services activities that fall within the scope of its existing QFCRA license. Its principal place of business in the QFC: Qatar Financial Centre, Tower, West Bay, Level 5, PO Box 14928, Doha, Qatar. This information
has been distributed by Deutsche Bank AG. Related financial products or services are only available only to Business Customers, as defined by the Qatar Financial Centre Regulatory Authority.

Russia: The information, interpretation and opinions submitted herein are not in the context of, and do not constitute, any appraisal or evaluation activity requiring a license in the Russian Federation.

Kingdom of Saudi Arabia: Deutsche Securities Saudi Arabia (DSSA) is a closed joint stock company authorized by the Capital Market Authority of the Kingdom of Saudi Arabia with a license number (No. 37-
07073) to conduct the following business activities: Dealing, Arranging, Advising, and Custody activities. . DSSA registered office is at Faisaliah Tower, 17th floor, King Fahad Road - Al Olaya District Riyadh,
Kingdom of Saudi Arabia P.O. Box 301806.

United Arab Emirates: Deutsche Bank AG in the Dubai International Financial Centre (registered no. 00045) is regulated by the Dubai Financial Services Authority. Deutsche Bank AG - DIFC Branch may only
undertake the financial services activities that fall within the scope of its existing DFSA license. Principal place of business in the DIFC: Dubai International Financial Centre, The Gate Village, Building 5, PO Box
504902, Dubai, U.A.E. This information has been distributed by Deutsche Bank AG. Related financial products or services are available only to Professional Clients, as defined by the Dubai Financial Services
Authority.

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 25
25
Additional Information

Australia and New Zealand: This research is intended only for "wholesale clients" within the meaning of the Australian Corporations Act and New Zealand Financial Advisors Act, respectively. Please refer to
Australian specific research disclosures and related information at https://www.dbresearch.com/PROD/RPS_EN-PROD/PROD0000000000521304.xhtml . Where research refers to any particular financial product
recipients of the research should consider any product disclosure statement, prospectus or other applicable disclosure document before making any decision about whether to acquire the product. In preparing
this report, the primary analyst or an individual who assisted in the preparation of this report has likely been in contact with the company that is the subject of this research for confirmation/clarification of data,
facts, statements, permission to use company-sourced material in the report, and/or site-visit attendance. Without prior approval from Research Management, analysts may not accept from current or potential
Banking clients the costs of travel, accommodations, or other expenses incurred by analysts attending site visits, conferences, social events, and the like. Similarly, without prior approval from Research
Management and Anti-Bribery and Corruption (“ABC”) team, analysts may not accept perks or other items of value for their personal use from issuers they cover.

Additional information relative to securities, other financial products or issuers discussed in this report is available upon request. This report may not be reproduced, distributed or published without Deutsche
Bank's prior written consent.

Backtested, hypothetical or simulated performance results have inherent limitations. Unlike an actual performance record based on trading actual client portfolios, simulated results are achieved by means of the
retroactive application of a backtested model itself designed with the benefit of hindsight. Taking into account historical events the backtesting of performance also differs from actual account performance
because an actual investment strategy may be adjusted any time, for any reason, including a response to material, economic or market factors. The backtested performance includes hypothetical results that do
not reflect the reinvestment of dividends and other earnings or the deduction of advisory fees, brokerage or other commissions, and any other expenses that a client would have paid or actually paid. No
representation is made that any trading strategy or account will or is likely to achieve profits or losses similar to those shown. Alternative modeling techniques or assumptions might produce significantly
different results and prove to be more appropriate. Past hypothetical backtest results are neither an indicator nor guarantee of future returns. Actual results will vary, perhaps materially, from the analysis.

The method for computing individual E,S,G and composite ESG scores set forth herein is a novel method developed by the Research department within Deutsche Bank AG, computed using a systematic
approach without human intervention. Different data providers, market sectors and geographies approach ESG analysis and incorporate the findings in a variety of ways. As such, the ESG scores referred to
herein may differ from equivalent ratings developed and implemented by other ESG data providers in the market and may also differ from equivalent ratings developed and implemented by other divisions within
the Deutsche Bank Group. Such ESG scores also differ from other ratings and rankings that have historically been applied in research reports published by Deutsche Bank AG. Further, such ESG scores do not
represent a formal or official view of Deutsche Bank AG.
It should be noted that the decision to incorporate ESG factors into any investment strategy may inhibit the ability to participate in certain investment opportunities that otherwise would be consistent with your
investment objective and other principal investment strategies. The returns on a portfolio consisting primarily of sustainable investments may be lower or higher than portfolios where ESG factors, exclusions, or
other sustainability issues are not considered, and the investment opportunities available to such portfolios may differ. Companies may not necessarily meet high performance standards on all aspects of ESG or
sustainable investing issues; there is also no guarantee that any company will meet expectations in connection with corporate responsibility, sustainability, and/or impact performance.
Copyright © 2023 Deutsche Bank AG

Deutsche Bank Research | Adrian Cox (+44) 20 7541 7775 | May 2023 26
26