1 of 25

GenAI, quantum computing, and the power of convergence

DECEMBER 2023

1

2 of 25

Topics for today's session

The value of GenAI for business

The convergence potential of quantum computing + GenAI

The roadblocks preventing it from scaling

1

3

2

Matt Langione
Managing Director & Partner
Boston Consulting Group
langione.matt@bcg.com

2

3 of 25

The power of generative AI (I/II)

Simulated exams	Score	Percentile
Bar Exam (MBE+MEE+MPT)¹	298/400	90^th
LSAT	163/180	88^th
SAT Reading & Writing	710/800	93^rd
SAT Math	700/800	89^th
GRE Quant	163/170	80^th
GRE Verbal	169/170	99^th
GRE Writing	4/6	54^th
USABO Semifinal Exam 2020	87/150	99-100^th
AP Art History	5/5	86^th - 100^th
AP Biology	5/5	85^th-100^th
AP Calculus BC	4/5	43^rd – 59^th

The value of GenAI

Generative AI is the most viral technology of all time…

… & the first to generalize human-level intelligence

0

25

50

75

100

125

150

0

400,000

800,000

1,200,000

# of days

Instagram

~75 days

ChatGPT

~5 days

Spotify

~150 days

# Users

GPT-4 Test Results

3

It's hard to imagine that anybody in this room hasn't heard of generative AI. It burst onto the scene, at least in the public sphere, one year ago when it became the most viral technology of all time, getting to 1M users in a mere 5 days after OpenAI released ChatGPT, an online conversation bot that seemed to instantly democratize access to the power of artificial intelligence.

Just like the graphic user interface enabled billions of non-programmers to harness the power of computers to digitize first-order analog tasks like typewriting and record-keeping a couple generations earlier, ChatGPT and its ilk provides a linguistic user interface that enables users, equipped with nothing more than plain language skills, to automate second-order cognitive tasks, like those many of us do at work every day.

Indeed, in all likelihood, it already outscores your lawyer on the Bar Exam or LSAT…

4 of 25

The power of generative AI (II/II)

What mood does the attached work of art convey? Why? What aesthetic choices does the artist make that contribute to the mood?

"The image portrays a sense of anguish and despair, which is most prominently conveyed through the central figure's open-mouthed scream. The swirling, tumultuous background further intensifies the emotional turmoil depicted in the scene. Here are some of the aesthetic choices the artist makes that contribute to the mood.

1. Color Palette: The artist uses a bold color palette with stark contrasts. The fiery red and orange tones in the sky evoke a sense of danger and urgency which is juxtaposed against the cool, eerie blue of the water below.

2. Brushwork: The swirling, almost chaotic brush strokes contribute to the unsettled, turbulent atmosphere of the scene. The lack of clear definition and the fluidity of the forms make the surroundings seem unstable and dreamlike, heightening the sense of disorientation and distress.

3. Expression: The central figure's facial expression is on of terror and desperation. The open mouth, wide eyes, and hands-on the sides of the face are universal signs of fear and horror. This expressive portrayal captures the viewer's attention and communications the emotional essence of the scene.

4. Composition: The curved, flowing lines of the landscape and the figure's elongated form create a sense of movement and fluidity, further contributing to the unsettling mood. The bridge and the two figures in the background provide a sense of depth and isolation emphasizing the central figure's alienation and despair.

5. Perspective: The exaggerated perspective, with the bridge receding sharply into the background, amplifies the sense of depth and distance, making the central figure appear trapped or enclosed within the foreground.

These elements work together to create a dramatic, emotionally charged atmosphere that encapsulates the existential angst and unease that is characteristic of Edvard Munch's iconic work, The Scream."

ChatGPT response

The value of GenAI

4

One may debate whether it's truly intelligent—whether it fulfills John Searle's criteria for "strong AI"—or not, but there's no debating that its outputs **reflect** contextual, semantic understanding of even the most human and unstructured inputs. Here we have Edvard Munch's "The Scream," an emotionally charged if perhaps unsubtle painting. I've uploaded the image to ChatGPT and asked about the aesthetic choices the artist made that contribute to the painting's mood.

Within 20 seconds it not only identifies the painting but outlines five choices, from color palette to brushwork to the use of perspective, where it makes the interesting observation that the bridge's sharp recession into the background "amplifies the sense of depth and distance, making the central figure appear trapped or enclosed within the foreground" as if there is no exit.

Let's not relieve the curator of his duties yet, but nor should we deny that this is real cognitive work requiring as much "general" intelligence as most of our day jobs.

5 of 25

GenAI today primarily used for enterprise operations

Source: June 2023 AI Adoption survey, N=125; BCG analysis

Value creation with GenAI today

FMs can be used to automate call centers, but we are going one step further, creating predicative call and text functionality that allows us to know ahead of time why customers are calling

We are using FMs in drug discovery to assist with gene sequencing and small molecule simulation to speed up pre-clinical studies and selection of leads for clinical studies

Operational efficiency use cases

(e.g. extracting clauses from legal documents)

Product & CX enhancement

(e.g. financial robo-advisor)

New products and business models

(e.g. GenAI drug discovery)

% of value creation by use case type

The value of GenAI

We are enhancing our chatbots with GPT trained on 30 years of documentation…it's like having our CSO sitting next to you when you're on a phone with a client.

$100B+ market size for tech providers by 2027

5

6 of 25

5 GenAI capabilities that drive value for enterprises

Tech. capabilities

Description

Illustrative use cases

Note: FM = foundation model

1. GenAI transformations can leverage multiple tech. capabilities (e.g., ChatGPT leverages content generation & creativity)

Interactive �& dynamic engagement of information, ideas, or questions between humans �& AI systems, responding to questions and generating appropriate responses

Chatbot for HCPs interactions
Internal chatbot for employees

UNDERSTAND

language / image

Extract structured knowledge from unstructured or semi-structured
data sources

Clinical data extraction from protocols or patient records

EXTRACT

knowledge

Summarize large amounts of information or text into shorter, more concise versions, that capture the key points of the content

Documents key take-aways
Content aggregator & syntheses

SUMMARIZE

& transcribe

Create specific types of content (e.g., text, images, videos, audio, code), and generate new and innovative ideas, concepts or designs

Meeting minutes writing & transcription
Marketing context localization
Innovative product concept generation

GENERATE

content

ANALYZE

data & text

Logical & reasoning process to make inferences
and draw conclusions, make informed judgments, derive new insights based on available information, data, or knowledge

Sentiment analysis

The value of GenAI

6

7 of 25

Model performance scales linearly with model size…

Expert systems

Machine Learning

Deep Learning

Large pre-trained (foundation) models
Generative, Multi-modal Neural Networks, Transfer Learning techniques

175 B

10k

0

10 M

Time

Power�model parameters

GPT-3

1.7 T

GPT-4¹

…

Ability to complete multiple tasks on text format

Ability to reason and control other systems

Ability to complete multiple tasks on multiple formats (texts, image, audio, video)

…but what made it possible?

The value of GenAI

7

8 of 25

The rise of GenAI is fueled by three primary factors

1

2

3

Algorithm advances ("attention is all you need")

Exponential increase in available training data

Exponential increase in compute power

The value of GenAI

8

9 of 25

Algorithm advances | Transformer models

Transformer Model Overview

First described in 2017 by Google researchers in a paper titled "Attention is All You Need"

Transformers are a neural network that can apply mathematical techniques called "attention" to detect how data may be related to each other

Transformers are now the dominant models used to "pre-train" foundation models

Source: Attention is all you need, Proceeding of NIPS (2017), pp. 5998-6008; https://blogs.nvidia.com/blog/2022/03/25/what-is-a-transformer-model/

"Attention" removes need for labelled data

Before transformers, neural networks were trained on data labelled by humans (e.g., this image is a cat), which is expensive and a huge limitation to training data availability.

With transformers and "attention", data labelling no longer needed as model automatically understands context. This allows model to be fed huge amounts of training data with no human labelling, which is key to scaling

"Attention" enables learning around context

With "attention" mechanisms, transformers can tease out subtle ways different parts of the data are related through self-learning (e.g., "cats" are related to "pets")

This enables the model to learn and understand context

"Attention" gives rise to generative content

Using both vast amounts of training data, and learning around context, transformers now had generative ability.

The model effectively connected distinct concepts to create new content (e.g., images, text, video etc.)

The value of GenAI

9

10 of 25

Digital data growth | 10 to 120 exabytes in 10 yrs�80% is unstructured, requiring "attention"

0

60

120

Stored Digital Data (exabytes)

1970

1980

1990

2000

2010

2020

Requires “attention mechanism” to avoid human labeling bottleneck

80%+

of digitized data

Unstructured�no data model

Structured�well-defined, easily-organized database information

The value of GenAI

10

11 of 25

Compute | Exponential increase via Moore's Law

Larger models perform better than smaller, pushing practitioners to build larger models…

Human ability to detect if AI generated news article

Error

Model parameters

100M

10B

100B

1B

50% line of random chance = human cannot distinguish AI generated news articles from human generated ones

Source: Brown Tom B. et al., 2020. Language models are few-shot learners; "In-Datacenter Performance Analysis of a Tensor Processing Unit", Google

1960

2000

1970

2010

1980

1990

2020

1,000,000

1

100

10,000

100,000,000

Year

Transistor Density (Transistors mm^-2)

Moore's law provides exponential advances in compute power

The value of GenAI

11

12 of 25

Each enabler also a roadblock for scaling GenAI

Transformer models are "greedy" and inefficient (as of now)

175B parameter model scored bottom 10% on bar exam, 1.7T model top 10%
$1B+ training costs for largest models

Roadblock

Evidence

We are running out of training data to feed the models

We are running out of compute resources to power the models

So far, each gen of model requires 10-100X more compute
Current models are trained on superclusters for months
The next iteration that can show visible improvement (like GPT3 to 4) may require a decade to train, or 100 superclusters

Roadblocks for GenAI

Error rates, currently at 20%, cut in half with every order of magnitude increase in training data
Available internet data in principle, incl. private data, permits only two more orders of magnitude

12

13 of 25

Key roadblocks | Will we run out of data?

…cutting error rates in half

(Tokens from the internet can increase from ~10¹² to ~10¹⁴ and reduce model error from ~20% to ~10%)

Source	Tokens in�current models	Tokens accessible�in principle
Internet	~10¹²	5 x 10¹⁴
Books	5 x 10¹¹	10¹³
Wikipedia (English)	6.5 x 10⁹	6.5 x 10⁹
Wikipedia (All)	2.5 x 10¹⁰	3.9 x 10¹⁰
Scientific papers	2.7 x 10¹⁰	1.5 x 10¹²
Twitter	Few	2.5 x 10¹³
Text Messages	0	10¹²/year
Youtube	0	4 x 10¹²

80%+ of model improvement due to amount of data vs. model design

Available internet training data can grow another 2 orders of magnitude …

Source: dynomight.net, BCG analysis

Roadblocks for GenAI

13

The data problem is one of quantity and of quality. The latest GPT was trained on ~10^12 tokens. If the scaling laws obtain, ever order of magnitude increase in data will lead to somewhere between a ¼ and 1/3 reduction in normalized total model error. What this means is that it will take another two orders of magnitude to get to human levels of error (which are about 8-10%). The problem is that the models have exhausted all of the public tokens on the internet.

You might object, but there's this massive proliferation of internet data being collected every day. That's where we come to the quality dilemma. The early data sources that fed the models were books from Google Books and Project Gutenburg, Wikipedia and the like. Much of the untapped sources are things like Twitter, text messages and Youtube.

14 of 25

Key roadblocks | Will we run out of compute?

Meta RSC AI supercluster

6,080 Nvidia's latest A100 GPUs

~1.9 exaflop/s (10¹⁸float-point operations per second)

If you were training on the largest AI supercluster…

~15 hours

~60 days

~16 YEARS

(or 600,000 GPUs)

GPT-3

GPT-4

Another 100X increase*

* GPT-4 uses ~100X computing power in training vs. GPT-3

Roadblocks for GenAI

14

15 of 25

Convergence possibilities for Quantum & GenAI

Quantum Computer x GenAI to work side by side

Maximize the power of GenAI and quantum computing in a hybrid workflow by using them to handle different tasks

Quantum Computing to solve challenges in scaling GenAI

Enhance performance with "quantum-inspired" classical algorithms
Run models on quantum computers / quantum data
Generate synthetic training data (with true randomness)

Enhance error correction
Improve quantum hardware design
Augment quantum software development
Speed up calibration process

GenAI accelerating progress in Quantum Computing

Convergence potential

A

C

B

15

16 of 25

Convergence | QC to resolve GenAI roadblocks

Transformer models are "greedy" and inefficient

Quantum-inspired techniques to improve efficiency (e.g., QNN, simulated annealing, tensor networks for LLM)

Roadblock

Use case

We are running out of training data to feed the models

Synthetic data with quantum random number generator

Using quantum data to reduce sample complexity

We are running out of compute resources to power the models

Run model on quantum machine leveraging vast state space

Example research in the field

A

Quantinuum (CQC) is researching quantum natural language processing to reduce the size of neural networks (e.g. training datasets for LLMs)

IonQ has partnered with FCAT to develop a quantum machine learning (QML) proof of concept for correlation in the finance industry using synthetic data from a quantum AI model

Zapata researchers explored substituting parts of MolGAN with variational quantum circuits (VQCs), resulting in quantum GANs surpassing classical GANs in performance

Convergence potential

16

The golden goose would be to redesign the transformer architecture either to run on quantum hardware, or leverage quantum-inspired techniques on classical hardware. Researchers at CQC/Quantinuum are looking at ways to reduce the size of neural networks this way. That would of course address the problem at its core.

Another way to approach this would be to address the data challenge. Synthetic data is often mooted as the answer to the problem of scarce high-quality data to train larger models. But synthetic data lacks intrinsic randomness, which allows a discriminator module to easily distinguish it from real data, therefore reducing its utility. IonQ is already leveraging quantum synthetic data in real world use cases in partnership with FCAT.

The final and perhaps most ambitious convergence mode is to actually run a model on quantum computer. This would supervene on or combine with some of what I've mentioned already because of course tranfsormers are not expected to run efficiently on quantum computers. Zapata is working on a number of research areas at the convergence of quantum and GenAI, but among them are substituting parts of MolGAN with VQC type models to improve performance and efficieincy.

These are speculative use cases, but ones I'd encourage this community to explore as the value creation potential is both massive and concrete.

17 of 25

Convergence | GenAI to accelerate QC timeline

Potential to impact fundamental QC algo research?

qecGPT Project¹: A framework called qecGPT has been proposed for decoding quantum error-correcting codes using generative modeling. This model employs autoregressive neural networks, specifically Transformers, to learn the joint probability of logical operators and syndromes

B

Source: 1. arvix.org H Cao, F Pan, Y Wang, P Zhang 2. GitHub, 3. Electronic Engineering Journal, BCG analysis

GenAI massively impacting classical programming today

Code assist

Benefit

Code assist tools help expedite processes like quantum code autocompletion for quantum languages like Cirq, Qiskit, etc
Developers using code assist tools coded up to 55% faster²

E.g. Copilot, Codey

Debugging

E.g. Tabine, PyCharm, Visual Studio Code

AI-assisted software may highlight errors in syntax and provide additional context for coding bugs
75% of users³ ranked a debugging tool (Visual Studio Code) as their most important tool

Anomaly detection

E.g. TensorFlow

AI-supported anomaly detection techniques help quantum by providing system monitoring (of qubits, etc) and identifying code inefficiencies

Convergence potential

17

Let's turn now to ways that GenAI could accelerate the quantum roadmap. Where the prior convergence vector was speculative, this one is not.

At yesterday's quantum summit, IBM announced that it was seeking to democratize access to programming in qiskit with codeassist features that enable users to program in plain language. Additionally they have started work applying GenAI to their transpilers and have seen very encouraging early results, including a 20-50% improvement in transpiler efficiency.

The bigger question is whether GenAI can eventually produce less assistive but more fundamental breakthroughs in things like error correction or base algorithms. The qecGPT project out of the Chinese Academy of Sciences is proposing just that. It's a general framework for generating the most-likely logical operators for any given syndrome with better decoding accuracy and computational complexity.

This is just a start, of course, but an initially auspicious one.

18 of 25

Registered chemicals

275,000,000

Entire chemical space - # of compounds:

1,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000 (10⁶⁰)

Approved medicines

Today ~10,000

Ultralarge chemistry databases

100,000,000,000,000,000,000,000,000 (10²⁶)

Illustrative drug discovery workflow

Completely inaccessible today…

Not even theoretically

C

Convergence potential

18

19 of 25

Registered chemicals

275,000,000

Entire chemical space - # of compounds:

1,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000 (10⁶⁰)

Approved medicines

Today ~10,000

Ultralarge chemistry databases

100,000,000,000,000,000,000,000,000 (10²⁶)

GenAI to generate potential chemical formulas that are likely synthesizable with desired properties

Quantum computer to verify exact properties with quantum simulation on atomic scale and find drug candidates

GenAI generated candidates

Tomorrow could be millions times more efficient

C

Convergence potential

Illustrative drug discovery workflow

19

This is the power of convergence, in which Generative AI can be used to create new molecular structures or drug formulas, learning from a vast dataset of existing molecular structures and using this knowledge to generate new molecules that have the potential to be effective drugs, considering various factors like the likelihood of synthesis, potential efficacy, and side effects while generating these structures.

Quantum computers, on the other hand, can be utilized to verify the exact properties of these newly generated molecules, simulating the behavior of molecules at the quantum level, allowing for the accurate prediction of the properties of new molecules, including their stability, binding affinity to targets, and potential side effects.

Then drug discovery is not just exponentially faster when you have a hypothesis, but you've actually exponentiated the process of hypothesis generation as well.

20 of 25

Our goal is to compress the next 250 years of chemistry and materials science progress into the next 25

Convergence potential

C

20

21 of 25

Traditional computers had many applications prior to the development of error correction

1937 | Atanasoff-Berry Computer solves systems of linear equations for astronomy research

1941 | British Bombe deciphers German Enigma codes

1944 | IBM Harvard Mark I simulates atomic reactions for Manhattan Project

1945 | ENIAC calculates artillery firing tables for the US Army

1950 | Hamming "error correction" codes are introduced

Why collaborate now

21

22 of 25

Government funding, and public-private collaboration around core tech critical ingredients

Research

Technology

Investment

Users

Research

Technology

Investment

Users

Why collaborate now

1951

1952

1953

1954

1955

1956

1953

IBM Harvard Mark I (1939-1944)

1941

1943

UNIVAC I (1950-1956)

Census Bureau was not a "passive" customer: defined specifications (e.g. size, tabulation time) and worked alongside researchers for six months on the design of UNIVAC and approved its readiness in 1951¹

1952

Source: 1. US Census Bureau

22

When you look at the hallmark achievements with computers in the 40s and 50s, it's not a story of heroic "go it alone" tech providers. It's a story about broad and intensive collaboration. The IBM Harvard Mark I came, of course, out of IBM and Harvard, but it was funded by the Navy. And when it was used in the Manhattan Project scientists from the Army Corps of Engineers sat beside researchers from Harvard and IBM to customize circuits for atomic simulations.
Let's look at UNIVAC. Again there was an Ivy League university, this time Penn, and a tech provider, in this case Remington Rand, but it was funded by an enormous grant from the US Census Bureau. What's remarkable about this is that practitioners from the companies on the right would visit Remington Rand. They would bring their data, describe their challenges and pain points. The machines were essentially co-authored. And my sense is that's how it has to work—upstack use case expertise in contact with downstack bit-level or qubit-level expertise.
Interestingly one name you won't see here is the one most famously associated with UNIVAC—and that's CBS. CBS had its election team working with UNIVAC deeply in 1951, and the computer was used on air to predict the results of the 1952 election. Gallup polls indicated a very narrow victory, 51 to 49 for Eisenhower. But UNIVAC, with only 5.5% of the vote in, predicted a landslide victory for Eisenhower, and was off by only 4 electoral votes. During the broadcast the CBS crew refused to announce the results of the simulation and told the world there was computer error. But when the results were in confessed that the computer had got it right. The papers the next day read "machine makes monkey out of man." More than a colorful story, this is actually one of incentives alignment. CBS didn't pay for the machine, but invested time and resources, dominated the ratings numbers, and Remington Rand wound up with all the customers on the right. Partnerships like this require creativity.

23 of 25

We are in early stages today but components in place

Research

Technology

Investment

Users

We must unify these efforts to scale effectively

In April 2023, IBM and Moderna launch partnership to explore the use of quantum computing and GenAI to advance and accelerate mRNA research & science.

Moderna kicks off the build of a quantum- and GenAI-ready workforce, while IBM explores how quantum technology can apply to mRNA research.⁴

Source: 1. J. Chem. Inf. Model. 2023, 63, 11, 3307–3318, 2. arvix.org H Cao, F Pan, Y Wang, P Zhang, 3. Global Quantum Intelligence, 4. IBM newsroom

May 2023: Fundamental research into GenAI and quantum computing convergence¹

July 2023: Researchers proposed a framework called qecGPT for decoding quantum error-correcting codes using generative modeling²

Governments have pledged $55B+ to quantum computing to date and over $200B by 2030.³

Why collaborate now

23

24 of 25

Key takeaways

GenAI and Quantum Computing independently have transformative value creation potential

Convergence opportunities include both (a) roadmap acceleration with QC advancing GenAI and vice versa, and (b) workflow transformation where they create exponential benefits working side-by-side

But the expertise and resources required for technology convergence on this scale overwhelms single-company efforts and even small partnerships

What it requires is concerted, integrated and multi-disciplinary consortium-led efforts to define goals, guide research and deliver results

24

25 of 25

Thank you

25