COATUE

The AI Revolution

Sri Viswanath

Vibhor Khanna

Yijia Liang

Coatue

群内每日免费分享5份+最新资料

300T网盘资源+40万份行业报告

为您的创业、职场、商业、投资、亲子、网赚、艺术、健身、心理、个人成长 …… 全面赋能！

关注公众号获取更多资料

添加微信，备
注“入群”
立刻免费领取
200套知识地图+最新研报

致终身学习者社群

行业报告/思维导图/电子书/资讯情报

收钱文案、增长黑客、产品运营、品牌企划、营销战略、办公软件、会计财务、广告设计、摄影修图、视频剪辑、直播带货、电商运营、投资理财、汽车房产、餐饮烹饪、职场经验、演讲口才、风水命理、心理思维、恋爱情趣、美妆护肤、健身瘦身、格斗搏击、漫画手绘、声乐训练、自媒体打造、效率软件工具、游戏影音……

扫码先加好友，以备不时之需

Disclaimer

THIS PRESENTATION DOES NOT CONSTITUTE AN OFFER TO SELL OR THE SOLICITATION OF ANY OFFER TO BUY AN INTEREST IN A FUND OR VEHICLE MANAGED BY COATUE MANAGEMENT,L.L.C. (“COATUE”).

ANY SUCH OFFER OR SOLICITATION WILL BE MADE ONLY AT THE TIME A QUALIFIED OFFEREE RECEIVES A CONFIDENTIAL PRIVATE OFFERING MEMORANDUM OR OTHER OFFERING DOCUMENT (“CPOM”) DESCRIBING THE OFFERING AND RELATED SUBSCRIPTION AGREEMENT.

IN THE CASE OF ANY INCONSISTENCY BETWEEN THE DESCRIPTIONS OR TERMS IN THIS PRESENTATION AND THE CPOM, THE CPOM SHALL CONTROL.

COATUE HAS PROVIDED THIS PRESENTATION SOLELY IN CONNECTION WITH ONGOING OR INTENDED DISCUSSIONS WITH THE PERSON TO WHOM IT HAS BEEN DELIVERED. IN THIS REGARD, THE PERFORMANCE FIGURES, AND CERTAIN OTHER INFORMATION INCLUDED IN THIS PRESENTATION, REQUIRE FURTHER EXPLANATION AND MUST BE DISCUSSED WITH YOUR COATUE REPRESENTATIVES. ACCORDINGLY, PROSPECTIVE INVETORS SHOULD NOT RELY ON THE INFORMATION IN THIS PRESENTATION ABSENT SUCH DISCUSSIONS.

NO SECURITIES OR SERVICES SHALL BE OFFERED OR SOLD IN ANY JURISDICTION IN WHICH SUCH OFFER, SOLICITATION OR SALE WOULD BE UNLAWFUL UNTIL THE REQUIREMENTS OF THE LAWS OF SUCH JURISDICTION HAVE BEEN SATISFIED. WHILE ALL THE INFORMATION PREPARED IN THIS PRESENTATION IS BELIEVED TO BE ACCURATE, COATUE MAKES NO EXPRESS WARRANTY AS TO THE COMPLETENESS OR ACCURACY NOR CAN IT ACCEPT RESPONSIBILITY FOR ERRORS, APPEARING IN THE PRESENTATION. NEITHER COATUE NOR ITS AFFILIATES ASSUMES ANY DUTY TO UPDATE THE INFORMATION CONTAINED HEREIN FOR SUBSEQUENT CHANGES OF ANY KIND AND THERE IS NO ASSURANCE THAT THE POLICIES, STRATEGIES OR APPROACHES DISCUSSED HEREIN WILL NOT CHANGE.

ANY PROJECTIONS, MARKET OUTLOOKS OR ESTIMATES IN THIS PRESENTATION ARE FORWARD-LOOKING STATEMENTS AND ARE BASED UPON CERTAIN ASSUMPTIONS. OTHER EVENTS WHICH WERE NOT TAKEN INTO ACCOUNT MAY OCCUR AND MAY SIGNIFICANTLY AFFECT INVESTMENT RETURNS OR PERFORMANCE.

ANY PROJECTIONS, OUTLOOKS OR ASSUMPTIONS SHOULD NOT BE CONSTRUED TO BE INDICATIVE OF THE ACTUAL EVENTS WHICH WILL OCCUR. HISTORICAL RETURNS ARE NOT PREDICTIVE OF FUTURE RESULTS.

THE INFORMATION PROVIDED HEREIN, INCLUDING, WITHOUT LIMITATION, INVESTMENT STRATEGIES, INVESTMENT RESTRICTIONS AND PARAMETERS, PROCEDURES, POLICIES, AND INVESTMENT AND OTHER PERSONNEL MAY BE CHANGED, MODIFIED, TERMINATED OR SUPPLEMENTED AT ANY TIME, WITHOUT NOTICE.

THIS PRESENTATION IS CONFIDENTIAL AND NOT INTENDED FOR PUBLIC USE OR DISTRIBUTION. THIS PRESENTATION HAS NOT BEEN APPROVED BY THE U.S. SECURITIES AND EXCHANGE COMMISSION OR ANY OTHER FEDERAL OR STATE REGULATOR.

YOU SHOULD CAREFULLY READ THE IMPORTANT DISCLOSURES IN THE APPENDIX.

COATUE

Key Topics

Where we are in AI today

→Al could break through the hype and improve our world

→ We believe open-source is the lifeblood of Al

→Al is transforming the tech ecosystem

→Coatue view: the best of Al is yet to come

We’re at Day 1 of AI…and riding on top of past waves

% US Technology Adoption

In the first half of 2023, the AI ecosystem exploded! (1/2)

ChatGPT weekly unique visits

NVDA 2Q Earnings

AI Basket Performance YTD

In the first half of 2023, the AI ecosystem exploded! (2/2)

% of trending Github repos in AI/ML

New AI applications over time

Number of models on Hugging Face

Adoption has been twice as fast with each platform shift

Halving of penetration time with new technology waves

AI has potential to drive the economy for years to come

of employees per \$1M of revenue (inflation adjusted – S&P 500 companies)

We’ve seen massive investment in AI

→ Cumulative funding in private AI companies since 2020

Valuation before public product launch

amples of the applicable statements or trends; do

We’ve backed AI founders across stages

# of Coatue AI portfolio companies based on round when Coatue first invested

opics
Where we are in Al today
AI could break through the hype and improve our world
We believe open-source is the lifeblood of Al
Al is transforming the tech ecosystem
Coatue view: the best of Al is yet to come

Key Topics

Coatue View: AI is not just hype

Characteristic of hype cycles

Value Accrual

Value accrual misaligned with investment

Capabilities

Overestimating timeline & capabilities of technology

Value Accrual

Lack of widespread utility due to maturity of technology

Coatue view

Comparison: Historic cycles like fiber and cloud represent two different outcomes for underlying infrastructure

Most investment in AI today is within the model layer, but it’s too early to declare who will be the AI model winners

Comparison: Self-driving cars finally arriving, but after 15+ years of work

AI is already useful within \~5 years. We are in early innings of adoption and expect models to improve, but AI regulation is likely and is a challenge to implement well

Comparison: Quantum computing hype was promising in theory but has not yet proven widespread practical utility

• AI already proving significant utility across domains

Historic Examples

1990s Fiber

Autonomous vehicles

Quantum computing

Comparison: Some enabling technologies become a public good

Fiber Infrastructure 1990s

Telco Co’s raised \$1.6T of equity & \$600B of debt

Bandwidth costs decreased 90% within 4 years

Most of these companies no longer exist today…

…While cloud infrastructure became a huge market

Cloud Infrastructure 2010s

Entirely new computing paradigm

Most were already public companies with resources to build out data centers

Google Cloud

Big 4 hyperscalers generate \~\$150B+ in revenue annually

Where did value accrue in the cloud stack?

Cloud Stack

Example Companies

Est. TAM

SaaS Apps

PaaS

CONFLUENT

Adobe

\~\$260B

\~\$140B

IaaS

ΛAzure

\~\$200B

Cloud Semis

AMDD

intel.

\~\$50B

% of total in stack1

40%

22%

30%

In AI, funding concentrated in model layer for now

Al Stack	Example Companies		Total funding	% of total in stack1
Apps	character.ai	replit	~$5B	17%
Models	SOpenAI	ANTHROP\C	~$17B	60%
Al Ops	Hugging Face	Weights & Biases	~$1B	4%
Al Cloud		databricksLambda	~$4B	13%

Al Semis	Cerebras	SSambaNova	~$2B	6%

The jury is out on which model companies will win

AI Models 2020s

Private AI model companies raised \~\$17B in venture funding since 2020

Open-source models are becoming more ubiquitous

ADEPT

ANTHROP\C

co:here

Outcome TBD

Levels of autonomy	Autonomous vehicles	Generative Al
L5	Fully autonomous	Superhuman reasoning & perception
4	15 Years Highly autonomous	Al autopilots for complex tasks
	Self-driving with light intervention1	Al co-pilot for skilled labor	~5 Years GitHub Copilot : replit
	Tesla autopilot	Supporting humans with basic tasks	图 ChatGPT 小 Sspeak
	Cruise control O	Generating basic content	Midjurney

Comparison: AI advancing much faster than previous waves

Has initial AI enthusiasm slowed down?

AI usage worldwide started flattening during summer

Early developer excitement waning and washing out AI “tourists”

Excitement in AI/ML on GitHub, measured by stars, has declined since April 2023

Serious AI builders remain

AI/ML commits have not declined as much since April 2023

ChatGPT usage has rebounded, new capabilities have shipped

ChatGPT weekly web visits worldwide since launch

We believe it’s first innings of enterprise AI adoption

\~60%+ of surveyed enterprises plan to adopt AI

Less than 10% have fully adopted; timeline is long

We are optimistic: AI is getting better, faster

Speed for models to reach human level accuracy on benchmarks has decreased

AI regulation may be more likely than most think

Initial polls suggest many Americans concerned about AI

EU’s AI Act is among first examples of regulation

Most models fall short of meeting requirements from EU AI Act, per Stanford study

Comparison: Is AI hype justified with real utility?

Quantum computing interest has steadily grown without real world proof points…what about Generative AI?

AI is delivering game-changing value

55%

Time saved for developers using Github Copilot

90%

Time saved from editing video on Runway

45%

Reduction in humananswered customer support requests

79%

AI chat rated higher quality vs. physician responses

AI poised to improve productivity significantly!

Example: Designing Marketing Materials

Analog 1950s

Months

Design on paper Iterate on physical copies

PCs 1960s – 1980s

Weeks

Design on local software Easy to iterate

Internet & Cloud 1990s – 2010s

Days

Design on web apps Instant collaboration

Generative AI 2020s + Beyond

Minutes

AI-created design Human prompts with text

Physical

Single-player

Manual

Digital

Single-player

Manual

Digital

Collaborative

Manual

Digital

Collaborative

Automated

GitHub Copilot makes programming faster

Over 8M downloads of Copilot extension

Copilot improving developer productivity over time

Companies have seen huge efficiency gains already

Experience from one Fintech company

AI is cheaper customer support

AI provides faster responses

AI makes customers happier

Source: Based on one company’s internal assessment data – not representative of all companies or guarantee other companies would achieve similar results. Customer satisfaction calculated by number of employees rating the interaction a 4 or 5 out of 5. Data compares against top 8 human employees, based on number of tickets participated in. Coatue analysis and opinion as of November 2023. For illustrative purposes only. There is no guarantee that Coatue’s views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend. See Appendix-Disclosures for important disclosures, including regarding projections and forward-looking statements and trends.

Key Topics

→Where we are in Al today

→Al could break through the hype and improve our world

We believe open-source is the lifeblood of AI

→Al is transforming the tech ecosystem

→Coatue view: the best of Al is yet to come

How we got here – AI is built in the open

Research

Openly Available Research

Research from academia & industry has driven advancements in AI

Collaboration using open-sourced state-of-the-art models has led to rapid pace of innovation

Community

Open Community

GitHub and Hugging Face underpin the open-source AI community

Participation in AI has been explosive

• There are varying degrees of openness across AI today

Models & Data

Companies realize the value of their own data, and model providers have become more secretive

8Meta

AI is a result of open research

Cumulative AI/ML publications submitted on Arxiv

Open collaboration accelerates innovation in AI

Illustrative launches of AI models over time (non-exhaustive)

The AI developer community has exploded!

→ Software developers are becoming AI engineers

Hobbyists are getting involved

GitHub has been a place to discover new AI/ML projects

Cumulative GitHub stars given to AI/ML related projects

Open-source AI is inflecting on Hugging Face as well

Number of times AI models have been downloaded from Hugging Face

However, AI is not always truly “open” (1/2)

Coatue’s open-source AI model checklist

However, AI is not always truly “open” (2/2)

Coatue’s open-source AI model checklist

Data is finally a new currency

Timeline of companies cracking down on data access in 2023

Key Topics

→Where we are in Al today

→Al could break through the hype and improve our world

→We believe open-source is the lifeblood of Al

AI is transforming the tech ecosystem

→ Coatue view: the best of Al is yet to come

The new AI-centric technology ecosystem

Foundation models are at the center of AI

Last 50 years was about building faster & faster “calculators”

CPU

Serial processing

“1 instruction at a time”

Chips got smaller & more powerful

Software

Based on instruction by programmer

Follows sequential programming logic

Does not require data

Computer or “Calculator”

But next 50 years will be about building super-intelligent “brains”

GPU

Parallel processing

Many calculations simultaneously

AI Models

Neural networks trained by data

Learns patterns from data

System makes decisions based on model rather than explicit instructions

“Reasoning” is opaque, not driven by programming logic

Brains

Much more than calculators!

Could become connected: Brain-to-brain network = AI Internet?

AI enables a new platform: Intelligence-as-a-service Intelligence is the next layer of innovation

Today

“[One of my] opinions on current LLMs: They are "reactive" & don't plan nor reason.”

laaS

Infrastructureas-a-service

PaaS

Platformas-a-service

aa

Softwareas-a-service

Intelligenceas-a-service

Augment

our capabilities

Future

Automate our workflows

“The right way to think of the models we create is a reasoning engine, not a fact database.”

Change our existing behaviors

character.ai

More research & headcount is not enough to win in AI

MSFT leads big research houses in publications

Number of AI/ML related research publications

OpenAI has shipped faster than larger peers

Talent has historically been concentrated at a few AI model hubs

Major AI model providers are “poaching” talent from one another

Talent inflows & outflows at major research hubs

Illustrative only: Represents both direct & indirect (>1 hop) flows of talent

Many talented people are now leaving to start new companies

Example AI mafias

Number of founders from AI research houses

Scaling AI model performance has been a focus of AI researchers

Strategies to scale

More parameters

Larger datasets

Focus of remainder of section

Open research questions

Will scaling parameters continue improving performance?

How much data is optimal for training models?

Coatue View

GPT-4 still “king”; scaling experiments likely continuing

Larger models need more data; extending data runway & optimizing quality crucial

More compute

Covered in next section on Cloud, DC, Semis

Can we reduce the compute costs of training & inference?

Can training for more epochs improve performance?

Not covered; waiting for more data points to emerge

Longer training

Training & inference is getting optimized; AI at the edge is emerging

Still an open question

Through 2022: We saw LLM parameters balloon!

→ Illustrative view of parameter sizes of large language models through 2022

2023: GPT-4 remains “king”…can we continue scaling?

Comparison of 2023 models across ELO ranking and MMLU benchmarks…GPT-4 is still the best

Higher ELO is better

What if there was a \$1B training-cost model?

Model Training Cost Over Time

How AI models are made: Data is a crucial component

Data is upstream in process of developing good models

Scaling models require scaling datasets

Number of tokens in training datasets increasing overall

Data scarcity is a potential wall to scaling models

High-quality text data could be exhausted soon, images & video have longer runway

Data quality is just as important as data quantity

Emerging evidence that training on pruned datasets can result in similar performance in language models

Performance on variety of language model benchmarks

cohere

Synthetic data can augment fine-tuning

Synthetically generated data can help clean or distill datasets for fine-tuning

Results: Synthetic data fine-tuning yields significant improvement compared to zero-shot attempts, though is not yet comparable to real data

Percentile performance on test

User feedback data is another way to improve performance Midjourney collects user feedback to continue refining models

Prompt /dungeons and dragons, female knight, of the rolling plains, full body, dark azure,victorian genre paintings, serene face, realistic depiction of light, golden light

Users choose which image to upscale (or generate a variant of), effectively choosing their preferred image

Midjourney V1 (Feb 2022)

V2 (Apr 2022)
→

V3 (Jul 2022)

V4 (Nov 2022)

V5 (Mar 2023)

Note: Midjourney improvement is partially due to collecting user feedback but is only one part of the equation!
V5.2 (Jun 2023)

There are many other opportunities to improve models!

Beyond scaling parameters & data, there are many angles to push on

Non-exhaustive, illustrative opportunity set

AI has potential to re-accelerate underlying infrastructure

This AI wave has been extremely compute hungry

ML models are doubling compute needs for training in months

Training compute has exponentially increased with model sizes

Massive inflection in compute needs for AI, \~70X increase from GPT-3 to GPT-4

Inference compute likely to dramatically outpace training

Ongoing inference likely to require much more compute than one-time training

AI compute has been primarily served by GPUs

GPUs are able to serve AI workloads much faster than CPUs

The demand for GPUs has only begun

Estimated ramp up for AI servers growing 25%+ Y/Y through 2026

What are the implications for this global shift in demand for GPUs?

Follow the GPUs: AI poised to have impacts across our economy

\$32B+

NVDA Revenue last 12 months

Power

Cloud

Data Centers

Semis

The AI wave may stress our power grid

\$32B+

NVDA Revenue last 12 months

2-3x

Electrical transformer price increases

3-5 Years

Leadtime to connect to the grid

3-5 GW

More data center power required ’23-24E

50%

Potential increase in power demand by 2026

More GPUs likely means more expensive servers

\$32B+

Higher Server Bill of Materials Due to GPUs ’23-24E

NVDA Revenue last 12 months

Note: all estimated numbers

Non-Al Server		10x	Al Server ~$20K Y
CPU	~$2K	10x	Al Server ~$20K Y
GPU			~$200K
Storage & RAM	~$5k	5x	~$20K y
Power Supply	<$1K	10x	~$5K 》
Networking & Other	~$5K	3x	$15k y
Server Cost	~$10K	26x	~$260K

Could AI workloads drive a cloud reacceleration?

\$32B+

NVDA Revenue last 12 months

Y/Y revenue growth rates for AWS, GCP, and Azure from Q2 2022 to Q2 2023

AI is the latest tech wave to inflect the semiconductor industry

	Mainframe		PC + Internet Smartphone		Public Cloud		Al
Time Period	Pre 1990	1990 - 2006		2007 - 2014		2014 - 2022		2022 onwards
Incremental Semi Spend	~$50B	~$100B		~$150B		~$250B		~$500B+ ？
Total Semi Spend	~$50B	~$200B		~$350B		~$600B		~$1.5tn+ ？
Key HW				Qualcom^		Google		ALam	AMDD
players		Microsoft		arm		amazon
								nVIDIA.	Micron'

The entire semiconductor supply chain has potential to benefit Semis value chain across manufacturing inputs to device assembly

Semi manufacturing inputs

Semi design and manufacturing

Device assembly

Materials

Capital equip.

IC design/embedded software

Fabrication, pkg. & testing

Assembly

AMDD

nVIDIA.

wistron

Ex: Silicon wafer

Ex: EUV Lithography

Ex: NVIDIA chip design

Ex: 3nm chip

End product: NVIDIA H100

Data center servers powering AI!

Demand for cutting edge GPUs has exceeded supply

H100 GPU prices on aftermarket are typically well above list price

Pricing of individual H100 80GB GPU units on eBay & Amazon, not reflective of cluster availability

Tesla H1OO8OGB NVIDIA Deep Learning GPUComputeGraphicsCard900-21010- 000-000

\$42,750.00

Tesla H100 80GB NVIDIA Deep Learning GPUComputeGraphicsCard

\$42,672.00

Nvidia H100-PCle-80GB Hopper H100 80GBPCleTensorCoreGPUAccelerator
\$39,995.00

NVIDIA H10O 80GB Tesla Deep Learning GPU Compute Graphics Card 900-21010- 000-000

\$45,000.00

Despite intense demand, AI compute costs have decreased

Could AI compute resemble cost curves in other sectors?

Cost of running GPT-3/3.5 down 90% in 5 months!

Models on edge could alleviate GPU shortage as well

Local models on Apple silicon becoming as fast as models running on GPUs

Tokens per second (higher the better) of LLMs

Will we reach AI’s Holy Grail?

Largest, most expensive models are still most performant….for now

AI has started to create a new ecosystem of tools

AI Ops is a new category of tools for AI engineers

Illustrative; non-exhaustive

Beyond labeling, data curation is becoming more important as engineers try to enhance model performance and alignment.

New tools enabling fine tuning and inference have become easier for the AI developer to use

00 Humanloop

Model training & development

Monitoring & evaluation

Emerging category of tools enabling developers to use LLMs more effectively; Vector DBs have exploded in popularity due to embeddings.

Securing the last mile deployment of LLMs through model evaluation, guardrails, and ongoing RLHF is supporting a new set of companies

Data curation tools will be critical for improving models

Data curation an emerging new category between data collection & labeling, and model training

Data collecting & labeling

Labeled data feeds ML models

Dataset curation (New category)

Preparation

Automatically clean and scrub data for errors

Analyze

Profile the dataset across various metrics (e.g. token length distribution, entity label distribution)

Curate

Filter data to include best examples for training; discard lowquality samples & identify human errors in labeling

Model training

Pre-training

Fine-tuning

RLHF

Curating datasets can reduce costs & improve quality. Training on \~30% of dataset can be sufficient! (from Cohere study)

非329977

Fine tuning models has become much more accessible Hugging Face Autotrain lets you fine tune open-source models on your own data in just a few clicks

Select type of task & upload training dataset

Hugging Face suggests suitable models

New project

Project name

Task

Text Classifcation (Binary)

89.4%

88.4%

Training files

IMDB_train.csv

Select model(s) and evaluate model training progress

ModelID	Loss	Accuracy
329982 pessimistic-aardvark	0.2462	0.9300
329971 baggy-woodpecker	0.2109	0.9216
#329970 this-leopard	0.2453	0.9189
#329981equatorial-raccoon	0.2212	0.9117
#329974 regal-termite	0.2606	0.9091
#329968 fearful-lobster	0.2440	0.9089
329975unlucky-cobra	0.2813	0.9022
329972 far-badger	0.2555	0.8981

LLM Ops & Vector DBs are a new enabling layer for AI apps

Tools are already being put into production

LLM Ops & Vector DBs enable retrieval augmentation

Can VectorDBs be standalone winners?

Only three DB types have significant public outcomes

Startups have tried to create new categories of DBs

OLTP CockroachDB

Observability

Real-time streaming DBs

Graph DBs

No large standalone outcomes yet…

Vector DBs

… Will vector DBs be different?

General purpose databases are being used as VectorDBs too

% of Supabase projects using pgvector

LMSYS

Problem: Model evaluation is broken today

Humans prefer Claude 1 to Claude 2…

Overall win rates (human preference) against other models

…But Claude 2 performs better on benchmarks

Performance on different benchmarks

ANTHROP\C

Evaluation is one of many ‘last mile’ tools needed for AI in production

Last-mile tooling can improve AI accuracy and alignment

= Example last mile tooling
This may help prevent hallucinations in critical sectors!
Industrials case study: AI applied to prognostics & health mgmt. (PHM)

We’re excited that AI Ops enables more developers to build apps

New AI Ops tooling is accessible to all developers

The application layer is where humans will interact with AI

We expect AI to become even more impactful and autonomous

We have seen this before: Copilot phase was short-lived in chess

Computer chess rating

AI adoption accelerating across creative modalities

Usage of Runway’s AI video tools over time

Midjourney Discord members over time

AI ramping up for design use cases in professional settings

Cumulative Tome signups

Otome → AI solar designs generated
aurora

AI is inflecting within software development tools too

AI projects on Replit (annualized)

replit

Cumulative VSCode installs

Massive apps created on both sides of S-Curves of prev. cycles

Application tech companies founded and US market cap along S-Curves

Incumbents gained most in smartphones: what about AI?

AI could drive incumbent price increases

Microsoft Office Copilot could be an 80%+ increase in ARPU for Office users

Per seat per month cost of Microsoft 365 E3 license

Is AI a game of kings or attackers?

Incumbent Kings

Data

Distribution & audience

Full support infrastructure

Large capital reserves

Slow and risk averse

AI Native Attackers

Fast and nimble

Creative and risk-on

No tech debt

Attract higher density of tech talent

Need for external funding

Fast moving incumbents > AI natives > Laggard incumbents

The biggest mobile startups unlocked new behaviors…

Mobile Technology

GPS
Mobile Camera
Mobile Internet
App stores

On Demand Access to Underutilized Assets

Uber

UDOORDASH

+\$215B Mkt Cap

Real-time Photo
Sharing &
Messaging

TikTok Instagram

+\$250B Mkt Cap

Mobile Fintech

affirm) Klarna.

Robinhood venmo

+\$80B Mkt Cap

AI startups should aim to create new behaviors too

Artificial Intelligence

“Most entrepreneurial ideas will sound crazy, stupid and uneconomic, and then they'll turn out to be right”

—Reed Hastings

AI autopilots?

Models that can reason?

Infinite gaming worlds?

Your invention here…

Example: Character AI creating new virtual companions

Users flocking to Character to chat with AI personas – brand new behavior!

character.ai

Example: AI as an autopilot could transform org structures

Previously, scaling businesses meant scaling headcount

With AI, scaling business means scaling compute!

Executive leadership

Senior management

Engineering

Product

Growth & Marketing

Sales & Support

Finance & Legal

Product & Engineering

Leadership

Operations
AI autopilot
Sales, support, and marketing

AI autopilot

Finance, legal

AI autopilot

There remains huge opportunity across modalities

Example opportunities

Open challenges

Code	Software synthesis	$600B+ size of global software1 market today	Hallucinations Code security & robustness at scale Integrations with existing SW deployment process
Audio	Al musicians	$26B+ size of music industry today	Copyright infringement with artists Incumbent distribution effcts more pronounced (e.g. Apple audiobooks, Spotify)
Search	LLM-based search	$162B+ Google Search revenue today	Costs to use agent-automated search Response latency Incumbent advantages in search data & distribution
Video	Text-to-video	$100B+ size of global movie & entertainment industry today	Photorealism Content moderation Compute costs at scale
Other (e.g. robotics)	Robotic perception	$35B+ size of robotics market today	Hardware costs and upfront capital outlays Unproven text-to-action workflow

AI could increase the opportunity size of all these markets!

AI has potential to unlock “the super app of the West”

Key Topics

→Where we are in Al today

→Al could break through the hype and improve our world

→ We believe open-source is the lifeblood of Al

→Al is transforming the tech ecosystem

Coatue view: the best of AI is yet to come

Coatue View: The future top coding language will be English

Today

Natural language to programming, debugging, and deployment

Future

From idea expressed in English…

Engineering

Private data

Copilot systems that suggest code

Emerging coding agents

Multimodality

Factory

With help of future AI systems…

And future companies?

Next gen

… To fully fledged application in minutes!

Future: Rumored Apple AI strategy

Coatue View: On-device AI will become more widespread

Will everyone have a powerful LLM in their pocket?
Today: MLC LLM allows you to run models on your phone

Coatue View: Scaling AI is an engineering challenge

Engineering headcount at OpenAI has increased over 3X since GPT-3 came out

Private data

Data Center

Coatue View: Private datasets can unlock new use cases

Imagine what AI can do with Proprietary Data!

English

Research

Public data

Unimodality

Everywhere

Engineering

Multimodality

Biotech

Next gen

Media

3D Models

Gaming

eCommerce

gohmonagohuon

NETFLIX

A AUTODESK.

SHEIN

Personalized medicine

Generate

shows ondemand

Engineer anything

AI shopping concierge

Coatue View: Innovation in multimodality is a new frontier

Today: GPT-4V is already creating new experiences for users

Data Center

English

Everywhere

Engineering

+ ChatGPT =

Research

Step by step written instructions if you get confused by IKEA diagrams!

Public data

Unimodality

Private data

Multimodality

Current gen

Next gen

Future: Lots of innovation within multimodal embeddings space!

Multimodal embeddings map relationships across text, image/video, and audio

Open questions:

Does training on multimodal datasets enable better reasoning capabilities for models?

How can we keep multimodal models safe to use?

What interesting emergent properties can we see from using the multimodal embeddings space?

How do we scale multimodal datasets to train better models?

Coatue View: New advancements in architecture coming

Cutting edge AI research emerging to improve Transformers

Future | New architectures: Hyena

Could Hyena architecture enable million-token length context windows?

Solving AI’s challenges has potential to unlock vast opportunities

Model output

Reducing Hallucinations

Discovery & novel insights

Model explainability

Modalities

Multi-modal models

Long-form video generation

Embodied robotics

Availability

On-device AI & lowcost deployments

Abundant hardware

SOTA models available for all

Enterprise deployment

Compliance guardrails

AI authentication & permissioning

AI ethics & security

+ more challenges yet to come...

Generally useful AI for all

If you are a founder innovating in this space, we would like to connect with you. Please reach out at ai@coatue.com

群内每日免费分享5份+最新资料

300T网盘资源+40万份行业报告

为您的创业、职场、商业、投资、亲子、网赚、艺术、健身、心理、个人成长 …… 全面赋能！

关注公众号获取更多资料

添加微信，备
注“入群”
立刻免费领取
200套知识地图+最新研报

致终身学习者社群

行业报告/思维导图/电子书/资讯情报

Coatue发布115页的最新报告《AI The Coming Revolution》（英）-2023-WN11