The Current State of

Artificial Intelligence
 Table of Contents

Chapter: Page:
Introduction 04
State of Chatbot Race 18
Reasoning Models 33
Business of AI 46
Agentic Systems 61
Multimodal / Voice 75
U.S.-China Primacy 87
Wrapping Up 99

2
 How to Read This Presentation

§ This presentation was designed to be read in chronological order, in one go, like a
flip book. Each section of this presentation builds on the prior and assumes no prior
knowledge about the discussed topic.

§ This presentation aims to provide a current overview of large language models,

including the emergence of advanced reasoning models and agents, as well as the
current state of AI businesses and their underlying dynamics.

§ By the end of this deep dive, you should understand how language models are
evolving under resource constraints and competitive pressures, what lessons we've
learned from this first wave of AI products, and how to identify the breakthrough
opportunities in this megatrend.

3
Introduction
 The human brain contains roughly 100 trillion neural connections,
working together to generate thoughts through patterns of activation.

Synapses adjust connection

strengths based on experience

Fire electric signals Recognize

across synapses patterns
A human learns

Neurons

5
 AI researchers took inspiration from the human brain, attempting to build
neural networks with computers that could "learn" through patterns in data.

Neural network adjusts

connection weights based on data

Process text Recognize

on the Internet patterns
A computer learns

Nodes

6
 The transformer architecture was a breakthrough that hit the sweet
spot between pattern recognition power and computational feasibility.

Transformer-Based
Neural Network

A neural network with

the transformer architecture
can both capture long-range
dependencies between words
and finish training at scale, even
if it ingests trillions of data.

7
 This breakthrough made it possible to compress most of the
publicly available knowledge stored on the internet into a language model.

Publicly available Transformer-Based

Neural Network ChatGPT
Internet data

Compresses
Internet data to
Fed into
become

8
 As these models were scaled with exponentially more internet data, researchers
observed that capabilities would suddenly "emerge" that weren't explicitly trained for.
Training Compute (FLOP) Required for Each OpenAI Language Model, 2017 to 2024

1.E+26

GPT-4
Write code
and reason

1.E+25
New abilities emerge
as models ingest
GPT-3
more Internet data.
Generalize using
GPT-2 limited data
GPT-1 Write coherent
1.E+24 Answer stories
questions

1.E+23
2017 2018 2019 2020 2021 2022 2023 2024

Source: Arxiv 9
 All the potential use cases of these capabilities and the pace of
advancement compelled people to invest more than a trillion dollars into AI.
Public and Private Investment in AI, Billions of U.S. Dollars, 2013 to 2024

$400B
$337B
$350B

$300B
Total invested over this period:
$1.29 trillion
$250B

$200B Public and private

investment
$150B $132B

$100B Private
investment
$50B

$B
2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023

Source: Stanford AI Index Report 2024 10

 And that trillion-plus-dollar investment fueled foundational models
and AI products, enabling rapid innovation and widespread adoption.
Private Investment in AI by Focus Area, 2023

Other $40B
AI Infrastructure/Research $18B
NLP Apps $8B
Data management $5.8B
Healthcare $4.2B
AV $2.5B Investments in
Fintech $2.1B companies building
Quantum computing foundational models like
$2.0B OpenAI and Anthropic.
Semiconductor $1.9B
Energy $1.9B
Creative content $1.8B
Ed tech $1.7B
Marketing $1.5B
Drones $1.0B
Cybersecurity $1.0B
Insuretech $0.8B
Legal $0.4B

Source: Stanford AI Index Report 2024 11

So what is the current state of AI?

12
 While AI has made prototyping easier, bringing AI products to market requires
extensive effort and engineering prowess to achieve production-grade reliability.

3x more
difficult
Easy

AI Reliability 70% 90% 99.9%

10x more
difficult

Source: Redhat E-book 13

 On top of that, new breakthroughs can dramatically change the
playing field and triggers a paradigm shift throughout the entire ecosystem.

Breakthrough Technology

DeepSeek-V3's breakthrough is an
open-source AI model that combines
efficient architecture and key
engineering advances to deliver top
DeepSeek-V3 performance at lower training costs.

Ecosystem Impact
Other model providers like While there may be much
OpenAI/Anthropic must adapt Model Hardware more demand for inference and
their API pricing or justify their hardware accelerators, higher
premium models with more Layer Layer efficiency could reduce overall
unique features. hardware requirements.

Cloud platforms may have to

Startups could build high- compete over inference efficiency
performance AI applications at Application Platform by optimizing around the cost,
lower costs, potentially leading Layer Layer speed, and energy consumption
to a new wave of products.
of running AI models.

14
 AI may be in the early innings of a megatrend that
may reshape society in ways we're only beginning to uncover.

Apps

Developers

Data
? ??
Infrastructure

AI Wave 1 (2022 to 20[xx]?) AI Wave 2? AI Wave 3?

15
 And some key questions remain…

Where will the What AI companies will

next killer app emerge? generate returns for investors?

16
 This deep dive will examine AI from infrastructure to application to help
you understand the dynamics forming in the early innings of this megatrend.

Apps

Developers

Data
? ??
Infrastructure

AI Wave 1 (2022 to 20[xx]?) AI Wave 2? AI Wave 3?

17
Read the full deep dive at
chamath.substack.com

18
 Future Deep Dives
Deep Dive Topics Premise Published

A Current State of Artificial How are language models evolving under resource constraints and competitive pressures? What lessons have we learned from this first wave of AI products? How do
Intelligence we identify the billion-dollar opportunities and killer apps in this megatrend?

How are the Magnificent Seven businesses structured and organized? What drives their revenue and bottom line? How are they allocating capital today, and are they
A Primer on The Magnificent Seven
positioned for continued outsized returns? How should we understand the Magnificent Seven in relation to the S&P 500?

Quantum Computing and its Potential What are the different areas of quantum computing research? What is the current state of quantum computing, and how many breakthroughs away is it from practical
Applications application? From first principles, why does quantum computing represent a paradigm shift from previous computing approaches?

Understanding Science and What is science? What is the right way to understand what science is and isn't? How has science evolved over the centuries as society has shifted and different forms
Evaluating Scientific Research of reasoning have emerged? How has the production of scientific research changed as universities have proliferated? How should we evaluate scientific research?

China’s Economic and Geopolitical What has driven China’s economic growth and is China positioned for continued growth? What is China’s geopolitical strategy and position? What is the state of its
Position technology industry and military capabilities? How should the U.S. position itself in response? Is conflict between the U.S. and China inevitable?

A Primer on Marijuana and What does scientific research say about marijuana and various psychedelics in terms of benefits and side effects? How do marijuana's complex interactions with
Psychedelics anxiety, creativity, and memory vary across individuals and contexts? How is the regulatory landscape evolving to accommodate medical applications?

Crypto Stablecoins and P2P What are the use cases for stablecoins, and why have stablecoins achieved product-market fit? Where is the high transaction volume coming from, and how are
Payments different stablecoin designs addressing technical and regulatory challenges? How should we understand and evaluate stablecoins as an asset class?

How have different atomic and molecular structures enabled different material properties? What physics explains why certain structures create specific properties?
A Primer on Materials Science
How do material choices that Apple and SpaceX make reveal about the interplay between aesthetics, engineering, and manufacturing?

What is Warren Buffett's investing philosophy? How do modern portfolio theory and models like CAPM work? How do we think about investments at Social Capital and
How to Invest and Allocate Capital
synthesize the thinking of key investors? Is today's investing climate different enough to warrant an approach different from the great investors of the past?

A Primer on Longevity and Aging What are the key areas of research when it comes to longevity and aging science? What are the most prominent longevity companies and where are they in terms of
Science developing products? What is metabolic dysfunction and its relationship to chronic disease? Is there consensus on what we should do to be healthy and well?

What is a battery and how do batteries work? Why do batteries need certain rare earth metals, and what is the cutting-edge battery technology and how could it be
A Primer on Battery Technology
game-changing to fields like electric vehicles and energy transition? Why has battery innovation followed a linear rather than exponential improvement curve?

The Current State of American What was the original purpose of college and how has that changed over time? Why do more people believe that college is not worthwhile? How did U.S. News & World
Universities Report and the proliferation of colleges and universities change what college is? Is it still worthwhile to go to college, and what subjects are most worthwhile to study?

19
 This document is provided for educational purposes only. Nothing contained in this document
is investment advice, a recommendation or an offer to sell, or a solicitation of an offer to buy,
any securities or investment products. References herein to specific sectors are not to be
considered a recommendation or solicitation for any such sector. Additionally, the contents
herein are not to be construed as legal, business, or tax advice.

Statements in this document are made as of the date of this document unless stated
otherwise, and there is no implication that the information contained herein is correct as of any
other time. Certain information contained or linked to in this document has been obtained from
sources believed to be reliable and current, but accuracy cannot be guaranteed.

This document contains statements that are not purely historical in nature but are “forward-
looking statements” or statements of opinion or intention. Any projections included herein are
also forward-looking statements. Forward-looking statements involve known and unknown
risks, uncertainties (including those related to general economic conditions), assumptions and
other factors, which may cause actual results, performance or achievements to be materially
different from those expressed or implied by such forward-looking statements. Accordingly, all
forward-looking statements should be evaluated with an understanding of their inherent
uncertainty and recipients should not rely on such forward-looking statements. There is no
obligation to update or revise these forward-looking statements for any reason.

This document also contains references to trademarks, service marks, trade names and
copyrights of other companies, which are the property of their respective owners. Solely for
convenience, trademarks and trade names referred to in this document may appear without
the ® or ™ symbols, but such references are not intended to indicate, in any way, that such
owner will not assert, to the fullest extent under applicable law, its rights or the right of the
applicable licensor to these trademarks and trade names.

Disclaimer