Code to Cognition: Overview &

Trends in Artificial Intelligence

Authors
Aviral Bhatnagar & Parth Chhaparwal
29th January 2024

1
1 Introduction to Artificial Intelligence

2 Artificial Intelligence Tech & Product Landscape

3 Potential Themes for Entrepreneurs

2
1 Introduction to Artificial Intelligence

2 Artificial Intelligence Tech & Product Landscape

3 Potential Themes for Entrepreneurs

3
 Introduction to Artificial Intelligence
• In embarking upon an exploration of AI products and their transformative potential, it is imperative to first delve into the

foundational concept of intelligence itself. The fundamental question: What is intelligence?

• At its core, intelligence has been characterized as the capacity to resolve complex challenges. Yet, this leads us to ponder

whether mere problem-solving, as exhibited by a calculator, qualifies as intelligence.

• What differentiates truly intelligent systems from mere computational tools? We unravel the nature of these complex

problems, their natural resolution mechanisms, and the potential for their replication in artificial systems.

• We then navigate through the history of AI, explaining how computational power and its associated costs have historically

impeded AI’s progress for decades.

• The history of AI illuminates the question: Why now? Why is this moment in history ripe for pioneering, investing, and

actualizing the potential of AI?

Defining Intelligence
Intelligence is the ability to solve complex problems – such as analyzing information,
creating new solutions, adapting to situations, and managing social relationships

Analytic Creative Practical Emotional

Intelligence applied Imagination, Change, adapt to, or Self awareness; self

to analyze a problems innovation & finding a new management; social
& arrive at solutions. problem-solving; Go environment in case awareness;
(IQ) beyond given data & of unseen problems relationship
create new management

5
Creating Intelligence
There are 6 outputs of intelligence created by the human brain

Physical components Outputs

Forebrain 1 Learning
: Memory
2 Linguistics
Midbrain 3 Vision
: Senses
4 Audition
Hindbrain
: Action 5 Coordination

6 Locomotion

6
Defining Artificial Intelligence
AI is a machine’s ability to create the outcomes of intelligence

Learning Machine Learning

Linguistics Computer Linguistics

Vision Computer Vision

Audition Computer Audition

Coordination Networks

Locomotion Robotics
7
Evolution of AI
Access to improved compute and funding has sped up AI adoption

1950 1956 1963

Turing’s Test Dartmouth Workshop Symbolic AI & Financing
First test for AI where where The term ”Artificial Intelligence” was coined • Evolution of Natural Language
indistinguishable answers of a machine & and defined as the field of study that helps • Evolution of If-Else Logic
human would deem the machine replicate human intelligence in machines • Increased funding at MIT
intelligent

1987 1980 1974

Second AI Winter AI Boom First AI Winter
Slowdown in funding due to • Evolution of 1st, 2nd gen computers, Slowdown in AI funding due to
• Limited adoption by commercial calculators • Lack of data, limited computing power
vendors • Use of data in decision making increases • Expensive and time consuming
• Tech was still not viable • Extensive funding in Japan on AI research

1993 2009 2018

Moore’s law Deep Learning & Big Data Artificial General Intelligence
Next generation processors provided Creation of artificial neural networks Foundation models like GPT3 by OpenAI
enormous speed; Speed and memory Emergence of Big Data management tools and Gato by DeepMind act as the first step
started doubling every 2 years to process enormous amounts of data to achieving artificial general intelligence

8
Creating Artificial Intelligence
Intelligence in machines is created by 4 pieces of hardware

CPU
Input Output
Memory Device
Device Control Logic
Unit Unit

Rules

Input data Rule stored in New input data

processed by Memory subjected to old
Logic Unit to processed via rule in memory
Analytic Creative Control Unit Practical and processed via
create a Rule
and store it in used to create control unit to
Memory Output create new rule

9
Creating Artificial Intelligence
Computers are imbued with intelligence using the following methods
Supervised Learning Unsupervised Learning Reinforced Learning

Learning from unclean and Learning by observing and

Explanation Learning from pre-labelled
unlabeled datasets interacting with the
data
environment
• Regression: Finding relationship • Clustering: Classifying datapoints • Deep Learning: Teaching your
between 2 variables into groups based on undefined brain to recognize things by
categories showing it a lot of examples
• Classification: Classifying
datapoints into a group based on • Association: Identifying patterns in • Policy Learning: Finding best way
pre-defined categories underlying data to do something using lots of
Techniques examples

x1 x1

x2 x2
Prediction of weather based Finding boundaries of a Learning to play chess by
Use Case
on atmospheric conditions product in an image winning/ losing

10
What will AI unlock?
AI makes us better at doing tasks, similar to other digital inventions

Improve Productivity Unleash Creativity Better Decisions

Reduces time, costs,

Giving life to your imagination Faster data analysis, insight
requirement of multiple people
without having artistic ability generation and decision making
on a single tasks

11
Is it the right time to invest in AI?
Enormous Models Accessibility and Ease of Use
Power of AI models outpacing Moore’s law Cheaper compute; Better UI/UX provided by LLMs

Legal Copywritin
g

R&D Marketing

Access and computation power of AI models has made AI output meaningful

12
Where we are today?
Technology Product Business
(2004 - 2018) (2018 - present)

Outpacing Moore’s Law Applications of LLM Technologies

~6 ~2
months years

Time to double AI Time to double

computation power silicon chips Yet to be discovered

Extremely powerful models Customer facing apps by R&D companies

13
AI Terminology

Machine Learning Technology allowing machines/computers to make decisions without being explicitly programmed

Artificial
Simulation of human intelligence processes by machines/computers
Intelligence

Artificial General
Theoretical AI that can behave exactly similar to a human
Intelligence (AGI)

Neural Network Computer model inspired by human brain with interconnected nodes of data connected in layers

LLMs ML model trained over large amounts of language data to produce meaningful output in a given context

GPT Generative Pretrained Transformer (GPT) is a type of LLM that uses natural language to generate human like text

Diffusion Models Type of ML models that predicts how information spreads from one coordinate to another – used in image generation

14
1 Introduction to Artificial Intelligence

2 Artificial Intelligence Tech & Product Landscape

3 Potential Themes for Entrepreneurs

15
 AI Tech & Product Landscape

• While looking at AI product landscape, it becomes essential to conceptualize the Foundation Model:

Standalone AI models capable of adaptation across a myriad of applications.

• This creates a market for essential "building blocks" tailored to these foundational models alongside the

"applications" that can be innovatively crafted atop them.

• The advent of models like GPT-3 and GPT-4 has catalyzed a surge in research and development efforts,

extending beyond linguistics to encompass other outputs of intelligence, including vision and audition.

• We then look at the 4-P framework of AI's impact, which encompasses a reduction in dependence on

people, improved productivity of individuals, increased programmability of processes, and enhanced

precision of their outputs.

Artificial Intelligence Investment Landscape
There are three kinds of products in the AI Product Landscape

Infrastructure & Tooling Foundation Models Applications

Hardware and software Standalone AI program Consumer and enterprise use

Definition required to run AI workloads deployable in AI applications cases of AI

Examples
- DallE

Limited right-to-win for Potential to build localized Tremendous potential for

Investment hardware plays out of India; plays; High R&D requirements; Indian AI apps to disrupt
Thesis Several opportunities in AI Monetization by building existing processes across
software applications using the model industries

17
 AI Infra & Tooling

AI Infrastructure and Tooling

Building blocks for AI assist in processes before, during & after model training

Data ML Process Output

Data Collection 4 NLP Tools 8
1
Orchestration & Deployment
Tools

5 ML Frameworks & Model Development

2 Data Preparation

9
6 Model Lifecycle Management (MLOps)
UI Frameworks

3 Vector Databases
7 Alignment and Safety
 Foundation Models

Foundation Models for AI

There are various foundation models based on the outputs of intelligence
Linguistics Vision Audition Robotics Reasoning

Type Text Images & Videos Audio Multimodal Tables & Vectors

Understanding
Understand & Understand & Understand & Software
patterns in
Purpose generate textual generate visual generate auditory controlling
quantitative and
patterns patterns patterns hardware
qualitative data

Diagnostics,
Language Improving sound
Example Object recognition Smart Homes Planning,
translation quality
Forecasting

Companie
s

19
 Foundation Models

Computer Linguistics: Underlying Technology

Large Language Models (LLMs) are AI models that can understand and
generate human language

LLM GPT4 Gemini Constitutional AI Inflection AI LLaMa

Parent
Company

AI Tools

General AI Assistant Code, Science, R&D ChatGPT Stronger emotion Open Source
Advantage
(Most powerful model) (Search Engine) competition and linguistics Equivalent

Power 1 Trillion+ 175 Billion+ 100-175 ~175 Billion 70 Billion

(Parameters) Billion

20
 Foundation Models

Computer Linguistics: Evolution

OpenAI has commoditized the technology to build computer linguistics
applications with several unicorns leveraging GPT4, Llama
Technology Product Business
(2004 onwards) (2018 onwards)

Applications of LLMs Yet to be established

Generative Pretrained
Transformer (GPT)

Type of LLM that finds

dependencies in text
arranged in sentences

21
 Foundation Models

Computer Vision
Models in computer vision are yet to achieve perfection with research
ongoing by 3 large players

Power
(Number of images 3.5 Billion 3.5 Billion 890 Million
used for training)

Widely available, can Free, can generate creative High quality, artistic
Pros
generate realistic images and original images images

Expensive, requires
Cons Slow, not prompt friendly Inconsistent results
coding ability

Example

(Prompt: A beautiful Sri lankan

woman wearing traditional clothes
half immerged in the Ganges river
looking at the camera with an
hypnotizing glare)

22
 Foundation Models

Computer Audition
Different Open Source Audio AI models are available

Speech Sing Audio

AudioGPT
Modules Text-to- Speech Style Speech Text-to- Audio
Text to Sing
(Open Source) speech Recognition Transfer Translation Audio Inpainting

Changing Converting
Converting Recognizing tone by Making Filling in
from one Turning text
Definition written text speech varying songs from gaps in
language to into speech
to speech coherency frequency & written lyrics audio
another
amplitude

Real-time Ariana Grande Donald Trump

dubbing while singing Insane speaking in 6
preserving voice in Punjabi languaged
and emotions
23
 Foundation Models

Robotics
R&D companies are building highly efficient robots with industrial use

Quadrupedal Humanoid Arm

Military Healthcare Logistics

24
 Foundation Models

Reasoning: Diagnostics, Planning and Forecasting

Multiple steps in business processes are getting automated by AI

Data Engineering Data Analytics Data Scientist

• Accessing, aggregating and • Business Intelligence and • Identifying irregularities &

processing raw data Visualization of data taking corrective measures
Processes
• Arranging organized data • Basic computations and • Predicting the future based
into a single repository calculations on customer on past trends
data

Tools Used

AI
N.A. Yet to be disrupted
counterparts

25
 Applications

AI applications: Emerging use cases

Time & Cost spent on
processes
High Healthcare
Manufa
cturing

Transport
& logistics
Auto
Banking and
Insurance
CPG & Retail
Travel
Software
Development

Media &
Content Creation

Low Margin for error

High Low

Media, Software Development, CPG and BFSI are low hanging fruits; Healthcare,
Manufacturing, Auto, Supply Chain are locked value pools
26
Size of Bubble: % digitization of workflow
 Applications

AI applications: Measuring Impact

AI applications can be identified on the basis of the impact they create

Acts as a proxy for numerical impact. More the number of people within a
People # of FTEs in industry
function and use case, more is the potential business impact created by AI.

# hours spent per FTE

annually
Acts as a proxy for value impact. More number of hours and more value of work of
Productivity an individual, more is the potential business impact created by AI.
FTE salary or wage per
hour

% manual work vs % Digitization of an individual’s workflow is necessary for AI adoption. If the

Programmability digitized individual is completely offline, it is hard for AI to create value.

Margin for error in use

case
Use cases having low margin for error, and high cost burden of an error are areas
Precision
where AI would create the maximum impact.
Cost of an error

27
1 Introduction to Artificial Intelligence

2 Artificial Intelligence Tech & Product Landscape

3 Potential Themes for Entrepreneurs

28
 Potential Themes for Entrepreneurs
• After defining the potential impacts of AI, we now dive deeper into the themes across the largest markets for AI applications:

software development, healthcare, banking, financial services and insurance (BFSI), consumer packaged goods (CPG) &

Retail, and Media & Creator Technology.

• In the context of India, a nation burgeoning with entrepreneurial talent, the prospects for creating AI-driven products are not only

global in scope but also hold immense promise for addressing uniquely Indian challenges.

• The development of Indian-language LLMs unlocks several India-focuses use cases, instrumental in extending the reach of AI to

rural India and democratizing access to AI and its benefits.

• Moreover, AI presents a pivotal opportunity to redefine the very essence of entrepreneurship. It fosters the emergence of novel

business models, go-to-market (GTM) strategies, and competitive advantages, thereby levelling the playing field for even the

most nascent yet talented founders to build in this domain.

 Interesting Themes in AI Applications

Software Banking and Consumer Goods & Media & Creator

Health
Development Finance Retail Tech

AI apps for Clinical Decision Personalization of Product Innovation Video applications

1 Virtualization and Making (Diagnostic, Products and Shopper built using
DevOps Monitoring, Treatment) (Personalized Offers) Marketing RunwayML

Biotech & Pharma

AI apps for Events, Demand & Inventory
(Drug Discovery, Loan Decisioning Real-time audio-to-
2 Notifications and Planning, Category
Predicting Chemical (Underwriting) audio dubbing tools
Messaging Management,
Reactions )

30
Interesting Applications in Software Development
Events & messaging, Virtualization & DevOps are high-value pool
Tech Stack Associated
Technologies Definition AI Feasibility AI Tools
Components Value / comp

JS, React, Angular, Developing interfaces High

Frontend HTML, CSS for input/output $75K

Programming Node, Java, Python,

Writing custom logic High
PHP $113K
Languages

App Spring, Django, Enabling standards for

$146K Moderate
Frameworks iOS, Java, Android building apps

Querying, managing,
Database MongoDB, MySQL
organizing data $146K High

Event & Kafka

Enabling 2-way real
High
$123K
Messaging time communication

Virtualization & Kubernetes, Making code efficient

$123K Moderate
DevOps Docker and organized

31
Interesting Applications in Health
Potential to augment Clinical Decision Making & Drug Discovery using AI
Pharmaceuticals; Biotech & Drug
Clinical Decision Making Physician Workflow
Discovery

Identification of patterns in
LLMS Authorization of treatment
protein sequences using GPT4

Financial clearance and claims Predicting outcomes of chemical

EHR, EMR, Medical management reactions
Proprietary
Journals, Clinical
Databases Images, Lecture Slides
Writing diagnostics, medical Analytics in drug discovery
coding, etc. (workflow) pipeline

Diagnostics
Personalized

Treatment

Monitoring

Strong product moat possible Difficult to establish right to win, tech moats High value play; Limited moats in India
32
Interesting Applications in Banking & Financial Services
AI will disrupt several repetitive and SOP based processes in fintech

Less Human First to be

Intervention disrupted by AI Product Personalization

Loan Decisioning Virtual Assistants

Wealth Planning

Investing Compliance

Fraud Support

Structured and Repetitive

33
Interesting Applications in Consumer Goods & Retail
CPG/Retail one of the first industries to accept ML; AI will follow

Product Development & Manufacturing &

CPG & Retail Merchandising & Sales Logistics & Delivery
Planning Procurement

Accurate Demand Low Inventory

Forecasting Trade Promotion Shorter Consumer
Key Output Effectiveness Lead Time
Increased Brand Loyalty Accurate Customer Insights

Co-Innovation Inventory Planning Trade Promotion

Key Category Mgmt Joint Logistics

Processes
Demand Planning Shopper Marketing Pricing

34
Legend: Simple ML; Complex ML; Difficult to Disrupt
Interesting Applications in Media and Creator Tech
Video emerges as the next big disruption in content creation using AI

Textual Content Visual Content Audio Content

Current Tools
Used

Specific AI content generation Applications of the world’s most

Interesting AI Real-time audio-to-audio
tools for blogs, email marketing, advanced foundational video
dubbing tools for creators
applications newsletters, tweets, etc. generation model, Runway ML

VH has invested in: VH has invested in:

Landscape
Competitors: Competitors:

35
The Rise of India LLMs
LLMs built for 750+ Indian languages to expand AI usage to rural India

India LLMs: India’s own foundation models Potential Use Cases

AI Tour Guides;, News Reporters, and Contact Center

People Agents communicating in regional language

Reducing language barriers in supply chains;

Productivity Verifying legal docs written in regional languages

OpenHathi
Programm- Automating bureaucratic processes;
ability Translating textbooks into various languages

Improved crop management delivered to farmers;

Tamil-LLaMa Precision Hyper-personalized diagnostics for rural India

36
Common go-to-market strategies in AI
GTM still being discovered; Build in Public, a special kind of PLG is popular

PLG Case Study: – achieving USD 2M ARR & 250K users within 10 months of launch

Establishing a constant, real time

Strategy feedback loop with users by providing
complete transparency on social channels

Channels
Used

GTM Output

37
Moats in AI
AI applications can win with proprietary data and distribution; Infra
and building blocks can win with a unique product

Infrastructure & Tooling Foundation Models Applications

Underlying
Can be proprietary IP Can be proprietary IP Usually no IP
Technology

Engineering
Should be high Should be high Usually limited engg. effort
Effort

Data
Independent of data Data should be easily available Should be proprietary / unique
Availability

Distribution Distribution should require

Usually low-effort digital Usually low-effort digital
Effort effort and channels

High engineering effort High engineering effort Proprietary data and locked
What should
and/or proprietary IP and/or proprietary IP distribution channels
be the right
to win?

38
The Ideal AI Founder Persona
Opportunity for young founders to create AI Infra and Middleware

Experienced founder Young founder building AI

building AI Applications Infra and Middleware

Startup Founder Google Startup Founder Sharechat

Founder | Google | Microsoft | Harvard | IITB Founder | Sharechat | IITK
Harvard IITK

Experience Experience
AI Security / AI Application AI Infra / AI Middleware
Stealth Founder Stealth Founder
Senior Software Engineer 2009-2013
Google USA ML Engineer
2020-2022
Sharechat
Software Engineer
2006-2009
Microsoft USA

Education Education
Harvard Business School, Boston
Masters of Science; CSE 2002-2006 Indian Institute of Technology, Kanpur
Bachelor of Technology; Material Science 2016-2020
Indian Institute of Technology, Bombay
Bachelor of Technology; CS 1998-2002

39