Over the past couple of years, we found ourselves investing in 7 AI and ML enabled companies, in areas as diverse as marketing, credit scoring, recruitment, fertility tracking and so on. It appears that we’ve been among the most active European investors in what most people today still view as a “theme”. Most importantly, more and more of our other portfolio companies are starting to adopt these technologies in order to make their products better.
What follows is a presentation that we gave to our LPs at our most recent investor day in February. We tried to give them a primer on these technologies: what they are ; why we are all talking about them now ; and how we, at Sunstone, are thinking about investing in those companies.
5. 5
AI is a 61-year old branch of Computer Science that uses algorithms
and techniques to mimic human intelligence
6. 6
The end goal of AI was (and still is) to build an Artificial Generalized
Intelligence holistically mimicking human intelligence.
Logical Reasoning
Perceiving the world
Navigating and moving
in the world
Moral Reasoning
Emotional Intelligence
Understanding Human
Language
Goal
7. 7
Machine Learning is one of several techniques to get computers to
perform sophisticated cognitive tasks. It focuses on giving computers
the ability to perform those tasks without being explicitly programmed.
Symbolic AI (e.g. Expert
Systems)
Probabilistic AI (e.g.
Search & Optimization)
Machine learning
Mathematical foundations
Algorithms and data structures
Artificial intelligence
Communication and security
Computer architecture
Computer graphics
Databases
…
Computer
Science
Decision Trees
Bayesian inference
Deep learning
Reinforcement learning
Support vector machines
Random forest
…
8. 8
The history of AI is a history of successive hype cycles about the
prospects of different techniques
Expert Systems
1980’s
Deep Learning
?
Markov ModelsConnectionism
2012
AI Hype Cycles and AI Winters
1960’s
…
1970’s
Source: Wikipedia ; Analysis: Sunstone
9. 9
Machine Learning is a particularly interesting technique because it
represents a paradigm shift within AI
Traditional AI techniques
Machine Learning
Data
Logic
Output
Ø Static – hard-coded set of steps
and scenarios
Ø Rule Based – expert knowledge
Ø No generalization – handling
special cases difficult
Ø Dynamic – evolves with data,
finds new patterns
Ø Data driven– discovers
knowledge
Ø Generalization – adapts to new
situations and special cases
Data
Output
Logic
10. 10
Example: excelling at playing the game of Go
Symbolic AI Mathematical/Statistical AI Machine Learning approach
“Let’s sit down with the
world’s best Go player,
Lee Sedol, and put his
knowledge into a
computer program”
“Let’s simulate all the
different possible
moves and the
associated outcomes at
each single step and go
with the most likely to
win”
“Let’s show millions of
examples of real life
and simulated games
(won and lost) to the
program, and let it learn
from experience”
11. 11
Machine Learning is particularly good at solving 2 types of problems
where other AI techniques fail
? ?
? ?
?
?
Tasks programmers can’t describe
Complex multidimensional problems that
can’t be solved by numerical reasoning
13. 13
In the past 5 years, we’ve seen unprecedented progress in solving
tough problems that defied our best efforts for 50+ years.
Unprecedented Progress
AI is Leaving the Lab and Being
Deployed in the Wild
14. 14
The confluence of 4 key factors is behind this new AI Renaissance
More Data
60 years of Research / Mature
Algorithms
More Computing Power
Open Source
Frameworks/Libraries
DSSTNE
PaddlePaddle
16. 16
We are seeing AI systems reaching equal to above human
performance at narrow tasks
Computer Performance
Human Performance
Time
Performance
we are here
Performance at Given Narrow Task Over Time
Source: Sunstone
17. 17
Google researchers built a ML model as good at diagnosing diabetic
retinopathy as human doctors (Dec 2016) – soon in production!
Source: http://jamanetwork.com/journals/jama/article-abstract/2588763
19. 19
Deep Learning models still need a lot of training data to reach
state-of-the-art performance (for now)
Significant risk of overfittingState of the art performance
Increased chance of good generalization
20. 20
Deep Learning models are excellent at mimicking training data, but
we’re still far away from building systems that “learn to learn” (for now)
Supervised Learning Unsupervised Learning
21. 21
Deep Learning models are excellent at performing narrow tasks but
we are still very very far away from generalized human-like intelligence
Déjà vu…
23. 23
AI/ML are the next major horizontal enabling technologies, just
like cloud, mobile or social. They will transform every industry and
make every product better
Infrastructure
Agriculture Education Healthcare Finance
Transportation
Legal
Industry
HR
Real Estate
Travel
Retail Advertising
SpaceGovernment
Energy
Solve complex multidimensional problems by looking
for answers in the data
(large productivity gains, close to zero marginal cost)
24. 24
…which is why a lot of money poured into companies focusing on AI
Data: Pitchbook ; Analysis: Sunstone
$194
$412 $507 $633
$1,982
$2,508
$3,247
$4,288
-
500
1,000
1,500
2,000
2,500
3,000
3,500
4,000
4,500
5,000
2010 2011 2012 2013 2014 2015 2016 2017* (ann.)
$M
Funding into VC backed AI companies ($M)
17x
25. 25
But investing in AI focused companies also has challenges –
Timing: it is increasingly difficult to filter signal from noise
Machine Learning / Deep Learning
Blockchain tech
VR
Brain/Computer
Interfaces
Conversational
UIs
Autonomous vehicles
Quantum
Computing
AR4D
Printing
3D printing
2-5 years
5-10 years
10+ years
Time to Plateau
Data: Gartner 2016 Hype Cycle
Gartner Hype Cycle 2016 (selected technologies)
26. 26
An anecdote: #RocketAI - how to create a completely fake AI company
“worth” $M in a few hours
Source: https://medium.com/the-mission/rocket-ai-2016s-most-notorious-ai-launch-and-the-problem-with-ai-hype-d7908013f8c9#.44sbmx7xf
RocketAI Launch Party Metrics at NIPS 2016
27. 27
Startups are competing against very aggressive incumbents that
have more $, data, and talent than startups can dream of
Geoffrey Hinton ; Fei-Fei Li ; Demis Hassabis
1.2BN MAUs
$19BN net income
Yann LeCun ; Joaquin Candela
2BN MAUs
$10BN net income
Andrew Ng
600M MAUs
$5BN net income
Hassan Sawaf
350M Active customer accounts
$2.2BN net income
Eric Horvitz ; Harry Shum
1.2BN office users, 500M LinkedIn profiles
$15BN net income
Ruslan Salakhutdinov
500M Apple users
$40BN net income
30. 30
Large incumbents are much better positioned to build broad
horizontal AI products and infrastructure. But startups can thrive in
vertical niches.
Solving broad AI problems: horizontal image/video/voice recognition, NLP, translation, AGI...
Solvingaparticularindustry
problem
31. 31
Data is a major source of defensibility. Access to a proprietary
dataset is a key component to build differentiated products.
More
Unique
Data
More
Accurate
Algorithm
Better
Product
Larger
Customer
Base
32. 32
… and getting paid to collect a proprietary dataset is even better!
33. Get in touch !
Yacine Ghalim
yacine@sunstone.eu
@yacineghalim