Building an AI Startup: Realities & Tactics

Matt Turck, FirstMark CapitalPeter Brodsky, HyperScienceSept 27, 2016

Building An AI StartupRealities & Tactics

MATT TURCK

Managing Director

Early stage venture capital firm

based in New York City.

FirstMarkCap.com

Largest data-focused monthly

event in the country

DataDrivenNYC.com

HyperScience is an AI company

based in New York, leveraging

unique technology to solve large

enterprise problems, starting with

back office automation.

HyperScience.com PETER BRODSKY

Co-Founder & CEO

AI is hot…

The

technology

is working…

Plenty of

hype!

Zeitgeist

VC money is pouring inA record $1.05B went into 121 private AI

companies in Q2, according to CB Insights

But the reality

from the

trenches is

different

How do you actually build an AI company?

| Positioning

| Product

| Petabytes

| Process

| People

| POSITIONING

Exciting times to build an AI startup!

Just One Small Problem

All the Big Tech Companies Got the Memo, Too

“Artificial intelligence would be the ultimate

version of Google.

We’re nowhere near doing that now.

However, we can get incrementally closer to

that, and that is basically what we work on.”

- Larry Page

CEO, Google, October 2000

They’ve Been Thinking About AI For A Long Time

They’re Now Betting the Farm On It

“We’ve been building the best AI

team and tools for years, and

recent breakthroughs will allow us

to do even more.

We will move from mobile first to

an AI first world.”

- Larry Page

CEO, Alphabet, April 28, 2016

They Can Hire

ALL THE AI ROCKSTARS

Andrew NgYann LecunGeoff HintonPeter Norvig

They Can Acquire The Best Teams

And They Have All The Data In The World

…Position

away from

them!

So What Do You Do?

V

E

R

T

I

C

A

L

However they’re not

going to tackle every

single vertical problem

HORIZONTALTech giants have a formidable

advantage when it comes to building

broad horizontal products

(image/video/voice recognition,

language translation) & infrastructure

(AI cloud)

v s

Enterprise vs. Consumer

• Tech giants, on the whole, are more focused on

consumer than enterprise

• Plenty of opportunities to deliver deep enterprise

solutions

• Fortune 1000 companies have large datasets!

Tools vs. Platforms

• Offering broad core technology (including

“strong AI”) is tricky, long-term, for any startup

• Giants may impact your business just by open-

sourcing some of their tech (TensorFlow)

• Focus on tools that solve specific customer

problems, including “last mile”

The HyperScience Experience

Broad AI technology that can be applied

to many problems

Decision #2

Back office

automation as first

beachhead

Decision #1

Focus on the

enterprise, particularly

Fortune 1000

| PRODUCT

Should your product be all AI?

Remember when I

said “the technology

works”?

LIES

It never works 100%

Sudden Perception that

Using Humans in AI = Failure

Sometimes You Need 100% Accuracy,

Sometimes You Don’t

Low Product Risk High Product Risk

Humans In The Loop: Avoid Disasters

Or Underwhelming User Experiences

Leveraging Human Users to Train the AI

AI in the Enterprise: Humans Needed!

• The “S Word”: Services.

VCs will scream in

horror!

• But reality of AI is that

services required for

successful deployment

in the enterprise

Build your Product with

Data Network Effects in Mind

Illustration Source: Moritz Mueller-Freitag,

”10 Data Acquisition Strategies for Startups”

Data Network Effects

Exemplified by Industry

Giants

Data Network Effects

Exemplified by Industry

Giants

But also available

to startups

| PETABYTES

Cold Start Problem

Usually have

to start the

flywheel here

Data Sets

Licensed data setsPublic data sets

Data Crawling

Web Crawling Real World Crawling

• While Tesla owners have driven around

100 million miles on Autopilot,

Anderson reveals that the fleet

Autopilot hardware-equipped cars has

collectively driven 780 million miles….

• Tesla basically turned its fleet of

vehicles into an incredible data

gathering asset for the Autopilot

program before enabling the software.

Electrek, May 2016

Data Capture Networks

Sense360's sensor-technology is on

more than 250 mobile apps and more

than 1.5 million devices in the US. Our

panel generates more than a terabyte

of anonymous sensor data every

single day and provides a detailed

view of more than 100 million

anonymous user visits a month.

Data "Traps"

• Consumer apps: Facebook, IoT products like Kinsa

• Enterprise apps: Slack >> Bots

• “Trojan Horse” side apps: Forevery (Clarifai)

Source: Clarifai, Recode

AI Training

| PROCESS

One Dirty Secret of AI

When it comes to successfully deploying AI in

the real world, half the battle has to do with

expectation management and social

engineering, not technological prowess

AI Engineers are a Novelty

in the Enterprise

Companies are trying to make sense of this strange

new type of vendors promising miracles

WHAT YOU THINK YOU LOOK

LIKEWHAT YOU ACTUALLY LOOK LIKE

How NOT to sell AI to the Enterprise

“Give us all your data, we’ll use

it to fine tune the algorithms in

our black box, and awesome

things will happen.”

Without real value provided upfront, that typically doesn’t work.

AI Social Engineering

• Help customers understand which problems can

be solved with AI (and which problems cannot)

• Assist customers in developing relevant testing

procedures and success metrics for AI

• Address any security or data privacy concern

| PEOPLE

One Big Misperception

• TensorFlow = doesn’t (yet) mean AI is now easy

to use!

• Not about slapping “.ai” after your startup name

• This stuff is really hard!

Need Deeply Technical Teams

• Need core machine learning talent – often PhD

level

• Also top engineers who can productize and

deploy AI

• Ideally, people who can do both!

• In most cases, CEO needs to be deeply technical

too

Rare Birds

• There is a very limited supply of such talent

• Big tech companies will pay millions just in sign up

bonus for a brand new PhD in deep learning

• Hard to attract top AI talent for a startup, but even

harder for a Fortune 1000 company

But Talent is Globally Distributed!

Recruit Globally

Deep Customer Focus

• Danger with very technical teams:

building “tech for the sake of tech”

• Focus on serving customers need to

be part of the team’s DNA

The HyperScience Experience

• 26 team members, only one “non-technical” to

handle sales (but he can code!)

• Half of the team is based in Bulgaria

• Secured customers very early in the life of the

company and built product closely with them

CONCLUSION

How do you actually build an AI company?

The Five P’s of AI

| Positioning

| Product

| Petabytes

| Process

| People

Now is the perfect time to

build an AI company!

“Do not throw

away your

shot!”