Chasm Waxing: A Startup, Cyber-Thriller (12 page)

“
Right now, a G-Master has
to watch every moment, of every game. The G-Master has to
ensure
the
Gamers don’t get stuck, and that the
cyber
-events in REALSPACE
are appropriately represented
in GAMESPACE. For example, what if the Gamers
collectively decided to seek treasure instead of thwarting a virus
in REALSPACE? Instead of fighting a monster in GAMESPACE, they’d be
filling their pockets with gold.”

“
Who could blame them?”
said Josh, sardonically.

“
That’s the problem. In
our
game
we call Castle Gecko
…
I mean Castle Chevaliers, we try to
handle this with Bitcoin rewards. But, Gamers aren’t forced to do
anything. It’s an open-world game. As G-Master, I might have to
intervene by spawning a non-player character or verbally chastising
them to get back into the fight.”

“
I
see,
” said Josh, “so your
client is Gecko Insurance?”

Becca went flush. “No, our
client builds castles,” she
replied,
with
a deadpan voice. “The point is, I could see using
your AI to replace our G-Master.”

“
I understand. I agree
that we can
definitely
take over the AI for REALSPACE when
you’re
dealing
with
cybersecurity
. And we can work further
on your G-Master replacement requirement. It does fit with my
vision of developing stronger artificial intelligence. Let me tell
you a little more about CyberAI.”

Josh stood up and moved towards Becca.
“May I?”

“
Sure.” Becca handed Josh
her coffee mug full of whiteboard markers, and the red
pen
in her
grasp. In the exchange, Becca and Josh’s hand touched—just for an
instant. Becca’s stomach tingled. All she could see were dimples.
She thought their touch
might
have passed a static electric
charge.

Becca struggled to get
back to her seat without drawing attention to herself. “You can
erase all of that.”
Phew, I’m glad my
voice didn’t squeak,
she
thought.

Josh
penned his own architectural diagram.
“I
t sounds like you’re very familiar with
AI. CyberAI is a comprehensive
cybersecurity
suite that touches
network security, application security, server security,
monitoring,
smart
availability throttling, and automated incident
response.”

Becca laughed. “How in the world do
you remember everything you do?”

Josh grinned. “I don’t.
The AI does.
Anyway
, one of my
special
focus areas was
mitigating insider threats—rogue system administrators and hackers
that game the system to get elevated privileges. If a system
administrator—like Edward Snowden—starts reviewing a lot of files
or searching in directories where
he’s never been
; CyberAI raises
alarms.

“
I’m trying to enhance
the
AI. I want it to
predict when someone will be a threat. The
General
really
likes that part. And
of
course,
we integrate with virus and
malware detection software. I’ve already told you about how we
ingest data from the web and social media sites. I’m also trying to
create an
AI-
bot that crawls the Deep Web. So that’s a broad overview.”
The Deep Web was the portion of the Internet that was not indexed
by search engines.

“
Our AI Kernel is at the
heart of all of this. We developed it to work with various AI
algorithms. Until last month, the AI Kernel focused on text—you
know with NLP—and machine learning. The algorithms operated on both
the log files and the web data.
Of
course,
the
perfect
NLP would read a book, and
understand
it as well as a human
reader.
My NLP is far from
perfect.”

Josh told Becca about his MIT
coursework and the genesis of CyberAI.

“
I used one-third of the
security log data to train my NLP machine learning algorithms.
Then, I used the other two-thirds of the
data
to see if I could detect the
attacks. I repeated this process of training and testing, training
and
testing,
until I was achieving good results. My CTO, Vish Kumar,
focused—”

“
Vish Kumar from
Graphica
Intelligence? He’s a rock star!”

“
Yeah,” replied Josh, “He
came to us after
Graphica’s
acquisition. Vish’s team is
superb
at packet inspection
and employing
graph
analytic
algorithms to identify threats.
His stuff is slick. It proactively identifies suspicious packets
from compromised hosts.”

Graph analysis used graph
theory to find
meaningful
relationships in social networks and other graph
data. Graph analysis of your Facebook friends or Twitter followers
could lead to discovering unexpected connections and
relationships.

“
So why are you in my
office?” said Becca, semi-sarcastically. “It sounds like you got
this all figured out.”

Josh huffed. “Hardly,
CyberAI
only recognizes
83% of cyber-threats. That’s too
low
. It doesn’t
do enough to minimize the need for human security administrators.
Now Vish’s stuff is killing it, but my
software isn’t improving in its recognition
capabilities.
And that’s just recognition.
I’m not moving any closer to where I
really
want CyberAI to go—prediction.
And beyond that, I want to enable discovery. I think the future of
AI is extending human intelligence, not replacing it.

“
I’ve gotten the last drop
of blood from improvements to NLP using standard machine learning
techniques. Last month, I decided to add another group of
algorithms to our AI kernel—deep
learning
. Do you know anything about
deep learning?”

“
Yes, but act like
I
don’t,
”
she replied, channeling her inner General Shields. Becca did draw a
line. She wouldn’t instruct Josh to pretend she was his
grandma.

“
Deep learning is a
particular
type
of machine learning. Deep learning algorithms break problems up
into many different layers. Statistical calculations
are performed on the data at each layer.
The information can be
images, sound,
text, graphs, and
so
on
. These calculations determine
the
essential
similarities, or features, for each layer.

“
Deep learning is inspired
by learning in the human brain. The multiple layers of a deep
learning algorithm
are
called
a,
‘neural network.’
Powerful
neural networks, trained by
deep learning, are behind some of the biggest breakthroughs in AI
in the last decade.”

“
Breakthroughs like what?”
asked Becca.

“
Breakthroughs like
self-driving cars. Autonomous
vehicles
must recognize
and react to obstacles and road conditions.
Breakthroughs like speech recognition in
digital personal
assistants.
Apple’s Siri, Microsoft’s Cortana, and Amazon’s Echo have to
understand
speech to answer questions and process commands.
Breakthroughs like Facebook’s software that automatically tags and
categorizes images without human help.”

“
That’s interesting. I
didn’t realize that deep learning was so pervasive.”

“
Who knew, right?” said
Josh. “Deep learning is becoming a new computing model. It’s going
to impact every industry. Not only is it going to allow voice
commands to be a new input for computing, but visual computing will
take off. Kids will play chess against the computer with a real
chess board.

“
Despite all its promise,
I’m struggling
to get
any meaningful impact
from
my neural network that
understands English. I’ve crawled the Internet with spiders, just
like the Atom search engine. I’ve assembled
an enormous
amount of text
to use as a training set. While my hardware certainly can’t compare
with Nucleus’ server farm, my dad was a big investor in NVIDIA. He
gave me two, first-generation NVIDIA DGX-1 deep learning
supercomputers
.

“
I used this data to train
the neural network to understand the text. Then, I applied the
neural network to the log files I’ve used in the past. When I
combined my old machine learning stuff, with new deep learning
algorithms; the AI’s inference results only improved .08%. That’s
why my demo for General Shields went so poorly.”

Becca carefully examined
CyberAI’s architecture. After some long moments in thought, she
said, “I
believe
you need a better training set.
You need a purer corpus of text that
serves as your ground truth.
My guess is
that there are too many semantic differences in your
Internet
text
.
Your neural network isn’t
understanding the
English well
enough.”

Josh looked impressed. “I’ve got a
pretty sophisticated annotation layer that allows me to add nuances
and labels to the corpora. But I agree; that’s where I need to
focus.”

“
Why not use the Bible?”
suggested Becca. “The Bible has a vast amount of text and many
different versions. They all carry the same ideas using different
words. That can help you with the semantics. If you ever want to
move beyond English, there’s a Bible
for
every language. Plus, you have a
massive amount of commentaries written in different time periods.
There’s a shared meaning between all of this corpora to help you
with semantics and labels. It’s like a
built-in
annotation layer. It’ll
give you many more
reliable
hooks.”

With a droll smile, Josh
replied, “The Bible? Are you some
kind of
Holy Roller? Hacker, computer
programmer,
hunter
, and Bible thumper?”

Becca scowled at Josh.

“
What? I’m just having
fun.”

She maintained her stare.
“No, I’m not a
Bible-thumper
. Far from it. I just
happen to know a lot about the Bible. My dad was a Pentecostal
preacher in Texas. He fed me the Bible for breakfast. He has
a vast library of
Bibles, Bible dictionaries, Bible handbooks, Bible
commentaries; he’s
really
smart
about Scripture.”

“
What’s a
Pentecostal?”

“
Pentecostals are
non-Catholics. Think Billy Graham, not the Pope. By that, I mean
they are Protestants. Pentecostals emphasize a more active role for
the Holy Spirit in their faith. So to them, God still speaks, God
still heals—things like that. A lot of TV preachers are
Pentecostals.”

“
Hmm,” replied Josh. “What
does your mom do?”

His question threw Becca
off. “My mom died of ovarian cancer when I was 10.”
Then God disappeared
,
she thought.

“
Oh, I’m
sorry,
” said
Josh. “My mom and dad divorced when I was starting high school, but
it’s nothing like losing your mom. I run the Jewish race—not
the
religion
—if you know what I mean.”

Becca didn’t catch
Josh’s
humor. She
didn’t want to talk
any more
about her mom, dad, or
religion.
“With all the variations
of text, I think you can use the Bible as a Rosetta stone to train
your neural network.”

“
I believe you’re right. I
think employing different translations of the Bible and the
commentaries as training data for my recurrent neural network makes
a lot of sense. I’d hypothesize that the text will be of higher
fidelity. It should
really
help improve my word vectors. I’m going to start
architecting and training the software tonight.

“
Once I get it trained, I
can see if the neural network improves the recognition of
cyber
-events.
Then, I can start thinking about replacing the G-Master. I’m
excited about that use
case. I
do
want to head in the area of stronger
AI. I’ve been trying to expand Vish’s vision beyond cybersecurity.
In time, I want to out-do the Atom search engine. I’d like
to
perform
discovery—not search. And I’d like to make it social
discovery
in a
VR world.”