00:00:00
Transcriber: Reihaneh Iranmanesh
Reviewer: Elisabeth Buffard
00:00:05
When most people think about AI,
00:00:07
they picture a sci-fi dystopian future,
with man versus machine.
00:00:12
Terminator,
00:00:14
Skynet, Black Mirror,
Blade Runner, Westworld.
00:00:20
But as someone who is working
00:00:22
on the most ambitious
AI projects in the world,
00:00:25
every day, I can tell you
that is far from reality.
00:00:29
To me, it’s the contrary of that.
00:00:32
AI enhances
and even supercharges humanity.
00:00:35
Let me explain why.
00:00:38
There are many reasons
why AI will never replace humans.
00:00:41
AI always has,
and always will, rely on humans.
00:00:45
That’s one of the reasons
00:00:48
that I was actually inspired
to start an AI company.
00:00:51
That and my background
have had a huge impact on me
00:00:54
and why I started Scale.
00:00:57
My parents were brilliant scientists
of Los Alamos,
00:01:00
who accomplished a lot
in advancing their field.
00:01:03
That inspired me to use
science and technology
00:01:06
to have a real impact on the world.
00:01:08
My dad was a physicist,
and my mom was an astrophysicist,
00:01:12
both at the top of their field,
00:01:14
who made meaningful contributions
to plasma fluid dynamics
00:01:18
and the beginnings of the universe.
00:01:20
Their scientific work will have
meaningful impact
00:01:22
on how we understand
and perceive our world.
00:01:25
And I wanted to work
on something as impactful
00:01:28
or even more impactful than that.
00:01:30
That’s why I decide
to become a programmer,
00:01:33
so I can make a difference in the world.
00:01:36
Growing up as a programmer,
00:01:38
despite how powerful computers are,
00:01:41
you quickly realize
how limited they are.
00:01:44
In particular, they lack
judgment and intelligence.
00:01:48
Programming is the art
of giving clear robotic instructions
00:01:52
to computers to accomplish
simple objectives.
00:01:56
It’s all black and white,
and there’s no gray area.
00:01:59
As I learned about AI,
it was clearly transformational.
00:02:04
It changed the game.
00:02:05
It equipped computers with intelligence,
and I knew I wanted to be deeply involved.
00:02:12
I was studying AI at MIT and slowly
became more and more excited
00:02:16
about all the potential applications of AI
for solving more nuanced problems.
00:02:22
For example, there was one class project
00:02:24
where I worked on applying AI
to human emotions.
00:02:29
The goal was to take picture
of human expressions
00:02:32
and ultimately identify and understand
the emotion through very subtle signals
00:02:37
in facial expressions.
00:02:39
Using AI, we built an algorithm
00:02:42
that was able to detect intent
with 80% accuracy and efficacy.
00:02:47
We were extremely proud of that.
00:02:49
It was the start of using AI to do
entirely new things using computers.
00:02:54
That’s when I realize
the implications of AI
00:02:57
and how it could tackle the gray areas
that involve judgment or intelligence.
00:03:02
You see, AI needs humans
to teach it individual values,
00:03:07
nudge it to find thoughtful outcomes,
00:03:10
and ensure that human intentions
and values are aligned with the AI.
00:03:14
It was a revelation.
00:03:16
Before, coding was like
a black-and-white film
00:03:20
versus watching in technicolor.
00:03:24
What’s more, AI has the potential
to take away the repetition in our lives,
00:03:28
meaning that new and fresh ideas
will matter more
00:03:31
and ultimately enable us to be more human.
00:03:36
But, to power AI,
you need powerful data,
00:03:39
which was especially hard to come by
at that time, in 2016, while I was at MIT.
00:03:45
I realized that nobody was building
anything with AI outside of school.
00:03:50
It’s unusual for MIT students
to not be building something.
00:03:54
Mechanical engineering majors
are building catapults in the lawn,
00:03:57
electrical engineering majors
are building robots,
00:04:00
and computer science majors are building
apps for their friends to use.
00:04:04
But nobody was building
anything using AI.
00:04:08
That’s when I discovered what
a bottleneck data can be
00:04:11
to building meaningful
and powerful AI systems.
00:04:15
You can't treat data as an afterthought.
00:04:18
Bad data or lack of data
results in bad AI.
00:04:22
I even realize this in my personal life.
00:04:25
I put a camera inside my fridge
to gather data,
00:04:28
to tell me when to refill my groceries
and what I needed to buy.
00:04:32
That’s when I realized
just how much data I needed
00:04:36
to actually be able
to successfully predict what to purchase.
00:04:39
There’s no way I could create enough data
00:04:41
to be successful
with the algorithms on my own.
00:04:44
But it did help me discover
that my roommate was stealing my food.
00:04:48
(Laughter)
00:04:51
At that point, I realized
00:04:53
that this was going to be
a pivotal problem for AI.
00:04:56
Building large-scale,
high-quality datasets
00:04:59
to power every single application.
00:05:02
This was the impetus
behind starting Scale:
00:05:04
quality data, to create
reliable AI outcomes,
00:05:08
requires human insight and guidance.
00:05:11
If you think about the core setup of AI,
the algorithms need data,
00:05:15
and data needs humans.
00:05:17
To ensure data is accurate,
00:05:19
an expert human is often required.
00:05:22
Only humans can understand
the context and nuance
00:05:25
to properly annotate
the data to be fed to algorithms.
00:05:30
Humans are the one who teach
the algorithms what to do.
00:05:33
They’re the ones making the decisions,
they guide them.
00:05:35
If something happens,
here’s what you should do.
00:05:38
And AI learns from that
and replicates it.
00:05:41
We are teaching the AI
our individual values
00:05:44
and nudging the AI
to find thoughtful outcomes.
00:05:48
Machines make mistakes.
00:05:49
We have to teach them and
incentivize them to tell the truth.
00:05:53
This is why teaching the AI human
intentions and values is so important.
00:05:57
It’s through this process
that we will ensure
00:05:59
that AI will have fair, ethical outcomes
in line with human values.
00:06:05
It’s this alignment
that we must solve for.
00:06:08
The constant alignment of AI to human
intentions will always require humans.
00:06:13
and human ideas and creativity
can actually matter much more,
00:06:17
with the power of AI behind them.
00:06:21
The long tale of real-world problems,
00:06:23
and the fact that there’s always
unknown unknowns
00:06:25
means that humans
will never be fully removed
00:06:28
from the AI development lifecycle.
00:06:30
For example, I remember back in 2016
00:06:33
when chatbots were first starting
to become a big thing.
00:06:36
It was right when we were starting Scale.
00:06:39
We were all thinking there's no way
to build a fully automated system.
00:06:43
There’re so many different conversations
that can have so many different pathways.
00:06:47
It’s hard to build AI systems that can
properly handle all these possibilities.
00:06:53
For chatbots to work, there’re humans
behind it who make the decisions once,
00:06:58
and from there, the chatbots
can replicate that over and over again.
00:07:02
That’s again why it’s impossible
for AI to improve
00:07:06
or change without human input.
00:07:09
Let’s take you to the front lines of AI.
00:07:11
The things that AI automates first
are not what you might expect.
00:07:16
An unintuitive example is the weather.
00:07:19
Humans have tried for many millennia
00:07:21
to crack the code
of how to predict the weather.
00:07:24
It’s especially hard for meteorologists
00:07:26
because there are
so many different small things
00:07:28
that can cause massive impacts
on the weather.
00:07:31
It's the butterfly effect.
00:07:32
Different elements react to one another
in unexpected ways.
00:07:37
There’re so many inputs
that affect the weather,
00:07:39
way more data than any person
would be able to comprehend on their own.
00:07:44
That’s why we need AI
to analyze the vast oceans of data
00:07:48
and provide more accurate, nuanced,
and comprehensive analysis.
00:07:52
At the moment,
00:07:53
AI can already provide extremely accurate
short-term predictions,
00:07:57
including for critical storms and floods.
00:08:00
So, it’s not what humans perceive
to be the simplest task
00:08:05
that AI will automate first,
but rather where we have the most data.
00:08:10
The use cases the brightest minds
are focusing on
00:08:13
are much more positive
than what you might think.
00:08:16
Much more so than Terminator or Westworld.
00:08:19
That’s again why I think AI
will be a supercharger for humanity.
00:08:23
Unlike the movies, AI developers
aren’t focusing their attention
00:08:28
on building replacements for humans.
00:08:30
They’re building tools
to help free up our time and energy
00:08:33
to focus on what human
can uniquely solve.
00:08:37
A good example about how AI
can be used in practice is health care.
00:08:42
According to the Association
of American Medical Colleges,
00:08:45
the United States could see
an estimated shortage
00:08:47
of between 38,000
and 124,000 physicians by 2034.
00:08:54
AI could save doctors’ time
with rogue tasks
00:08:57
and ultimately enable them to serve
more patients and help more people.
00:09:02
Health care is full of repetitive tasks
which are right for AI.
00:09:07
When a patient is sick,
they go through all kinds of tests
00:09:10
which produce all sorts of data:
blood tests, imagery,
00:09:14
lab results, X-rays, etc.
00:09:18
Doctors then analyze all this data
to make decisions about a case.
00:09:23
AI can analyze all this data proactively
00:09:27
and go through a list of possibilities
00:09:28
by cross-referencing
against all prior data in cases.
00:09:32
It can identify when something isn’t right
long before a physician can
00:09:37
and flag it to a physician,
if it requires more attention.
00:09:40
With AI,
doctors are still integral to the process,
00:09:44
but it takes less time to get a diagnosis.
00:09:46
You have to wait several weeks
00:09:48
for your case to go
from one doctor to another.
00:09:50
The AI will supercharge,
finding a diagnosis faster.
00:09:55
Similarly, in the field of drug discovery,
it’s all about using complex data:
00:10:00
experiment data, patient data,
00:10:02
protein simulations and far more
00:10:05
to guide a more efficient process
00:10:08
of solving diseas
through new drugs and compounds.
00:10:11
Recent advancements in AI
00:10:13
have dramatically sped up
the scientific process
00:10:16
by allowing us to process and make us
of more data than ever before.
00:10:21
Another good example, and potentially
more concrete, is the Russia-Ukraine war.
00:10:26
We've all seen the images of tanks
lining up ready to enter Kiev.
00:10:31
AI can help assess satellite imagery
with superhuman speed and precision,
00:10:36
so Ukrainian forces can respond faster.
00:10:39
At Scale,
00:10:41
we’re using our platform
to do damage assessment
00:10:43
in key areas affected by the war.
00:10:45
We’ve rapidly analyzed
over 2,000 square kilometers of Ukraine,
00:10:50
identifying over 370,000 structures,
00:10:53
including thousands not previously
available via other datasets.
00:10:58
We focused on Kiev, Kharkiv and Dnipro,
00:11:01
in which we provided some data
directly to government and users.
00:11:05
We also made the data publicly available
to the broader AI community via Scale.
00:11:11
We can also use this data
to maximize allocations of resources,
00:11:15
people or commodities.
00:11:17
It’s clear satellite data
can be extremely useful
00:11:20
in these types of situations.
00:11:22
Thanks to satellite data,
AI can analyze if planes or tanks
00:11:26
have been moved from one place to another.
00:11:28
This is called change detection.
00:11:30
Algorithms can constantly be monitoring
for this kind of data,
00:11:33
and if it notices a change or movement,
00:11:36
it will alert a human
to further investigate,
00:11:38
otherwise known as predictive modeling.
00:11:41
AI can also help us understand
the economic impacts of war.
00:11:45
We can use AI to track
farmland in Ukraine
00:11:47
and measure the agricultural damage
in real time.
00:11:51
Ukraine is a major food supplier
for much of the world.
00:11:54
Understanding these impacts
is absolutely critical.
00:11:58
In conclusion, AI is not
something to be feared,
00:12:03
but it’s a tool that can be used
to better understand…
00:12:06
that needs to be better understood,
00:12:08
and has the potential
to transform our lives for the better.
00:12:11
AI enables us to make use and sense
of massive amounts of data
00:12:15
that has historically been
beyond human capacity.
00:12:18
It allows us to add
intelligence and nuance
00:12:20
to automated systems
that will dramatically improve humanity.
00:12:24
Areas like health care and agriculture.
00:12:27
This then allows humans
to do what they do best.
00:12:29
Take this information,
put it into context with sensitivity,
00:12:33
to strategize and act in a timely manner.
00:12:36
AI is a supercharger for humanity.
00:12:39
When AI is better than humans,
00:12:41
it makes humans better.
00:12:43
AI will automate repetitive tasks
00:12:45
that don’t require
constant human judgment or creativity,
00:12:48
which frees us up to explore
and focus on fresher, newer ideas.
00:12:53
AI will enable us to be
even more creative and more idea-driven,
00:12:57
which I personally find
incredibly exciting.
00:13:00
It allows us to embrace
the generative aspects of human nature,
00:13:04
so we can run faster with ideas
00:13:06
and build better
and more powerful solutions
00:13:08
to the world’s biggest problems.
00:13:11
That’s why I believe
that human-led AI is the path forward,
00:13:15
and I’m proud to usher all of us
into a future with human-led AI.
00:13:18
Thank you.
00:13:19
(Applause)
