00:00 - 00:05
There's kind of an illusion with
generative AI. "This promises to be the viral
00:05 - 00:07
sensation that could completely
00:07 - 00:09
reset how we do things."
00:09 - 00:13
According to all the headlines, it's on
the brink of solving all business problems
00:13 - 00:13
automatically
00:14 - 00:17
with the slight side effect of displacing
huge amounts of the workforce.
00:18 - 00:21
It seems so amazing. It's potentially
a panacea.
00:23 - 00:25
It's hyperbole. It's hype.
00:26 - 00:28
What we get with generative AI
is extremely
00:29 - 00:31
but it's not going to run the world.
00:32 - 00:34
It does have the ability to
create efficiencies,
00:35 - 00:36
but it's more limited.
00:36 - 00:40
Whereas predictive AI, which is older,
00:41 - 00:44
very much still has great amounts of
untapped value.
00:45 - 00:49
I'm Eric Siegel. I'm the co-founder and CEO
of Goodr AI,
00:49 - 00:54
the founder of the Machine Learning Week
conference series, and the author of "The
00:54 - 00:55
AI Playbook:
00:55 - 00:58
Mastering the Rare Art of Machine Learning
01:00 - 01:01
Deployment."
01:05 - 01:09
I became fascinated with the concept of
artificial intelligence as a kid in the late
01:09 - 01:12
seventies and then in the early eighties.
01:13 - 01:18
my education led me to machine learning,
and I've been in the field since nineteen
01:21 - 01:23
Kasparov, after the move c4,
01:24 - 01:25
has resigned."
01:25 - 01:30
Now I was sort of semi-horrified with the
AI hype for a few
01:30 - 01:35
decades, and it just got a lot worse in
recent years because of generative
01:35 - 01:38
AI. It's going to feed that frenzy because
it's so
01:38 - 01:39
seemingly human-like.
01:40 - 01:43
Generative AI, something like chatGPT,
01:43 - 01:45
a large language model, it is capable
01:46 - 01:48
of communicating about any topic
01:48 - 01:53
and often giving responses that seem to
understand what you're saying.
01:53 - 01:55
And I grant that on some level,
01:56 - 01:57
it has captured
01:57 - 02:01
understanding and the meaning of words and
phrases and sentences and paragraphs.
02:01 - 02:02
But I believe
02:03 - 02:07
that the difference between what it can do
and what humans can do is going to
02:07 - 02:09
become increasingly apparent.
02:13 - 02:16
Generative AI is sort of correct often
02:17 - 02:20
only as a side effect. When people say
"hallucinate," they're like, "Well, look. It
02:20 - 02:22
just makes things up."
02:22 - 02:23
What impresses me
02:24 - 02:28
is that it actually gets things right
sometimes because it's only working on that
02:28 - 02:32
low level of detail, the per-word level,
which results in that sort of seemingly
02:32 - 02:33
human-like capability.
02:34 - 02:37
There's a big difference between that
impressive capability
02:37 - 02:41
and the potential value. It's certainly
valuable for writing first drafts.
02:42 - 02:46
So it'll produce a first draft of a letter
you need to write or a syllabus or
02:46 - 02:51
something like that. But you can't trust
it blindly. You have to proofread
02:51 - 02:52
everything that it gives you.
02:53 - 02:55
That actually, in a way, makes it less
03:00 - 03:04
The whole point of computers is to
automate. Right? It does things really fast. And
03:04 - 03:09
to the degree that we can actually trust
it well enough to do things automatically,
03:09 - 03:13
that ultimately helps the economy. It
helps the efficiencies of the world.
03:13 - 03:18
Predictive AI, that's the technology you
turn to when you want to improve your
03:18 - 03:20
existing largest scale operation.
03:21 - 03:24
It does have the potential to enjoy the
benefits of autonomy.
03:25 - 03:30
So predictive AI or enterprise machine
learning, that's the technology that learns
03:30 - 03:32
from data to predict in order to improve
03:33 - 03:37
any and all of the millions of decisions
that make up large-scale enterprise
03:38 - 03:42
And these are the things that make the
world go round. So predict who's going to buy in
03:42 - 03:46
order to decide who to contact with
marketing, which transaction is most likely to
03:46 - 03:49
be fraudulent to decide which transactions
to block or
03:50 - 03:53
audit, which train wheel is most likely to
fail in order to decide which one to
03:53 - 03:56
inspect. It's not just train wheels. The
New York Fire Department does that to
03:56 - 04:00
predict which buildings are at most risk
of fire to triage and prioritize
04:00 - 04:00
inspections,
04:01 - 04:05
or which healthcare patient should we
take another look at before discharging
04:05 - 04:08
because they're predicted very likely to
be readmitted to the hospital?
04:09 - 04:11
All of these predictive applications
04:11 - 04:13
are a form of prioritization
04:14 - 04:19
and the computer is systematically making
those decisions over and over again real
04:19 - 04:20
fast, fully autonomous.
04:22 - 04:26
So we have data. We give it to machine
learning, which is the underlying
04:27 - 04:31
It generates models that predict, and
those predictions improve all the large-scale
04:31 - 04:32
operations that we conduct.
04:34 - 04:36
Predictive AI is so applicable
04:36 - 04:37
across industries.
04:38 - 04:42
Let's take the delivery industry. UPS is
one of the biggest three delivery
04:42 - 04:45
companies in the United States, and
they actually
04:46 - 04:49
streamlined the efficiency of
their deliveries
04:49 - 04:50
by predicting
04:50 - 04:51
tomorrow's deliveries.
04:54 - 04:55
That makes such a big difference
04:56 - 05:00
that in combination with another system
that actually prescribes the driving
05:01 - 05:02
to this day, UPS enjoys
05:03 - 05:07
savings of three hundred and fifty million
dollars a year and hundreds of thousands
05:07 - 05:09
of metric tons of emissions.
05:09 - 05:14
So this is how it works. When they have to
start planning and then loading the
05:14 - 05:18
trucks in the late afternoon or early
evening so that it'll be ready the next
05:18 - 05:20
morning, they have incomplete information.
05:21 - 05:26
What they don't know is some of the
packages that are still coming in later that
05:26 - 05:29
night. So what they do is they augment the
known information, which is that they
05:29 - 05:33
already have a bunch of packages in hand
that they know are meant to go out tomorrow
05:33 - 05:34
morning for their final deliveries.
05:35 - 05:37
And they'll augment that with tentatively
05:39 - 05:40
predicted deliveries
05:40 - 05:44
by applying a predictive model for each
potential delivery address and saying, "Hey,
05:44 - 05:47
what are the chances that there'll be a
delivery there tomorrow?"
05:48 - 05:49
Now they have a more complete
05:49 - 05:52
picture of all the deliveries needed
for tomorrow.
05:53 - 05:57
They can do a better job planning and
loading the packages overnight so that when
05:57 - 06:02
the trucks go out in the morning, they'll
have relatively optimal routes that don't
06:02 - 06:04
take too many miles of driving, too
much gasoline,
06:05 - 06:06
too much time of the drivers.
06:07 - 06:10
Now some of those predictions will be
wrong, but they're confident enough that the
06:10 - 06:13
completeness now actually overweighs
06:13 - 06:15
some of that uncertainty.
06:15 - 06:20
This is what you need to do if you want to
improve existing large-scale operations.
06:20 - 06:21
You need to work with
06:21 - 06:22
probability.
06:23 - 06:25
Assign a number. How likely is
this outcome?
06:25 - 06:27
Here's the thing.
06:27 - 06:31
It doesn't make a difference how good the
number crunching is unless you act on it.
06:31 - 06:35
It's not intrinsically valuable. The value
only comes if you actually
06:35 - 06:38
deploy it and change your
existing operations.
06:46 - 06:47
We have this incredible
06:48 - 06:51
seemingly human-like capability of
generative AI,
06:51 - 06:56
which in one sense, I think is the most
amazing thing I've ever seen.
06:57 - 07:00
underlying the excitement is the idea that
we are moving
07:00 - 07:03
steadily towards and potentially very near
07:05 - 07:06
Artificial General Intelligence,
07:07 - 07:12
which is a computer that can do anything a
person can do. It's this feeling
07:12 - 07:15
of a computer, kind of, coming alive,
like Frankenstein,
07:15 - 07:18
which we see over and over again in
science fiction movies.
07:19 - 07:20
In the real world,
07:21 - 07:26
I do not believe we're going to fully
replicate humans anytime soon or that
07:26 - 07:29
we're actively making progress in that
direction. That is a recipe
07:30 - 07:31
for mismanaged expectations,
07:32 - 07:33
otherwise known as hype.
07:34 - 07:36
The antidote to hype is simple.
07:37 - 07:38
Focus on concrete value.
07:39 - 07:42
Discover whether you're using generative
AI or predictive AI.
07:42 - 07:44
Determine a very
07:46 - 07:50
credible use case of exactly how
this technology
07:50 - 07:53
is going to improve some kind of operation
in the enterprise
07:54 - 07:55
and deliver value.
07:55 - 08:00
If you want to just sort of explore how
close is it to the human mind and why you
08:00 - 08:03
think it might be getting there, that's
kind of a philosophical conversation,
08:03 - 08:06
and that's great. But if you're talking
about, sort of, improving efficiencies of
08:06 - 08:10
operations that make the world go around,
I think we should be a lot more practical
08:11 - 08:12
and less pie in the sky.
video to article