Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:06
Welcome to Practical AI. If
0:09
you work in artificial intelligence, aspire
0:12
to, or are curious how
0:15
AI-related tech is changing the world,
0:17
this is the show for you.
0:20
Thank you to our partners
0:22
at fly.io, the home of
0:24
changelog.com. Fly transforms
0:26
containers into microvms that run on
0:28
their hardware in 30 plus regions
0:30
on six continents. So you can
0:33
launch your app near your users.
0:35
Learn more at fly.io. Welcome
0:42
to another
0:44
fully connected episode of the
0:46
Practical AI podcast. This
0:49
is a fully connected episode where we
0:51
keep you connected with everything
0:54
that's happening in the AI community,
0:56
all the interesting and crazy news
0:58
out there, and hopefully a few
1:00
things that will help you level
1:02
up your machine learning game. My
1:04
name is Daniel Wightnack. I am
1:06
the founder and CEO at Prediction
1:08
Guard, and I'm joined as always
1:11
by my co-host Chris Benson, who
1:13
is a principal AI research engineer
1:15
at Lockheed Martin. How are you doing, Chris?
1:17
I'm doing great today, Daniel. How are you
1:19
doing? I am
1:22
doing well mentally,
1:24
a little bit less physically.
1:26
I ran a half
1:28
marathon yesterday, which was really
1:31
exciting, and the first sort
1:34
of running type event that
1:36
I've done personally. And
1:38
I have to say, my training was
1:41
going good for a while. I would
1:43
say the last couple months was not
1:45
going as well. And so let's
1:48
just say that I'm in a good amount of
1:50
pain today, but self-inflicted,
1:52
I guess. It is. I'm
1:55
sorry. I sympathize. I have done a
1:57
couple of half marathons, but it has been a while
1:59
since I've done it. I've done them and
2:01
I know that at the end of those I was
2:03
indefinitely You sound much better than
2:05
I did afterwards. I gotta tell you well, I've been
2:07
in bed most of the day Since
2:10
you and I can see each other but listeners
2:12
can't I will I will report that you look
2:14
very well for someone who just I have marathon.
2:16
I I look terrible at the time. I am
2:19
sitting in a chair not moving. So Yeah,
2:22
that's key. Excellent Well, I guess
2:24
someday it will be doing half
2:26
marathons and there'll be things like
2:29
robots running along beside us Maybe
2:31
powered by artificial general intelligence and you
2:34
know, they'll have their own I'm presuming
2:36
We don't have to compete against the
2:38
robots I'm I'm hoping because you
2:41
know, I I don't think I would
2:43
do very well or maybe I'll just
2:45
have some sort of automated or augmented
2:49
Knees or legs put in and I
2:51
can cyborg the Marathon,
2:53
you know, they've long had Meniscus
2:56
is the is the cushioning in your
2:58
knees. They've long had meniscus transplants,
3:00
but maybe they'll have like robotic
3:02
You know intelligent meniscus and like
3:05
it springs you up, you
3:07
know I'm a off or something like that
3:09
Nobody you'll have that edge and they'll have
3:11
to detect it then, you know the competition,
3:13
you know for everything being equal Who
3:16
knows where we're going on that but
3:18
you know, yeah speaking of autonomous systems
3:21
And you know in the spirit of
3:23
robots and stuff. I thought
3:25
I would kick us off by talking about The
3:29
I've been keeping track of kind of an ongoing
3:31
news story, but it popped up in the last
3:33
week or so Which is the
3:36
x62 a Vista, which is
3:39
it's a project that the Air Force has
3:42
been leading with a number of companies and
3:45
For full disclosure Lockheed Martin my employer
3:48
is involved though I personally have absolutely
3:50
nothing to do with this and
3:52
my information is only what's available publicly So
3:55
I just wanted to give my my
3:57
disclosure there before we got into but I've been following
3:59
the news stories on this because it
4:01
is super cool. It is an F-16
4:04
Fighting Falcon fighter plane,
4:07
which are, they've been around for a long time.
4:09
They're actually 50 years old this year, but
4:12
it's gone through multiple ownership. Lockheed Martin is
4:14
the owner of the F-16 now. And
4:17
it's kind of one of those for NATO
4:19
countries, kind of standard baseline
4:22
fighter planes. The
4:24
reason it's an X-62A versus
4:26
an F-16 in this case
4:29
is it has been enabled
4:32
with a fully autonomous AI autopilot
4:34
that's not only designed to fly
4:36
the plane, but flies the plane
4:38
in combat. And they have been
4:40
doing simulated tests for the past,
4:42
roughly the past year. I don't
4:44
have all the dates in front
4:47
of me and stuff. But
4:49
this last week it made a new splash because
4:52
in addition to the usual
4:54
human test pilot, which sits in the
4:56
cockpit but does nothing, they have manual
4:59
controls to override the AI. But on
5:01
all the tests, they have not needed the
5:03
test pilot to do anything because the AI
5:05
autopilot is so darn good. And this past
5:08
week, the Secretary of the United States Air
5:10
Force also flew in the
5:13
cockpit. It has two seats and flew in the front
5:15
seat with the test pilot in the back seat. Neither
5:18
human touching any controls while
5:20
they did a simulated combat
5:23
scenario in the sky with
5:25
other airplanes flying against,
5:27
you know, in a human controlled
5:29
airplane, human controlled airplanes against
5:31
other test pilots flying combat
5:34
scenarios. And rumor,
5:36
according to what the news reports
5:38
are, everything has just gone flawlessly.
5:41
It performs exceptionally well.
5:43
And it's just, you
5:45
know, it's one of those moments in time
5:47
where you realize this stuff, it's, you know,
5:49
we, we talk about models and often our
5:51
models are, you know, just in the cloud
5:54
and we're using them on apps and things
5:56
like that. But this is a type where
5:58
you have a model that It is in the
6:01
lingo out on the edge. It
6:03
is controlling an advanced piece of
6:05
machinery to a very high degree
6:07
of performance. We
6:10
kind of had the moment with Tesla cars doing
6:12
full auto, but now we're talking about
6:14
some of the most sophisticated
6:17
aircraft in the world, not
6:19
just little drones, but big
6:21
full-on fighter planes being
6:23
flown as well as any human or better
6:25
than any human fighter pilot in the world.
6:28
What do you think of that? I've talked to him for a
6:30
while, but I'd rather take him with
6:32
just the moment. It's really interesting in a
6:35
number of ways. I was thinking back to,
6:37
I guess it was last
6:40
month when I was in Boston and I
6:43
got to stop by the MIT
6:45
Media Lab for an event. They
6:47
had a panel with some various
6:49
luminaries. One of the panels was
6:51
an investor panel. They
6:54
were all talking. Some of the
6:56
questions were, of course, related to various
6:59
things about AI. It was an AI-focused
7:01
event, but I was struck by one
7:03
of the comments about this
7:06
next wave of innovation in AI.
7:09
The panelist was basically saying
7:11
that the days of just
7:13
being an innovator in AI as a model
7:17
builder, as a
7:19
foundation model builder, are in
7:22
some ways over. What's
7:24
really interesting now is embedding
7:27
AI everywhere in the physical
7:30
world. At the
7:32
edge, here's an example of that
7:34
happening in an airplane, of course,
7:36
but there's certainly other things happening
7:39
in the civilian space as well with
7:41
AI assistance in the retail
7:44
environment. Also, of course,
7:46
in cars and that sort of
7:48
thing. Retail
7:50
environments or manufacturing environments,
7:53
agriculture, machinery,
7:56
all of these sorts of things where AI is
7:58
going to be involved. embedded in all
8:00
of these physical spaces. That
8:02
brought up that in my mind as I
8:05
was thinking back to that event. But then
8:07
also thinking here, I
8:09
know you've made some comments
8:11
before being a pilot yourself,
8:13
just a civilian aircraft pilot
8:16
about the AI systems that
8:18
already exist, for example, for
8:21
commercial airliners and
8:23
other systems that actually can even
8:25
now do better in many
8:28
ways than human pilots. But
8:30
then there's always that, I guess,
8:32
fear on people's parts where it's
8:34
acceptable for a human to make
8:37
a mistake in such a scenario
8:39
because they could potentially be punished. Of
8:41
course, in air flight, maybe they wouldn't
8:44
survive if they made a mistake, which
8:46
would be really unfortunate. But
8:48
for a machine to make a mistake in
8:50
such a scenario is sort of unforgivable because
8:52
the machine shouldn't make a mistake. So
8:54
there's kind of this double standard that's
8:56
happening. Do you see that shifting or
8:58
changing at all with some of these
9:01
recent developments? I think it'll take longer
9:03
in the commercial airspace. And just to
9:05
address one quick thing, to the best
9:07
of my knowledge at this moment, there
9:09
are no AI systems authorized by the
9:11
FAA in the United States to fly
9:13
commercial airliners. But there's a lot of
9:15
interest in testing about those kinds of
9:17
systems that are out there. There's even,
9:19
I may be wrong about this, but
9:21
I believe it was MIT that has a
9:23
system that is designed for
9:26
that. It's not been deployed in production. It's
9:29
kind of an open system for
9:31
airliner navigation and such. But
9:33
there's a lot of work in this area. And certainly on
9:35
the military side, there's lots
9:37
and lots of constraints. So I don't
9:40
wanna represent it as like, oh, you
9:42
can do whatever you want. There's tons
9:44
and tons of gateways you
9:46
have to earn your way through in the testing.
9:49
But there is definitely a full
9:52
on interest in military circles
9:54
and defense circles about
9:57
using AI in just about
9:59
every conceivable. use case that you might
10:01
want to come up with on the ground in the air under
10:03
sea in space, you name it everything and
10:06
That's without getting sidetracked I spend a lot
10:09
of time in those scenarios in my day
10:11
job away from the podcast But
10:14
many things in the military
10:16
world are classified and you can't really talk about on
10:18
one of the really cool things about The
10:21
x62 a program is it's being done in the
10:23
light of day It's a
10:25
news story every time something news happens and you can go
10:27
and and search it and
10:29
find all sorts of Information
10:31
about it. It's interesting over
10:33
time if you over the last
10:35
few years I am
10:37
one of those people because I've seen
10:40
this a lot as a
10:42
pilot and as just a Non-pilot
10:44
I will trust myself to AI
10:46
autopilots and trust my family's lives
10:48
if it were to come to that Because
10:50
they're so darn good that I've seen them
10:52
back as far back as a
10:54
DARPA event that was public on YouTube in 2020
10:58
It was a simulator But the AI pilot
11:00
beat one of the best fighter pilot instructors
11:02
in the world an Air Force instructor Yeah
11:04
the equivalent of what people would know is
11:07
top gun in the Navy and just Amolish
11:09
the poor guy and that
11:11
was four years ago now and over four
11:13
years ago And so we've come you know,
11:15
that's the prehistoric times in AI, you know
11:17
in the way we think of AI So
11:20
I really do Think that
11:22
we're we're crossing some thresholds now
11:25
and really the thing that will hold us
11:27
back is the public becoming comfortable
11:30
enough to really You know embrace
11:32
the technology as that and I think one
11:34
of the before I draw
11:36
to an end and I'm not picking
11:38
on Boeing But the Boeing, you know
11:40
problems with the 737 max, which is
11:43
not an AI system They are automated
11:45
systems, but they're not AI systems Has
11:48
really shaken the public's trust
11:50
in automation in aircraft and
11:52
airliners And so there's
11:54
that will slow things down but
11:56
you know someday when we do
11:58
have FAA approved systems and the airliners
12:01
that we're all flying every day,
12:03
I think that we will be orders
12:05
of magnitude safer than we are with even
12:07
seasoned airline pilots today. I'm so sorry as
12:10
a pilot to say that to you pilots
12:12
out there. I don't mean that. I have
12:14
many good friends who are
12:16
in that occupation, but that's just the way AI
12:18
is. It's quite amazing. If
12:32
you're anything like me, you have a
12:35
certain tendency to put things off until
12:37
the very last minute. Seeing
12:39
the dentist, going to the doctor,
12:41
home improvements, that never ending chore
12:44
list of yours. And
12:46
while most of the time it works out
12:48
just fine, the one thing in life that
12:50
you really cannot afford to wait on is
12:52
setting up term coverage life insurance. You've
12:55
probably seen life insurance commercials on TV and
12:57
thought, yeah, I'll look into that later.
13:00
No, later doesn't come. This really
13:02
isn't something you can wait on. Choose life
13:04
insurance through a ladder today. Here's
13:06
what we love about ladder and why we allow
13:08
them as a sponsor. They are 100% digital. No
13:12
doctors, no needles, no paperwork, when you
13:14
apply for $3 million in coverage or
13:16
less. Just answer a few questions about
13:18
your health in an application. Ladders
13:21
customers rate them 4.8 out of
13:23
five stars on Trustpilot and they made
13:25
Forbes best life insurance 2021 list. You
13:28
just need a few minutes and a phone or
13:30
laptop to apply. Ladders smart algorithm
13:32
works in real time, so you'll find out
13:34
if you're instantly approved. No hidden fees, you
13:37
can cancel any time. Get a full refund
13:39
if you change your mind in the first
13:41
30 days. Ladder
13:43
policies are issued by insurers
13:46
with long proven histories of
13:48
paying claims. They're rated A
13:50
and A plus by AM best.
13:53
Why since life insurance costs more as
13:55
you age now. Yeah, right now. Now's
13:58
the time to cross it off your list. So
14:01
go to ladderlife.com-practical-ai
14:03
today to see
14:05
if you're instantly
14:07
approved. Again that's
14:09
ladder.com-practical-ai. L-A-D-D-E-R
14:14
life.com-practical-ai. Well,
14:35
Chris, one of the things that I was
14:37
thinking about when you were bringing up this
14:39
story about the X-62
14:42
autonomous testing was
14:44
one of the comments you talked
14:46
about was the sort of regulations
14:48
and guardrails around the testing that
14:50
it's also happening in open. There's
14:53
regulations, especially in the airspace
14:55
about testing these vehicles and
14:57
that sort of thing. I
15:00
was remembering back I had a conversation
15:02
with breakfast
15:04
with a group that just
15:07
came out here to Purdue
15:09
University where I'm located. The
15:11
company is called Wind Racers and
15:14
they have sort of
15:16
commercial autonomous drones that
15:18
are really kind of mid-sized
15:21
drones that do like
15:24
mail remote or
15:26
rural mail routes or something like that. If
15:28
they send mail in the UK, they have
15:31
drones that take mail out to all
15:33
of these different islands in the UK
15:35
that need mail deliveries and that sort
15:37
of thing. But then also there's the
15:40
chance to use these for disaster
15:42
relief or humanitarian aid and
15:45
that sort of thing. I know
15:47
one of the things that they talked about
15:49
was just the struggle in finding ways to
15:53
test autonomous drones,
15:55
especially in the airspace,
15:58
to actually make some significant progress in
16:01
the R&D and testing and all that,
16:03
you actually have to be able to
16:05
take flights over significant distances and that
16:07
sort of thing. And here
16:10
you see, you know, these tests happening
16:12
on the military side. I know there's
16:14
differences kind of civilian
16:16
and government with the ability to
16:19
test things and availability of airspace
16:21
and all of that, but how
16:23
do you, as a pilot, maybe
16:25
you're maybe more familiar with some
16:27
of these regulations than
16:29
the rest of us are, how
16:31
do you see this technology being
16:34
able to develop over time with
16:38
such restrictions around testing and how
16:40
could that be eased up in
16:42
a reasonable way without
16:44
undue, you know, issues
16:47
and danger and that sort of thing.
16:49
Cause obviously if you have drones flying
16:51
over populated areas, that is
16:53
definitely an issue. But
16:56
at some point there's going to have to be
16:58
a drone fly over a populated area.
17:00
Indeed. And so, and to start off with,
17:02
I certainly am not an expert in that.
17:04
I have some very loose familiarity with the
17:07
process. Military, they have their own
17:10
dedicated air spaces. There's military airspace
17:12
and especially, it's all
17:14
over, but especially out West, places like Edwards
17:16
Air Force Base and a number of others
17:19
where you have literally, you know, hundreds of
17:21
square miles that you
17:23
can do testing in. And obviously there's
17:25
a long history of that already since
17:27
the dawn of flight. The FAA is
17:29
very aware, you know, of
17:32
the need to innovate on this. And so
17:34
they basically, you have to apply
17:36
for what you're trying to do and
17:38
show them that you've done
17:40
due diligence from the engineering safety, you know,
17:42
all the concerns about that. And
17:45
basically I follow a lot of aviation
17:47
news. So I've kind of read about a number of
17:49
these programs that have come into being,
17:51
and then they give you a little bit of
17:53
a leash and you can kind of, you have
17:55
to kind of earn your way through a number
17:58
of gateways, you know, where you're going. you
18:00
where you successfully do something in very
18:02
small scale, very small scope, and increase
18:04
your way into it. But it seems
18:06
to me that that is happening more
18:08
and more. And in some cases, if
18:10
there is a military utility
18:13
to doing that, then there
18:15
can be coordination also with
18:17
military and taking advantage of
18:19
military airspace to have more
18:21
room, things like that. So it
18:24
seems though, though, obviously, government
18:27
agencies are not the speediest
18:29
things, typically, that there are
18:31
opportunities for even private businesses
18:33
and stuff to get some support
18:36
in that way. They know it's coming. Yeah,
18:39
this is probably something
18:41
we could refer people back
18:43
to our previous episode
18:45
with Jake and others.
18:48
It's unlikely that we'll be seeing
18:51
the skies filled with weaponized
18:53
autonomous drones doing whatever they want.
18:55
There's a lot of
18:58
hopefully responsible people thinking about
19:00
these things. But the
19:03
main interesting piece here is
19:05
both on the commercial side
19:07
and on the military side,
19:09
the ability to increase safety
19:12
and decrease people, human
19:14
pilots being in dangerous
19:16
situations, I think it
19:18
seems to be the focus of a lot
19:21
of this. Now, you
19:23
know, there's probably all of those out there
19:25
that can imagine all sorts of scenarios of
19:28
misuse and all of those sorts of things.
19:30
But there's also in our
19:32
previous conversations with people, at least
19:34
I have some hope
19:36
that there's some reasonable
19:38
people and thoughtful people that
19:41
are part of these programs.
19:43
Yeah, just at risk of sounding
19:45
like an apologist, I point out to
19:47
people there are a lot of safeguards
19:50
to that point. I work in
19:52
defense. I come home. I
19:54
Mostly work from home, but I have my
19:56
family and my dog and everybody else who's
19:59
doing this whether. In the military or
20:01
whether they're civilian supporting that they have
20:03
their family and their kids and all
20:05
that. So on the notion that there's
20:07
like the dark military minds behind the
20:09
closed doors are in my experience of
20:11
fiction we all you know when we
20:13
get on the phone or even for
20:15
your business thing, we're talking about the
20:17
same things that everybody else talks about,
20:19
you know, the weekend and what? You
20:21
know, if my dog wasn't feeling well
20:24
and my kid with stay home from
20:26
school or whatever and so I'm very
20:28
encouraged and that way or it's normal.
20:30
People running these are and they have different
20:32
motivations obviously are depending on where the rat
20:34
and what organization there with but it's one
20:36
of their things that I get worried about
20:39
with a I going forward but that's not
20:41
one of of. Yeah, I'm might
20:43
refer people back to our episode
20:45
leading the charge on a I
20:47
A National Security with a General
20:50
Jackson a Ham. Really good episode
20:52
to yeah retired Us Air Force
20:54
So if you want to get
20:56
a sense of someone on the
20:58
was sort of leading the charge
21:01
on the inside for a good
21:03
long time then I would recommend
21:05
that episode from being a civilian
21:07
myself who is good to have
21:09
a chat with him. Yeah General
21:11
Shanahan is. Who's now retired is
21:14
both are. That was a recent episode
21:16
as we record this and was also
21:18
the original hard charger for a I
21:20
in the military an hour and is
21:22
in a A uniquely he still considered
21:24
even though he's retired to be one
21:26
of the top experts and influencers. Have
21:28
a hope people check that out. Yeah,
21:30
well. I don't know if this was
21:32
widespread news, but I thought it would
21:34
be a cool thing to highlight for
21:36
people. You. Know you're talking about
21:38
kind of this further testing and
21:41
I'm sure some of that testing
21:43
on the autonomous vehicle side involves.
21:45
Standards. And best practices and
21:48
frameworks. All of that's necessary
21:50
to really advance. A technology
21:52
from are in the to produce
21:54
type and otherwise and I think
21:57
that we're seeing also some of
21:59
that. On the. Enterprise.
22:02
Ai generative, Ai side of
22:04
things. So this last couple
22:07
weeks I was. Informed
22:09
about this project which is
22:11
now a project at the
22:13
Linux Foundation. In the project
22:15
is called the Open Platform
22:17
for Enterprise A I just
22:19
abbreviated to O P A
22:21
which seems like an unfortunate
22:23
and awkward acronym. The homicide
22:25
Her I don't I was
22:27
trying to think like, how
22:29
do I it via over
22:31
here Either that or know
22:33
that. See you avoiding the
22:35
obvious A high school. A
22:39
of doing it. The. I've been bought
22:41
the greatest of acronyms, but ah yeah,
22:43
the Linux Foundation has this ai and
22:46
data foundation, so if you're not familiar
22:48
with the Linux foundation, you can look
22:50
it up. But. This.
22:52
Enterprise Open platform for
22:54
enterprise Ai. Is. A
22:56
very collaborative initiative it seems
22:59
and just some of the
23:01
companies involved are kind of
23:03
was them out. Not all
23:05
of them but just to
23:07
give you a sense includes
23:09
Intel and any scale cloud
23:11
era data stacks, Domino Data
23:13
Lab hugging face many yo.
23:15
Zola, X A bunch of
23:17
different. Companies. That probably
23:20
your familiar with, certainly
23:22
ones that we've talked
23:24
about on this show
23:26
and. There's. A few
23:28
interesting elements of this.
23:31
Open platform for. Enterprise.
23:34
A I but the
23:36
general goal I think
23:38
is to enable and
23:41
facilitate. Or the way that
23:43
they frame it as aims to facilitate
23:45
and enable the development of flexible. Scale.
23:48
Of Bullets and Ai systems.
23:51
That. Harness the best open
23:53
source innovation from across the
23:55
ecosystem. And. That's kind
23:57
of vague in terms of the.
24:00
The where they're going with this, but
24:02
I think if you look sort of
24:04
a little bit deeper I think there's
24:06
some really interesting things of where this
24:08
could lead one as they recognize certain.
24:11
Common. And developing.
24:14
Archetypes. Or I'm. Main.
24:16
Use cases where people are using
24:18
two hundred of a I, for
24:21
example, the rag work flow retrieval,
24:23
augmented generation work flow and they're
24:25
kind of take that rag work
24:28
flow and are creating blueprints for
24:30
the various pieces that are involved
24:32
in a an industry standard kind
24:35
of advance. Rag work for not
24:37
just stay naive rag work flow
24:39
that you might play around with
24:42
on your laptop or something that
24:44
could be deployed in the Enterprise.
24:46
And so they have some blueprints or
24:48
kind of architecture type of things. I
24:50
think they'll be more of that that
24:53
will be developed and then those architectures
24:55
or blueprints have certain components within them.
24:57
For. Example: Ah, retrieve or system
24:59
and abetting model or guard
25:02
rails for models or fine
25:04
tuning systems are a vector
25:06
database and then if you
25:08
follow the link to the
25:10
get hub related to the
25:12
o p a project appear
25:14
project whatever you wanna call
25:16
it. I notice some really
25:18
interesting kind of a few
25:21
categories of some things that
25:23
aren't quite complete there yet
25:25
but that they're building and
25:27
public. And those are
25:29
both examples of implementing these
25:31
sort of reference implementations of
25:33
industry standard ways of going
25:35
about doing certain things so
25:37
like Sat with your dogs,
25:39
cogeneration assistance the you can
25:41
plug into. Visual. Studio Code
25:44
Document Summary Visual: A question
25:46
and answer. Am those reference
25:48
implementations include open source ways of
25:51
doing these different things in a
25:53
kind of industry standard way. Another
25:55
one, as they have, it seems
25:58
like they're developing a. These
26:00
of micro open micro services that
26:02
could be plugged and to do
26:05
various of these components. and then
26:07
finally I said of evaluations. So
26:09
they have a repo valuation benchmark
26:12
and scorecard. Targeting. Performance on
26:14
throughput and latency accuracy Unpopular evaluation
26:16
harnesses for safety hallucination Other things
26:18
like that, so there seems to
26:21
all of that put together. I
26:23
know there was a little bit
26:25
rambling, but it seems like they're
26:27
kind of focus here on these
26:29
blueprints. Reference. Implementations of
26:32
things represented in those blueprints,
26:34
and then industry kind of
26:37
enterprise level evaluations for performance
26:39
and issues within these systems,
26:42
that sort of thing, So
26:44
this definitely seems encouraging. To.
26:47
See a lot of collaboration on this and
26:49
see the support from the Linux Foundation. Yeah.
26:51
I mean with the Linux Foundation being
26:54
in a one of the most reputable
26:56
open source organizations in the world certainly
26:58
the top few, it's really important that
27:01
finishes like this come into being in.
27:03
The reason is that in the business
27:05
world I know you and your company
27:07
and I certainly is. I'm talking to
27:10
people in different companies, everyone out there
27:12
is trying to find their own way
27:14
and to implementing. Jenner they I solutions
27:16
and how do you put it together?
27:19
had you architect it? I have my
27:21
own thoughts. Around that and and I
27:23
know the company I work at has
27:25
its own thoughts around that and I
27:27
end up talking to people of different
27:29
organizations and they're struggling with many of
27:31
the same problems, but they come to
27:33
their own solutions you know of based
27:35
on however their team wants to approach
27:37
it and as we know from other
27:40
in a before Gen Vi and even
27:42
before A I came along, it's an
27:44
early point in every growth. You know
27:46
development of every are you know whether
27:48
in software, anything else, where you have
27:50
everyone kind of going off. and doing
27:52
their own thing but they realize that that
27:54
itself or while it might solve the immediate
27:56
it's they need it creates a whole new
27:58
set of problems as the after grow and
28:00
integrate with other organizations. So seeing
28:03
what the open platform for enterprise AI
28:05
has to offer, it looks
28:07
very promising. And I would, I
28:09
would encourage organizations out there to
28:12
take a look at it and whether
28:14
you adopt it or not, maybe it
28:16
helps frame how you're choosing to solve
28:18
problems in a way that might make
28:20
situations you're in down the road that you're not
28:22
thinking about yet a little bit easier to cope
28:25
with. Well, Chris, as we
28:27
kind of look back to the last sets
28:29
of newsworthy AI stuff happening
28:33
in all over the place, both
28:36
in terms of large language models, Gen
28:38
AI and not Gen AI,
28:40
one of the themes recently that
28:42
it seems like has been
28:45
happening and kind of in getting
28:48
into its prime is
28:51
video generation. I
28:53
don't know if you've been following this
28:55
sort of stuff, but I
28:57
know that there was, I saw
28:59
something from Microsoft. I saw
29:02
something from Alibaba. I think,
29:05
of course, there was the open
29:07
AI video generation stuff. There's
29:09
been things from runway ML
29:12
and yeah. So what are
29:15
your general thoughts on where all of
29:17
this video generation stuff is happening or
29:19
is going? I had a couple of
29:21
thoughts there. I don't think it should
29:23
surprise anyone at this point who's following
29:25
the industry. You know, when we
29:28
were doing our thoughts for 2024 last
29:30
year, we were
29:32
talking about this would surely come next, you
29:34
know, because we were willing to still imagery
29:36
and stuff and the rate that we're seeing
29:39
things progress from a quality standpoint,
29:41
you know, when it's going so fast,
29:43
you know, it was not long ago that
29:46
open AI released Soros. That
29:49
wasn't long ago at all. And we were kind of
29:51
going, wow, look at, you know, it's
29:53
here and look at this first thing. And now
29:55
there are many options available after just a few
29:57
weeks. And I think I've
30:00
been somewhat amused to look at
30:02
the reactions in public about people
30:05
and the concerns about
30:08
safety and you know deep fakes
30:10
being so much better now
30:12
in two thousand twenty four than they were a year ago
30:14
right now. We're gonna have to
30:16
adjust and take it in and recognize
30:18
the utility and come up with some
30:20
safeguards for it. I guess it
30:22
was kind of obvious to us and
30:24
those of us who are following this weekend and
30:26
we got that we be here and so now
30:28
we're here. I'm waiting to see
30:30
some of the more interesting creative productive things that people
30:32
are going to put this to I'm
30:34
really looking forward at this point
30:37
to seeing some utility coming
30:39
from it that's meaningful. And
30:41
yeah, just so people can go
30:44
out there and look at these
30:46
things. One is called Vasa one,
30:49
which is the one from Microsoft
30:51
research. And
30:54
the kind of tagline there is
30:56
life like audio driven talking faces
30:58
generated in real time. This was
31:00
an interesting one. It kind of
31:02
almost reminded me of the
31:05
sort of videos that I've seen from
31:07
Cynthia and these other companies that kind
31:09
of help create talking heads
31:12
essentially for marketing videos or
31:14
training videos, the sort of
31:16
thing. And very
31:18
impressive stuff there might
31:21
have seen something going through on Twitter,
31:23
LinkedIn with, you know, people always try
31:26
to make the Mona Lisa face talk
31:28
and I felt I was one of
31:30
their examples that they had, which,
31:33
you know, that seems to be a sort of
31:35
given that you try that if you're working in
31:37
this space. And the most recent one wasn't actually
31:39
anywhere close to being the best stuff.
31:41
I saw that maybe
31:45
a week ago and it was pretty cheesy, but
31:48
I mean we're truly arrived in 2024. If
31:50
you can have video now, Certainly
31:52
at least talking head video that is
31:54
indistinguishable from a person, you would be
31:56
very if you were to put, you
31:58
know, compare it. Have two or
32:01
three people and have two or three
32:03
I generate ones. mix them up and
32:05
have people choose which ones are which
32:07
of. I know that I probably could
32:09
not do that successfully. a number. I
32:11
might get lucky and pick one or
32:14
two, but we're getting there and so
32:16
far as I really am curious to
32:18
see how these are put it that
32:20
like beyond the novelty of it's of
32:22
seeing I'm finally arriving. After talking about
32:25
this stuff for a while, I really
32:27
am curious to see how people use
32:29
them for. You know we like to
32:31
talk about Ai for good. I really
32:33
want to see instead of people worrying
32:35
about this are strictly about the security
32:37
concern which is legit. I'd like to
32:39
see some people do some amazing things
32:41
for it. Ah that is gonna benefit
32:43
people in humility at large and I'm
32:45
excited to see this use cases and
32:47
if anybody out there has something please
32:50
point us to it can cite those
32:52
are these cases on Wait to see?
32:54
Yeah and the one if if people
32:56
are searching from Ali Baba's is called
32:58
email or I guess he is. He
33:00
am oh I assume emo
33:03
guy or Ali Baba's email
33:05
and Vasa. From Microsoft if
33:07
you wanna if you wanna take a
33:09
closer look. It kind of seems to
33:11
me chris like a time when. You.
33:14
Know when Dolly came out the first
33:16
one and then there was. It was
33:18
like Dolly stable, diffuse and and that
33:20
the just seem to be the snowball
33:23
really quickly of image sooner. he said
33:25
things. It seems like we're in a
33:27
similar cycle right now with the video
33:29
to nourish and. Stuff.
33:32
And then eventually you know it'll be integrated
33:34
into our chat interfaces and other things that
33:36
I don't think it's going to be long
33:38
at all to get to that point. I
33:40
think we're going to be amazed at how
33:42
fast as get integrated there are because every
33:44
time they keep building on themselves in your
33:46
we have. The one thing we've noticed over
33:48
the last two years is the acceleration in
33:50
the development. I'm we will say something will
33:52
come out and the next year and then
33:54
it comes out two months later and you
33:57
know couple a times said well we predicted
33:59
a prefer hot. The tiring on that?
34:01
I think it's gonna happen pretty darn
34:03
quick. And and to illustrate that there
34:05
was not specific to the suitcase hugging
34:08
face, announce this past week that they
34:10
had crossed over the one million mark.
34:12
There's One million The I Models hosted
34:15
hugging Face Yes, Congratulations to Ah, Hugging
34:17
Face and and the team there that's
34:19
amazing All those you know. it wasn't
34:21
that long ago where they were nowhere
34:24
close to a million, but they keeps
34:26
accelerating and so who hit and they'll
34:28
hit ten million in. No time, I'm
34:31
sure. But on to your point earlier
34:33
that I think it's not just going
34:35
to be seeing these new technologies coming
34:37
out where we're looking at the said
34:39
the demo but I think for like
34:42
the second half of Twenty Twenty Four
34:44
and into Twenty Twenty Five will be
34:46
such a huge push and getting models
34:48
integrated into real world scenarios. You know
34:50
what we would like to say as
34:53
at the edge in all sorts of
34:55
different contacts and are those that's really
34:57
quite honestly what I'm excited to see.
35:00
Is if instead of just a talking head
35:02
with the audio that's indiscernible I I want
35:04
to see that ah and some good contacts
35:06
that are in places that were not used
35:08
to sing on that make a big difference
35:10
and said that will be a pretty cool
35:12
for me that will be of a cooler
35:14
milestone than just sing the demo upfront. Yeah.
35:17
It does seem like that
35:19
There's some big possibilities in
35:21
even spaces like education and
35:24
other places where hey you
35:26
have some tax contents, you
35:28
have some sort of cure
35:31
a son in place but
35:33
creating. Very. Much I'm
35:35
appealing and. Realistic
35:38
looking, Educational content
35:41
that would fit a certain scenarios
35:43
cause there's tons of sort of
35:45
self study stuff online. Are
35:48
some of it has better
35:50
video quality. Than. Others, but
35:52
also some of it's at a
35:55
certain level. that's. You. Know
35:57
if you have one set of
35:59
content a perfect. The records may
36:01
be a video course or something
36:03
that lasts. Me: You'd have to
36:05
watch it for an hour every
36:07
day for many weeks, maybe. But
36:09
if you can read purpose some
36:11
of that content to. Answer
36:14
questions and create engaging, Courses.
36:17
In. Different. Shorter forms
36:19
or four different age levels and
36:21
that sort of thing and some
36:24
of that was able to still
36:26
be video, still be engaging, but
36:28
not take a huge amount of
36:30
video production. To. Create, which
36:32
is very expensive and time consuming. I could
36:34
see a lot of possibility there is probably
36:36
many others Id love to hear. From.
36:39
Our Listeners: If they have ideas about
36:41
this, we'd love to hear about them.
36:43
Inner Slack Channel if you wanna join
36:45
or elsewhere. Just. To illustrate that
36:47
for a moment or and we talk
36:50
we know we talked about education use
36:52
cases many times. those are and how
36:54
it intersects with traditional education in a
36:56
like i have a daughter in middle
36:59
school. Ah and also you know things
37:01
like continue education for grown ups You
37:03
know that are continuing through this ever
37:05
changing world that are the you know
37:08
constitutes our careers but it's very easy
37:10
to leap from. You know the Vassar
37:12
example that we're talking about with the
37:14
Talking Face is being generated Real? Time
37:17
as they know and thinking. Every
37:19
kid in school in addition to
37:21
you know potentially it as we
37:23
as Things are friends, you sing
37:25
forward and we still have traditional
37:27
educational paradigms that most kids are
37:29
involved in. But maybe every kid
37:31
has their own other own personal
37:33
teacher in addition to a classroom
37:35
teacher. And that personal teacher explains
37:37
the mass in a way that
37:39
that student understands compared to the
37:41
student next to them, when you
37:43
get a lot of personalization and
37:45
support that way that. would be wonderful
37:47
to see that and so kids aren't
37:49
left behind and if you don't understand
37:51
the with teachers explaining it you don't
37:53
have to struggle because you already have
37:55
your personal assistant says there's many many
37:57
thousands of use cases along these lines
38:00
So that's the kind of thing that I'm pretty excited about
38:02
for the future. Cool. Yeah Well
38:04
as we kind of draw things to
38:06
a bit of a close here We
38:09
normally tried to provide a learning
38:11
resource for people in these fully
38:13
connected episodes and I want
38:16
to share one today We've been doing
38:18
a bit of experimentation of our own
38:20
Chris with these practical AI webinars These
38:23
I think what we've been calling them gen AI
38:26
mastery. So we've done two at this point One
38:29
related to text to sequel and
38:32
one related to private chat
38:35
UIs and I think
38:37
it's been a good experience so far at
38:39
least to motivate us to do do it
38:41
a bit more And we're
38:43
really trying to make these
38:45
webinars a live good
38:48
learning experience for
38:50
people and And something
38:52
where we have some hands-on, you
38:54
know A visual component with some
38:56
hands-on that you don't kind of
38:58
get in just the audio podcast
39:00
scenario so we we do
39:02
have another one of these planned and I
39:06
would highly recommend that that
39:08
you go to tiny URL
39:10
comm slash gen AI mastery
39:13
three tiny URL comm
39:16
slash gen AI Dash mastery three and
39:18
we'll put that in the show notes
39:20
as well and sign up for this
39:22
next one It's gonna be about multimodal
39:25
AI and we're finalizing the guests But
39:28
I already I think I know who they're
39:30
gonna be and it's gonna be a sort
39:32
of rock star They're helping
39:34
us learn about multimodal AI doing
39:36
cool things with video as we've
39:39
been talking here but also imagery
39:41
and Kind of tying together text
39:43
prompts in there as well for
39:46
kind of multimodal rag sort of
39:48
systems So if you're interested
39:50
in that definitely sign up, it's gonna
39:52
be it's gonna be a great experience So we'll have
39:54
that link in the show notes and look forward to
39:57
seeing everyone there Yeah, it's a lot of fun to
39:59
do this session because it's live real time
40:01
and everybody can see everybody else in
40:03
the chat and there's real time communications
40:05
as we're doing them make
40:08
it pretty special. Yep. All right,
40:10
Chris. Well, it's been fun. I hope you
40:12
can enjoy the rest of your weekend and
40:14
we'll talk to you soon. Take it easy,
40:16
Daniel. All
40:25
right, that is Practical AI for
40:27
this week. Subscribe now.
40:29
If you haven't already, head
40:32
to practicalai.fm for all the
40:34
ways and join our free
40:36
Slack team where you can hang out
40:38
with Daniel, Chris and the entire change
40:40
log community. Sign up
40:43
today at practicalai.fm
40:45
slash community. Thanks
40:47
again to our partners at fly.io,
40:49
to our beat freaking residents, Breakmaster
40:51
Cylinder and to you for listening.
40:53
We appreciate you spending time with
40:55
us. That's all for now.
40:57
We'll talk to you next time.
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More