0:00
So, what the heck does Palanteer
0:03
>> Left click, right click, left click.
0:05
Magically, it becomes a detection.
0:08
>> This demo by the chief AI officer for
0:10
the Department of War is perhaps the
0:12
closest look we've gotten at Maven Smart
0:16
The AI powered software at the front
0:18
lines of global conflicts today. The big
0:21
bottleneck in modern warfare is the time
0:23
it takes to target your adversary. And
0:25
in the first 12 hours of the Iran War,
0:27
the US struck nearly 900 targets and
0:30
over 13,000 targets in 38 days. This is
0:33
a massive increase over the way things
0:36
used to be. And central to that is Maven
0:38
Smart System. I want to walk you through
0:41
how this thing actually works and then
0:42
show you how much of these capabilities
0:44
we can actually build using commercial
0:46
and off-the-shelf tools minus of course
0:48
the missiles at the end because the
0:49
applications for situational awareness
0:51
go way beyond the battlefield. And
0:53
because TED is over, this is exactly why
0:55
I've been heads down building God's Eye
0:57
View into something bigger than a
0:59
YouTube series. And if you've been
1:00
asking me when you can actually use this
1:02
stuff, you're going to want to stick
1:03
around till the end. So, let's get into
1:08
All right, so there's this military
1:10
concept called the UDA loop. Observe,
1:12
orient, decide, and act. This is
1:14
honestly what you do when you wake up in
1:16
the morning and decide to go get your
1:17
coffee or eat a burrito in the evening.
1:19
The key point though is your decision
1:20
cycle needs to be faster than your
1:22
adversary because if you do that, they
1:24
can basically never catch up. They can't
1:26
orient. They can't plan. They're always
1:28
in react mode. And the whole idea behind
1:30
Maven is to compress that entire loop.
1:32
And we're going to walk through every
1:33
single step so it's crystal clear. Okay.
1:35
First is the backstory. This was really
1:37
interesting to get into because to
1:38
understand why Maven exists, you need to
1:40
understand what came before it.
1:42
Adversarial targets were tracked in
1:44
Excel spreadsheets. PowerPoint was used
1:46
for mapping network connections between
1:48
all of them. Google Earth was there for
1:50
zooming in and out. As one officer put
1:52
it, "We've killed more people on
1:54
Microsoft Office than you would ever
1:56
imagine." I mean, just think about it.
1:57
The modern commander with all the crazy
1:59
sensor systems at his disposal is going
2:01
into battle with Microsoft Office and
2:03
Google Earth. Maven was created to fix
2:06
all of that. Basically, Google Earth for
2:08
war, but with AI telling you what's
2:10
actually on the screen. All right, so
2:11
here's what they built. What you're
2:12
looking at are feeds from satellites,
2:14
drones, signals, intelligence, prior map
2:16
data, all layered together on a globe.
2:19
You can see which aerial assets are in
2:20
the area, as well as different target
2:22
designations. Now, we're seeing a drone
2:24
view. You can see the vector outlines of
2:26
the buildings that are highlighted. With
2:27
a couple clicks, you load up your, let's
2:29
say, road network data, so you can see
2:30
how exactly to make it to that point.
2:32
And then, when you zoom in, you see all
2:34
these dots. These are basically
2:35
detections made using computer vision.
2:37
And you'll notice there's a number
2:38
associated when you hover over them and
2:40
nominate things to the board. That
2:42
number is its stable identifier. The
2:44
idea being that you can tell this is
2:45
indeed the same card that let's say
2:46
you're seeing in a satellite view versus
2:48
a drone view versus CCTV. Then you can
2:50
take these detections and nominate these
2:52
to a literal cananban board like a task
2:54
tracker where you can drag cards from
2:56
todo to done. These vertical columns
2:58
each have their own different process
2:59
and they essentially represent what
3:01
different teams would have been doing.
3:02
Once you've taken these detections, you
3:04
can take advantage of other workflows to
3:06
figure out what is the right asset to
3:08
task to execute that target. You can
3:09
choose different criteria. You can
3:11
optimize different metrics like time to
3:13
target, how much fuel, munitions,
3:14
distance, etc. And from all of that, it
3:17
gives you a recommendation that a human
3:18
actually approves a plan which it goes
3:20
about and executes or as the chief AI
3:22
officer put it in order to get to our
3:24
desired instinct, actually closing a
3:26
kill chain. So basically the Department
3:28
of War loves this because Maven is
3:30
taking something that used to take a
3:31
room full of analysts, right, to go
3:32
generate that course of action for them.
3:34
Now that entire complicated pipeline can
3:37
all be collapsed down into one software
3:39
that all these teams use to make the
3:40
stuff happen. The AI starts recommending
3:42
which asset should prosecute that
3:44
target, which drone, which missile
3:45
system, which aircraft based on
3:47
proximity, payload, rules of
3:48
engagements, and then you approve that
3:50
plan. So you go from like detection to
3:52
strike decision in one system. And just
3:54
for a second, imagine that this has been
3:56
happening thus far in Microsoft Office
3:57
and Google Earth. Now, let's map this to
4:00
the UDA loop. The sensor feeds that are
4:01
coming in, right? Like the satellites,
4:03
aerial, drone, all that stuff. That's
4:05
the observation. The system is fusing
4:07
and classifying what it sees. That's
4:09
orientation. The canban board and the
4:11
course of action generation, that's
4:12
decision. And then finally, actioning
4:14
the target. That's the acting bit. We're
4:17
going to come back to this demo
4:18
throughout the video cuz each section
4:19
breaks down one step of this loop. Now,
4:21
here's the part that's super complicated
4:23
and controversial. There's pretty good
4:24
consensus based on reporting that the AI
4:27
that's doing this like natural language
4:29
intelligence queries inside Maven,
4:31
that's Claude, built by Anthropic.
4:33
Anthropic actually was one of the first
4:35
to deploy large language models in
4:37
classified military settings. In fact,
4:38
the Washington Post reported that Claude
4:40
is central to the US campaign in Iran.
4:42
But Anthropic obviously drew a line
4:44
which Daria got a lot of flack from the
4:47
administration for. basically said, "You
4:48
can use claude as long as it's not used
4:50
for fully autonomous weapon systems or
4:52
mass domestic surveillance." The Trump
4:55
administration replied by declaring
4:56
anthropic a supply chain risk. Their
4:58
argument was, "Hey, you're a contractor.
5:00
You provide us the tool. You have no say
5:02
over how it's actually used." And when
5:04
you combine this with other reporting
5:05
that seems to suggest that it's actually
5:06
clawed 3.5 sonnet that's actually being
5:09
used in the system, it really makes you
5:11
think like, are these models actually
5:12
good enough to deploy in the wild? So,
5:14
the AI company that first brought their
5:16
large language model to a classified
5:17
setting, that's a key part of this Maven
5:19
smart system is now at the center of
5:21
this battle with the government. And
5:22
based on recent reporting, it seems like
5:24
a deal can still be struck. But the
5:26
tension is very real. And we'll come
5:28
back to this cuz it matters.
5:32
All right, so Maven is obviously only as
5:34
good as what it can see, the underlying
5:36
data that is made available to it. So,
5:38
let me show you what that actually is.
5:39
Obviously, you've got crazy spy
5:41
satellites in orbit, keyhole stuff, the
5:43
new things that SpaceX is building.
5:45
You've got synthetic aperture radar,
5:46
which we've extensively covered in the
5:48
past. Both military and commercial
5:50
satellite systems like ISI and Capella
5:52
Space that fire literal radar pulses
5:54
through cloud through darkness that
5:56
return 25 cm resolution imagery of the
5:58
ground. We've covered this previously.
5:59
Very dual use tech, the same stuff to
6:01
figure out if a landslide's going to
6:03
happen or if a bridge is about to break,
6:04
is also critical to these kill chains.
6:06
Now, here's the stuff that most people
6:08
don't know about. wide area motion
6:10
imagery or whammy. They started off with
6:12
a system called Gorgon Stair. This is
6:14
mounted on those classic MQ9 Reaper
6:17
drones and basically it uses 368 cameras
6:20
stitched together into a 1.8 gigapixel
6:23
composite image so that you can
6:25
persistently surveil an entire small
6:27
city from 25,000 ft. Think of this like
6:29
Google Earth Live. If you've actually
6:31
seen the movie Enemy of the State
6:32
featuring Will Smith, this program was
6:34
actually inspired by that movie. That's
6:35
freaking wild to me. So the idea is once
6:37
you've got this persistent image of
6:38
absolutely everything, when something
6:40
bad happens, let's say an IED goes off
6:42
in Afghanistan, you can then rewind to
6:44
see where that car originated from.
6:46
There's an even more advanced version of
6:48
this called Argus is that can cover 36
6:50
square miles with enough resolution to
6:52
track individual pedestrians. This way,
6:54
as long as you have an airplane or a
6:56
drone, you know, orbiting over a city,
6:58
you can figure out exactly what's
6:59
happening on the ground in the past, in
7:01
the present, and of course the future as
7:03
well. So, if we go from wide area to
7:05
narrow areas of monitoring, that's what
7:07
you might usually associate with drone
7:09
footage, essentially FMV or full motion
7:11
video. You've got these Reaper drones
7:12
that are capable of taking these live
7:14
feeds directly into Maven for real-time
7:16
object detection and classification. And
7:18
what's interesting is that these drones
7:19
can also now operate in GPS denied
7:21
environments, right? Like which Iran
7:23
obviously is and other conflict zones
7:25
across the region. They're constantly
7:26
getting jammed and spoofing GPS, right?
7:28
You can't rely on GPS in those
7:30
scenarios. If you remember, we've talked
7:31
about visual positioning systems in the
7:33
past before where you can take a camera
7:35
image, match it to a prior 3D model to
7:37
figure out exactly where that photo is
7:39
taken. Drones use this technology, too.
7:41
And Vantor's got something called the
7:42
Raptor system that matches that onboard
7:45
camera feed against their precision 3D
7:47
terrain model. So, Vantor's used a 30cm
7:49
satellite imagery to build this coarse
7:51
3D model of the world. And that becomes
7:53
the key for it to figure out exactly
7:56
where every drone photo was taken.
7:58
Suddenly, the drone doesn't even need
7:59
GPS. It knows where it is just based on
8:02
the terrain. Now, as you probably saw
8:03
with God's Eye View, one or two layers
8:05
by themselves are only so interesting.
8:07
But when you start putting in different
8:09
sources of intelligence together, that's
8:11
where you get that classified grade
8:13
picture. Signals intelligence is
8:15
obviously a huge part of how these
8:16
systems work. I'm talking intercepted
8:18
communications, electronic emissions
8:21
integrated alongside the imagery. We've
8:23
talked about RF satellite constellations
8:25
in orbit like Spire that do dark vessel
8:27
detection, right? like vessels that are
8:29
turning off their transponder. In
8:30
addition to that, you have entire shadow
8:32
fleets, right, that are not just turning
8:34
off their AIS, but they're actively
8:35
masking or offiscating their presence.
8:38
But there's another source of data that
8:39
solves that problem, too. And that's
8:41
called advertising intelligence. So,
8:43
sure, you're a shadow fleet. You're out
8:45
in the wild. You've actually repainted
8:47
your ship and you think you're
8:48
scot-free. But turns out the people on
8:50
board have a cluster of devices that
8:52
have Candy Crush installed. You can use
8:54
that ad network data to geollocate where
8:56
that ship is. That's the world we live
8:58
in. You can offiscate all other signals,
9:00
but the fact that you've got some folks
9:02
on board that like playing Candy Crush
9:04
means you will be pinpointed. And then
9:05
there's a secret one that's been
9:07
reported lately, the RQ80, America's
9:10
most classified stealth reconnaissance
9:12
drone. It was publicly exposed for the
9:14
first time during this war when it
9:16
emergency landed at an air base in
9:17
Greece this past March. It's a flying
9:19
wing that's believed to operate at
9:20
60,000 ft plus for 24-hour autonomous
9:23
missions. And it's believed to have been
9:25
tracking Iran's mobile missile
9:27
launchers, the kind of targets that move
9:28
and hide. You need that persistent
9:31
overhead coverage to catch them. Now,
9:33
does this have a SAR system on board? Is
9:35
it something more like whammy? We just
9:37
don't know, but I suspect it's a
9:39
combination of modalities. Now, of
9:40
course, on top of all of this, you layer
9:42
in commercial satellite imagery from
9:43
like Vantor, Planet Labs, you've got the
9:46
maritime AI tracking data. All of it
9:48
gets fused together into one picture.
9:50
That's the observe step. Maven basically
9:53
can see everything. Now the question is
9:55
how do you make sense of it?
10:00
This is where we go back to left click,
10:01
right click, left click. This is orient,
10:03
decide and act happening in one
10:05
interface. Now watch what happens when
10:06
you do this at scale. Obviously I've
10:08
been covering Operation Epic Fury deeply
10:10
on this channel. If you've seen those
10:11
God's Eye View videos and our Hermos
10:13
coverage, you know the timeline already.
10:15
Now, Epic Fury kicked off Feb 28th, and
10:17
in the first 12 hours, the US struck
10:19
nearly 900 targets across Iran. But
10:22
perhaps the most dramatic single
10:23
operation was Car Island. On March 13th,
10:25
the US hit over 90 military targets
10:28
simultaneously. I'm talking naval mine
10:30
storage, missile bunkers, air defense
10:32
systems in what Trump called one of the
10:34
most powerful bombing raids in the
10:36
history of the Middle East. Critically,
10:38
the oil infrastructure was deliberately
10:40
left intact. So, how do you coordinate
10:42
90 plus simultaneous strikes? Suddenly
10:44
software like Maven earns its keep.
10:46
Orient, decide, act all premputed using
10:50
the latest sensor data and executed and
10:52
tasked in parallel. Now here comes the
10:54
critical trade-off, right? Like is there
10:55
a cost to speed? Is there a point at
10:57
which Maven's accuracy surpasses that of
10:59
a human analyst? And if it does, there's
11:01
still going to be a gap and at a large
11:03
enough scale that gap is going to
11:05
manifest itself in targets getting hit
11:07
that shouldn't have been hit. I think
11:08
this is where we're also going to have
11:09
to contend with how this technology is
11:11
deployed in the wild. Right? The
11:12
Department of War right now is a very
11:14
clear stance. There are no fully
11:15
autonomous weapon systems. Everything
11:17
has a human in the loop. But think about
11:19
self-driving vehicles. At some point,
11:20
self-driving gets accurate enough where
11:22
it's better than the average human
11:24
driver. But if you add up enough of
11:25
those drives, people are going to die.
11:27
And what happens when the systems make
11:29
incorrect decisions? Like why was that
11:31
incorrect decision made? Is it the model
11:33
maker like Claude or Anthropic that's at
11:35
a fault here? Is it the sensor data
11:37
perhaps that it was made on? or even
11:38
underlying maps data. For example, the
11:40
underlying map data that you have your
11:42
AI digest incorrectly classified a
11:45
school, whose fault is that? Now, look,
11:46
there are plenty more questions to get
11:48
in here. But the broader question
11:50
remains, when you're processing
11:51
thousands of targets in 24 hours, what
11:54
gets missed and who's at fault there?
11:56
And then the question gets heavier. Once
11:58
these systems get good enough to justify
12:00
not having a human in the loop, what
12:02
happens when you're processing thousands
12:03
of targets in a short period of time?
12:05
And to be clear, this isn't just
12:06
Palanteer. Like Android has lattis,
12:08
right? Very similar concept, different
12:09
platform. Take a bunch of these sensor
12:11
systems together and create a fused
12:13
operational picture of the battle space
12:15
and have AI handle the orient and decide
12:17
steps. But this common operational
12:19
picture isn't just useful on the
12:21
battlefield. This is the same kind of
12:23
platform. The same Palunteer technology
12:25
actually runs the UK's NHS CO data
12:28
system. It optimizes logistics for the
12:30
world food programs. It runs Airbus's
12:33
supply chain. Oil and gas giants use it
12:35
to create incremental revenue
12:36
opportunities. It catches money
12:38
laundering for banks. I mean this is the
12:40
same udal loop, the same fusion layer
12:42
just being applied to different
12:48
So when you pull all these capabilities
12:49
apart, the same pieces that the defense
12:52
sector is using can be mapped to
12:53
civilian capabilities that exist right
12:55
now. So observe you've got the sensing
12:57
layer, you can buy satellite imagery.
12:59
SAR imagery is also commercial. You've
13:01
got capernicus and if you want to go low
13:03
resolution, satellite and plane tracking
13:05
data is effectively free. The more
13:06
expensive thing is AIS ship tracking
13:08
data, but that too is accessible. In
13:10
countries like Japan, you have a
13:12
plethora of CCTV camera feeds. You can
13:14
fly your own drones. You can buy robots.
13:17
You've got social media adtech data. All
13:19
of this is stuff that you can either
13:20
access or purchase. Then you have the
13:22
Orient step. This is classification and
13:24
fusion. This is exactly what computer
13:26
vision and large language models already
13:28
do. Let's say you take your own personal
13:30
security camera and give it to Meta
13:32
Segment anything model. It can figure
13:33
out, hey, here are the objects in the
13:35
scene. Then you give it to Claude or
13:37
Gemini to tell you what those things are
13:38
doing and it can infer the context of
13:40
everything else that's going on. It can
13:42
check to see, hey, you were supposed to
13:43
get an Amazon delivery and that is
13:45
indeed the person that's now showing up.
13:46
Basically, you can build your own AI
13:48
layer that makes sense of all the data
13:50
that's at your disposal. And this
13:52
technology is available to everyone. I
13:54
mean, the more I use my own like Ring
13:56
camera, I am so frustrated by how
13:58
limited this thing is and the fact that
14:00
it has to go to some centralized server
14:01
to do such basic things for me annoys
14:03
me, especially when I know the AI tools
14:05
at my disposal are capable of so much
14:07
more. And then finally, you have decide
14:09
course of action generation, right? So,
14:11
for the military, that means
14:12
recommending a weapon and an asset. For
14:13
us, it's things like what should you
14:14
actually be paying attention to? What
14:16
changed? What's anomalous? and have the
14:18
system surface that automatically. Not a
14:20
barrage of detections like unknown
14:22
person detected at this time stamp that
14:25
doesn't tell me anything and then act in
14:27
Maven. That's kinetics or deploying
14:29
human assets. For us, it's decisions.
14:31
It's knowing what's happening and being
14:33
able to respond to it in near real time.
14:35
Now, I've already been building this and
14:36
here's the road map. The first
14:38
application is God's eye view. You've
14:39
seen this tool evolving and I've been
14:41
building it to fuse open source
14:42
intelligence into a common operational
14:44
picture. what's happening in the world
14:46
visualized the same way these military
14:48
systems do it but with publicly
14:50
available data. You've seen it in my
14:52
coverage of Epic Fury and Straight of
14:53
Hermuz. And clearly y'all love it
14:55
because it's the most unbiased
14:57
perspective we can have without the
14:59
usual color of traditional media. And
15:01
let me tell you, that's getting a lot
15:02
more capabilities because now I've got a
15:04
team. The second application is
15:06
situational awareness for your own life.
15:08
Your own data, your own cameras, your
15:10
own feed, your own context fused into
15:13
the same kind of unified view, not
15:15
surveillance. I'm talking about
15:16
awareness of what's happening in your
15:18
sanctum sanctorum. You've got the map of
15:20
the world and everything that's publicly
15:22
available and then your own island that
15:24
you control. Want to share it with
15:25
others? Go for it. You want to keep it
15:27
private, you can do that, too. Private
15:29
by default and optionally sharable with
15:31
others. I mean, the cool part is you can
15:33
buy all of these fancy sensors. The AI
15:35
is commercial, too. The only piece that
15:37
doesn't exist is the fusion layer, the
15:39
thing that pulls all of this together
15:41
into one picture. And that's what I'm
15:42
building. Observe, orient, decide, act.
15:45
That's how all these platforms work. And
15:46
honestly, that's how great
15:48
decision-making works in general. You're
15:50
probably just doing it slower without
15:52
all of this technology. So, that's where
15:54
I'm going next. God's Eye View is the
15:55
window to the world that we all deserve
15:57
to see. I'm going to keep monitoring the
15:59
global situation as it evolves, getting
16:00
sharper, faster, and soon something that
16:02
you can actually use yourself. And then
16:04
there's the second surface, Argus. The
16:07
same framework turned inwards to your
16:09
world, your data, your context, so you
16:11
can create automations around the
16:12
physical world that you can't even
16:14
imagine today. So, this is what I hope
16:16
you'll get clarity about what's
16:18
happening in the world, but then also
16:20
your world. And by the way, if you can't
16:22
wait for me to release this, so many of
16:23
you reached out and I've responded to
16:25
you giving you the ingredients so you
16:27
can at least go make the viewer
16:28
yourself. Just point your clanker of
16:30
choice at my Substack post and they'll
16:31
do a really good job. If you take my
16:33
YouTube video transcripts in addition to
16:35
that, it'll take you rest of the way
16:36
there. But if you are willing to be
16:38
patient, I've got some really exciting
16:40
updates coming for you very soon. Blavo
16:42
signing off and I'll see youall on the