00:00hi and welcome to the a 16z podcast I'm
00:02Hannah and today we're talking about AI
00:04and automation in the context of
00:06national security given the nature of
00:08today's conflict situations how did
00:11these technologies change how we protect
00:12lives in those conflict situations and
00:15also how is AI shifting power dynamics
00:18joining me is Gregory Allen fellow at
00:20the Center for new American security and
00:22co-author of the Belfer Center report on
00:25AI and national security
00:26Gail LeMond chief marketing officer of
00:29shield AI and the author of the
00:30dressmaker of care Khanna and Ashley's
00:33war both of which dealt with post 9/11
00:35conflicts and Ryan sang the CEO and
00:37co-founder of shield AI we're all aware
00:40that the nature of warfare and national
00:43security today is no longer hundreds of
00:46thousands of men on a field facing each
00:48other and nor is it simply nuclear races
00:51so what are today's conflict situations
00:53actually what do they really look like
00:55the United States has been at war
00:57continuously now basically since 2001 in
01:00a non-stop situation and those wars or
01:03the majority of their time frame have
01:05been in counterinsurgency operations
01:07which does involve boots on the ground
01:09and it's very people intensive I've been
01:12doing that for a long time we're talking
01:14about wars that are increasingly fought
01:16in urban environments settings where it
01:19is increasingly difficult for
01:20servicemembers to distinguish between
01:22civilians and combatants and where
01:24combatants are basically using that
01:26strategically Special Operations has
01:29been asked to do a great deal in these
01:31post 9-11 conflicts 2016 was actually
01:34the first year in which Special
01:36Operations combat deaths outnumbered
01:38those of conventional forces and think
01:40about that is less than 5% of the entire
01:43United States military less than 1% of
01:46this country has fought 100 percent of
01:48its Wars for 17 years if you think about
01:51the numbers it really points to the
01:54mission set these are not nameless
01:56faceless people whose lives are caught
01:58up in these conflicts and I think that
02:00technology has tried but not necessarily
02:03really succeeded in catching up with
02:06what is happening on the battlefield so
02:08if we sort of telescope out how is that
02:11reflected in the overall US national
02:13strategy for dealing with conflict
02:15what's the relationship between
02:17on-the-ground missions to the
02:18overarching military strategy
02:21the most recent u.s. national security
02:23strategy anticipated a shift in posture
02:26for the United States military from the
02:29type of wars that we have been fighting
02:30and we have now a lot of experience
02:32fighting from the types of conflicts
02:34that they want to focus on preparing for
02:37for the first time in a long time the
02:39United States did not name terrorism as
02:41the top national security threat facing
02:43the nation instead great power conflict
02:46right is large-scale exact okay so the
02:49primary sort of named competitors or
02:52potential adversaries in the national
02:54security strategy would be China and
02:56Russia and if you look towards the
02:58preparations that the militaries are
03:00thinking for and their long term
03:01strategy both for organization and for
03:04acquisition that's really where they are
03:06gearing towards but at the same time we
03:08are still west into the first category
03:10exactly so we are struggling to try and
03:13make this transition without actually
03:15transitioning even in these new or
03:18potential conflicts of the future we're
03:20seeing the same uncertainty so if you
03:21look at Russia moving into the Ukraine
03:23in the past couple years there was a
03:25huge amount of deception and uncertainty
03:28in terms of what was actually happening
03:29on the ground yeah let's talk about what
03:31that really looks like as it plays out
03:34what we know what we don't know from
03:36both kind of Mission Control and also
03:39immediately on the ground like what's
03:40the information flow like currently and
03:42what are the tools that are providing
03:44that information like the number one
03:46challenge is getting eyes and ears in
03:48the right places so we have very
03:50advanced technologies in terms of
03:52satellites high altitude platforms and
03:54we also have very brave young men and
03:57women that are willing to get very close
03:58to the information in order to collect
04:00it but the challenge with these
04:01approaches is that the quality of
04:03information is not at a standard where
04:05you can say with certainty what's
04:06happening on the ground either because
04:08of the way that it's collected or
04:09because the limited amount that might be
04:11available and that's because it's from
04:13satellites it's either because it's from
04:15satellites and then the availability of
04:16resources on the ground is quite limited
04:19you know you give an example of the kind
04:20of information that we might have and
04:22the other kinds that we do not so just a
04:24very simple example is what's indoors
04:26versus what's outdoors
04:27right people spend the vast majority of
04:29their time first of all their macro
04:30trends towards urbanization and then on
04:33top of that you know people live inside
04:35they don't spend their days sitting out
04:36in the middle of a field with all of
04:38their activities in open view and so
04:40kind of at a very basic level we can't
04:42see the vast majority of the world in
04:45terms of where people and things are
04:47located I would add to that a
04:49distinction between the amount of data
04:51that we collect and the amount of data
04:53that we analyze right now for instance
04:55on just drone platforms alone more than
04:5795% of the data that is collected is
05:00never viewed by anyone ever and that is
05:03simply because we are collecting far
05:05more data than we have humans that are
05:07able to analyze it so there's one sensor
05:09in particular that is capable of
05:11observing and basically an entire city
05:14at one time but the problem with this
05:16sensor is that there's not enough humans
05:18to be watching it all of the time and so
05:20really its primary use cases as a time
05:23once a ie D improvised explosive device
05:26goes off we then look at that footage
05:29and rewind it to the day before and say
05:31okay who must have planted that
05:32explosive because we don't have enough
05:34people to watch the sensor to see the
05:37explosive being planted but once it's
05:38gone off then we know where to look we
05:40can carry footage at a gas station exams
05:42that gets robbed exactly so this
05:44analysis shortage is really a human
05:46bottleneck within the US military there
05:48are literally thousands of people whose
05:50primary job is to watch drone footage
05:53and analyze it for information that is
05:55relevant to the conflict at hand or us
05:57national security but there's far more
05:59data than you could ever hire enough
06:01individuals to go after there are two
06:03ways that I think artificial
06:05intelligence can really make a big
06:06impact one is just the helping
06:08pre-process that information before it's
06:10presented to people and there's a great
06:12opportunity there and there's some
06:13programs within the DoD that are focused
06:15on doing that the second is actually
06:17just getting better information in the
06:19first place so when you don't have
06:20somebody on the ground and you can so
06:22let's say for example you're looking for
06:23a person and you cannot actually see
06:25that person because the satellite is so
06:27far away exactly and so just getting
06:31higher quality information in the first
06:32place can dramatically accelerate or
06:35reduce the amount of information that
06:36you even need to go through to make a
06:38positive identification of that of
06:40or object or thing that you're looking
06:42for you're really talking about
06:44providing mission-critical information
06:46at mission-critical moments and how do
06:48you get the right information to the
06:49right person at the right time in a
06:51moment when it can really make a
06:52difference so if we're seeing rooftops
06:54how does that play out right now there's
06:56some amount of analysis that is
06:59conducted ahead of time to try to give
07:01the best possible picture to these these
07:03people before they go on their operation
07:05but in many circumstances when people
07:07are asked to conduct operations there is
07:10still a good amount of uncertainty and
07:12so brave men and women are asked if for
07:14instance going to buildings not knowing
07:16whether or not they're booby-trapped
07:18whether or not they're walking into an
07:20ambush or a number of other possible
07:23risks there is a huge opportunity to
07:25improve the gap of uncertainty that
07:27exists yeah yeah because they're just
07:29imagine it's 2018 know it and people are
07:31walking into still walk into a black
07:34dark room and have no idea what's on the
07:37I mean that does sort of feel like rkx
07:40it's a shocking gap right I mean in some
07:42ways when you think about what is
07:44possible versus what is right now more
07:47often times a reality than not so what
07:50is AI doing with that specific kind of
07:52immediate unknown and relaying
07:55information in the moment that's a good
07:58question because clearing buildings of
08:00threats has been one of the most costly
08:02missions for US forces in terms of human
08:04life and one of the most costly missions
08:06for civilians in terms of human life
08:07since post 9/11 we've applied an
08:09artificial intelligence to a drone
08:11that's able to fly through buildings and
08:13and basically in a completely autonomous
08:15manner it looks for people inside of
08:18those buildings and so rather than an 18
08:22year old or 19 year old being the first
08:24person through a door you can throw a
08:26robot inside which will provide a very
08:28clear picture about what's going on in
08:30the inside similar to people listening
08:32have seen the movie Prometheus where
08:34they've met these robots that explore
08:35caves and you really can make a
08:37difference to get in terms of mission
08:38effectiveness right I think we have to
08:40remember the stakes when you think about
08:42the consequences of not having that
08:44information we have seen over and over
08:46people walking into booby trapped
08:49compounds or compounds where people who
08:51understood when US forces were coming in
08:54but left things for them very explosives
08:57for them to find and what happens when
09:00you don't have that information ahead of
09:01time is you know you can have really
09:03tragic loss of life and that kind of
09:06thing is happening all that every single
09:07night and do think we're a country that
09:09has gotten perilously far away from what
09:12we've asked of people that's an
09:15interesting point that it's bringing you
09:17back to a very immediate this is the
09:19information that they need right now can
09:22you break down the technology that makes
09:24that possible that wasn't before so for
09:26a long time we've had machines that
09:29could significantly outperform people
09:32and their ability to execute mechanical
09:34functions provided they were repetitive
09:35or well constrained and what AI
09:37represents is now the ability to allow
09:39machines to apply them to a larger
09:43spectrum of activities and so for us or
09:46the way that we think about the tech
09:48stack is in terms of a decomposition of
09:50intelligence and what does it mean to be
09:51what we would consider a resilient
09:53intelligent machine that breaks down
09:56into two buckets of things so the first
09:58is what we call perception action and
10:00cognition and the second we call
10:03introspection adaptation and involvement
10:06so let's start with the first perception
10:08action cognition perception is the
10:10ability to look around the world and
10:11understand what you are seeing and it
10:14could be through a machine a camera but
10:16it doesn't necessarily need to be
10:18looking at the physical world it could
10:19exist in the cyber domain it could exist
10:20in the electronic domain but kind of at
10:22a fundamental level if there are objects
10:24in these locations and I recognize what
10:26they are and for shield is it purely
10:28visual no we use a combination of
10:31cameras light ours and radars and many
10:35other sensors to help machines navigate
10:37the world to get into the areas they
10:38need to in order to collect the
10:39information the environments that our
10:41machines operate in are relatively
10:44challenging in the sense that they're
10:45very there's a lot of dust a lot of
10:48unstructured obstacles battlefields are
10:51dynamic at the same time they're people
10:53and so on so you need a lot of
10:54complementary sensors in order to ensure
10:58reliability how about things that a
11:00human would walk in and notice like
11:01smell I smell cat robots can carry
11:04similar sensors that don't necessarily
11:08they can look for things that you
11:09wouldn't be able to see explosive
11:11residue for example it's a beyond human
11:14perception there how does that work
11:16their chemical sensors there are hyper
11:18special cameras that can see things
11:20beyond what our eyes or normal cameras
11:22would be able to see and so that they
11:24actually do have superhuman sensing yeah
11:26capability but the key is to turn that
11:29from pixels and data into actual
11:31understanding that the machine can use
11:33because for a long time we've actually
11:35had the ability to apply these sensors
11:36but it always came back to humans
11:38needing to assess the information in
11:40order to determine courses of action
11:41which is no good if you're walking into
11:43a doorway right then that's correct
11:45so perception is this notion of here's
11:46where everything is in the world
11:47cognition is given my prior experiences
11:50and what I want to achieve this is what
11:53I should do and then action is just I
11:55affect the world or move myself in the
11:58way that I need to in order to take
12:00whatever step I decided to do and we
12:02just go through this perception
12:03cognition action loop over and over and
12:05over again as people and machines do the
12:07same thing now in order to achieve
12:09really advanced levels of performance
12:10there's another component and we would
12:12call this loop the introspection
12:14adaptation and involvement loop so
12:16introspection is the notion of what are
12:19my capabilities or what is my health and
12:21so this could be something as simple as
12:23what's my battery life or it could be
12:26something more complex such as I know
12:29that I'm good at doing X and therefore I
12:32can behave optimally in these
12:34circumstances and I know that I'm bad at
12:36Y and therefore I'm going to spin up
12:39millions of simulations to become better
12:41at Y and what are some of those X's and
12:43Y's it could be I know that if I try to
12:46fly through a doorway that is 24 inches
12:50across I know that I can do that very
12:52well if I know that I need to coordinate
12:55the exploration of a village with 100
12:57other robots and the communication
13:00network is going to be jammed too the
13:02whole time I might not know how to do
13:04that well today I need to solve this
13:08and so then adaptation is given my
13:11awareness of my capabilities how can I
13:13change what I'm doing or change
13:15something about the circumstance in
13:16order to improve the likelihood that I
13:18succeed and finally evolve meant
13:21that given an encounter with a situation
13:25enough times the machine then becomes
13:27very good at it it feels like a very
13:29granular immediate level of all this
13:31large-scale AI and machine learning
13:34playing out into moment by moment I'm
13:36walking through a door what am I gonna
13:37see right before I get there but you're
13:39gathering this incredible amount of
13:41information about the spaces about the
13:43context about the environments about all
13:45kinds of things that humans aren't even
13:48picking up on are there other uses that
13:50are less immediate when you're doing all
13:53this information gathering that you can
13:54see this information playing out in sort
13:56of longer-term ways in terms of either
13:59on-the-ground conflict like this or
14:03ways that you can see that information
14:05being used beyond on-the-ground
14:08decision-making an awful lot goes on in
14:10a conflict zone and there's a ton of
14:12different diverse types of machines in
14:14the environment and sensors in the
14:16environment the United States military
14:18has outfitted itself with an
14:20extraordinary diversity of sensors and
14:22they are collecting an unimaginable
14:24amount of data but most of that data
14:27just goes into cold storage never to be
14:29seen again because there aren't enough
14:31people to analyze it and to drive
14:33insights from it now with advanced
14:36machine learning we are for the first
14:38time really seeing an opportunity to
14:40make use of datasets that historically
14:42would live dormant so the archives are
14:45suddenly newly useful right there's two
14:48types of data here one it would be the
14:49sort of data that the United States
14:50military knows that it wants to collect
14:52which might be like intelligence or
14:54reconnaissance imagery like satellite
14:57imagery or drone based imagery but
14:59there's also this whole diversity of
15:01data of what is going on within the
15:03mission within the platforms that we are
15:05using for instance any kind of flight
15:08scenario what occurred with the airplane
15:11while and was flying while I was
15:12executing the mission that data is not
15:14normally saved or archived in a way that
15:17would be accessible to an algorithm
15:20trying to learn about what happens when
15:22we fly military aircraft in general so
15:24what kind of use would that information
15:26be applied to what will it actually
15:28change the opportunities there are
15:30really interesting for applying AI to
15:33enhance our training and simulation
15:35because we can learn more about the
15:38truth of the types of situations that we
15:40encounter and then create simulations
15:41based on that truth upon which to train
15:43and also to think about our strategy and
15:46tactics and our organizational
15:48efficiency that goes from the whole
15:50spectrum of military logistics to
15:52enhancing fuel efficiency all the way
15:55down to getting into the nitty-gritty of
15:58combat operations and thinking through
15:59how do we reduce casualties and loss of
16:02life on our side and then how do we also
16:04reduce unintentional casualties and loss
16:07of life on the other side right now when
16:09we see a building and US troops are
16:11receiving fire from that building we
16:13have to make the decision do we take the
16:15easy way out which would be to call it
16:17an airstrike and topple the whole
16:19building or do we recognize that there
16:21might be non-combatants in that building
16:23that we don't know about and do we
16:25choose the harder choice of going in on
16:27the ground and that's often used against
16:30the United States so if you look at
16:31Syria the last stand of Isis in the town
16:34of Raqqa you had Isis really using human
16:37shields so you cannot leave for two of
16:40this building we have floor four of this
16:42building and we know that that will keep
16:44you here and that will protect our lives
16:46because we're endangering yours the
16:48shift in tactics from the early days of
16:50the Iraq war to the more
16:51counterinsurgency strategy that we saw
16:53really throughout the sort of second
16:55half of that conflict was all about the
16:57United States is saying that we believe
17:00we need to take the higher risk and
17:02endure higher casualties because winning
17:05the support of the local population and
17:07showing them that we absolutely care
17:09about their lives and quality of life as
17:11we are engaged in this conflict is
17:13crucial and so what I think is very
17:15exciting about artificial intelligence
17:17is can we still make the hard choice to
17:20not call in an air strike or not call an
17:22artillery but can we use technologies
17:25such as robotics such as AI enhanced
17:28sensors in order to reduce the risk when
17:31we make that hard choice of loss of life
17:33on our side but also again unintentional
17:35loss of life among the civilian
17:37population you know I think tying
17:38together this whole idea of current
17:39conflicts and future conflicts you know
17:42as we look at great power conflict there
17:45is a sense that the post-war rules
17:48order is facing increasing threat
17:51secretary mattis calls it the greatest
17:53gift of the greatest generation is this
17:55rules-based order we've all lived in and
17:57now kind of take for granted and I think
17:59you know we all think it's free we think
18:02that it is possible to ignore it and we
18:04also think that it's permanent right and
18:06the truth is it's none of those three
18:08things did a question about we're
18:10collecting all of this data what do we
18:11do with it I think that there's
18:13absolutely huge opportunity to improve
18:15human understanding to enable the best
18:17possible decisions and we should view
18:19that as kind of the the critical use of
18:21the data so let's fast forward 50 years
18:23we'll battlefields be predominantly
18:25composed of machines or predominantly
18:27people I think it's reasonable to
18:28believe that that certainly there will
18:30be far more machines in the future than
18:32there are today and so therefore this
18:33data represents the opportunity to train
18:36these machines to reach the level of
18:38capability that they need in order to
18:39protect national security and global
18:42stability we think a lot about taking
18:44that data not only deriving human in
18:47sight but deriving machine insight so
18:49that it can continually evolve and
18:51advance its performance what does that
18:53look like our chief science officer nei
18:54took a quadrotor that doesn't know how
18:56to fly that has the most advanced
18:58controllers designed by people ever and
19:00in a period of a few days the quadcopter
19:03just through its own experiences learned
19:05to reach boundaries of performance that
19:08far exceeded what could be realized by a
19:10controller designed by humans and this
19:14was notable for a couple of reasons one
19:16the learning was lifelong and it was
19:18doing it unsupervised a lot of times the
19:21challenges with these machine learning
19:23approaches are you worried about them
19:25learning the wrong thing and therefore
19:27you have people kind of in the loop
19:28cross-checking whether or not the
19:30machines are learning the right thing
19:31and was able to do this and continues to
19:33be able to do this to learn basically
19:34forever from its experiences and the
19:36things that it learns are within
19:37performance boundaries so the humans are
19:39still setting all those boundaries and
19:41it's just continually honing and
19:43learning new things within those we
19:45believe in the role of having humans in
19:46the loop and allowing them to learn
19:48particular skills within performance
19:50boundaries but still having people there
19:53with final authority on what they
19:54actually do is a key concept but also
19:57it's finding its own boundaries in terms
19:59of what's physically possible okay
20:01given environmental constraints given
20:04its own health and so on and it's able
20:06to transfer that learning to every other
20:08robot in the fleet immediately and is
20:09also able to transfer that learning to
20:11machines that have different
20:13computations sensing and actuation
20:14constraints and each of those machines
20:16are able to introspect identify the
20:18differences between themselves in the
20:21learning machine and only take lessons
20:22that are relevant to them if you have
20:25performance guarantees or boundaries for
20:26the system and you know that all the
20:28learning will take place within the
20:29performance guarantees people can
20:32anticipate everything that it will do
20:35ultimately they might not know how it's
20:36going to get there but they know the
20:37behaviors will be bounded I mean you
20:39talk about robots fighting words
20:41eventually but what you're also
20:42describing is a technology for humans to
20:45make better decisions no I think that's
20:46a critical point I think people want to
20:48go to what robots are going to be able
20:50to do tomorrow but I think we really
20:52come back to what can we do to protect
20:53lives today the conversation until they
20:56are is really always about the idea of
20:58getting the best decisions getting the
20:59best information and making sure that
21:01you're creating the most knowledge I
21:03think it's really important when people
21:05think of artificial intelligence systems
21:07that they also think of them as
21:10information and intelligence gathering
21:12tools and that's often lost in the
21:15discussion about AI and national
21:16security right when we go to a place of
21:18Terminator you're not thinking about the
21:20information right it's Hollywood version
21:22versus the battlefields reality and we
21:25just did that here we're talking about
21:28performance guarantees in the context of
21:30collecting information and immediately
21:32it jumps to oh my goodness AI is this
21:34terrible thing and it's these machines
21:36are just learning to collect information
21:38better which will save lives
21:39historically on the battlefield new
21:41technologies change power dynamics so
21:45from small like I have a boner oh you
21:48have a rock I have a gun you don't have
21:49a gun all the way up to like I have a
21:52nuclear missile and you don't I hear the
21:55immediate way that AI is changing that
21:56power dynamic on the ground going into
21:58unfamiliar environments are there bigger
22:01ways that it will shift a relationship
22:03between developed or undeveloped
22:04countries or different players in
22:06conflict advanced AI techniques come
22:09from a long history of the military's
22:11use of automation on the battlefield so
22:13the first aircraft auto pie
22:15was developed with the use case of
22:18military aircraft in mind when I was in
22:20the 1920s so we've had autonomy exactly
22:24the 1920 yes absolutely but I think
22:26what's interesting is that this is all
22:27with traditional software programming
22:30architectures with a very long list of
22:33if-then statements ultimately all of
22:35which were typed in by some human and
22:37what's what's different now is the use
22:39of machine learning whereby two a sort
22:42of oversimplified since the system is
22:44programming itself based on exposure to
22:47examples indeed all right humans aren't
22:49necessarily labeling the features
22:51anymore well they might be labeling the
22:52training data but they are not in as
22:54many words programming the system in the
22:56traditional sense so the military has
22:59this whole series of verification and
23:01validation procedures that it has
23:03developed for traditional software how
23:06do we know that our automatic systems
23:08are auto pilots or our heat-seeking
23:10missiles or anything that we do that
23:12uses software is going to do what we
23:14want it to do well we have evaluation
23:16procedures for such software but that's
23:19for traditional programming
23:20architectures right machine learning is
23:22a new program in architecture and we are
23:25pretty optimistic over all that we think
23:27that this can actually enhance safety
23:29but that's not an inherent feature of
23:31the technology today electricity is by
23:34far the safest way to light your home
23:36far safer than using candles but that's
23:39not an inherent feature of electricity
23:40it's very it's very easy to start a fire
23:43using electricity and in the early days
23:45of electricity they started a lot of
23:46fires and so right now I would say the
23:48military is in my view doing the right
23:51thing which is its early use cases of AI
23:54are far removed from the use of force
23:57and in fact are in non safety critical
24:00applications such as data analysis we're
24:03not so certain that all other countries
24:05on earth are going to abide by that
24:07there was a recent headline in defense
24:10one Russia to the United Nations don't
24:13try to stop us from building killer
24:15robots and I wish that was a joke
24:17headline but that's actually a pretty
24:19accurate summary of what Russia said at
24:21the most recent UN meetings on
24:23autonomous weapons so we're doing this
24:26within a global security context
24:28renewed great power conflict and other
24:30countries see artificial intelligence as
24:33a way to close the gap between their
24:35militaries and that of the United States
24:36and not everybody is functioning under
24:38the same conceptual framework there so
24:40how is the policy community responding
24:42to this how is it actually perceived
24:44right now you do see this concern among
24:47the policy community will the folks that
24:50we are up against have the same ethical
24:52framework and I do think that is a
24:54question that will be facing
24:56policymakers in the future it's actually
24:59pretty easy to policymakers to say oh AI
25:02is quite interesting I think the
25:03challenge is persuading them of the
25:05scale of the importance of this
25:08technology they're hoping that they can
25:10mostly do things the same way they've
25:12always done them with a few tweaks here
25:14to update for the new technology no this
25:16is a complete revolution it will take
25:18decades to unfold but it will be on the
25:19same scale as the invention of aircraft
25:21right for the early new paradigm it's a
25:24whole new paradigm national security and
25:27if you think you can get by with the old
25:29rules I mean the old approaches that led
25:32to success in the Cold War those just
25:34aren't gonna apply here so it's really
25:36the scale of recognizing how much change
25:39is required and how much investment is
25:40required to realize that change there's
25:42a kind of pacing question in that which
25:44is do we'd want people going slowly to
25:46try to understand this enormity or do we
25:49move ahead quickly there's a tension
25:51the most recent defense budget basically
25:53said let's buy a ton more weapons of
25:56dwellings that have already been
25:58designed and politically that's
26:00incredibly popular right because those
26:02weapons are built in congressional
26:03districts all over America it's very
26:05easy to say let's just spread the money
26:08around but in my view that's the
26:10equivalent of Kodak in 1991 radically
26:13increasing its investment in film
26:14cameras right you're buying a lot of
26:16stuff that is probably going to be
26:18obsolete in the not-too-distant future
26:20and so what I wish the military was
26:22doing was thinking more on the
26:24modernization question investing more in
26:26research and development and preparing
26:28themselves for the AI revolution the
26:30United States has the software talent to
26:33make a difference in this conversation
26:35so I do think it's important that we
26:37keep in mind that it's not all doom and
26:39gloom and there's actually real
26:41protection that this
26:42technology can offer one thing that I
26:43think is very interesting about AI in
26:46contrasting with previous technological
26:48revolutions is that the source is very
26:50much in commercial industry I am NOT
26:52breaking some like security clearance or
26:55classification rules to tell you this
26:56here but there is no super secret
26:58government lab with like advanced AI way
27:01better than commercial industry the
27:02government the military they are behind
27:05commercial industry and is this the
27:07first time we've seen that swap
27:09it's very unfamiliar territory yeah for
27:11the US military as a commercial start-up
27:14working with government how do you see
27:15them responding to these technologies
27:17coming from outside instead of within I
27:19think there's recognition that this is a
27:21new paradigm and so I think that's why
27:24you've seen di UX is probably the most
27:26predominant example of the DoD
27:28attempting to respond to the new dynamic
27:30mxpeg feel like b2b which is a pretty
27:33wonderful thing when you think about it
27:35a lot of the legacy defence companies
27:37they sort of have two competitive moats
27:39one is their experience in aircrafts or
27:42boats or whatever and the second is
27:44their familiarity with the government
27:45contracting process which historically
27:47is incredibly painful China has just
27:50announced that they were investing 2
27:51point 1 billion dollars to open up a new
27:53AI Research Center that is all
27:56consistent with their strategy of
27:57military civil fusion that's very
28:00different than in the United States
28:01where the Department of Defense in the
28:03national security community generally is
28:05at a very hard time persuading big
28:08Silicon Valley tech companies that they
28:10should devote the time and effort to
28:12help the DoD think through an AI that's
28:14why the DoD is so excited to work with
28:16startups because they don't have the
28:17legacy that some of these bigger
28:19technology companies do that's really
28:21interesting thank you very much for
28:23joining us on the a 16-0 assed thank you
28:25it's great to join you thanks it's great