Can AI See Better Than Humans?

But you and I have done, you know, we've crossed this path before about high dynamic range cameras and I was super impressed by some of the stuff that you showed me. Can you tell us a little bit more about that? What's happening there?

Yeah. What do you think about the data center uh impact?

All right. So, let's um let's explore some other areas. We talked a little about driving with sensor fusion and so on. Do you see them going into other industries? I mean, I'm working in construction, for example. I see there's opportunities there. What are you seeing from your perspective?

I mean, I I have a Tesla and I turn it on to full self-driving to go, you know, 200 miles on the freeway and for a freeway environment and even during, you know, a small town, it does really well.

The final question, which is

AI, is it hype or help?

It's absolutely.

Good day and welcome to the show. I'm your host, Pete Cooper, and our topic is AI hype or help? Today's guest is Chris Silsby. Now Chris is a lifetime professional in image science, image sensors, cameras, and now AI and how it relates to that topic. Now he and I share that very common background and interest in vision systems and cameras. So this is going to be a very interesting discussion.

I hope you enjoy it and let's get into it.

Good day Chris.

Hi. Hello.

Welcome. Thank you for having me on.

Thank you. You know, I've been chasing you. Maybe not actively chasing you, but thinking about having you on the show for for a while. Um, and I know we're going to get into some really interesting stuff because there's a great crossover. Um, so you and I go back to I think it was 98 when we first met. I was with Nokia and you were with a company called Agulant. Um,

we started working on cameras way back then. Do you remember what people used to say about putting cameras in phones way back then?

It depend on who you ask, but I mean uh they had uh conferences where they presented that older population said there's there's no way cameras would replace people's uh primary camera

and or cameras and cell phones would replace a primary camera. But the younger generation, you know, the 18 to 24 year olds, a large percentage of them thought, you know, someday I can see that my primary camera will be in my phone.

And sure enough, you know, as a as that younger generation

grew older um and the camera's improved now that we see that everybody's almost everyone's using their

Exactly. As a primary camera.

Yeah.

I I just remember this sound bite and I quote to everyone and that is that people used to say why would I need a camera in my phone?

Yeah, exactly. Remember when we first did a study that that uh everybody talked about what they would use a camera for and uh it turns out that you know just a VGA image would be enough for 80% of the cases of what people would use it for. You know like hey honey is this what I'm supposed to buy at the store you know or um just trying to capture our memory.

Yeah.

So,

yeah.

So, let's get into that, but let's first of all introduce you, Chris. Um tell us a little bit about your background and what you're doing now.

Yeah, I started at HP um in the integrated circuit business division. So I spent a lot of time um learning semiconductor processes and I worked in a variety of of roles there with as a process development manager, as a product engineer, as a yield manager um and photoiththography engineer um and then moved to the sensor uh image sensor design role in a similar time frame when we met

at the end of 90 1999 00

and uh that's what I've been doing for the last 25 years.

Uh and recently moved to an applications development role which is more of a systems level automotive systems level role

uh looking how to integrate image sensors with the automotive system.

Yeah.

So if the audience doesn't know what the image sensor part is every digital camera now has an image sensor which is the silicon that sits behind the lens. Um,

yeah.

Does all the kind of the work of converting the light photons into electrical signals, right?

Mhm.

Yeah. And so, have you always been in that area um of the world? You're in um Oregon, right? Have you always been there or did you move around a bit?

Yeah.

Well, I moved to Colorado for a couple years just because HP was

um had a fab startup there and we were transferring some processes

um to Colorado and then overseas and so a little adventure with the family to Colorado and then I moved back and currently in Oregon.

Nice. Nice part of the world that um the way you live. I come up to see you a couple of times and

it's beautiful.

Very it's a little wet I guess. Um but

oh that's really nice.

Mhm. Yeah. Yeah.

Yeah. So um what are you doing now? I mean what's what's interesting now with what you're doing? I mean and we've seen I guess let's let's educate everyone about what's happened.

We started off with together with these tiny little sensors right

you know VGA as you said which is only 360, 000 pixels. Um, and I can remember also these conversations that VGA would be enough. And also it was I guess we constrained ourselves by thinking well you know you got to fit it into a small phone and back then I mean phones were small.

Yeah. Yeah.

They were like getting we still sell you know I'm focused on camera uh for automobiles. So cameras in automobiles whether it's a backup camera or a forward facing camera or a side camera. Um that's been our focus. It's been my focus for the last 15 years really.

Um, and the we still sell VGA cameras uh mostly for backup cameras, right? But for a lot of people u a VGA uh camera, you know, is 640x 480 um columns and and rows is uh is enough to see what's behind you and avoid running over somebody. Um but the latest generation is 8 million pixels right and u the 8 million pixel 8meg cameras which is like a 4K resolution right 4, 000 columns

um those 4K cameras is that's pretty typical now in the the highest end cars

and for forward facing and in some cases even around the car

um cars have maybe 10 to 15 cameras uh now

wow And eight eightmeg is usually a forward facing. There's usually one to three cameras are forward facing.

And then it's either 8 meg or threemeg um around the rest of the car

um capturing the environment. And

in in some cases it's just a surround view function where people want to see what's around the car when they're parking or um you know when they're sur looking to to move the car. But in other cases, uh, we we have customers that are actually using all those cameras to make decisions about, um, whether to change lanes and or, you know, whether to move the car out of the way if there's a threat event. So,

um, it's there's a a lot of processing that's going on now in the in recent cars.

Yeah, I think we're all familiar, I think, with the backup camera.

I have one of those in both my cars, but I don't have any other cameras other than that. Um,

and then I guess the Tesla example is probably the most notori has the most notoriety in the sense that there's been a march towards full self-driving. Um,

which hasn't necessarily gone to the schedule and the timeline. And my perception there is that it's it's very very hard to make the system as reliable as and as perfect as we expect it to be. And um that's not that we're um trying to be as good as humans. We're actually trying to we're trying to make the hardware significantly better than what humans can do. And that to me is the big challenge there.

I think the you know the better than humans is uh is possible right because humans tend to focus have in a narrow view and just right in front of them even even our own visual system um your brain is mostly processing the center field of view of of what you're looking at

um then you have to move your head around to really gather the rest of the information where where car has 10 15 cameras and it's all feeding into an ECU that can process all that information at the same time.

Um, so it's it's its potential is there, right, to to be better than humans for making better decisions, you know, not turning and changing lanes when someone's trying to cross the road and running over somebody or not pulling out when there's a car coming quick coming fast, you know,

and maybe running a a stoplight, but you don't see it.

Um, so I think the potential is definitely there. Um, and but like you said, it's those corner case conditions. It's the it's the understanding intent and where people where other cars are moving. Um, what's going on in an environment if it hasn't been something that it's been trained before,

right?

And it has to come up with some its own uh

causal reasoning, you know, for what's going to happen next. And how does

um these different elements that have come into the scene, how are they going to behave? If it hasn't been trained, then it's has a difficulty, you know, deciding what to do.

There was a case with the Tesla that went under the uh

under the truck um

and completely,

you know, decapitated the guy. Um it found out that he was not paying attention and he was relying on it uh completely. And I think this goes back a few years.

So obviously they fixed that problem, but it was the fact that the truck was white um and it kind of blended with the sky a little bit and um it was a new edge case that um that the camera couldn't have camera system couldn't have detected. And I wonder if that's a sort of a sign of a fundamental flaw with uh vision systems is that if they don't know that there's something there, they assume nothing's there, right? And that's perhaps not the most

it could be. Yeah. And those older systems, the older cameras, I mean, that's that was that camera is now 15 years old, right?

That that had those issues, but also radar was involved and the radar was was uh imaging under the car, under the truck.

Oh, okay. And so, um, reliance on two different systems, radar and cameras, and the fusion of that information,

um, is also important for making decisions. And, um, I I I think at the time Tesla went away from radar temporarily and then

then just use cameras. Um, and then I think now more recently, they've added radar back in

and have improved the the fusion decision- making. Um but those those corner case conditions they're not so much limited by the sensors anymore, right?

I mean the the the latest sensors are um significantly higher dynamic range and significantly u able to avoid saturation even in high

bright scenes.

So in that dynamic range thing that's uh that's the cases where you have like the sun um yeah you know in in the scene white truck in the sun. Yeah.

Yeah. And there's a dark corner over here. Like with a human, you'll look at you look away from the sun and you'll look to the dark corner. Your eyes will adjust and you'll focus in that area or naturally turn your head away. But a camera can't turn its head. You can't turn away. So it has to deal with the whole scene.

Exactly. Yeah.

And it doesn't see things that are in the dark because it's been

blinded. But you and I have done, you know, we've crossed this path before about high dynamic range cameras and I was super impressed by some of the stuff you showed me. Can you tell us a little bit more about that? What's happening there?

Yeah, I mean the we're talking about cameras now that historically uh a digital camera could only hold maybe 100, 000 electrons. You know, even even the movie cameras that that were used to to make movies in films, they had big pixels and maybe hold a 100, 000 electrons.

Um, the cameras now that we're shipping

um and just announced, for example, an 8meg camera that that can capture over 10 million electrons in every pixel.

Wow. And if you can capture 10 million electrons in a pixel with the read noise less than an electron, right? It's 0. 56 electrons read noise, we're talking about, you know, over 130 dB, um, you know, over 21 stops of of of photography in every pixel.

Um, it's, you know, it's extremely powerful to to capture everything that's needed to to avoid making the wrong decision. And at that point, now you're now your challenges are just um teaching the system um how to get past the just what it's trained for, right?

How do you how do you teach the system um to be more than just a child that's experiencing everything for the first time? You know, how do you set up the the next layers of of uh of of reasoning and and intuition about

what what's the right thing to do, right?

And that that is kind of the limitation, I think, of current systems is that

is that they're

mostly based on training.

And I know that the neural network training um a lot of the companies are adding on some structured um decision- making on top of the neural network

to kind of give it some guard band and some constraints and um

because you don't want the AI system to either hallucinate or um make some intuition assumptions that that uh based on the structure that you might learn as a human going through life.

I think we underestimate just the huge amount of contextual we gather as we

um

yeah I think that um I want to pick up on that because I was at a conference only a couple of weeks ago this is in construction talk a little bit about that and one of the professors from university says he doesn't believe in uh that the the path we're on with just more and more uh large language models is the path to artificial intelligence. He said that that's a pattern recognition machine, but it has no bearing in uh the real world, the physics of the real world. He says what we need is something else.

He doesn't know what that is

that can interpret the real world like a physics engine or something like that. He didn't even use that. So I just made that up.

That kind of incorporates itself into the to the AI that builds up the logic. So you can't make you know completely irrational judgment about

what's that he said that's the path forward, but he doesn't know how that or whatever how that's going to work out. And is that even really AI? Is it something else? I I mean, you and I have been trained in science and it's it's deterministic. It's uh it's algorithmic. It's created on on mathematics and the mathematics is has a result which is black and white.

Um and AI doesn't work that way. And but the hybrid of the two is something that we as engineers I guess we have our pattern recognition. We deal with people, we have all that neural network kind of stuff, but then we also be able to bring in the real world physics and the maths and um maybe that's uh maybe that's the path forward for, you know, the self-driving um too. You know, you can't hallucinate something that doesn't doesn't can't possibly exist. That couldn't be an elephant flying through the sky.

Right. Right. Yeah.

And

you know and I there's there's interesting results that are happening as mathematicians I heard I just read recently that that a bunch of mathematicians converged at Berkeley uh in order to study uh whether AI could solve some of the most difficult mathematical uh um questions and mathematical uh solutions and they were challenging the the AI system that they were working on um and it was also learning as they were going

and what after the end of this week of of uh challenging AI and the the system getting better and better, they they came away um you know with with the most of the people that were interviewed with sort of a profound new new view of where things are headed. Right?

It's not quite there yet, but they they see the potential for um these AI computers to have significant role in in mathematical theoretical uh solutions and

and sort of um breaking new ground and

that this is just an example of one type of computer that's you know focused for a very specific type of use and and and there, you know, these AI tools are trained based on patterns and and mathematical rules, etc.

Um, but in general, like you said, I think it's it's these layers of expertise on top of each other

that end up um converting what looks like a new one-year-old child that's looking around and trying to experience the environment and crawling and,

you know, putting stuff in its mouth and, you know, is it seems like these AI tools are they're they're really good at what they're trained to do. Um but there's there's still some corner case um

contextual learning that we still need uh to get added to these systems to be fully

um uh not only

safe and but for for a wide range of conditions but um to be able to handle new new things that happen, right?

Some some new uh corner case condition uh that it hasn't ever seen before or been trained how to deal with. Right.

Yeah. Exactly. So you you mentioned that you know we touched on the whole thing about Tesla moving out of radar and back into radar. And I believe one of the changes is they changed the the uh to millimeter wave in other words high frequency um radar which meant that it was more granular and more able to do things.

Uh and so it's back to the whole idea of sensor fusion which is taking cameras and other things and and from the work that I've done in the field I've I've noticed that when it's safety critical systems they tend to have sense of fusion like um forklifts running around uh warehouses they tend to have some LAR. When I was in Texas at this uh show these those who cars they had cameras but they also had those those LAR systems. So,

do you think we still need to have sensor fusion or or do you think there's a near-term future for just uh for just cameras?

So, um the nice thing about LAR is that it gives you a depth uh result for every pixel.

M

and if you think about what decision making you have to mi make and what uh analysis of the scene uh that that you're going to process having depth immediately um and accurately simplifies the the decision-m

right because you already know

where an object is and how far away it is and and so um there's some simplification.

Um

Tesla's presented an example in their in some of their uh YouTube updates or some of their conferences that that they get depth because you know from parallelism

um with two different cameras they you know they get a depth information the AI it system itself will tell them you know this how far an object is

um because of the parallax I mean not parallelism but parallax um And so the the that's more processing needed to get the same information.

Um but it's still possible.

You know what I mean? Um

so the the question is how much easier is it for Whimo with several maybe three or four LAR uh sensors around the car all getting depth information.

It's it's likely s more simple for them to make decisions about that are um that are every frame that they gets us information accurate as far as where things are, right?

Whereas um it takes some more latency and processing power uh to get some reference from the parallax. So

okay,

I think it helps Whimo with you know to make faster decisions and and some of the processing is probably easier. So I was

you to take it to a different direction. I was going to go. I was going to go in the direction.

Well, if you've got the LAR information

and it sees something like there's definitely an object there, you can override the whole system and say, well, look, if you see something I'm going to run into,

apply brakes no matter what the vision system says.

Yeah. Yeah. Exactly. And and I think the you know, from parallax, you can get the same information,

right?

Um

Right. Um, I mean, you you look at a shark that a hammerhead shark, you know, that's got the sensors. They they've done tests and, you know, it can see a quarter um, you know, on the surface of the uh,

the floor, you know, at

100 meters or 50 meters or something like that, right?

So it's it's uh their sensors from echolocation are extremely accurate because of the the parallax and um you know so I think the the capability is there but but that's taking processing power from an you know overall box unit for you know 30 to 50 watts or whatever that's available in the system to do the processing you're taking up processing power to get that information.

Interesting. So

I don't know that someone's done a lot of analysis of you know how much of that is of the total processing power how much is that impacting the final system. I don't really know that but it is a little bit easier if you already have depth.

Yeah.

So I mean I think about parallax in terms of you know humans and the eyes right they're there in apart

and it's very good at arms length right? If you want to reach out, you're a monkey, you want to reach out and grab the branch, it's really good at that. But as soon as it gets to sort of 30 or 40 meters away, it's very hard to gauge that distance. Now, with a car, you can go a little wider because the width of the car,

but at some point, you you're not going to get the resolution you need. And of course, cars go so much faster,

but maybe there's enough trade-off with the width and then the depth to to to get that to work.

Yeah. And the, you know, the Tesla system is very good.

Um, I think the, you know, now that we have cameras that don't saturate, you know, in various scenes, we're driving into the sun and with with trees with, you know, or dark regions in a tunnel. Um, they're so good now that that I think a camera system is, you know, is is very good and can be trained for those corner case conditions. I think

the the next level of, you know, what do you do with with something you haven't seen before? I think both both Whimo and Tesla have to deal with that. Um,

and um, you know, we're going to hear about corner case conditions for both of those car systems and both those companies

as as we move forward.

Yeah.

Um, I think one thing that uh has really resonated is that and initially people thought in the industry, well, you know, we're replacing something that's really not very good. There's 35, 000 people die in the United States every year from car accidents and so people are going to be happy if we can reduce that to 20, 000 or 15, 000 um with cars, right? And that's not really the expectation, right?

The expectation is it that it's like an airplane and that you know it's it's any accident has to get investigated and and the these car systems are expected to be a thousand times better than a human and I think the expectation is that there's less than instead of 35, 000 if everybody was driving autonomous cars the expectation is there's less than 10 you know deaths in the United States from these autonomous cars because because now it's a corporation killing in your family or killing killing someone that you know

not not a person, right? Which is fallible and and uh you know it's possible and we're all we all make mistakes, right? But

yeah,

expectations are are still really high.

So that's kind of surprised people a little bit. I think

um

I would say the the almost the legal side of things which is a bigger problem than the technology. Um the who do you blame? Like you can't okay that that AI instance which was on there did the wrong thing let's delete that

instance you know who cares that instance doesn't care because it doesn't have a consciousness but a human being if they make a mistake they can go to prison and there's a lot of consequences and

there's this kind of I guess this equity that we have um in our society that if you do something wrong you have to pay you know the consequences of that But

AI doesn't have any consequences.

So it has to go up to the corporation

and then the corporation has to bear the liability.

Um and that's a faceless thing. You know, it's that there's no person to blame. And I think it's a challenge for us to maybe this is one of the very first areas where we're starting to see AI um replacing or taking society to a new level. Like if we go back to the jungle, I was in my my monkey troop, you know, if I didn't like the guys over there, we all gang up. I mean, that they're in trouble, right? That that one per that one monkeyy's in trouble. That was very basic, you know, social fabric. Now we're so abstract from that.

I mean, you know, we were global company, global where uh, you know, I know thousands and thousands of people, you know, I can't possibly operate in that way. Now we got AI again throwing a whole new challenge. Um, I don't know the way forward, honestly. Do you think there's a future for self-driving cars? I guess that's that's the question, right?

Right. Well, um the capability is there um to uh to grow and learn to to be capable um of significant improvement over the number of deaths that we are seeing in the United States right now. Right.

Yeah. So, I I do think that that if everyone was driving the most advanced car like a Whimo car, right?

If everyone was driving behind the wheel, um but letting that system drive um and was only there for, you know, occasional, hey, you know, something's going on, can you take over? That the number of deaths in the United States would wouldn't be 35, 000, right? It would be less. It would be significantly lower.

Yeah. Um, so I I do believe that they're better than humans with with the expensive uh systems that are out there. And and Wayne really isn't I don't think targeting trying to put all these expensive systems on every car. I think they're really targeting um the ride share market.

If you've seen, you know, they have rid share in San Francisco and and uh hundreds of of people are are using this ride share. Thousands, you know, people are using this ride share regularly. I just rode in one uh a couple weeks ago. Um and it's it's really interesting, you know, having no one in the driver's seat

um and using an app to pick you up and drop you off at a certain location and

um and it's it's pretty amazing um

you know what it can do. Yeah,

but they they can afford the significant number of sensors and the significant sensing elements um because it's a it's running 24 hours a day, you know, it's a

um ride share application.

It's not like somebody has to to buy

you can invest in the capital to get a return on it.

Right. Right. I see this the way I kind of put frame this is I say well look I don't think AI at the moment full self-driving is as good as my best F1 driver formula one driver

on on a good day in good conditions when he's just on point

for sure

versus

uh somebody who's maybe elderly uh impaired with alcohol late at night poor conditions. Mhm.

AI AI is definitely better than that, but it's not as good as that. Right.

And maybe that's the easy in. Right. And you kind of pointed to it, you know, that the ride share where I know I'm not really shouldn't be driving.

I don't feel that I'm capable of driving for whatever reason.

I can choose to take um a self-driving vehicle, right?

That is a safer option than me driving home. Yeah. That's

Yeah. Um and you know the typical um human driver has an an attention span and an emotional level up and down um and distractions.

There's so many different things that contribute Yeah. Contribute to accidents. Besides the chemical, you know, large percentage of of accidents are due to, you know, chemical related. Um, so the AI can can help those systems. There's there's a lot of stuff that doesn't get reported, right?

I I've I was driving on a a fivelane highway and there was a Tesla in the fast lane up ahead of me and lots of cars filling all the lanes. And this this car saw an opening and he was was weaving in and out of traffic and he saw an opening from the right side to get all the way across diagonally to the fast lane and it looked he was going to hit the Tesla who was in the fast lane already right

um and had an opening in front of it and so he's coming at a 45 degree angle behind the Tesla moves over to the side of the road on the shoulder and the car takes its place and then zooms ahead and passes everybody. Um, and I don't think the driver could have seen this car coming at him.

Uh, I think the Tesla uh, identified the car, got out of the way, and then came back into the lane after the car, you know, took off. So, and there's nobody that's going to record that and report it and, oh, this, you know, somebody's life got saved or it avoided a

multiple call pileup, you know, because of an accident. Um, that doesn't get reported anywhere, right? So those kind of things um are happening and

and avoiding accidents, but you're not going to hear about them.

Well, I guess maybe Tesla can can capture that stuff, can't it? With through through their cameras and

That's true. Yeah,

they can they could probably promote that

and promote that whole

idea.

Now, you were probably at CES recently. I was not see the convention center there and they've got this uh this Tesla loop Vegas loop they call it using Tesla cars. I was surprised to see that there's no self-driving. It's all

drivers and

it's reliability probably right. Yeah.

Well, I thought to myself, if they can't do it in this closed environment with the things going down tunnels, how are they going to do it out in the wide world? Well, I mean, I I have a Tesla and I turn it on to full self-driving um to go, you know, 200 miles on the freeway and for a freeway environment and even during, you know, a small town, um it does really well.

All right, so let's um let's explore some other areas. U we talked a little about driving with sensor fusion and so on. Do you see it sort of vision systems going? I mean, this is kind of where we're going with this conversation. Do you see them going into other industries? I mean, I'm working in construction for example. I see there's opportunities there. What are you seeing from your perspective?

Oh, absolutely. I mean, there's significant interest and already there's uh um autonomous driving dump trucks and uh you know uh large equipment. Um there's quite a few cameras on farming equipment, right?

Minimize and notice my video is kind of messing up, but there's there's cameras on farming equipment to reduce the amount of um chemicals that get sprayed on on crops, right? So, analyze, look for weeds and only spray if there's a weed, for example. Right.

I saw that a couple of years ago. We were at CES and I think it was

I think it was John Deere um had a booth and they were they these massive massive computers

and they were able to put the exactly the right amount of uh insecticide or pesticide or herbicide down against and around the the specific

uh in case we think well that's labor saving but it's also resource saving.

So it's extremely good at um metering out the particular insecticide or herbicide so that you really lower the amount that you need but also you don't have this you know this too much of it going out and getting and I guess infecting the soil. So

Exactly. Yeah. And the drainage and whatever extra doesn't get you know absorbed and processed by the plant doesn't get drained to the to rivers or to somewhere that that in the water supply.

Who would have thought that AI would have an environmental impact which is positive.

I know right. It's pretty cool.

It gets a lot of negative press right with all the data centers and the stuff that's going on.

Yeah. Yeah. Yeah.

What do you think about the data center uh impact?

Uh I think yeah I I believe there's no choice. You know if we don't do this we fall behind. Everybody is doing it. Uh I guess it's a necessary challenge. I'm going to say evil. Um I think it's going to definitely impact our energy usage. Um we have been on a trajectory since the industrial revolution or even before that of using more and more energy per capita and I don't see that changing. I think if you cast yourself into the future, the energy consumption of every single person is just going to continue to go up. So the challenge to me isn't uh to turn off data centers.

The challenge is to find ways to to to better create energy and use it more efficiently. Look, I'm not going to turn off my use of AI. I'm not going to stop, you know, jumping into ChatGPT and I'm going to be more and more demanding on it and everybody I speak to on this podcast and pretty much who's active in business or technology is using it more and more. So, we're all in there. Um, whether we like it or not, I don't know. What do you think, Chris?

Yeah. No, I mean the right now there's it's the integration is in servers and a lot of it's in tools that we use on daily basis. Um um and people are using it for code. Probably the biggest advantage is with coding.

Uh you know people that are writing software um they get significantly u more efficient by having AI tools help them write code. Um, but I've seen the application in all kinds of areas at work, right? And

whether it's from having uh, you know, you're trying to save multiple images from Outlook and they all the images have the same name because you transferred it from your phone. They all say image. Jpeg. You know,

pilot for example can generate a PowerShell that that'll download those images, change the name for each one so they don't override each other and just provide a solution. Um that that this is an example of the kind of things AI can help people do.

And there's there's all kinds of examples like this where it's it's acting as a tool,

but it's very uh directed, very specific uh requests.

Um

and it is like a calculator, right? It's like an advanced calculator that's helping you achieve what you want to achieve.

Um and um it's making people more efficient. Yeah, I think it's uh it is a tool. I I think I made a commitment this year to use it much more than I had been in the past and to to commit to several hours uh at least several hours a week.

And that's paying off because I find the more I use it, the more value I can extract from it, the more tools I turn on, the more ways I I think um I think in the past I would have probably spoke to more people about my ideas whereas now I tend to speak more to the uh the chat bots as a kind of way to bounce ideas back and forward. Um I'm starting to wonder if I'm too reliant on it, you know, or maybe I'm trusting it too much.

Um

just I've just noticed my workflows have shifted, you know, away from

away from what I used to do towards that. And it also builds up a history of you too. So it knows from your context. Oh, you actually were talking about this over here.

It's completely different chat and incorporating that in. They're starting to know me. And I think that's extremely powerful.

Yeah.

Right. And that's one of the things I that I'm interested in is um the fact that that this a lot of people that are developing this agent sort of in a way or this ChatGPT um um learning about themselves and what they ask and what they're interested in. Um that's a lot of it's on servers, right? And um

really I'm interested in uh the next wave of people having this capability on their own computer separate from a server.

So, you know, the Peter Cooper agent that learns and anticipates what you want and and uh and you teach it maybe your own ideas and your own direction and and nuance and that that becomes yours, not you know what somebody else owns,

right? Uh and and that's this is an area where people are just starting to realize hey you know I'm I'm helping develop uh this capability for the rest of the world or or for you know on a server but how do I develop that capability so that I can own it or if I want to you know move to to a different source or

y

add some capability that that that it's mine and

keep it private as well.

Yeah.

And also I guess you have the you have more scope to customize. So you know I'm kind of limited with ChatGPT for example by the amount of memory it will allow me to take.

I might say that I want you to take more memory. I'm prepared to pay for that and I want you to know everything about what's on my computers and everything I say

and uh as long as it's all private uh I want you to know me in depth and help me um

as a very very close partner. I think that I think you've got a really good idea there. I think that's got some real potential.

Yeah.

And then you know it's like you know you're developing your own children and uh guiding and uh as as systems get more capable um you know you're developing those layers of of constraint or context um I I just remember when uh as an example right that my sister adopted a um a young boy. And the, you know, there's when you meet a child, there's a certain history and background that that, you know, that you would expect from um one environment that someone's been through and then then you meet a child that's been through a different environment and and you notice a difference, right?

And so that that environment that that a child grows up in and what they learn, what they see, it significantly shapes who they are and how they process things and

right

you know the constraints around what they do and decisions they make and I think AI is the same way right I mean

what we how we build the layers and how we develop that context

becomes more and more important and um you know that that'll be more and more interesting as these systems get more powerful.

That's an interesting concept.

I haven't wrangled with that one, but it's like, you know, the twins, the nature versus nurture where twins are separated at birth and they come back and they analyze, you know, how different they were based on the different environments.

And uh I think the other thing is you're talking about boundaries. I think you mentioned that word boundaries. You can say, well, in every relationship, you can create boundaries, right? You know, I would I don't want you to call me outside of my work hours. You know, I want you to call me during work hours. That's that's my time when I'll I'll deal with that. That's a boundary.

We could do the same thing with AI, right?

We could say, well, I want you to answer my emails, but

only my professional emails, not my personal emails, or something like that. You could do a lot of creativity around

building that AI model as a person and then treating it as a person and then creating all the relationship boundaries and and the limitations and liberties that come with that. It's really interesting.

Yeah. Well, Chris, this is we've covered a lot of really interesting ground. Um, I want to switch back to the vision systems because that's where your expertise is. Where do you think this is going um in the future?

Well, just like we were talking about the, you know, the personal agents, the vision systems are adding these layers of uh um, you know, uh, casual reasoning, you know, where they're they're linking the scene that they have with uh, with prior history and with the physics of how fast is something moving. Is it where is it headed? Is it headed towards the car? And and you know, I I remember a statistic um looking at data in local in our county, for example, right? And um 71% of the car accidents that resulted in death were related to lane lane departure changes, right? And and it makes sense because if you're departing lane, you're probably either going to hit a tree or hit an oncoming car.

And and head headto-head collisions are probably really really dangerous. And uh

um this is one of the things that the AI systems are really good at, you know, because they can they know where the lanes are. They know when a car is, you know, coming at them, where it's headed. Um and so you know it's the the constraint of of uh where they can where they need to drive. This is something that that AI systems is really good at.

So I think there's definitely um improvement in the death rate right of the of humans in cars uh in the future as far as where they're headed. You know I think that's one thing is that they're going to save a lot of lives. So, and

yeah,

go ahead.

Well, we've we've seen from some of the OEMs that have analyzed the the death rates of of with cars that have their AI systems and cars that don't, right? And there's there's a definite improvement in in the percentage of death death rates.

Anyway, interesting. You can ask a question.

Yeah. You know, I'd be interesting to see that because I can see that somebody that's driving a a Tesla's probably got a few other safety things that, you know, so on and so on, whereas there's, you know, maybe there's other factors is what I'm saying that just relate to that. So, I'd be very interested to see whether it's causation or correlation.

That could be. That's true.

I I get,

but it would be good to dig into that. What I was going to give you one more sort of technical question. Do you think there's a future in um multisspectral or multi-sensing? We talked a little bit about LAR in the sort of in that sort of sense, but is there a future in not just a human vision system, but actually trying to see more of the spectrum um and then using that as part of

interesting system? You know, some of our customers um they they focused initially on um I want to analyze every single object.

Um and so they they focused on classification and color was really important to them,

reflectance and and they built this huge database of 10, 00ands, you know, 10, 000 plus objects. Um and and the one of the uh reasons that they gave was for example if a water bottle flies off a car, you know, a truck at you, do you slam on the brakes? Do you veer left and right? You know, is it going to kill somebody or is it just a water bottle? And uh

you know I I thought it was interesting and and a good first forier into what do you do with with all this information?

Um

those those systems they're they're expensive. It takes a lot of processing power to to get that information.

But for for in general that's where everybody started, right? And I think the object classification has because the people have spent the last five to 10 years focusing on that that is sort of a solved problem in a way right I think they have a they they're doing a really good job detecting what what is this object um the harder the next harder thing is you know the all the different decision making and the next layer of you know what do I do with this new scene that's happening And

um that's kind of where everyone's focusing now.

Yeah. So knowing the material multisspectral can give you some

understanding of the material that that it is.

So example with the bottle, it's a plastic bottle which is empty versus a brick or a stone or something piece of pipe which is metal and that could come right through the windshield.

Right. Yeah. That helps the decision making.

Well, this has been a really good conversation, Chris. We're going to have you back. But before we finish, we have to give you the final question, which is

okay.

AI, is it hype or help?

It's absolutely helping. Um,

and the we're we're in the infancy, right? So,

like most new technologies, there's a hype of uh people imaginations of like what the possibilities are.

Um, and those things will evolve, but it it's it's not going to happen next year, you know, or or in two years, right? Um,

but we're we're things that people are imagining

those were in Isaac Azimoff's robot books, right?

You know, 20, 30 years ago. Uh, you could read about all this stuff and and uh imagine what it could look like.

Um, you know, 10 years from now though, uh, as the the contextual uh, relationships and the processing power that you can put in the size of a box or

for 50 watts that doesn't drain an EV car battery, you know, that'll be 10 times more powerful and and that 10 times more power is going to be uh, adding this contextual interaction and and these the physics that you're talking about. Um and so it it you know that that's when sort of new new sort of uh capabilities are are going to surprise people um to the point where they they trust it and they they believe it's going to be safe, right?

I'm not sure everybody's there right now because of the corner case accidents and the things that happen. I'm not sure everybody trusts it other than Whimo which has significantly number of sensors and and the processing power to avoid accidents. Um you know the the trust is still developing. So yeah it's definitely a help.

Yeah definitely help. All right Chris well let's end on that. Thank you so much. As I said we we're gonna have to dig in again

to some more deeper topics but uh

bye for now. Thanks.

All right. Thank you. Bye.

Wow. What a great discussion I had with Chris. We touched on some amazing topics.

You might not know or you may that full self-driving is becoming a reality. So on the cars we're seeing them now and we understood a little bit of the challenges and how they've been overcome. Some cars have 12 cameras. They have very high dynamic ranges means they can literally see into the sunlight.

But also the processing power that's required has become you know affordable in terms of just the low wattage so he can run it off you know a car and so really fascinating this is all emerging it's going to happen very very soon in the future uh Chris also talked about some of his beliefs about where um the LLMs are going like ChatGPT and he thinks there's a new frontier coming where we have more private um um chats and private information and we talked about you know how that's going to evolve the boundaries around that and how useful that could be uh for us and it's literally setting up a relationship with your AI um agent on your own computer. Lots more stuff. Very interesting.

Look, if you want to know more about Skillion, please click the links in the description. If you like this, please subscribe, hit the like button. I hope that helps and thanks for watching.