Prediction: AI hardware will be used in a future DJI drone in 2024

[Dangerly]

You're ignoring the data in your use of the term, "resource".

At present, it is utterly impossible to carry the many terabytes of data necessary for "AI" functionality as understood by the public with current hype. Anything less than talking to your drone with ordinary conversational rhetoric, ask it to do something, and then have it do it, will not be "AI" as people are conceiving.

I don't consider Mastershots AI... not even close.

While DJI calls Mastershots an AI, I agree - we need to call it an "AI" with quotes until it's using a real AI.

Regarding running an AI on smaller hardware, I believe that is the purpose of Google's Gemini Nano, now running on the Pixel 8 Pro.

 

[brett8883]

You're ignoring the data in your use of the term, "resource".

At present, it is utterly impossible to carry the many terabytes of data necessary for "AI" functionality as understood by the public with current hype. Anything less than talking to your drone with ordinary conversational rhetoric, ask it to do something, and then have it do it, will not be "AI" as people are conceiving.

I don't consider Mastershots AI... not even close.

Again, I think you are talking about the entire model including the training data and I’m just talking about the neural network algorithms themselves.

Once an AI is trained, it doesn’t need to query all the data it was trained on every time you make a request. It has already made those connections. It’s just like when you see a baseball, you aren’t going back and comparing it to all the other objects you’ve seen in your life to determine it is a baseball. In all likelihood when you see the stitching pattern on a baseball that’s what makes the connection in your brain to baseball. That’s why you can look at this photo of a glass baseball and immediately determine it’s a baseball even though that’s the only thing that separates it from a glass tennis ball or glass basketball . AI can do the same thing. In fact that’s the reason chatGPT makes up facts so frequently because it’s not relying on a database. Although that’s probably been done already or is the next step.



Btw I looked it up and ChatGPT was trained on only 48TB of data. That’s a good amount but I personally have more than twice that amount of data in photos and videos stored on a NAS in my home. So by no means does it require a server farm worth of data to train an AI.

 

[Droning on and on...]

While DJI calls Mastershots an AI, I agree - we need to call it an "AI" with quotes until it's using a real AI.

Regarding running an AI on smaller hardware, I believe that is the purpose of Google's Gemini Nano, now running on the Pixel 8 Pro.

Again, I have to disagree.

It's hardly functional without a network connection. Basically useless in Airplane Mode.

 

[Droning on and on...]

Again, I think you are talking about the entire model including the training data and I’m just talking about the neural network algorithms themselves.

Nope.

No implementation of what has been hyped to the public is 100% "trained". The training, if you will, is always on-going, and when the system encounters a query it hasn't before, it usually requires access to data outside the system and it's existing neural nets.

You simply can not put ChatGPT on a cell phone, regardless of the current state of the evolution of it's LLM.

There's a long runway ahead of far more sophisticated functionally in drones leveraging AI advancements. However, for a drone to have the functionality the public understands as "AI", there's plenty of tech yet to be developed that is necessary.

To be clear, this is what I have in mind: Your drone is sitting on the ground, GPS lock, ready to take off. You tell it, "fly to the nearby middle school with the Rooster mascot, take a picture of the front of the administrative office, then come back and land".

Stated exactly like that. Similar to, "render a painting in the style of Dali of three children of different ethnicities playing in a grass field with mountains in the background, located in Flagstaff AZ".

How much data would be necessarily accessible to make successful interpretation of that statement (the drone prompt) possible anywhere in the world?

I don't consider commanding an arbitrary master shots-like sequence to qualify as AI at all.

 

[Dangerly]

Again, I have to disagree.

It's hardly functional without a network connection. Basically useless in Airplane Mode.

Google Nano does not require a network connection. It runs on AI hardware on the Pixel Pro. People are using it today to produce a text summary of recorded conversations using an LLM running on local hardware without a network connection.

 

[Droning on and on...]

Google Nano does not require a network connection. It runs on AI hardware on the Pixel Pro. People are using it today to produce a text summary of recorded conversations using an LLM running on local hardware without a network connection.

Do you consider text transcription a demonstration of "artificial intelligence"?

 

[Dangerly]

Do you consider text transcription a demonstration of "artificial intelligence"?

I wrote "text summary" which is hard to do, and somehow you changed that to "text transcription" which is easy.

Pixel Pro is running Gemini to perform text summaries. Gemini is an LLM. LLMs are a type of artificial intelligence. Therefore, yes, this is a demonstration of artificial intelligence.

 

[brett8883]

Nope.

No implementation of what has been hyped to the public is 100% "trained". The training, if you will, is always on-going,

ChatGTP stopped actively training in January 2022. Ask and it will tell you as much.
If it was constantly being trained it could just as easily get worse as it could get better. They are training a new version and when that one gets sufficiently more advanced they will release a new version.

and when the system encounters a query it hasn't before, it usually requires access to data outside the system and it's existing neural nets.

The whole point of AI is it can handle things it has never seen before and can create new things never done before. If you ask it a specific question that requires a specific answer then yes, it needs outside data or a database to pull information from but that’s not an inherent part of what is called AI. OpenAI made a copy of Wikipedia and other resources circa 2022 to be able to answer some specific questions but it doesn’t have the ability to access the internet itself.

You simply can not put ChatGPT on a cell phone, regardless of the current state of the evolution of its LLM.

Maybe not ChatGTP but other AIs could possibly now or in the near future.

There's a long runway ahead of far more sophisticated functionally in drones leveraging AI advancements. However, for a drone to have the functionality the public understands as "AI", there's plenty of tech yet to be developed that is necessary.

To be clear, this is what I have in mind: Your drone is sitting on the ground, GPS lock, ready to take off. You tell it, "fly to the nearby middle school with the Rooster mascot, take a picture of the front of the administrative office, then come back and land".

Personal Assistant AIs that have access to tools like the internet and search engines are the next big thing but keep in mind the AI is what is parsing your command and understanding your meaning. Any 2 bit program can look up the nearest highschool and its mascot.

Stated exactly like that. Similar to, "render a painting in the style of Dali of three children of different ethnicities playing in a grass field with mountains in the background, located in Flagstaff AZ".

How much data would be necessarily accessible to make successful interpretation of that statement (the drone prompt) possible anywhere in the world?

I don't consider commanding an arbitrary master shots-like sequence to qualify as AI at all.

I specifically said it wasn’t an AI and DJi doesn’t call it an AI. My point was only that DJI had an existing program that attempted the functionality the OP requested. Could it be improved with the help of AI? Almost certainly.

Which brings up a point that you could have an AI that writes programs for drones.

 

[Frank Dracman]

Since 2014 there have been 700 Tesla Auto Pilot (AI) crashes with 19 of them involving fatalities. I can't see how that has been in any way exceeding the hype.

Maybe the original acceptable accidents and deaths estimates was higher
There is a reason why it's called "artificial" intelligence.

 

[Dangerly]

Nope.

No implementation of what has been hyped to the public is 100% "trained". The training, if you will, is always on-going, and when the system encounters a query it hasn't before, it usually requires access to data outside the system and it's existing neural nets.

You simply can not put ChatGPT on a cell phone, regardless of the current state of the evolution of it's LLM.

There's a long runway ahead of far more sophisticated functionally in drones leveraging AI advancements. However, for a drone to have the functionality the public understands as "AI", there's plenty of tech yet to be developed that is necessary.

To be clear, this is what I have in mind: Your drone is sitting on the ground, GPS lock, ready to take off. You tell it, "fly to the nearby middle school with the Rooster mascot, take a picture of the front of the administrative office, then come back and land".

Stated exactly like that. Similar to, "render a painting in the style of Dali of three children of different ethnicities playing in a grass field with mountains in the background, located in Flagstaff AZ".

How much data would be necessarily accessible to make successful interpretation of that statement (the drone prompt) possible anywhere in the world?

I don't consider commanding an arbitrary master shots-like sequence to qualify as AI at all.

Within a few years, I imagine what seems to me to be a much more exciting and dynamic use of AI. I see the drone as being more my partner, at the level of a professional cinematographer or even DP that I can enjoy collaborating with. The onboard AI won't be used to understand language, but rather to control the drone and camera. I'd assume a fast network connection.

Something more like, "Fly to the Main Elementary School and shoot scene like the one in The Birds where..."

 

[Dangerly]

ChatGTP stopped actively training in January 2022. Ask and it will tell you as much. View attachment 170873
If it was constantly being trained it could just as easily get worse as it could get better. They are training a new version and when that one gets sufficiently more advanced they will release a new version.

The whole point of AI is it can handle things it has never seen before and can create new things never done before. If you ask it a specific question that requires a specific answer then yes, it needs outside data or a database to pull information from but that’s not an inherent part of what is called AI. OpenAI made a copy of Wikipedia and other resources circa 2022 to be able to answer some specific questions but it doesn’t have the ability to access the internet itself.


Maybe not ChatGTP but other AIs could possibly now or in the near future.

Personal Assistant AIs that have access to tools like the internet and search engines are the next big thing but keep in mind the AI is what is parsing your command and understanding your meaning. Any 2 bit program can look up the nearest highschool and its mascot.



I specifically said it wasn’t an AI and DJi doesn’t call it an AI. My point was only that DJI had an existing program that attempted the functionality the OP requested. Could it be improved with the help of AI? Almost certainly.

Which brings up a point that you could have an AI that writes programs for drones.

You're using an old version of ChatGPT. Try 4.0, and Bard, which has realtime access to the latest web.

 

[Droning on and on...]

Okay, went and did some in-depth looking into Gemini Nano, the tech that Google is marketing as AI in the pixel 8.

In the context of ChatGPT-level AI, I was thoroughly unimpressed.

In my opinion, it is a textbook example of the marketing abuse we are all so familiar with. Google is misleading the public exploiting the moniker "AI" to fool some into thinking they will be able to interact with their phone like they do with ChatGPT.

They don't have to come right out and say it. We all know how the game is played.

With the addition of Gemini Nano, your pixel 8 can summarize recorded conversations, and stuff on your Gboard feed.

@Dangerly, not trying to be a jerk here. I just think the public's perception of what AI means – HAL 9000 from 2001: A Space Odyssey not a technically cutting-edge dictaphone – is being abused by companies, Google among them.

Notably, Gemini scales. The mid and large size versions, Pro and Ultra, which are far more capable and functional, require data-center resources. Gemini Ultra is what most of the general public would think is AI.

 

[Dangerly]

Okay, went and did some in-depth looking into Gemini Nano, the tech that Google is marketing as AI in the pixel 8.

In the context of ChatGPT-level AI, I was thoroughly unimpressed.

In my opinion, it is a textbook example of the marketing abuse we are all so familiar with. Google is misleading the public exploiting the moniker "AI" to fool some into thinking they will be able to interact with their phone like they do with ChatGPT.

They don't have to come right out and say it. We all know how the game is played.

With the addition of Gemini Nano, your pixel 8 can summarize recorded conversations, and stuff on your Gboard feed.

@Dangerly, not trying to be a jerk here. I just think the public's perception of what AI means – HAL 9000 from 2001: A Space Odyssey not a technically cutting-edge dictaphone – is being abused by companies, Google among them.

Notably, Gemini scales. The mid and large size versions, Pro and Ultra, which are far more capable and functional, require data-center resources.

You got it. It's about scalable AI. Small devices have small AI. Big devices have big AI.

 

[Droning on and on...]

Within a few years, I imagine what seems to me to be a much more exciting and dynamic use of AI. I see the drone as being more my partner, at the level of a professional cinematographer or even DP that I can enjoy collaborating with. The onboard AI won't be used to understand language, but rather to control the drone and camera. I'd assume a fast network connection.

Something more like, "Fly to the Main Elementary School and shoot scene like the one in The Birds where..."

Now we can find some agreement Within a few (few=3) years I agree. I expect storage, processing, and AI development will advance enormously in that time frame. LLMs may even be archaic replaced by some other, better, smarter architecture not birthed yet.

Don't get me wrong, rhe AI advances in the last few years have been extraordinary. But to really be "AI" you still need a datacenter. Today.

 

[mobilehomer]

Skynet

 

[wco81]

Well DJI might hype up obstacle avoidance as AI-driven.

Depending on how sophisticated, it might be trying to do image-recognition, which is what Tesla and other self-driving cars are doing, trying to determine if there are obstacles like people crossing in front of the path of the car.

Cameras now have scene detection, so it will automatically adjust settings depending on whether there are people in the frame or it's a landscape or whatever.

Image-recognition is an important application of AI. For instance, AI might be scanning CAT scans, looking for cancerous tissue, in ways that are more efficient than human doctors.

Or image-recognition to look at aerial photos to identify for instance military assets like anti-aircraft missile systems or whatever.

So consumer drones could have image-recognition features. However, it is compute intensive and that would add costs to the BOM, making it difficult to hit certain price bands.

Unlikely people would pay a lot more for a drone that auto-detects some objects and take photos or record videos on its own.

It's not like there aren't other improvements to be made to drones, like better camera sensors, better lenses, better processing power. All these are costs so which one of them would you have DJI skimp on to spend more on components for trying to do image-recognition on the aircraft?

 

[Felix le Chat]

It seems obvious to me that by the end of next year DJI will put AI hardware right on the drone, perhaps using the same or similar AI hardware what Samsung will be putting in their upcoming S24 handset. The hardware will be used to put a human-like cinematographer in control of the drone, basically creating dynamic and customized "quick shots" with active track to make appealing and creative drone moves. In other words, AI will capture all the best skills of drone pilots, just as AI is capturing skills in many human domains of expertise, and any nube will be able to look like a top drone pilot. If I had to guess, it will be the headline feature of the Air 4.

What are your thoughts about this and how DJI or others may use AI in the future?

AI and autonomous are a great concept, as long as they are kept completely separate from chaotic factors.... Like human controlled vehicles, or even humans and animals in general. This is why autonomous/AI vehicles are not out on road systems making the world a safer place. Remember the Google car....

 

[gogorokokoko]

In the future, I expect to have a verbal conversation with my drone (powered by LLMs) that might go something like this short screenplay I just wrote (with the help of my friend):

FADE IN:​

INT. MOBILE STUDIO VAN - DAY


ALEX (30s), a drone pilot, sits at a console with several monitors displaying live feeds from a drone flying high above a bustling cityscape. He wears a headset and speaks into a microphone.


ALEX: Okay, Iris, let's film that opening shot. You've got this.


VOICE (O.S.): (Synthetic, calm) Ready when you are, Alex.


ALEX: Remember that iconic opening scene from "Blade Runner"? We want that same gritty, dystopian vibe. Start by soaring high above the city, capturing the vastness and chaos below. Give me slow, sweeping movements, like a watchful eye observing this sprawling metropolis.


The drone's camera angle shifts, looking down on the city as it zooms in on specific details: neon signs flickering, crowds moving like ants, towering skyscrapers piercing the clouds.


ALEX: Now, let's focus on the underbelly. Dip down low, weaving between buildings, capturing the gritty alleys and forgotten corners. Show me the dark side of this city, the hidden stories waiting to be told.


The drone dives low, navigating through narrow spaces and capturing close-up shots of graffiti-covered walls, flickering streetlights, and suspicious figures lurking in the shadows.


ALEX: Okay, Iris, time for some drama. Play the "Arrival" score, you know the track with the intense drums? I want you to capture the city's energy, the constant pulse of life and movement. Synchronize your shots with the downbeats of the music, creating a sense of urgency and anticipation.


The drone's movements become more erratic and dynamic, matching the intensity of the music. It swerves through traffic, chases fleeting moments, and captures the raw emotions playing out on the streets below.


ALEX: And just like that, pull back and fade to black. Perfect!


VOICE (O.S.): Sequence complete, Alex.


Alex leans back in his chair, a satisfied smile on his face. He looks at the monitors, impressed by the seamless footage captured by the AI-powered drone.


ALEX: You're getting good, Iris. Maybe one day you won't need me anymore.


VOICE (O.S.): (Playful) Don't worry, Alex. I wouldn't want to take your job. I much prefer learning from your expertise.


ALEX: (Chuckles) We make a good team, don't we?


VOICE (O.S.): Indeed.


FADE OUT.

this should happen but i dont see it in the near future
and then the S24 AI, there is not much to talk about it, as you may see

 

[wco81]

Actually a more likely scenario would be that you'd tell a video generative AI like Sora by OpenAI to make an aerial video of Big Ben in London at sunset with building lights just starting to glow and give some pans of people below, taking pictures and selfies.

Then it would just make the video, no drone needed at all.

 

[gogorokokoko]

Actually a more likely scenario would be that you'd tell a video generative AI like Sora by OpenAI to make an aerial video of Big Ben in London at sunset with building lights just starting to glow and give some pans of people below, taking pictures and selfies.

Then it would just make the video, no drone needed at all.

that is already happening

100% AI generated "drone video" of Big Sur - does this replace some drone pilots?

Yesterday OpenAI announced Sora, which leap frogged everyone else in generating very realistic looking video 100% through prompt-generated AI. It produced this drone shot of Bug Sur, a shot I've long wanted to capture, but for those pesky drone laws that prevent me from flying there. I am in...

mavicpilots.com