Hey there,

Let's talk gen AI

Oh great, another post about AI. Yes, I know, I know. This post won't be so much about maps, but about creating content online, teaching and trying to convey excitement about a topic as well as getting things right.

If you've been online over the past few days, you've probably heard about OpenAI's Sora. If not, the TL;DR is that OpenAI just created a text to video model, with results being mind blowing. Here's an example:

Taking a step back first

I want to write about this because I'm overall unsure what I feel about it. I don't know if I have any new thoughts on this topic, but it kept me up at night last night, so here we go

I first got into satellite imagery a few years ago when deep learning computer vision first starting blowing up and saw an opportunity back around 2017 as some were starting to apply it to Earth Observation images, and it felt like a really exciting field.
Fast forward a few years, in 2021 in the middle of the pandemic I start Minds Behind Maps wanting to have long thoughtful conversations with people building the modern maps of our world. I started learning about the industry but also filming, editing, animating, telling better stories. And so last year I wanted to combine all of these skills into making videos for a wider audience, explaining what satellite images & modern maps are today.

For the past year I've been putting a lot of my effort towards producing podcast interviews and videos explaining satellite images and mapping. 6 months ago I published a video explaining how satellite images work at a high level, and as I've teased before I'm working on a new similar video coming... soon.

I've shared some behind the scenes of how these are made on Patreon. I put a lot of effort into making sure as much as possible as the things I show on screen reflect real data, that I use the real models & correct archival footage for the satellites I show

Here's an example, the title sequence for the satellite image intro video looks like this:

The satellite you see fly by, that shows up all across the video is a Landsat satellite (not sure which one exactly, doesn't matter as much, I wanted to get an Earth Observation satellite at least). Here's what this looks like in Blender by the way:

Making these videos has been interesting experience for me, it makes 2 worlds I want to be in come together:

• The data scientist in me wants things to be accurate. My job is to make sure that when a pixel has a specific value, I know how confident I am in those values. Things need to be accurate and we need to be sure how accurate they are. I've worked on monitoring floods, and now focus on measuring tree height during my day job.
• Visual Effects is about making things look nice. Telling story, specifically in video isn't just about making things accurate and true, it's also about conveying an emotion and a feeling. As I've learned more about how others make their movies, videos & documentaries, I've realised how much thought goes into making things look good rather than always being 100% truthful. Sometimes it's switching up the order an interview rendered to the order it was spoken to make it more impactful, adding visual effects to enhance the feeling conveyed. On that last point, this video series on No CGI is just invisible CGI is great food for thoughts.

I've had to think about these two things together, I want to tell more stories about what can be done with satellites & maps but I also want to get people to feel something. How do I bring those two things together?

One aspect where most people don't realise things are completely made up for example are sound effects. If sound effects are done well, most people never think about them. Putting together sound effects is about finding a sound that feels good matching what's on screen. This is a great video on the process.

Okay, so what does this have to do about AI?

I feel like generative AI makes this problem so much more complicated. I want to be in the business of explaining physical phenomena with some of the coolest sensors humans have ever built. I think getting things as correct and factual as possible is important. And I'm not sure where factual-but-entertaining stuff is going.

I think this OpenAI just killed all the stock imagery & footage, and I think that's (mostly) fine. A lot of stock footage is used to represent some vague notion of a specific concept. You're talking about Japan? Slap a stock footage video of Mount Fuji. You're talking about how awesome your company is? Slap a smiling group of people in suits putting their thumbs up.

But that approach couldn't get you make cinematic movies, or explain how the Saturn V rocket works. Now, you can type "Internal function of the Saturn V upper stage" and Sora will give you something, and it might look really convincing. It is true? ¯\_(ツ)_/¯

This is ChatGPT all over agin, except for video now. Writing is something most humans on the planet know how to do. Video isn't.

Don't get me wrong, I'm really excited about tinkering with gen AI to find new ideas:

Maxime Lenormand @MaxLenormand My ChatGPT use over the past 3 months has been: - 5% image generation for mood board & quick idea generation - 10% text based idea generation - 85% making little bash scripts that automate so many tiny things 6:41 PM • Jan 18, 2024 0 Retweets 11 Likes Read 1 replies

​

But I just don't know where this is going when making well thought-out, researched and accurate pieces of media (Journalism is in a bit of a crisis at the moment).

When I want to make a video showing specific satellite images, I know where to get specific imagery, of a specific sensor and a specific date, so I know what I'm showing. I know that because it's my job to find data I want. But how do I show that to the viewer? How do I tell them "hey, you can probably trust this a bit more, because look at all this other work I've done and I actually know what I'm talking about"?

Reputation I guess

I think there's a difference between wanting a quick answer to a question and going on an emotional roller coaster.

I don't care how I get the answer to what the current population of Sri Lanka is, I just want to know the answer.

Back to my Saturn V question, I don't want people like Smarter Every Day, Everyday Astronaut or Techniques Spatiales to go away. I want more of people like them to make amazing, thoughtful stuff. I want us to lower the barrier to entry so more people can contribute. But I don't want that to happen at the expense of false information and especially not convincingly wrong stuff.

Right now a lot of the generative AI tools we've seen have helped on point number 2 I mentioned earlier: making things look good. I'm worried that this will take out the people working in point 1, making sure things are accurate enough.

My guess is that reputation is going to become more and more important; that if Destin from Smarter Every Day shows something on screen, I'm way more likely to believe it, because I trust he put in the work.

What I don't know if it there will be an environment online to sustain these types of work in the face of such drastic simplicity to make things that look good.

Excitement

I also am excited about the potential of tinkering around with tech like this. I think most technology ends up creating new tools. That means new ways of working. I'm excited to see people try out new ideas and make bigger projects that what they may otherwise have been able to before, be it for video and movies but also video games, web apps, you name it.

I ask all my podcast guests how they would describe themselves. I usually describe myself leaning towards techno-optimism and wanting to empower people by teaching, spreading knowledge and wanting to lower barriers of entry. I hope gen AI can help empower creative people, but I'm not sure if it won't just completely cut them out.

Making an AI-resilient Internet

I'm equal parts confused, low key freaked out, and excited at the same time.

I haven't even touched on the copyright aspect of all this, which I would like to see figured out, but I have no idea what that would look like. The engineer in me is fascinated; the one that wants to lower the barrier of entry to creative work is thrilled; the part that wants people to make sustainable careers making stuff to educate and entertain people is really confused and frankly quite scared.

This is the part of the post where I regularly nudge you towards my Patreon. I'll do it again this time, mostly because as I said when I had Jed Sundwall on the podcast, the future is more uncertain than ever for how people putting things online will get paid:

So if you do want to help me making informed, well resourced and as-much-as-I-can factually correct stuff, please consider supporting my work on Patreon, and finding out how to support all the people who's worked you've valued.

Related conversations

• I just interviewed Gilberto Camara on the podcast. He was the Director of INPE, Brazil's National Institute for Space Research from 2005 to 2012. It's a long conversation but if you want to make impact with satellite images, I think this is a worthy conversation. More related to this conversation, there's a moment where he talks about Sentinel compared to Planet imagery. Even if Planet data were free, Gilberto would still use Sentinel for deforestation mapping.
• ​I also had Thomas Ager on to talk about radar images, SAR, and how totally counter intuitive they are. I'm a bit biased but I think this is also totally worth listening, Tom is probably one of the best speakers on SAR. Towards the end of the conversation I ask Tom what he thinks about machine learning for SAR. I'll leave you to listen to that section

I'm curious to see how this post will age in say 5 years time. Time will tell, but in the meantime thanks for reading and as always, let me know what you think!

​
Cheers,
Maxime

​Read previous issues of the newsletter​