|
Curious Reader!
Welcome to this week’s ChatGPT Curious companion newsletter.
What you came for is below, and you can CLICK HERE to listen to the episode if you decide you prefer earbuds to eyeballs.
Happy reading!
In this episode of ChatGPT Curious, we break down OpenAI’s newest release: Sora 2, the text-to-video model now living in its own app, Sora. From how diffusion models work to what the new invite-only platform looks like, we’re covering the tech, the ethics, and the implications. You’ll learn how Sora 2 generates videos, why it’s a big deal (for better and worse), and what it means for the future of AI-generated media.
What Is Sora 2?
Sora 2 is OpenAI’s text-to-video diffusion model. You type in a prompt, and it generates a fully animated, photorealistic video clip…with sound!
- As of now, clips are limited to 10 seconds, but that’ll likely change soon…because everything changes every day with these models.
- It accepts both text and image inputs, and can interpret camera motion cues and environmental details.
- This is one of those rare times when prompting skill actually matters, and small details can dramatically shape the results.
How Diffusion Works (Without Getting Lost in the Math)
Sora 2 is a text-to-video diffusion model. WTF does diffusion mean?
When you hear the word “diffusion,” think physics, aka the way particles spread out over time, moving from areas of high concentration to low concentration.
Example: a drop of dye dispersing in water.
Here’s what that means regarding AI video generation:
- Training Phase:
- The model takes a clear image or video and gradually adds noise, turning it into visual static. Think about taking an image on tv and diffusing it by gradually turning it into static or “snow”.
- The first part of training is turning clear images into snow (adding noise), and the second part is learning how to reverse that process to create a clear image from snow (removing noise).
- Worth noting, it’s not the image itself being diffused, but the data that represents it.
- Generation Phase:
- Video generation does not start from a blank canvas.
- Think of it more as starting with a canvas that has everything on it, but it’s scattered and random and chaotic, and the video is generated by organizing the chaos (denoising).
Analogy time:
Remember that scene in Willy Wonka where Mike Teavee (the cowboy kid) gets turned into a cloud of colorful dots and teleported through the air into a TV?
- The training phase is Mike getting scattered into dots and the computer learning the logic of how all the dots used to fit together..
- The generation phase is reassembling Mike from those dots.
That’s diffusion: Starting with randomness and teaching the model to re-create order.
From Sora1 to Sora 2
This new model is named Sora 2, which means there was a Sora 1 that perhaps you didn’t even know about. It lived inside ChatGPT’s sidebar and was rolled out in December 2024.
- It produced videos that you could 100% tell were AI.
- I’ll admit, I never used it. I’m not a fan of AI video content, and I didn’t have much use for it.
Sora 2 was rolled out on October 1, 2025, and is an entirely different beast.
- It’s no longer part of ChatGPT, and instead is housed within a standalone app called Sora.
- Why is it called “Sora”? I literally have no idea. I did some searching and all I found was that “sora” means sky in Japanese, which might hint at big picture thinking or wide-open possibilities, but there was nothing concrete about what that name was chosen.
Inside the Sora App
Right now, Sora is invite-only, which means that someone who has access has to send you an invite code. I was invited by my guy Forrest (no idea where he got his invite from, I don’t ask questions), so I do have an account.
- Each user gets four invites. I sent one to Lex, which means I have three left. Reply to this email if you want one. First come first served.
The Interface:
- Inside the app it looks almost exactly like TikTok, an endless feed of vertical videos. Except this time ALL the videos are AI generated, which has earned it the nickname, “SlopTok”.
- To generate a video, users simply input a prompt (just like with ChatGPT), literally anything they want to make a video of, and hit enter
- It accepts both text and image inputs, and can interpret camera motion cues and environmental details.
- Videos are currently limited to 10 seconds
- The captions on each video vary, but some account users share the prompts they used in the captions, which is actually pretty interesting, and highlights my earlier point about the importance of prompt quality.
- The app also has a feature called ‘cameo’, which lets users insert themselves (or anyone else who’s opted in) directly into videos. (Scary shit if I’m being honest.)
Big Issue:
- Sora’s opt-out policy is backward. Instead of asking for permission to use your data, it assumes consent by default.
- You can toggle off “Improve model for everyone” under Data Controls to stop training use, but there’s no clear way to stop content generation use. At least not that I could find.
- All in all, the data policy feels sketchy and I anticipate a lot of lawsuits coming down the pipe.
A Benefit to the Standalone App
Having Sora separated from ChatGPT might actually be a good thing (if we must have it at all) as it clearly distinguishes AI videos from real videos. If it’s on Sora, you know it’s AI.
Worth noting, you are able to download the videos and then upload them to other social media platforms. Folks are already doing this, and although the videos have Sora watermarks, many people have no idea what Sora is, and are responding to the videos as if they’re real.
Concerns About Sora 2
Real talk, I don’t think any of this is a good thing. I mean, yes, it’s incredible that we have the intelligence and technology to be able to do such things, but using it to make a video of a horse skateboarding? Not really sure
A few other things to think about:
1. Compute Cost
- Video generation is incredibly energy-intensive.
- The usual argument that “AI uses way less energy than streaming” doesn’t hold here — this is streaming plus creation.
- Even a 10-second video takes significant computational power, and we’re just getting started.
2. Deepfake Potential
- Sora 2 blurs the line between real and fake like never before and has basically given anyone with a phone the ability to create a deepfake.
- We’re talking about hyper-realistic content — think Martin Luther King Jr. talking about video games or Michael Jackson stealing fried chicken.
- Even if someone proves that a video is fake later, once it’s been seen, the damage has been done.
3. The Speed of Progress
- The leap in quality has been staggering.
- These clips went from “clearly AI” to “Wait, that’s not real?” in less than a year. Who knows what things will look like in another year.
Quick Summary
- Sora 2 is OpenAI’s newest text-to-video generator.
- It lives in its own TikTok-style app called Sora.
- Currently invite-only.
- Creates SUPER realistic videos, with sound, and works via diffusion.
- I foresee many lawsuits and more bad things than good, but here we are.
Bottom line: I’m not telling you to go use it. I’m curious, you’re curious, and I just want to make sure you know what’s going on in the world of AI.
How I Used ChatGPT This Week
Each episode I include a section where I briefly discuss how I, or someone I know, has used ChatGPT that day/week.
This week I used ChatGPT in two different ways that I’d like to share:
1. To Generate a Prompt for Sora
Yes, I made a video. I wanted to see what it could do. No, it’s not a video of me. I made a simple beach scene and utilized a prompt that I asked ChatGPT to create for me. You can click the link below to check out the video.
2. To Repair Our Wine Fridge
Here’s how it went down:
- Our wine fridge started glitching out of the blue.
- Lex called a repair guy. He showed up (2 hours late), fiddled with it for about an hour, then quoted us $500 to fix it.
- I told her to give me a shot first.
- With ChatGPT’s help, I learned how to use a multimeter, diagnosed the problem, and ordered a new power board (which it found for me).
- The power board arrived, I swapped it, and the wine fridge now runs good as new!
ChatGPT didn’t do the fixing, but it helped me diagnose the problem, which is, IMO, one of the best ways to use it. Once you know what’s wrong, you can cross-reference with YouTube, Reddit, and your own logic.
Important caveat: Always push back when using ChatGPT and don’t just take things at face value.
Alrighty, that’s it for today’s episode. As always, endlessly grateful for you and your curiosity.
Questions, comments, concerns, additions, subtractions, requests? Hit reply or head to the website (chatgptcurious.com) and use that contact form. I’d love to hear from you.
Catch you next Thursday.
Maestro out.
AI Disclaimer: In the spirit of transparency (if only we could get that from these tech companies), this email was generated with a very solid alley-oop from ChatGPT. I write super detailed outlines for every podcast episode (proof here), and then use ChatGPT to turn those into succinct, readable recaps that I lightly edit to produce these Curious Companions. Could I “write” it all by hand? Sure. Do I want to? Absolutely not. So instead, I let the robot do the work, so I can focus on the stuff that I actually enjoy doing and you get the content delivered to your digital doorstep, no AirPods required. High fives all around.
Did someone forward you this email?
Click here to join the ChatGPT Curious newsletter family.
Stay curious.
|
