👨‍🚀 Consistent AI videos with these prompts

Hello AInauts,

Welcome to the latest issue of your favorite newsletter! A bit OpenAI-heavy today, but we also have some best practices for AI video in store:

• 🤔 Does OpenAI reads and saves everything?

• 🎥 Consistent AI videos with Google Veo-3

• 🤑 OpenAI launched o3-Pro and o3 will be significantly cheaper

Here we go!

🤔 OpenAI reads and saves everything?

Today we start with a somewhat more philosophical topic that is currently occupying us.

OpenAI is fighting back against having to save all chats.

The background: A US court is forcing OpenAI to save ALL conversations - even if you delete them manually!

Why? Because the New York Times is suing for copyright infringement and wants to prevent "evidence" from being deleted. The consequences of this now affect hundreds of millions of ChatGPT users!

Somehow, we realize that OpenAI is "reading along". However, it doesn't feel good when all chats are saved…

Important note: Enterprise & API customers are exempt from the court order. Anyone using ChatGPT in an enterprise context or via the API should therefore be in the clear.

But there's also a positive side effect of OpenAI analyzing conversations!

OpenAI has just published its latest Threat Intelligence Report. And it reads like a spy novel...

What has happened in the last three months:

• 10 different malicious cyber operations have been detected and stopped.

• The attackers come from China, Russia, Iran, Cambodia and the Philippines.

• They range from fake job schemes and disinformation campaigns to malware development.

• A Russian hacker even used ChatGPT for Windows malware debugging 😲!

The crazy thing is that these threat actors are becoming increasingly professional. One Chinese network used ChatGPT to generate hundreds of comments for a single TikTok post to simulate fake engagement. Another network from Cambodia sent text messages to random UK numbers and promised £360-500 a day for liking social media posts.

In such cases, it seems like a good thing that OpenAI is having an eye on it and intervening to prevent worse.

Our take: This is the acid test for AI trust!

On the one hand, surveillance protects us to some extent from cyber criminals.

On the other hand, there are already countless open source models that cannot be monitored in this form and are certainly being used for similar neferious activities.

Sam Altman is now proposing an "AI Privilege" - something similar to doctor-patient confidentiality.

Honestly, the man has a point. If AI is our digital therapist, consultant and creative partner, maybe even a doctor, it should be treated as a confidential relationship ...

If you're looking for alternatives to ChatGPT that work locally, in your own cloud or at least via API, we wrote about a solution some time ago.

🎥 Consistent AI videos with Google Veo-3

Google Veo-3 is currently probably the best AI video model. It has finally been made available via API last week and can be used in tools such as Krea and Replicate.

Last week, the Replicate servers were overloaded, and it took ages to get a video. Now it is much quicker, which is why we're also writing about it here. You can try it out quickly on Replicate without a subscription.

But beware: AI videos are expensive. An 8-second video currently costs around 5 bucks.

How to create consistent, longer videos

The big problem with AI videos is that the clips are usually only 5–10 seconds long.

If you want to make longer videos, you have to create several videos and stitch them together.

The issue with this is that it is very difficult to keep characters and scenes consistent across several videos.

First of all: We have not yet found a final solution - but we have found two ways that you can try out to get more consistent results.

1) Use a seed

With Replicate, for example, you have the option of using the seed ID of a video to create the next one.

But as you can see from the picture above in comparison with the GIF, the scenes are similar, but neither the studio nor the presenter are exactly the same.

But still, they are very similar ...

2) Detailed, consistent prompts

A second method that currently works quite well for us is to use extremely detailed prompts with exact descriptions, and to keep these descriptions consistent across the different scenes.

It is quite time-consuming, but fortunately, the AI helps us with this. Here is an example prompt that turns various scenes into detailed and consistent prompts for you:

{Enter your list of scenes here}

Above, you can see the list of scenes we need for my AI-generated video. We're using Google Veo 3, which supports both video and audio output.

BUT: Veo 3 has some clear limitations:

Each video can be a maximum of 8 seconds long (so you have to prompt accordingly).

Each scene is generated in isolation—that is, Veo 3 has no idea what the last scene looked like, what the characters looked like, or what happened in the previous clip.

That's why you have to describe every single detail EXACTLY. You can't leave anything to chance. Describe appearance, clothing, surroundings, mood, camera angle – everything. Repeat these descriptions from scene to scene to maintain style and continuity.

Your task: Write a series of prompts that I can give to Veo 3 to generate an 8-second clip each. I will edit the clips together later to create a consistent longer video.

IMPORTANT:
- No ambiguity.
- No open interpretations.
- Each prompt must be completely self-contained.

Procedure:
First, think about the implementation for at least 25 paragraphs in the <thinking> tag.
Then output the finished prompts in numbered order in the <prompts> tag.

Prompt tips from Google

If you haven't experimented with Google Veo-3 or AI Video yet, here are a few important prompting terms to use, straight from Google:

• Image composition: "close-up", "two-shot", "over-the-shoulder shot", ...

• Lens & focus: "macro lens", "shallow depth of field", "wide-angle lens", ...

• Genre & style: "Science fiction", "Romantic comedy", "Action movie", ...

• Camera movement: "Zoom", "Dolly shot (tracking shot)", "Tracking shot (tracking)", "Pan", ...

Example prompt:

Close-up (composition) of melting icicles (subject) on a frozen rock face (context) with cool blue tones (mood), zoomed in (camera movement), preserving the details of the dripping water droplets (action).

🤑 OpenAI launches o3-Pro and o3 becomes significantly cheaper

Finally, a short piece of news that adds further confusion to OpenAI's model madness: OpenAI has just launched its smartest reasoning model yet: o3-Pro!

We can already see it in the OpenAI Playground, and it will soon be rolled out in ChatGPT. As always, the model is even smarter, even better and feels like it has already overtaken all living humans.

It "thinks longer" than other models and delivers extremely reliable results.

BUT: Before you start chatting with o3-Pro, be warned.

It's slow. It's incredibly slow. And expensive. Ben Hylak from Raindrop warns: A simple "Hi" can cost $80 and take 20 minutes. 😲

But there is also good news for the regular o3 model - it will be almost 80% cheaper, from $10/$40 to $2/$8 per million tokens (87% cheaper than the old o1-pro!). The o3-Pro is priced at $20/$80.

When and how should you use o3-Pro?

As already mentioned: the new model is expensive and slow. But infinitely smart, making it a good choice for your most complex challenges. It's also great that it can handle all tools and MCP connections.

But never waste short prompts on the o3-Pro model! Instead, go really in-depth and pack an infinite amount of context into your prompt. Then the result will be really good!

We find the following comparison of a prompt anatomy for o3-Pro very good. The difference between o1 and o3-Pro is that it can handle much more context and has tool access. Have fun trying it out!

We made it! But no need to be sad. The AInauts will be back soon, with new stuff for you.

Reto & Fabian from the AInauts

P.S.: Follow us on social media - it motivates us to keep going 😁!
X, LinkedIn, Facebook, Insta, YouTube, TikTok

Your feedback is essential for us. We read EVERY comment and feedback, just respond to this email. Tell us what was (not) good and what is interesting for YOU.