🇨🇳 vs. 🇺🇸 - New language model beats OpenAI o1?

AI-HOY AInauts,

Welcome to the latest issue of your favorite newsletter. The AI world is on fire this week!

Here are two really cool updates to try out:

• 🇨🇳 vs. 🇺🇸 - New language model beats OpenAI o1?

• 💥 AI video game changer: Consistent characters

• 🤔 OpenAI Stargate Project?

Here we go!

🇨🇳 vs. 🇺🇸 – New language model beats OpenAI o1?

We don't really care where AI innovations come from.

But since our social feeds are full of political stuff about the Trump campaign, tariffs etc. these days, let's use the whole thing as a hook.

The AI network community is once again celebrating a new language model from China, which is on a par with or even better than OpenAI o1 in many benchmarks.

We are of course talking about DeepSeek, which is creating hype with its new R1 reasoning model.

The politically exciting thing is that China, which is otherwise so closed, is now leading the way in the field of open source models - and the open USA is doing everything it can to remain closed and shielded with OpenAI.

Why is it so exciting?

R1 is an open source reasoning model. Similar to OpenAI's o1, it thinks about a given task in several steps in order to arrive at the best solution.

There are already several versions of the model. Even the smaller ones with 1.5B parameters are very powerful and can even be run locally on your computer thanks to open source.

It is also interesting to see how R1 was trained. It was not simply more data that was used (as is often the case with others), but reinforcement learning.

In other words, the model is trained until it arrives at the correct solution and thus becomes better.

This is comparable to AlphaGo. Here, the machine has also learned to play the game of Go better than humans by playing against itself. And not by analyzing more games between humans.

What we find most exciting, however, is the price: DeepSeek-R1 costs only a fraction of OpenAI o1 and is, believe it or not, around 95% cheaper… 🤯

You can try it out here:

The easiest way is with DeepSeek directly. However, we still prefer to use OpenRouter, sometimes in conjunction with LobeChat (activate in the settings under AI Service Provider).

What did our tests reveal?

It has to be said that DeepSeek R1 is really good in most of our tests.

If you've been reading us for a while, then you know our analysis tests, such as our classic leads analysis.

It didn't work the first time, but on request:

Here, OpenAI o1 is still the only model that gets such analyzes right first time and does not let itself be dissuaded.

Interestingly, Grok 2 also seems to have improved and now also gets the numbers right first time. ... (and we can expect Grok 3 soon, so it remains exciting!)

💥 AI video game changer: Consistent characters

We have a great update on AI video that will please a lot of AInauts... especially those who play around with AI Influencers or create videos with the same people all the time will celebrate this update.

But first, let's start at the beginning:

The problem with AI videos is that you can create cool videos with a text prompt. But the AI decides anew each time, for example, what the people in the videos look like...

For example, you can't create a video of exactly YOU based on just a text prompt.

So far, there have been two hacks here:

1) Create video from image

You simply generate a static image of the scene and then enter it into one of the video models as the first frame.

This brings the image to life, so to speak. But the videos are usually not as good and realistic because the AI has a harder time bringing existing scenes to life instead of creating new ones.

The second problem is: often the person only stays consistent for the first two seconds and then morphs into someone else.

2) Create a video with a similar character and then use FaceSwap

A second option is to create a video with a detailed description of the desired character and then swap the face using tools such as FaceFusion.

The problem here is that it's a bit time-consuming, takes longer and the face fits - but the rest of the body is often not 100%..

Consistent characters in videos with just one click

But now to the solution: Hailuo AI has launched an ingenious new feature for its Minimax video model.

It's called Subject Reference.

This means: you just upload a picture and create a video based on a prompt - and Hailuo makes sure that your person is in the video.

Maybe you've already seen our AI influencer AIlexa. If not, here she is:

You can now use any of these photos as a reference for a video.

Important: Only the person is used, not the surrounding picture. You describe these details in your prompt.

So here is a video of AIlexa on a walk through Tokyo:

How it works

The best thing about it: it's super easy to implement.

We recommend Replicate for this (as so often), you can find the model here.

You simply enter a prompt, upload your reference image, and off you go!

And for everyone in our AI + Automation Accelerator Community: thanks to Replicate, the whole thing is also easy to use in automations and can be conveniently controlled via API.

Spoilers for Hailuo: Multiple characters in one video are rather difficult 😉 …

🤔 OpenAI Stargate project?

The newsletter has already become extensive again, so we'll finish with a brief piece of "good-to-know" news.

OpenAI and partners have just pulled a project/company called "Stargate" out of the hat. It's a $500 billion monster to massively expand the US AI infrastructure. For comparison, that's twice as much as the Apollo program cost (adjusted for inflation). And yes, this is the biggest private AI investment ever.

OpenAI makes the AI magic, SoftBank takes care of the financing, NVIDIA supplies the chips and Oracle builds the systems.

It starts immediately in Texas, with a $100 billion investment, because everything-is-bigger-in-Texas is not just a motto there, but a way of life.

And we owe it all to Trump, of course. Just kidding - the project has been in the pipeline for a long time. But the new president took the opportunity to present it to the media as part of his agenda - video here...

500 billion in four years sounds like a lot. According to Elon, it is still unclear exactly where the money will come from - which was immediately denied by Sam 😉.

Let’s get the popcorn ready 🍿 …

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 - that motivates us to keep going 😁!
Twitter, 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.