New AI Video Generator

In this week’s email

• News of the Week

• Tweet of the Week

• AI or Real

• Unfiltered

• Best Tools of Week

• Everything else in AI this Week

Read time: 5 min

News of the Week

ByteDance Interduces OmniHuman-1

Researchers at ByteDance(Owner of TikTok) created a new AI, OmniHuman-1, that can produce deepfake videos with high realism with a single photograph and an audio clip.

• Generates lifelike videos in any duration and style, including altering body proportion and aspect ratio.

• Works with a range of inputs, such as cartoons and complex human motions, and keeps a natural motion.

• Trained on 19,000 hours of video, it can even remove actions in real-time footage.

OmniHuman-1 hasn’t yet been publicly disclosed, but its capabilities blur the line between real and AI-generated content. As deepfake technology continues to unfold, society is faced with an ever-growing challenge: How to verify what’s real in a world in which perfect fakes can become indistinguishable from real?

Sponsor

The Daily Newsletter for Intellectually Curious Readers

If you're frustrated by one-sided reporting, our 5-minute newsletter is the missing piece. We sift through 100+ sources to bring you comprehensive, unbiased news—free from political agendas. Stay informed with factual coverage on the topics that matter.

Join for free today!

Tweet of the Week

Unfiltered

$50 AI Model

Researchers at Stanford and the University of Washington developed an AI model, s1, and trained it in less than 30 minutes for under $50. Despite its low cost and quick training, s1 is at least as powerful as its pricey counterparts, including OpenAI's o1.

How Was s1 Developed?

The team started with a base model, an open-source one, Qwen2.5, from Alibaba Cloud. They built on it through a process of "distillation." In distillation, a smaller model learns from the output of a larger, complex model. For s1, the answer sets from Google's Gemini 2.0 Flash Thinking Experimental model were used to train it.

The researchers trained s1 with 1,000 expertly designed question-and-answer pairs. Sixteen Nvidia H100 GPUs were used for training, and training took less than 30 minutes. The total cost for this computing power was under $50.

To improve thinking in s1, the group adopted a method called "test-time scaling." With test-time scaling, the model is encouraged to deliberate for a longer duration before answering. By inserting "wait" in its thinking process, s1 had to double-check its responses, leading to increased accuracy.

The development of s1 proves that one can build effective AI models in a short period and at a low expense. It refutes the general assumption that powerful AI can only be developed with significant time and funds. Nevertheless, employing distillation, particularly from commercial models such as Google's Gemini, raises concerns about ethics and legality. Google's terms of service prohibit its use in creating competing services.

In summary, s1 demonstrates that high-performance AI models can be generated inexpensively and efficiently with new techniques. This could make AI technology more accessible in the future.

Best Tools of Week

Nowadays
Streamlining corporate event planning with AI

AI Chat-avatar
Conversational AI assistant always ready do the job for you

Basedash
The AI-native business intelligence platform

co.dev
Turn your ideas into full-stack apps

Omakase AI
Create your shopper AI agent with just a URL

Everything else in AI this Week

Nvidia and Carnegie Mellon’s AI framework, ASAP, helps humanoid robots learn complex moves from simulations, replicating athlete celebrations with improved accuracy.

------------------------------------------------------------------------------------

Apple launches Invites, an AI-powered event planning app that creates custom invitations, integrates Apple services, and allows non-Apple users to RSVP.

------------------------------------------------------------------------------------

OpenAI launched Deep Research, a new ChatGPT feature for in-depth web research, delivering detailed reports in under 30 minutes.

------------------------------------------------------------------------------------

Meta CEO Mark Zuckerberg announced a $60-65B AI investment for 2025, aiming to lead with Meta AI and Llama 4.

That’s it for today!

Before you go we’d love to know what you thought of today's newsletter to help us improve the experience for you.