🤖 Video Models - Geometry Recovery

🤖 Video Models - Geometry Recovery

This article discusses the evaluation of video models, suggesting that consistent geometry recovery should be prioritized over accurate physics predictions. The focus is on a fundamental aspect of model evaluation.

Key Points:

• Prioritize consistent geometry recovery in video model evaluation.

• Accurate physics prediction is a secondary evaluation criterion.

• Consistent geometry ensures the model's foundational understanding of the scene.

🔗 Resources:

• Alan Melling ↗ - Expert opinion on video models

Image

🤖 3D Object Tracking - Inverse Neural Rendering

This article summarizes a research paper on 3D object tracking without training data, using inverse neural rendering to fit a scene graph to an image.

Key Points:

• Recasts 3D tracking as an inverse neural rendering task.

• Fits a scene graph to an image for optimal explanation.

• Generalizes to unseen scenarios.

• Achieves 3D tracking without relying on training data.

🔗 Resources:

• Nature Machine Intelligence Paper ↗ - Research paper on 3D object tracking

Image

💡 Software Development - Vibe Coding Critique

This article critiques the concept of "vibe coding," arguing against its lack of structure and maintainability.

Key Points:

• Vibe coding lacks essential software development practices.

• Automation of the fun parts leaves tedious tasks for humans.

• Sustainable development requires standardized processes for all stages.

💡 Historical Narrative - State Control Concerns

This article expresses concern over state control of historical narratives, citing a specific instance involving the Smithsonian Institution.

Key Points:

• State control of historical narratives is detrimental.

• "Aligning messaging" with political agendas creates a "ministry of truth".

• This practice undermines the integrity of historical record.

🚀 Video Generation - OpenAI GPT-OSS 120B

This article demonstrates the use of OpenAI's GPT-OSS 120B to orchestrate a full video creation process using a single prompt.

Key Points:

• Generates a video from a single prompt.

• Uses open-source tools for image and video generation.

🚀 Implementation:

1. Use a single prompt to generate an image of a Labrador.
2. Employ Flux.1 Krea Dev and LTX Fast to generate a simple video.
3. The entire process is orchestrated by gpt-oss.

🔗 Resources:

• Flux.1 Krea Dev ↗ - Image generation tool
• LTX Fast ↗ - Video generation tool

Image

🚀 Image and Video Generation - Minimal Agent Orchestration

This article describes a system using a minimal agent to orchestrate image generation and subsequent animation, emphasizing its open-source nature and accessibility.

Key Points:

• Image generation followed by animation orchestrated by a minimal agent.

• Uses a full open-source stack.

• Accessible even with limited GPU resources.

🔗 Resources:

• Hugging Face Spaces ↗ - Provides GPU resources.

💡 Content Creation - Iterative Approach

This article advocates for an iterative and fast-paced approach to content creation, similar to software development.

Key Points:

• Frequent content releases enable rapid iteration and learning.

• Embracing potential embarrassment is crucial for progress.

• Direct association of creator with content increases risk aversion.

✨ PhD Defense - Successful Completion

This article announces the successful PhD defense of Tobias Kirschstein under the supervision of Prof. Justus Thies.

Key Points:

• Successful completion of PhD studies.

• Deep gratitude for the supervisor's support and trust.

🔗 Resources:

Image

💡 Model Evaluation - Real-World Application Focus

This article argues against relying solely on benchmarks for evaluating AI models, advocating for real-world applications.

Key Points:

• Benchmarks are inadequate for comprehensive model evaluation.

• Real-world tasks offer a more truthful assessment.

• Evaluation should move beyond pre-defined human notions.

🚀 Delta Program - Updates and Voting

This article provides updates on the Delta program, highlighting an upcoming workshop and a call to vote for its finalists.

Key Points:

• Upcoming workshop on game building.

• Finalist status in the Delta program.

• Call to action to increase kid's access to the program.

🚀 Implementation:

1. Participate in the upcoming workshop.
2. Vote for the Delta program finalists.
3. Support the program to help more kids.

🔗 Resources:

Image

⭐️ Support

If you liked reading this report, please star ⭐️ this repository and follow me on Github ↗, 𝕏 (previously known as Twitter) ↗ to help others discover these resources and regular updates.