The hosts expressed concern about Apple's seemingly cautious approach to AI, with Addy noting, "If you are one step behind today, you're ten steps behind tomorrow." They discussed how Apple might be opening itself to disruption if it doesn't innovate more aggressively in AI, particularly as wearable technology advances.
Joey pointed out that Apple has historically never been first to market with new technologies, preferring to refine existing concepts. However, both hosts agreed that in the rapidly evolving AI space, this strategy might be riskier than usual.
4D Gaussian Splats: Dynamic Scene Capture Takes a Major Step Forward
The conversation shifted to an exciting development in computational photography and scene capture: the arrival of 4D Gaussian Splats. Addy provided a clear explanation of this technology, which represents an important advance in how dynamic scenes can be captured and reproduced.
"Gaussian splats have been around for a couple of years. It's basically AI-powered photogrammetry," Addy explained. "It's somewhere in the middle between machine vision, traditional photogrammetry, and AI inference."
Key points about 4D Gaussian Splats:
What They Are: Gaussian Splats are point clouds with color information and gradients that can represent 3D objects without traditional 3D rendering
The 4D Breakthrough: The addition of time as the fourth dimension allows for capturing dynamic scenes rather than just static environments
Technical Approach: The system references past and present frames to optimize the point cloud solution, similar to how video compression works
Current Requirements: Approximately 20 cameras are needed to capture enough data, making this still primarily a professional tool
Processing Demands: Content must be processed after capture rather than in real-time, currently limiting some applications
The hosts emphasized the potential applications in filmmaking and virtual production, noting that this technology could eliminate the need to build complex digital twins of real-world environments. As Addy explained, "You don't have to build complex real-world things if you can just capture it well enough."
They clarified an important misconception: this technology does not extract 3D information from a single 2D video. Multiple camera angles are required to generate the data needed for the 4D Gaussian Splat reconstruction.
Innovative AI Research: Video-Based Memory and Faster Image Generation
The final segment covered two interesting AI research papers with potential implications for how AI systems store and process information.
Storing AI Memory in MP4 Files
The hosts discussed a novel approach called Memvid that uses video files to store AI memory and knowledge bases.
"Unlike traditional vector databases that consume massive amounts of RAM and storage, Memvid compresses your knowledge base into compact video files while maintaining instant access to any piece of information," Joey explained.
This approach represents creative thinking about data storage for AI systems:
The Concept: Storing text information within video file formats rather than traditional databases
Potential Benefits: More efficient memory usage and potentially expanded knowledge windows for large language models
Technical Hypothesis: Addy suggested this might create a "pseudo RAM with the video" that allows for more efficient memory management
Research Applications: Could help make large AI models more manageable for researchers without requiring massive computing resources
"This is such outside-the-box thinking," Addy noted, suggesting this kind of innovative approach is what companies like Apple should be pursuing.
Contrastive Flow Matching for Image Generation
The final topic covered a technical advancement in image generation called Contrastive Flow Matching that could make training image generation models substantially more efficient.
The paper claims to achieve:
Addy explained that traditional image generation involves many steps of adding noise and then removing it, with all processes funneling through one neural network. The contrastive approach keeps these processes separate, leading to more efficient training.
The hosts noted that while smaller, more efficient models are valuable, particularly for mobile devices, they typically require large models first. "You can't get to these more efficient routes or these smaller models without having the big models first," Joey pointed out.
Conclusion: Practical Impacts for Creators and Technologists
This episode of Denoised highlighted several developments that media and entertainment professionals should monitor. Apple's design-focused updates and cautious AI approach suggest the company is taking a measured path while competitors make bolder moves. Meanwhile, 4D Gaussian Splats offer a glimpse of how dynamic scene capture could transform production workflows, potentially eliminating the need for complex digital twins in many scenarios.
The AI research papers demonstrate how quickly the field is evolving, with novel approaches to data storage and model training that could make powerful AI tools more accessible and efficient. For creators and technologists working in media production, these developments represent both new capabilities and challenges as the industry continues to transform.
The views and opinions expressed in this podcast are the personal views of the hosts and do not necessarily reflect the views or positions of their respective employers or organizations. This show is independently produced by VP Land without the use of any outside company resources, confidential information, or affiliations.
