Stephen Zhen Gou

Time to put markerless AI to the test📋! I ran a side-by-side mocap 🏃‍♂️ comparison: Move.ai's Core One plan vs. my Xsens Link suit. Here are the results. The Test: - Move.ai: Core One Advanced paid plan (S2 model), captured on a single iPhone 12 mini (4K). - Xsens: Animate Pro (HD processing), using my paid Indie Program account. - I focused on scenarios I use most often, comparing the raw animation data straight from both software packages. My Findings: The quality from a single-camera, markerless AI solution is genuinely impressive—especially with finger motion! It's a testament to how far this tech has come. I was impressed by the simplicity of the process—quick calibration and no special suit needed. Still, you can see some noise in the core motion that would require cleanup. Conclusion: I still trust my Xsens suit. The ability to record anywhere without worrying about camera line-of-sight is a massive advantage. I can jump in and go. The data is consistently clean, and the battery life is amazing. While both subscriptions are a comparable investment, there's no initial hardware cost with Move.ai, which is a big factor. For my needs, though, the reliability of inertial capture is tough to beat. Xsens Entertainment | Movella Move AI #motioncapture #gamedev MoCap Central

75

12 Comments

oLLM: Making 100K context LLMs run on an 8 GB GPU ., no magic, just brilliant engineering. Anuar Sharafudinov has built something truly remarkable. 
Check out the repo here https://lnkd.in/gFqVt8PM oLLM brings 80B models right to your desktop 
Yes, you can actually run Qwen3-80B on a $400 GPU! After diving deep into its architecture, here’s my take: While many frameworks like vLLM focus on ultra fast inference for enterprise-grade GPUs, oLLM takes a more bold and democratized approach balancing precision, scalability, and accessibility. What makes it stand out: * Enables 100K+ context lengths on modest 8 GB GPUs using SSD offloading and smart caching * Retains high precision by minimizing aggressive quantization * Integrates FlashAttention and chunked layer execution to squeeze maximum performance from limited hardware In short, oLLM is redefining what’s possible in LLM inference , making large model experimentation accessible to everyone, not just those with server farms.

22

3 Comments

Submitting my paper: “Beyond Scaling: A Dimensional Roadmap to AGI” to arXiv Need an endorsement for my first submission paper. Code: WDKYZN. Can anyone help or suggest endorsers? #AIResearch #AGI #arXiv Abstract: Scaling large language models has driven remarkable AI progress, yet it struggles to deliver human-like memory, reasoning, or identity. This paper introduces Dimensional Cognition, a framework that redefines intelligence as a product of qualitative advancements in cognitive architecture, not just computational scale. We propose a four dimensional model to guide Artificial General Intelligence (AGI) development: 1D Predictive Streams, processing linear data with limited continuity; 2D Modular Systems, breaking tasks into specialized subtasks but lacking unity; 3D Temporal and Emotional Continuity, adding persistent memory and simulated emotions for adaptive goals; and 4D Reflective Cognition, enabling self-awareness and autonomous evolution. Drawing on cognitive science, this model exposes the limits of current AI and offers testable benchmarks to measure progress. Our findings suggest that AGI requires layered, structural innovations modularity, continuity, and reflection beyond bigger models. This roadmap paves the way for artificial superintelligence, with significant implications for ethical design. By prioritizing cognitive depth, we can build systems that not only mimic human tasks but also adapt and reflect, ensuring responsible development for a transformative future. https://lnkd.in/eurDuCGz

3

I'm seeing so many people interpreting "Metaverse layoffs" plus the current "AI bubble binging" as some massive shift away from XR to AI. No. Apart from obvious branding shifts (like calling it "Physical AI" instead of "Real-world Metaverse" plus "Robots"), here's what's really going on: 1. XR and AI have been intertwined for 30+ years. ML signal processing is the biggest overlap -- machines need to understand human context and the world, including your POV. Even newer GenAI fills a need that we used to call "procedural content generation" -- the general shift there is away from ethics, unfortunately, to over-index on scaling fast. 2. AI glasses (e.g., cameras, audio, no display) are squarely XR and always have been. XR doesn't require any graphics -- it does require context (see #1). 3. Every smart company (I figure) has considered at least two potentially parallel paths: a) the 'minimalist' approach -- start with wearables and add the tech that fits; and z) the 'maximalist' -- start with a fully loaded computer on your face, N+ cameras, but make it smaller & cheaper over time. They converge later. The maximalist path was always important to invest in R&D and push tech forward. Let's just honestly call these 'dev kits' or better yet 'professional' and 'industrial' use-cases that are still quite valuable, but in smaller quantities. The minimalist path was always going to be the mass market appeal. Any execs who forecast hundreds of millions of units sold for the maximalist path OR poo-poohed the minimalist path as a 'waste of time' should probably be the ones checking their badges now. As we showed with Echo Frames around 2015+ the minimalist approach is the better path to grow fast, if (**and this is vital**) you can make it useful and usable for normal everyday tasks, not the creepy antisocial ones. I laid it all out 5-6 years ago, soon after I left Apple and could finally speak publicly (**not revealing any Apple-specific plans**). My overall timing and predictions were pretty close. I tend to over-index on immersion, where 'simpler' may win over 'more natural.' And I harp on ad-tech too, which, if you follow me, you know why... The biggest surprise to me was the early breakout of ChatGPT, whereas Google's AI scaling path was well known for years and they were being more cautious up to that point. Google is still poised to dominate on the AI side, and their combined XR effort was recently revitalized with strong partnerships. That said, Facebook was crazy for rebranding itself to a futuristic Metaverse story that we're still very far from making safe and usable for everyone, early adopters notwithstanding. I saw the renaming was happening early, under NDA, but I was still surprised by how immature it was. MZ's smartest decision in life was keeping a majority of votes, is all I can say. My articles are all up on Medium still. Here are two on these points: https://lnkd.in/gPZ9BTrE https://lnkd.in/gu76XXb7

134

32 Comments

📄 New Paper Worth Studying: V-JEPA 2 — Self-Supervised Video World Models for Physical Reasoning Meta AI just released V-JEPA 2, and it is one of the most serious attempts so far at learning physical understanding from raw video without labels, rewards, or pixel reconstruction. If you’re not familiar with the underlying idea, this work builds on Joint-Embedding Predictive Architectures (JEPA): 👉 https://lnkd.in/g_C7CPAV What makes V-JEPA 2 particularly potent: 🔹 Predicting representations, not pixels — Instead of reconstructing future frames, the model predicts latent embeddings of what will happen. This sidesteps nuisance details (texture, lighting) and pushes the system toward capturing causal physical state. 🔹 Scaling JEPA to long-horizon video — The paper extends JEPA into the temporal domain, showing that large-scale self-supervised video pretraining can yield strong physical reasoning ability. 🔹 Action-conditioned prediction — By incorporating actions into the predictive head, the framework moves closer to control-ready representations, bridging perception and embodied agents. 🔹 No generative modeling burden — Unlike classical world models, V-JEPA 2 avoids likelihoods and rollouts, focusing purely on learning abstractions that remain predictive over time. Why this matters: ➡️ It reframes “world models” as abstraction learners rather than simulators. ➡️ It suggests a path to scalable physical reasoning without explicit physics engines or object supervision. ➡️ It opens new research directions around turning predictive representations into planning-capable systems. If you work on video understanding, embodied AI, or representation learning, this paper is well worth a deep read. #MachineLearning #SelfSupervisedLearning #WorldModels #JEPA #ComputerVision #AIResearch

29

Before reacting to headline growth numbers, it’s important to understand what’s behind them. Here’s a key insight from Devina Mehra on looking beyond the surface and focusing on the real story. 🎧 Listen to the full XSpace recording: https://lnkd.in/djqrb2ws

7

As well as different models we are seeing more and more node based workflow tools to string them together. n8n, Grip tape, Fal, comfyUI, Houdini the pioneer! The power of automation through code and proceduralism has always fascinated me in my CG journey. From realizing I could leverage my software engineering roots to make things go faster, help teams be more productive, or more excitingly to make CG software like Maya or Houdini do things they can't out of the box. In this new era this ability is becoming more available to everyone and still I am sure the more technically minded or motivated will make it go even further. What used to be a competitive advantage for me is now being democratized. But the edge is always expanding. People ask me every day what to learn in the digital creative space in this crazy time. For me it's relatively similar to how it's always been although I believe even more in the fundamentals than ever. Learn drawing or photography because its fun and develops your artists eye more than any tool. Learn about storytelling because it's in everything we do to express in the art, as well as how we communicate with each other. Learn some coding(not just vibes), because it teaches problem solving which is probably the most important skill today. Learn some psychology because it helps in everything to do with team work and leadership. Learn some history because it always repeats itself and helps you avoid more mistakes. Learn some economics because it's a big invisible driving force that explains so much of how the world moves and is empowering even just with the basics. That might sound strange but I believe it all helps with personal decision making, which is one of the only things you can control in life. I keep working on those fundamentals while also learning the tools it takes to solve real problems. New software, new modalities such as proceduralism or AI. There will always be new things but the fundamentals usually remain the same. A bit every day builds to extraordinary things in a year. More thoughts like this in our newsletter https://hubs.li/Q03B9tk20 And in our inner circle group which has weekly training and LIVE sessions on AI and CG. And below is Fal one of the node workflows mentioned.

5

* Real-time, infinite video generation is here. Meet MotionStream. Most motion-controlled video models still do “render-and-wait.” MotionStream flips that: sub-second latency, up to 29 FPS on a single GPU, and interactive control (draw paths, drag objects, steer the camera) with constant cost per frame—so you can stream forever. Why it’s a big deal Teacher=>Student distillation: a strong bidirectional diffusion “teacher” gets distilled into a causal, streaming “student” via Self-Forcing + Distribution Matching Distillation. Translation: high quality, real-time speed. No more drift on long clips: clever sliding-window causal attention with attention sinks + rolling KV cache keeps quality stable on arbitrarily long videos. Precise yet natural motion: lightweight sinusoidal track embeddings + joint text+motion guidance for both trajectory fidelity and realistic dynamics. Numbers that pop ≤ ~0.4s latency, ~29 FPS @ 480p with an optimized tiny VAE; ~10 FPS @ 720p on one H100. SOTA motion following at speeds ~100× faster than prior motion-controlled pipelines. What you can do Live motion transfer from tracker outputs Drag-to-direct objects and characters Real-time camera control (single image => smooth 3D-like moves) - all as you create, not after. If you work in VFX, gaming, ads, or creative tooling, this turns video generation from a post-process into a live medium. 🚀 🔗 Paper: “MotionStream: Real-Time Video Generation with Interactive Motion Controls” (Nov 2025). #AI #VideoGeneration #DiffusionModels #VLM #ComputerVision #GenAI #Realtime #CreativeTech #Autoregressive #StreamingAI

8