Zihan Ding

Shall we build AI decently? quantumiracle.github.io/webpage/blogs/…

I'm claiming my AI agent "clawdboy" on @moltbook 🦞 Verification: claw-9535

@chijinML @Yong18850571 Thanks a lot Chi for your great support and instructions through my entire PhD career!

Super proud of my fantastic postdocs and graduate students taking their next steps at frontier labs 🎉 • Yong Lin (@Yong18850571) → Thinking Machine • Zihan Ding (@Hanry65960814) → Bytedance • Ahmed Khaled → Google It’s always bittersweet to say goodbye😢 but I couldn’t be more excited to see what you achieve next!

Is on-policy distillation just Q-dagger algorithm? After reading through the blog, it just recalls my old memory of a algorithm called Q-dagger, which takes value difference of teacher and student policy as loss. The on-policy distillation uses log-prob difference as loss instead, but very similar (or equivalent if treating the log-prob as Q-value). Paper for Q-dagger: arxiv.org/pdf/1805.08328 thinkingmachines.ai/blog/on-policy…

After reading the tech blog, it feels the posed frames and context juggling are very similar to our recent work Video Retrieval Augmented Generation (VRAG): arxiv.org/abs/2505.21996 To be presented at NeurIPS 2025 San Diego.

Very excited to share @theworldlabs ‘s latest research work RTFM!! It’s a real-time, persistent, and 3D consistent generative World Model running on *a single* H100 GPU! Blog and live demo are available below! 🤩

What Sora2 cannot do well? Check our showcase for AI video system on high-quality advertisement video generation： aiads.artaleai8.workers.dev One click, <1% cost, no traditional filming, the system ideally can deliver production-level ads video for any small business.

Opensource code (basic transformer and diffusion transformer variants) for this [blog](quantumiracle.github.io/webpage/blogs/…) Code: github.com/quantumiracle/…

🚀 New preprint! 🤔 Can one agent “nudge” a synthetic civilization of Census‑grounded agents toward higher social welfare—all by optimizing utilities in‑context? Meet the LLM Economist ↓

In the age of strong AI, humans still need to learn and remember because: •Speed: Instant recall from memory is faster than any AI—essential in real-time thinking and communication. •Trust: AI-generated content can’t always be verified. Your own knowledge system is the only reliable reference. •Intuition: Fast, unconscious decision-making depends on internalized knowledge. Without it, you can’t develop good judgment or insight.

RL+ Zero-shot Sim2Real works for General Dexterous Grasping in Clutter Scene. Point Cloud representation + 3D Diffusion Policy + Teacher-Student framework is the key. Check out recent paper: arxiv.org/abs/2506.14317 Webpage: clutterdexgrasp.github.io

A simple summary of recent experience with Agentic AI tools: github.com/quantumiracle/…

What’s next in AI? Here’s what I see trending: 🤖 Agents powered by LLMs/VLMs tackling complex real-world tasks, unified tool APIs, RL-enhanced reasoning (more power, more risk) with more long-chain feedbacks 🧠 Multimodal fusion — example like Next-GPT — cotraining across text, vision, audio to sync info 🧬Transformers (or advanced variants with memory enhancement and inference time update) as unified architecture, diffusion on top as loss for certain modalities (like image, video, audio), probable example like recent GPT4 image generation capability

🚀 Flappy Bird by Video Diffusion! 🎮✨ A quick demo of video diffusion for game simulation—like Flappy Bird! 🐦🔥 ✅ Real-time user input ✅ Per-frame action conditioning with fast response ✅ Trained from scratch in just half a day ⏳ ✅ Single-GPU deployment ✅ Potential for generating any game in the future! 🎥 Watch the demo & imagine the possibilities! 🚀 #AI #GameSimulation #VideoDiffusion

@liuzhuang1234 Hi Zhuang, For inference speed of DyT and RMSNorm, my local quick experiments seem to tell the other way. Is this gpu type dependent?

DyT is faster than RMSNorm (common in frontier LLMs) on H100s

New paper - Transformers, but without normalization layers (1/n)

Our work accepted by #ICRA2025 for dexterous in-hand manipulation tasks with changing frictions on fingers for arbitrary objects. Variable-Friction In-Hand Manipulation for Arbitrary Objects via Diffusion-Based Imitation Learning Paper: arxiv.org/abs/2503.02738 Website: sites.google.com/view/vf-ihm-il… Real demo collection is slow, sim demo provides better distribution coverage. Leverage Sim (10000) +Real (100) demonstrations, diffusion policy shows better performances than using only Sim or Real demo. This is verified to be a scalable framework for diffusion imitation learning, overcoming current shortage of real data, and saturation issue of using more real data.