Wenbo Ji
Master’s student and researcher at TUM, working on video diffusion, world models, and robotic manipulation.
I write technical notes on research and engineering, with occasional reflections on work and life.
-
When Execution Gets Cheap, What Still Compounds?
What remains scarce when AI makes execution cheap—and why wealth should include ownership, dignity, and the freedom to learn, care, belong, and choose.
-
Human Motion as a Control Interface for Interactive World Models
How 2D pose, SMPL bodies, and tracked head-and-hand motion become causal guidance for responsive video world models rather than offline animation controls.
-
Reasoning Effort Is an Inference-Time Systems Knob
Reasoning effort changes hidden token use, latency, cost, context pressure, request residency, and serving capacity across the LLM inference stack.
-
From Human Video to Robot Action: Five Interfaces Across the Embodiment Gap
Five interfaces that turn human video into robot supervision: latent actions, interaction tokens, digital twins, contact topology, and robotized demonstrations.
-
AI Infrastructure Is Not Plumbing
Why AI infrastructure determines whether experiments are trustworthy, inference is economical, and AI products remain reliable under real workloads.
-
Streaming Autoregressive Video Generation
A research review of streaming autoregressive video generation: causal distillation, self-forced rollouts, visual memory, few-step sampling, and world models.
-
Embodied AI World Models
A practical map of embodied world models covering prediction targets, action grounding, robot data, uncertainty, planning, and closed-loop deployment.
-
A Token's Journey Across the LLM Stack
A unified diagnostic model connecting token computation, GPU kernels, KV-cache memory, request scheduling, latency, and throughput across LLM inference.
-
The Life of a Request Inside vLLM
How vLLM-style serving engines coordinate prefill, decode, continuous batching, paged KV cache, scheduling, speculative decoding, and streaming.
-
Why Transformer Performance Is Mostly a Matmul Problem
How Transformer linear layers become tiled GPU matrix multiplications shaped by memory movement, tensor-core utilization, batch shape, and arithmetic intensity.
-
The Life of a Token Inside a Transformer
How token IDs become contextual representations through embeddings, normalization, attention, MLPs, positional encoding, KV cache, and logits.
-
LLM Inference Is a Full-Stack Systems Problem
A full-stack map of LLM inference across Transformer computation, GPU kernels, KV-cache memory, request scheduling, and streamed token delivery.
-
From π0 to π0.7: A Tutorial on Open-pi and Robot Foundation Models
How Physical Intelligence's open-pi models turn language and vision into continuous robot actions, then add context, memory, feedback, and steering.
-
Work Is Important Enough Not to Become Everything
A reflection on The Good Enough Job and why meaningful work should develop craft and responsibility without becoming the sole measure of a life.
-
The Age of Auto Research, and Other Certainties
Where AI scientist systems automate bounded experimental search, where their verification loops fail, and why human framing and judgment remain consequential.
-
On 3D, Video World Models, and the Approaching ImageNet Moment of Perception-Action Learning
How 3D structure and video world models could support an ImageNet-scale shift from isolated vision tasks toward causal perception-action learning.
-
Humans Are Imperfect, but Worthy of Love
A reading of The Stories of Ibis about machine love, human imperfection, and why coexistence may depend on acceptance without complete understanding.