Wenbo Ji
Master’s student at TUM and research intern at Agile Robots SE. I write technical notes on video generation, ML systems, and embodied AI, with occasional reflections on work and life.
-
Same Skeleton, Different Boundaries: How Open-Source AI Inference Frameworks Are Built
A source-level comparison of vLLM, SGLang, TensorRT-LLM, LMDeploy, vLLM-Omni, SGLang-Omni, verl, slime, OpenRLHF and AReaL at November 2026 releases.
-
When the Harness Becomes a Platform
Skills, hooks, plugins, MCP, ACP, and SDKs extend a coding agent in different ways. Learn what each one carries and when a task actually needs it.
Harness Engineering: How Coding Agents Turn Decisions into Work · Part 6
-
More Agents, More State to Own
When subagents help a coding agent: split research, use worktrees for separate edits, track budgets and cancellation, then test the final combined patch.
Harness Engineering: How Coding Agents Turn Decisions into Work · Part 5
-
Who Gets to Run This Command?
How a coding agent decides which commands may run: instructions, permission rules, approval prompts, and OS sandboxes, shown with one blocked package install.
Harness Engineering: How Coding Agents Turn Decisions into Work · Part 4
-
A Transcript Is Not the Context Window
How coding agents choose what the model sees: code search, compaction summaries, AGENTS.md or CLAUDE.md files, and memory checked against current files.
Harness Engineering: How Coding Agents Turn Decisions into Work · Part 3
-
A Patch Is a Protocol
Why AI coding agents' edits fail, and how an edit tool should handle duplicate matches, patch errors, partial writes, and the regression test that follows.
Harness Engineering: How Coding Agents Turn Decisions into Work · Part 2
-
From Model Response to Agent Execution
How a coding agent's loop runs tool calls, handles cut-off responses, errors, and cancellation, and decides when a task is done, traced through one bug fix.
Harness Engineering: How Coding Agents Turn Decisions into Work · Part 1
-
Robot-Use Agents: When Frontier Models Leave the Screen
Robot-use agents use language models, skills, physical prompts and verified trials together, with constraints from contact, latency and hardware protection.
-
When Execution Gets Cheap, What Still Compounds?
What wealth can mean as AI makes execution easier: how ownership and bargaining power shape gains, and why learning, dignity, and real choice still matter.
-
Human Motion as a Control Interface for Interactive World Models
This article compares 2D pose, SMPL bodies and tracked head-and-hand motion in interactive video world models. It examines limits from timing, viewpoint and scene persistence.
-
Reasoning Effort Is an Inference-Time Systems Knob
Reasoning effort can change internal work, latency, cost and cache occupancy. Compare tasks before you select low, medium or high.
The Life of a Token Across the LLM Stack · Part 5
-
From Human Video to Robot Action: Five Interfaces Across the Embodiment Gap
Compare five methods for learning robot actions from human video. They include latent actions and contact-aware retargeting, with different tasks and evaluations.
-
AI Infrastructure Is Not Plumbing
AI infrastructure can change numerical results, inference cost and reliability. Examine kernels, memory, scheduling and research checks together.
-
Streaming Autoregressive Video Generation
This article examines causal video rollouts, teacher-student mismatch, persistent visual memory and online control with a limited sampling budget.
-
Embodied AI World Models
Robot world models connect prediction and action. This guide examines representations, planning, policy learning, aligned data, uncertainty and deployment.
-
A Token's Journey Across the LLM Stack
Find LLM inference bottlenecks with one worked example. Calculate KV-cache bytes for each token and request capacity in 8 GiB. Examine why long outputs keep cache occupied.
The Life of a Token Across the LLM Stack · Part 4
-
The Life of a Request Inside vLLM
This article gives engineers an introduction to vLLM serving. It compares batching, cache paging, prefix reuse and speculative decoding.
The Life of a Token Across the LLM Stack · Part 3
-
Why Transformer Performance Is Mostly a Matmul Problem
Most dense Transformer arithmetic is matmul. GPU matmul speed depends on tiling and data movement. Small decode batches are frequently memory-bound.
The Life of a Token Across the LLM Stack · Part 2
-
The Life of a Token Inside a Transformer
A Transformer maps token IDs to next-token logits. This article examines embeddings, attention, MLPs, RMSNorm, RoPE, GQA and the KV cache.
The Life of a Token Across the LLM Stack · Part 1
-
LLM Inference Is a Full-Stack Systems Problem
A slow LLM response can have different causes. This article examines prefill, decode, GPU kernels, KV-cache memory and request scheduling.
The Life of a Token Across the LLM Stack · Overview
-
From π0 to π0.7: A Tutorial on Open-pi and Robot Foundation Models
The tutorial examines Physical Intelligence’s open-pi work: continuous action chunks, robot transfer, context, memory, feedback and real-time control.
-
Work Is Important Enough Not to Become Everything
A reflection on Simone Stolzoff’s The Good Enough Job: caring about craft and ambition without asking employment to define your worth or your whole life.
-
The Age of Auto Research, and Other Certainties
This article compares tests for research proposals, experiments and claims in automated research. A better score alone cannot prove discovery.
-
On 3D, Video World Models, and the Approaching ImageNet Moment of Perception-Action Learning
This article examines 3D geometry and video world models for perception-action learning. It gives proposed ImageNet-style transfer tests for robotics.
-
Humans Are Imperfect, but Worthy of Love
A reading of Hiroshi Yamamoto’s The Stories of Ibis: machine love, human flaws and respect for minds whose experience we cannot fully understand.