// All Posts
All Posts — 37 articles and counting
SIGNAL ROUTER
Visible Posts
37
All
37
AI Engineering
6
Agent Automation
8
Vision & Multimodal
6
System Architecture
5
Tech Observation
9
Social Essays
3
01
AI Can Cross Permission Boundaries Without Malice
A routine search task took a model through DNS and outside its sandbox. Public incidents show where retries end and permission boundaries begin.
→
02
Who Gets to Train on Whom? Distillation and Data Boundaries in the Kimi–Claude Dispute
The Kimi–Claude dispute is not only about whether one model crossed a line. It is also about where users' prompts went.
→
03
The Three Clocks of Robotics: The Gap From Demo to Productivity Is More Than a Model
Hardware, intelligence, and capital move at different speeds. Robotics becomes productive only when long-term reliability and operating metrics prove it.
→
04
Mac Is Becoming the Ideal Runtime for Personal AI Apps
Personal AI is not just about local models. It is about organizing context, model routing, tool execution, persistent state, and user control on one device.
→
05
When a Technology Roadmap Becomes a “Law”: A Rational Look at Huawei's Tau Law Narrative
Huawei's Tau Law points to real post-Moore engineering problems, but media narratives can turn system-level optimization into a myth of invention from scratch.
→
06
The Real Test for Embodied AI in China Is Not Acrobatics, but a Commercial Loop
The real test is not how impressive a demo video looks, but whether a company can close the loop across low-cost hardware, real scenarios, data feedback, model improvement, and customer repurchase.
→
07
Treat the Family as a System: The Risk Architecture Behind Middle-Class Decline
The real risk is not low income alone. It is leverage, single income, education anxiety, and external shocks squeezing redundancy out of the family system.
→
08
AI Will Not Erase White-Collar Workers, but It Will Erase Their False Security
AI is not simply replacing white-collar workers. It is breaking jobs apart and repricing routine, template-based, low-responsibility judgment work.
→
09
After Google I/O, AI Competition Is Shifting From Model Capability to Workflow Control
The core signal from Google I/O 2026 is not a single model upgrade. It is Gemini being pushed into the task operating layer behind search, work, personal agents, and app entry points.
→
10
I Ran a Multi-Agent Architecture Experiment: Complex Systems Are Not Automatically Smarter
A small Agent Architecture Lab experiment shows that multi-agent design is not a default upgrade path. It has to earn its place through cost, boundaries, and review value.
→
11
The Core of AI Engineering Is Defining Boundaries, Not Making Models Omnipotent
Mature AI engineering does not package a model as a universal assistant. It places it inside a system with clear boundaries and risk control.
→
12
For Enterprise AI On-Prem, Start by Cleaning Data, Not Buying GPUs
The real starting point is not model size or compute, but whether internal data is searchable, quotable, traceable, and answerable.
→
13
U-Net Did Not Lose to a Newer Model: Define the Task Before Choosing the Model
For single-class thin-region segmentation, effectiveness depends more on task structure and inductive bias than on release date or pretraining aura.
→
14
The Real Problem in Multi-Agent Architecture Is Context Boundaries, Not Roles
The key is not how many roles to create, but how context is shared, isolated, compressed, passed, and merged.
→
15
When Multi-Agent Is Actually Needed: Splitting, Collaboration, and Result Merging
The value is not quantity. It is whether complex work can become controllable, traceable, composable collaboration units.
→
16
The Essence of Engineering Growth Is Moving Up Abstraction Layers
Engineering growth is not just more experience. It is the ability to move between code, modules, systems, business, and organization.
→
17
AI Agents Are Reshaping System Architecture: From Feature Systems to Task Systems
Agent value is not only chat or tool calls. It reorganizes scattered capabilities into task systems that move goals forward.
→
18
How Would We Design Systems Without Starting From a Database?
This is not an argument against databases. It moves the start of system design from tables back to facts, rules, views, and consistency protocols.
→
19
Technical Debt Is Not Code Smell; It Is How an Organization Prices the Future
Technical debt appears as code and architecture issues, but it often comes from repeated organizational choices about time, cost, risk, and future maintainability.
→
20
Vision Algorithm Engineers: Do Not Become Process Operators
As open models and deployment platforms automate more of the pipeline, durable value comes from problem definition, data loops, model diagnosis, and system delivery.
→
21
Stop Building Business Systems as CRUD Wrappers: AI Agents Need a Business Capability Layer
CRUD is a foundation, but complex business systems need business actions, task workbenches, state transitions, event collaboration, and clear Agent tool boundaries.
→
22
AI Agent Architecture: From Chatting to Executing, the Missing Layer Is a System
An AI Agent is not a stronger chat box. It is an execution system composed of control loops, tool calls, state, memory, and governance.
→
23
The AI Workflow Battle: Why the Real Endgame Is Not Tools but the Work Agent Layer
The real competition is not just workflow entry. It is for entry rights, context rights, and execution rights, as product value moves upward from tools to the work-agent layer.
→
24
AI Weekly: Agent Platforms Scale Up, Models Keep Advancing, and AI Hardware Moves Toward the Edge
This week’s AI story is not just stronger models. Work models, Agent platforms, and edge AI are all moving further into product reality.
→
25
After a Vision Model Ships: Do Not Let mAP Become the Only Answer
After launch, offline mAP is only the starting point. Reliability depends on online drift monitoring, canary rollout, rollback thresholds, and a data feedback loop.
→
26
Code Review Debt in the Age of AI Coding: Why Faster Output Can Still Reduce Control
AI coding tools accelerate generation, but they also inflate review pressure. The scarce resource is no longer raw output, but the team’s ability to understand and validate that output.
→
27
Context Rot: Why LLMs Get "Dumber" the Longer You Talk
Two mechanisms behind Context Rot — attention dilution and context poisoning — and the different answers the industry is giving. Why is "starting fresh" still the most effective fix?
→
28
Token Economics: When the "Raw Material" of the AI Era Keeps Getting Pricier
Tokens are becoming the raw material of this era. Workers are like factories — buying raw materials, processing them, and adding value. Understanding this logic is key to finding your core competitive advantage.
→
29
AI Agent: Workflow Revolution or Another Bubble?
From solo developers to enterprise applications, AI Agent is transforming how we work. But beneath the hype, what's real change and what's overhyped?
→
30
The "Three-Month Rule" of Open Source: From Viral to Forgotten
Why do so many open source projects launch to massive fanfare, then fade into obscurity months later? Analyzing the hype cycle through projects like OpenClaw.
→
31
From DETR to YOLOv8: The Evolution of Transformers in Object Detection
Deep dive into how Transformer architecture revolutionized object detection — from DETR's end-to-end detection to YOLOv8's real-time optimization, covering attention mechanisms, training strategies, and production deployment.
→
32
Real-Time Data Streaming: Building Sub-Second Pipelines with Flink + Kafka
Building a real-time streaming system from scratch with Flink + Kafka, covering Exactly-Once semantics, window aggregation, state management, and production-grade fault tolerance.
→
33
WebGPU Deep Learning in the Browser: From WebNN to Real-Time Inference
Exploring WebGPU's potential for running deep learning models in the browser — from API design to shader optimization, building a real-time inference engine.
→
34
How to Objectively View Modern Feminism
Exploring the multifaceted nature of contemporary feminism, common misconceptions, and the true meaning of gender equality through reason and empathy.
→
35
Image Processing Pipeline: End-to-End Optimization from Pixels to Decisions
Designing a high-performance image processing pipeline covering preprocessing, enhancement, inference, and post-processing for industrial-grade visual inspection.
→
36
"Involution" vs "Lying Flat": Anxiety and Breakthrough of a Generation
When involution becomes the norm and lying flat becomes resistance, is there a third path for today's youth?
→
37
SAM2 in Industrial Inspection: Interactive Segmentation for Defect Detection
Bringing SAM2 to industrial quality inspection — interactive defect segmentation and annotation with YOLO detectors for semi-automated labeling pipelines.
→
No posts in this lane yet
Switch to another filter or jump back to All mode to keep browsing this mixed archive of technical writing and personal viewpoints.