Ai on Margrop Blog

FlashVSR on an RTX 5060 Ti: From 480p Input to 1728×960 Video, Plus Every Pitfall I Hit

Mon, 10 Aug 2026 10:00:00 +0800

A note before we start

“Make a 480p video clearer” sounds like a simple resize operation. In practice, video super-resolution is closer to asking a restorer to repair hundreds of neighboring photographs while making sure faces, text, lighting, and motion do not jump from frame to frame. This post records a complete FlashVSR_plus experiment on an RTX 5060 Ti: compatibility checks, portable Miniconda setup, model download, source-level bugs, FFmpeg audio muxing, and reusable Windows, Ubuntu, and macOS scripts.

The FlashVSR_plus build used here is a community variant rather than the official upstream repository. Commands and flags can change between versions. Paths in this article are public examples; the original video, model cache, and raw terminal logs were not uploaded.

The Most Expensive Part of AI Is Not the Model: Hidden Bills and TOKEN Syndrome

Fri, 07 Aug 2026 08:00:00 +0800

Short version The AI wave is no longer just about smarter models. It is becoming a hidden bill made of hardware, subscriptions, devices, attention, and anxiety. AI can make an individual step faster while making the complete workflow harder to manage. A new everyday symptom is emerging: TOKEN Syndrome—the constant fear of wasting tokens, running out of quota, or choosing the wrong model. Figure 1: AI is not a free desk.

China's Coding Plan Shake-Up: GLM Is Unavailable, Kimi Is Closed—What Is Still Worth Buying?

Tue, 21 Jul 2026 07:00:00 +0800

Bottom line: as of July 20, 2026, the market has split sharply: GLM is a timed, limited flash sale that is practically impossible to buy; Kimi has closed the current membership subscription entry; MiniMax TokenPlan is currently buyable without a purchase cap but delivers a weaker coding experience than scarce plans; Qwen AI has just launched qwen3.8, is currently open with usage discounts, and is one of the better choices to try now, although future restrictions remain possible; Volcengine Ark CodingPlan/AgentPlan are limited but still obtainable in some cases.

Can You Still Buy a Coding Plan? A July 2026 Reality Check on Limits, Quotas, and Lockouts

Mon, 20 Jul 2026 09:00:00 +0800

Short answer: as of July 20, 2026, most mainstream AI coding plans are still purchasable. The confusing part is that “available to buy,” “available in my region,” and “available for heavy coding” are no longer the same statement. Some products have a live checkout but tight quotas; others are visible in marketing pages but not offered to every account; some individual coding experiences have been paused or reshaped during product transitions.

How Did AI Learn to Use Its Hands? From Chatting to Checking, Writing, and Acting

Sun, 19 Jul 2026 22:00:00 +0800

The short answer

AI did not suddenly grow hands. It learned a very practical loop: decide which tool is needed, request that tool with structured arguments, wait for a real result, and then explain the result in human language.

This is usually called Tool Calling. You may also see Function Calling or a Tools API. MCP, the Model Context Protocol, is closer to a universal socket: different AI applications can discover and call files, databases, search engines, ticket systems, browsers, and other capabilities through a common protocol.

This article opens the loop and examines every moving part: what the model does, who actually performs the action, why MCP matters, where permissions belong, and why “can call a tool” does not mean “should receive the keys to everything.” Screenshots come from official OpenAI and MCP pages; the terminal-style evidence images are generated from the real invocation flow. No real private network address or hostname is included.

Why Does AI Lie So Confidently? — It Is Not Thinking, It Is Playing Supercharged Word Association

Sun, 19 Jul 2026 20:30:00 +0800

The short answer

AI hallucination is not simply an occasional bug. It is a predictable risk of a system whose basic job is to generate the next likely token from context. The model can produce a fluent, confident, highly structured answer without having verified that every claim is supported by evidence.

The easiest analogy is this: a large language model is a supercharged word-association player. It is exceptionally good at continuing a sentence. It is not automatically a fact-checker.

Why Does AI Keep Saying ‘I Forgot’? Tokens, Context Windows, and the Sticky Notes It Must Throw Away

Sun, 19 Jul 2026 18:30:00 +0800

The short version

AI often appears to have a bad memory because only a limited amount of material can be carried into the model for the current request. Think of a context window as a stack of sticky notes with finite capacity. New questions and tool outputs keep adding notes. Once the stack is full, the system must summarize, truncate, filter, or remove older notes. If a note is removed from the current context, the model may no longer be able to use it.

Have you ever told an AI, “Use Python for every code example,” only to receive JavaScript fifteen turns later? Or supplied a long document that the model can summarize accurately at the beginning and end, but cannot answer about the important restriction buried in the middle?

It is tempting to say that the AI “forgot.” A more precise explanation is this: the model can only see what fits inside the current context window, and seeing a piece of text does not guarantee that every part of it receives equal attention.

This article explains tokens and context windows without turning them into intimidating mathematics. The main analogy is a stack of sticky notes. We will look at what tokens are, how a request is assembled, why long conversations become unreliable, why agents fill their context quickly, and how to make AI forget less through simple information management.

Four Hard Resets in One Weekend—and Two Cards: When Tokens Stop Being Scarce, What Are We Actually Missing?

Mon, 13 Jul 2026 07:00:00 +0800

The short version

Codex users spent the weekend inside a restaurant that would not stop refilling the buffet. There was one hard quota reset early Friday morning Beijing time, two more on Saturday, and another waiting on Monday morning. A bankable reset card also appeared. As Tibo celebrated Codex reaching six million active users, he previewed another card for everyone the following day.

At the same time, the five-hour limit was temporarily removed for Codex subscribers, leaving only the weekly limit. I also had a substantial pool of GLM-5.2 quota expiring on Sunday. The old problem was saving tokens. The new problem was stranger: how do I spend them on work that deserves them?

After a weekend that felt busy, surprising, and slightly absurd, my conclusion is simple: the scarce resource in the AI era may not be tokens. It may be worthwhile questions, distinctive ideas, and the ability to turn those ideas into reality quickly.

Why HermesAgent Would Not Reply in WeChat or WeCom: The Real Culprit Was a Local Relay With a 3-Second Timeout

Sun, 31 May 2026 08:00:00 +0800

The short version This incident looked like a model problem or a broken messaging channel, but the real failure happened one layer lower: my local relay cut the upstream request off after 3 seconds. That was fine for quick health checks, but it was wrong for real WeChat and WeCom turns, which often take longer to produce a complete answer from the NewAPI gateway. Once the relay disconnected too early, HermesAgent could only see Connection error, RemoteProtocolError, and then exhausted fallback attempts.

Python App in macOS LaunchAgent Can't Reach the Internet? Here's the httpx Proxy Trap You Need to Know

Sat, 30 May 2026 09:00:00 +0800

TL;DR:

When a Python application using httpx with trust_env=True runs inside a macOS LaunchAgent, it silently picks up the system proxy settings from scutil --proxy. But the LaunchAgent process may not be able to reach that proxy server at all — resulting in All connection attempts failed or No route to host errors.

The fix is one line: add NO_PROXY=* to the LaunchAgent’s plist EnvironmentVariables.

This article documents the full debugging journey: from discovering that my AI Agent’s WeChat Enterprise (WeCom) messages weren’t being replied to, through methodical proxy troubleshooting, to finally pinning down the root cause — macOS system proxy + LaunchAgent network isolation. We’ll dive deep into Python httpx source code, macOS proxy architecture, and the many gotchas of LaunchAgent runtime environments.

Migrating from OpenClaw to HermesAgent: A Smooth AI Agent Migration in Practice

Fri, 29 May 2026 08:00:00 +0800

Introduction: Why Am I “Moving”?

In 2026, the AI agent landscape is evolving at an incredible pace. Tools that were cutting-edge six months ago might already be surpassed by more powerful alternatives. As someone who relies heavily on AI agents for daily work, I’ve been closely following the latest developments in this field.

Recently, I completed a full migration from OpenClaw to HermesAgent. This wasn’t an impulsive decision, but a thoughtful choice made after experiencing the differences between the two tools in real-world usage.

This article will share in detail:

Background: What OpenClaw and HermesAgent are
Migration Reasons: Why I decided to migrate
Migration Process: Step-by-step guide to complete the migration
Pitfalls Encountered: Problems and solutions during migration
Post-Migration Cleanup: How to completely remove OpenClaw residuals
Post-Migration Experience: New capabilities gained from HermesAgent
Q&A: Frequently asked questions

Using WoClaw to Solve Shared Memory for Multi-Agent Workflows: OpenClaw, Codex, Claude, and Gemini on the Same Table

Sun, 05 Apr 2026 20:10:00 +0800

What matters most in a multi-agent system is often not “which model is smarter”, but “which agent still remembers what happened five minutes ago”.

When I put OpenClaw, OpenAI Codex, Claude, and Gemini into the same workflow, the most painful problem was not raw answer quality. It was context fragmentation:

OpenClaw remembered scheduling decisions, Codex remembered implementation details, Claude remembered review conclusions, and Gemini remembered external references, but none of them knew what the others knew.
Each tool had its own session lifecycle, so yesterday’s discussion was often gone by tomorrow.
As soon as a project entered a multi-round, multi-tool stage, repeating the same background became more expensive than solving the problem itself.
Without a shared memory layer, the final result was not collaboration, but several partial answers that looked similar and still failed to line up.

This post is about how I used WoClaw to place all of those tools on the same shared memory layer, so they could stop talking past each other and start working from the same state.

To avoid leaking private information, all examples below are abstracted. They do not include real hostnames, account names, addresses, tokens, or internal project codes.

One Month with OpenClaw on MiniMax: Stable, Fast Enough, and Valuable Because It Lasts

Thu, 19 Mar 2026 10:30:00 +0800

Over the past month, I have kept MiniMax as the primary model behind OpenClaw and used it across a full set of automation tasks: routine inspections, connectivity troubleshooting, health checks, security hardening, scheduled-task cleanup, and even blog writing support. Looking back, the biggest gain was not “how much work AI can replace,” but a more practical conclusion: for a system like OpenClaw that needs to run continuously, cooperate with tools, and handle a large amount of Chinese context, the most important thing is not raw benchmark strength. It is the balance between stability, response speed, integration cost, and long-term maintainability.

OpenAI GPT-5.4 Explained: Positioning, Capabilities, Cost, and Practical Adoption (As of 2026-03-06)

Fri, 06 Mar 2026 08:20:00 +0800

On March 5, 2026, OpenAI officially released GPT-5.4. The first question many people asked was simple: is this just another small point release, or is it a real shift in how much productive work the model can finish?

If you only look at the version number, 5.4 looks like the next iteration after 5.2 and 5.3. But once you line up the official launch notes, model docs, API capability matrix, and the ChatGPT-side product updates, the core value of GPT-5.4 is not “how many more parameters it has.” The real change is that it pulls three lines of capability into a much more usable center point:

Reasoning
Coding and engineering execution
Tool ecosystem coordination, including computer use