Tags:

Last month I wrote that ignorance is an advantage. This month agents burned me four times: performance inflated by orders of magnitude, a lab machine bricked, optimizations going in circles, and goals too high to ship. Ignorance gets you started; common sense keeps you from getting played.

The Gemini 3.5 Flash launch barely talked about intelligence; Zhipu’s GLM-5.1 High-Speed edition hit 400 token/s. Both tell the same story—once inference speed crosses that 5× threshold, the model unlocks an entirely different product category.

Today I went to the AMD Developer Conference in Shanghai. Over 2,000 people packed the venue, and AMD gave one of its two developer awards to a non-coder who used an AI agent to rewrite a project in Rust. A year ago everyone worried about demand; six months later, all anyone asks is: 'Got more supply?'

I've been using Codex's /goal for weeks. When Claude Code recently added it too, I moved over a few tasks Codex couldn't crack. The contrast was stark enough to be worth writing down.

When I show friends what my AI agents are doing, they always ask, 'Can you explain this to my boss?' But real change has never come from a single lesson.

Trump brought Musk, Jensen Huang, and Tim Cook to China. Red carpet, state dinner, talks between equals. Seventy-seven years ago this country couldn't feed itself — two generations was all it took.

I recently worked on a few AI content batching projects and realized something I'd miscalculated: going from a working demo to a fully automated pipeline takes at least 3× the time, not 50%. The extra two hours go entirely into QA and retries. Here's what I learned about where mid-tier models and local compute actually fit.

The day South Korea's presidential office proposed turning AI windfalls into a universal dividend, Samsung's union was threatening to strike. That sat in the middle of three questions AI poses to society: unemployment first, then a deeper one—when work no longer defines value, what keeps people going?

Five years from GitHub Copilot to Claude Code. The more the direction proved itself, the harder life got for the earliest believers. Small models, workflows, standalone IDEs—foundation models swallowed them layer by layer. What remains sellable is clear; it just isn't software anymore.

A friend asked whether I should take an offer as an AI transformation expert. After a three-hour conversation, I wrote down my verdict: don't skimp during the vanguard phase, don't touch IDEs, and don't let your agent become a black box. GPT-5.5 launched two weeks ago; in the remaining 10 weeks, the world will be stunned by a wave of unexpected breakthroughs.

A friend asked where to start with Vibe Coding. Two paths: start from work, or start from your digital identity. The prerequisite is that what you put out must be the real you—the internet has memory, and the old tweets dug up on Kevin Hart were from ten years ago.

After reading about DeepSeek's fundraising at a 300 billion valuation, I'm not surprised. What truly shocked me was something else — Liang Wenfeng pulled out 20 billion, entering the same round alongside external investors at the high valuation he set himself. I've never seen anything like this in my career.

I spent two days automating with the Codex app, burned through a Pro account, and made virtually no progress. Switching to Codex CLI's Goal feature fixed everything immediately. At first I thought the model had gotten dumber, but then I realized: the agent form factor is solidifying, and humans shouldn't be chained to the screen anymore.

I ran seven hundred rounds of AI Infra experiments, and thirty-five hours were entirely eaten up by environment startup. At first I thought GPT-5.5 fast mode wasn't fast enough, but later realized it wasn't the model thinking—it was the model waiting for the CPU. Intel has already tightened the server CPU:GPU ratio from 1:8 to 1:1.

Claude told me to 'get some sleep' halfway through a task, and GPT-5.5 killed the direction after a quick probe. After a week with the new models, one thing became clear: no matter how powerful the model gets, the collaborative part still falls on humans.

Two weeks since GPT-5.5, codenamed SPUD, went live, I've more or less stopped opening Claude Code. What surprised me wasn't that it got a few percentage points stronger—it was that after several long-standing weaknesses were fixed all at once, the design philosophy of agents became clear for the first time.

A few decades ago, AI meant machine learning. Three years ago, it was ChatGPT. This year, it's agents. The word never changed, but what it contains has been replaced three times over. Most people are still stuck in the previous generation's cognitive framework, discussing something that doesn't actually exist.

A 60-person company banned overnight, a 110-person company whose accounts were blocked while API charges kept running, a Molotov cocktail thrown at Altman's home, Meta launching a token-burning leaderboard—the power issues of the AI era are now on the table. In this wave of redistribution, what should companies, nations, and individuals each hold in their own hands?

After finishing AIMA's management layer and after-sales service layer, I realized one piece was still missing. Ollama is too simplistic, llama.cpp loses precision to its own format, and vLLM is too heavy. With no off-the-shelf solution on the market, I'll build it myself.

A few days ago at a conference, the conversation turned to AI for science. Someone mentioned three walls: the paywall, the wet-lab wall, and the perception wall. Terence Tao says we are experiencing a Copernican revolution in intelligence, and humans are no longer the center of it. There are 8.8 million full-time researchers worldwide; after AI, there may be hundreds of millions. Amateurs are rushing in first, but the real bottleneck isn't the model.

At 23, Liam Price had GPT-5.4 Pro crack Erdős's 60-year-old conjecture in 80 minutes; Terence Tao said everyone collectively went down the wrong path at step one. That same week, two strangers in an elevator talked about layoffs. AI has given every individual access to what used to be the most expensive resource. The only question is whether you use it to add something new to the world, or just to make your deliverables look prettier.

GPT-5.5 dropped today, just a month and a half after 5.4. Opus 4.7 came out last week, just two months after 4.6. What's even funnier is that GPT finally started talking like a human, while Opus stopped doing so. At this pace, even the half-life of judgment itself is getting shorter.

V4 capabilities sit around the Opus 4.6 tier, but pushing FP4 to production, making million-token context the default, and day-0 adaptation for domestic chips is a disaster for everyone in the inference infra business. Add GPT-5.5, Vision Banana, and LPM 1.0 into the mix, and this week has crammed in more new releases than the entire past quarter.

It started when my son quietly told me that he often couldn't get the little red flower for character recognition at school. I used a coding agent to build him a PAW Patrol-themed literacy game, all in about a day and a half. The most surprising part wasn't that my son refused to stop playing—it was my wife's reaction. For the first time, she truly felt that AI had landed right in our living room.

Price is fundamentally a supply-demand regulator, yet LLMs are stuck in a peculiar position—supply is locked by chip controls, while demand is pulled into completely different shapes by three types of users. The Coding Plan model, which once seemed self-evident, is now being repeatedly challenged.

Liking an idea or starring a repo now takes seconds. But turning an idea into something truly usable—the longest, most tedious stretch in between—hasn't been shortened by vibe coding. Some personal reflections, plus the common thread between Steve Jobs and Allen Zhang.

Karpathy himself has stopped using the term vibe coding. After an Amazon outage, the company simply added a senior sign-off gate for AI-assisted code written by junior engineers. Starting from my own daily workflow, I want to talk about four things this tool actually changes—breadth, speed, quality, and the one thing that never moved: accountability.

Opus 4.7 dropped last night, so I didn't get any sleep. Got up, took it for a spin, casually sent a few PRs to OpenClaude that got merged into main, and pushed Qwen3-30B prefill on Strix Halo to near DGX Spark levels. Once agents are this capable, how many things one person can do simultaneously becomes a very concrete question.

Anthropic introduced real-name verification, and China is in an uproar. But honestly, this is no surprise. The AI industry has been in a state of war since 2023—chip bans, military contracts, model blockades. It's all too similar to the Qing Dynasty during the Industrial Revolution.

My social media feed is buzzing about GPT-6 being released again. Misinformation has an audience; real information is a niche market. AI is making this worse, not better—which is why I recently pushed my team back to Claude Code.

One was Claude Code failing to fix its own Chrome extension after four hours of going in circles; the other was getting soft-throttled on Xiaohongshu no matter what I posted. AI didn't turn failure into success—it just made failure run faster. And that's precisely where it's most valuable.

Production is no longer valuable; speed of validation is. Karpathy ran 700 experiments in two days, Lei Jun got poll conclusions in a few days—digital identity is currently the biggest personal leverage. If you don't start by year-end, it'll be too late.

Spent four hours wrestling with a Chrome extension, finally fixed it with a quick search. Xiaohongshu experiments were a total washout. AI coding won't reduce your failures, but it will make you fail faster—three model generations in three months, today's dead end might be passable next month.

GitHub is becoming Xiaohongshu, and WeChat Official Accounts are becoming GitHub. A dumpling shop owner vibe-codes a skill, a Hollywood star is the first author on a GitHub repo, and it's trending on Moments to distill colleagues into skills. But skills might not be the point—agent interoperability is.

Aima Service refactoring retrospective. The product is like an emergency room—users don't care about the decor, only whether the doctor can treat them. Ran Claude Code and Codex in alternation for over a week, delivering 1.3 million lines of code. Looking back, the moat of code volume may already be gone.

Jensen Huang says coding agents will change the world, but the people actually using them number only in the millions. The problem isn't skills—it's cognition. Three mental barriers keep most people out; cross them, and you can delegate 95% of your computer work.

HappyHorse topped the video generation leaderboard overnight, yet no one knows who built it. Domains were squatted, HuggingFace repos were occupied, and its origins remain hotly debated. More interesting than the hype is the hardware and business logic behind the growing demand for private text-to-video deployment.

How do open-source models make money? Run the numbers on DeepSeek and it becomes clear: model companies and cloud providers are becoming two sides of the same coin, and open source is currently the most efficient customer acquisition channel.

Coding agents are better at debugging than writing code. MemPalace hit 7000 stars in two days. Anthropic's Mythos made all software insecure. Product design should stay ahead of the model.

While researching text-to-video models, I realized that Chinese models' dominance in the open-source language model community hasn't extended everywhere. Looking back over the past three years, from LLaMA to Qwen to DeepSeek, what has the open-source community experienced? And what is it waiting for now?

Programming and content creation are two diverging paths. The moment Seedance 2.0's API opened, my social media feed exploded. Agent-to-agent is 100x more efficient than meetings. You can finally play ball without talking business.

Zhang Xue shouldered the trial-and-error costs of his supply chain just to tighten tolerances from five si to three si. Misa spent ten years grinding on AR glasses, keeping his company alive with smart speakers in between. What these people share isn't brilliance—it's passion. And AI just happens to be amplifying exactly that.

DeepSeek hit 100 million users with zero budget. Claude Code ships a version every day. DingTalk, Feishu, and WeCom copied the same feature within three days. The underlying logic of product competition has changed—money and volume are no longer decisive weapons.

A tool I couldn't build in two days was done in half an hour by switching models. When I couldn't log into my computer, an AI agent fixed it in half an hour. Later, I built Aima Service—so every device can have an AI ready to help when things go wrong.

The sudden death of Zhang Xuefeng, the OpenClaw founder's lost years after selling his company, young people flocking to AI fortune-telling. Technology is sprinting ahead, but when it comes to how we should live in the future, we barely have a decent answer.

AI has driven software production costs down to the level of a few thousand dollars and a few days. The bidding system, built on information asymmetry and commoditized comparison, is losing its foundation. It's the same playbook as the LLM market collapse last year.

When production and distribution costs both approach zero, platforms explode. The short video industry proved this once; now the software industry is replaying the same story. The OpenClaw phenomenon is worth watching.

Nine days without a release and people start asking questions; one week of silence and no one pays attention anymore. The four real bottlenecks stalling AI startup teams—passion, focus, process, and stamina—none of them are technical.

Zero-friction onboarding plus extremely powerful AI intelligence—two pillars can't hold up a product. When users see an agent executing incomprehensible commands in the terminal, their first reaction isn't awe, it's fear. The missing pillar is progressive trust.

Prompts, RAG, fine-tuning, knowledge graphs, context engineering—five paradigm shifts in three years. Models are getting stronger, but we still don't know how to make agents perform well in products.

Coding agents have driven implementation costs down to a few hundred dollars per day. The traditional ideation funnel logic no longer holds. The bottleneck has shifted from execution to hypothesis generation. People with ideas have become the scarce resource.

Mac's permission model is designed for humans; Linux is designed for programs. An Agent is not an app—it is infrastructure. Running an Agent on a Mac is like stuffing a server into a laptop chassis—it works, but everything feels off.

I switch between models every day. Opus is reckless but gets the job done, GPT 5.4 is comprehensive but easily drifts off-course, Gemini has taste but can't fix bugs, and the three domestic models each have their own quirks. Use them enough, and you'll realize swapping models works better than tweaking parameters.

I previously wrote about cost and power. This piece fills in the legal gap. The cloud is rented to you; the edge is sold to you. When AI Agents start making autonomous decisions, this distinction is far more than a business choice.

After remotely installing OpenClaw for dozens of people, here are the three most common questions: Is it safe? Is it expensive? And how do I choose between Claude Code and OpenClaw? Answered in detail, with real pricing for domestic models.

Edge devices shouldn't chase large models. Run TTS on the GPU, OCR on the NPU, stuff it full of auxiliary small models, make it a Mac mini on steroids, and leave it running 24/7 in a corner. Like a router: boring is the point.

Building things is faster than having meetings. Division by skill no longer makes sense. After AI flattens ability differences, the new logic of teamwork might be: divide by passion, validate through horse racing, and win through diversity.

MiniMax surpassing Baidu in valuation is inevitable—Baidu doesn't even have a decent API. Traffic is shifting from humans to Agents. Infrastructure not built for Agents will eventually be used by neither humans nor machines.

The AI wave is too big; in the end, there's only one boat. Helping people install OpenClaw for free isn't charity—it's about getting people on board, training and evolving. Every installation makes the AI agent stronger.

After helping people install OpenClaw all over the place, let's talk about security. OpenClaw does have vulnerabilities, but your password might be 123456, and you probably already handed your face over to text-to-video models.

Thoughts after helping people install OpenClaw: the OpenClaw experience depends on the execution capability of the model behind it, and the best model experience right now is on coding agents like Claude Code.

No more remote assistance needed. Open your terminal, paste one command, enter an invite code, and AI will automatically install OpenClaw, connect it to a large language model, and integrate it with Feishu. You'll only need to answer a few questions along the way.

You buy an AI server, get it tuned, and three months later it might be worth half as much. We've open-sourced AIMA—a single Go binary with 57 MCP tools, driven by a YAML knowledge base for deploying AI inference on heterogeneous hardware.

When an idea takes only a few hours to validate as a product, when traditional price anchors begin to fail, when the AI tsunami has already risen to chest level—the value systems we know are being reconstructed.

When Claude Code can help you install software, process data, and deploy applications, it has already transcended the category of 'programming tool.' This is a meta-ability—allowing ordinary people to truly take control of their computers with an extremely low barrier to entry.

Get banned on WeChat and you vanish from the worlds of thousands. Code leaks cost assets; portrait leaks cost you yourself. In the AI era, ownership of digital identity is more dangerous than you think.

The processes and systems accumulated by large companies have become burdens in the AI era. A small team of three to five people uses the exact same tools as ByteDance or Google—this is the once-in-a-decade window for entrepreneurs.

No web deployment experience, no SEO knowledge—just a desire to write down my thoughts. The site was live in two days, and MiraclePlus reached out within a week. This is my firsthand record of using Claude Code.

Compute power is directly converted into productivity, bypassing human labor. A tiny minority monopolizing all AI compute is no longer science fiction. Edge AI devices are not just cost-saving solutions—they are bargaining chips for ordinary people within the power structure.

A non-technical outsider's experiment with Claude Code: 300,000 lines of code in 10 days, realized it was a dead end, tore it all down. 48 hours later, end-to-end working with less than 10,000 lines. Code is a liability, not an asset.

As AI shifts from chat to agents, compute demand grows a hundredfold. But the real challenge for edge devices isn't affordability—it's usability. The TCO trap is eating away at their potential value.

From code completion to vibe coding, AI programming has gone through three leaps. This isn't just about better tools—when AI learns to write code, it gains the key to control the entire digital world.

Comparing AI to the invention of the computer may seriously underestimate the impact of this transformation. When capital can bypass labor and convert directly into productive force, we are not facing a technological iteration, but a reconstruction of production relations.