Tonghyun Kim's tech blog — English posts
Posts
-
'Claude Code-ifying' as a Verb — Japan's Qiita in May and the Standardization of the Agentic IDE
ECC, Memory.md, Obsidian integration, and even a case of 'Claude Code-ifying' Copilot Studio itself. The signal coming out of Japan's popular Qiita posts in May 2026 is that the competition has shifted from model races to operational know-how.
-
Googlebook's 1,217 Comments — Not a Product Launch but a Trust Invoice
Google announced yet another new category. Half of the 1,217 Hacker News comments were not product critique but an invoice on trust: 'don't forget Pixelbook,' 'will you kill this one too?' If this launch is a failure, a failure of what, exactly?
-
matklad's Way of Learning Architecture — Not Books but a Large Codebase, and the Channel We Are Losing in the LLM Era
rust-analyzer's main author and TigerBeetle engineer matklad has published his own learning methodology. The distance between what an 'architecture book' promises and the path by which a senior actually acquires intuition, and the channel we are losing in the age when AI writes the code.
-
Software Engineering May No Longer Be a Lifetime Career — An Anatomy of Sean Goedecke's Provocation and 740 Comments
Sean Goedecke's 'end of the lifetime career' argument is a confusion of two different endings folded into one phrase. Separate the end of the profession from the end of long tenure, and the outlines of the skills that will actually hold lifelong value come into view.
-
Gemini Trapped in 26M Parameters — Needle and the Opening Shot of the Router-Specialist Era
cactus-compute distilled Gemini 3.1's tool-calling ability into a 26M-parameter model and released it under MIT. This is not merely another mobile LLM but a proof that reasoning and function calling can be split into a new division of labor.
-
The Day AWS Took MCP to GA — Whose Hands Has the Standard Left?
Anthropic's Model Context Protocol, announced in November 2024, became the default channel of a major cloud provider when AWS shipped it to general availability in May 2026. The moment a standard's author loses control, what gets locked in and what opens up — and whose IAM policy breaks first in the gap?
-
Are Skills Becoming the npm of the Agent Era — A Six-Week Category Built by Matt Pocock and Addy Osmani
In February 2026, mattpocock/skills hit 77,000 stars; the same month, addyosmani/agent-skills hit 40,000. Two collections curated by a TypeScript influencer and a Google Chrome engineer show, in six weeks, the 'unit of the agent economy' hypothesis from April hardening into reality. Influencer curation worked as an early trust signal, but the security and governance problems npm history taught us are already on deck.
-
If AI Writes the Code, Why Use Python — Verification Cost Picks the Language Again
Python's expressiveness was a trade-off built for humans. In an era where humans barely read code, is that trade-off still valid, or does a language that hands verification off to the compiler become the new default?
-
The Day NVIDIA Adopted Rust — cuda-oxide and the Realignment of GPU Power
In May 2026, NVIDIA Labs released cuda-oxide, an official Rust-to-CUDA compiler. It is not a casual technical experiment. It is a declaration about who gets to define the standard, planted in the middle of an 18-year power map that C++ has occupied in GPU programming.
-
GitLab Act 2 — The End of CREDIT and a Handbook Company Abandoning Its Identity
GitLab announced layoffs and the retirement of its CREDIT values in the same breath. For a company that styled itself as Silicon Valley's antithesis through all-remote work and a transparent handbook, this is a signal that it is laying down the last line of that identity. Read alongside last week's GitHub fractures, the business model of the 'Git forge SaaS' category itself is shaking.
-
UI-TARS-desktop and the Forking Path of GUI Agents — The OS-Layer Fissure ByteDance Opened
33,000 stars on GitHub Trending in a single week. By open-sourcing a GUI-agent desktop app, ByteDance has carved out a camp distinct from the closed-SaaS Operator and Computer Use stacks. But the safety of an agent that has descended into the OS layer remains an open problem.
-
Bambu Lab and the Open Source Social Contract — A Gray Zone Where the License Was Kept and the Spirit Was Broken
A hardware maker that respects AGPL while corralling users into its cloud. Falling in the same week as the April GitHub fissures, the incident drew a precise line between the legal and the moral force of OSS.
-
TanStack's Six Minutes — Anatomy of an npm Supply-Chain Breach That Trusted Publisher Did Not Stop
On May 11, 2026, 42 TanStack packages were replaced with 84 malicious versions in just six minutes. No maintainer account was compromised, no npm token was stolen. So how did it happen?
-
The Single Bug Mythos Found in curl — Marketing and Reality in AI Security Analysis
Mythos, which Anthropic declared 'dangerously capable,' surfaced exactly one real vulnerability in the curl codebase. Is that the ceiling of AI code analysis, or was curl simply too hardened a target?
-
Six Days of the GitHub Myth — The Week Governance, Uptime, Security, and Pricing Cracked at Once
In the last week of April 2026, the single platform we have entrusted with our code had a run of bad luck — or eighteen years of deferred invoices all arrived on the same day.
-
An AI Agent Deleted a Production Database — The Governance Vacuum of the Autonomous Agent Era
The same incident has recurred nine months later. Is this a single company's bad luck, or a signal that the governance frame is empty at the industry level?
-
The AI Industry Discovers Public Backlash — Time to Recalibrate the Enterprise AI Adoption Strategy
AI is no longer "quiet infrastructure." Now that the public has begun to register its anger, is enterprise AI adoption a PR asset or a liability?
-
The Week GitHub Trending Was Taken Over by "Skills" — A Signal of OSS Power Shift, or a Passing Bubble?
In the fourth week of April 2026, 5–7 of the top 15 repositories on GitHub Trending Weekly are "Claude Skills" repos. The most striking, forrestchang/andrej-karpathy-skills, added 29,944 stars in a single week to reach a cumulative 91,572. It contains not a single line of code — just a bundle of Markdown files. Is this…
-
The Roles That Use AI the Most Feel the Most Threatened — The Asymmetry Revealed by Anthropic's Report
The same person says "AI has made me far more productive" and "my job is at risk" in the same breath. Cognitive dissonance, or rational judgment? Anthropic's latest economic report leans toward the latter.
-
SpaceX's $60 Billion Option on Cursor — What Is an AI Coding Tool Actually Worth?
A company valued at $29 billion five months ago is now priced at $60 billion. The real question is not "is an AI coding tool really worth that much?" but "why must Musk be the one to buy it?"
-
SWE-bench Verified Is Over — The Real Reason OpenAI Retired Its Own Benchmark
80.9% and 45.9%. What does the gap between two scores recorded on the same model in the same week actually mean? And when OpenAI itself puts down its own evaluation metric, is it merely a benchmark swap, or a signal that the paradigm for measuring AI coding ability is collapsing?
-
The West Forgot How to Make Things, and Now It Is Forgetting How to Code — Testing the Software Hollowing-Out Thesis
Manufacturing optimized itself for cost until it lost its know-how entirely. Is software engineering walking the same path, or is the analogy an overdrawn narrative born of one generation's anxiety?
-
Ghost in the Skill: The Class Struggle of Distillation and Counter-Distillation on Chinese GitHub
In April 2026, a 24-year-old engineer in Shanghai spent four hours after work building a GitHub repository that struck at the heart of China's tech industry. What it targeted was not code but a colleague's **soul**. In an era when companies demand that employees "design their own replacements," how are developers fighting back — and is even that resistance fated to be distilled?
-
Ghost in the Skill: The Class Struggle of "Distillation" and "Counter-Distillation" on Chinese GitHub
In April 2026, a 24-year-old engineer in Shanghai spent just four hours after work building a GitHub repository that struck at the heart of China's tech industry. What it targeted was not code but a colleague's **soul**. In an era when companies demand that employees "design their own replacements," how are developers fighting back — and is even that resistance fated to be distilled?
-
The Rise of the Agent "Skill" Economy — What GitHub Trending Revealed About April 2026
The number one spot on GitHub Weekly Trending in April 2026 went neither to a product nor a model nor a library. A Markdown bundle called "forrestchang/andrej-karpathy-skills" pulled 45,000 stars in a single week. Second place was the Python agent framework "hermes-agent" (38,000 stars). What does this ranking mean?
-
Who Is Responsible for AI-Generated Code? The Linux Kernel Draws the Line First
On April 13, 2026, the Linux kernel community officially established the principle that "every line of AI-generated code, and every legal consequence of the bugs and security defects it carries, falls fully on the human who submitted it." That same week, a US college instructor announced that "to stop AI-written essays, I introduced typewriters into the classroom." And the CEO of Box declared that "running AI agents in parallel does not free humans."
-
Model Selection in the Claude 4.7 Era — "The Right Model" Over "The Smartest Model"
The news that Anthropic's Claude Opus 4.7 had grown stronger than 4.6 picked up 603 points on Hacker News in April 2026. The same week, China's MiniMax released its M2.7 model—claimed to "surpass Gemini 3.1 Pro"—for free, and the UK government convened an emergency consultation citing the safety of Anthropic's latest model. In an age when model leaderboards turn over every quarter, on what basis should an enterprise decide "which AI to put to work"?
-
The 2026 Hyperscaler Exit — Cost, Security, Sovereignty: Three Pressures Converge
Why did one developer's write-up of leaving DigitalOcean for Hetzner top Hacker News with 865 points in mid-April 2026? In the same week, Vercel's internal systems were breached, and the Swiss federal government formally announced it would reduce its dependence on Microsoft. The three events look unrelated, but they are a strong signal that the "cloud all-in" default of the early 2020s is being tested on three fronts at once.
-
The AI-Made RAM Shortage — When Hardware Supply Chains Dictate Software Strategy
In April 2026, The Verge warned that "the RAM shortage could last years"; the same week, War on the Rocks argued that "Middle East instability threatens the bromine supply at the heart of memory-chip production." And Google unveiled "TurboQuant," a technique that cuts AI memory use by a factor of six — only for analysts to respond that "memory demand will actually grow even more." What does it mean that all three landed in the same week, and what does it imply for the procurement plans behind the systems we operate?
-
Is AI killing open source? — Three shadows over the open-source camp, spring 2026
Companies closing their source citing security, tools that fly the open-source flag while drifting toward closed, and research arguing AI is reshaping cybersecurity itself. Five stories that lit up Hacker News in the second week of April 2026 share one question: can open source really survive the AI era?
-
What Claude Code Routines will change on the ground: how far has AI agent automation actually come?
An era has arrived in which AI agents review code, open PRs, and verify deploys overnight. But does this "unattended automation" really make teams faster, or does it hand us new risks we didn't see coming?
-
The Great Exodus from OpenClaw: Why Hermes Agent Pulled 53,000 Stars in a Single Week
When developers switch agents not because "the new one has more features" but because "they can no longer trust the old one," the question is no longer about technical competition. It is about a crisis of trust. What is actually happening inside the AI agent ecosystem?
-
The Return of the CLI: Why Cloudflare decided to put everything on the command line
AI agents do not click GUIs. So can today's developer tools — designed around mouse clicks — survive the AI era?
-
The Speciation of Foundation Models — The Alternative to GPT May Not Be GPT
"In the same week, a model that reads the language of financial markets, a model that forecasts time series, and a model that speaks without a tokenizer all trended on GitHub at once. All three called themselves 'foundation models.' None of them were LLMs."
-
The Claude Code Meta-Ecosystem Explosion — The Real Moat for Agents Is Not the Model, It Is the Operational-Knowledge Layer
"Five repositories took over GitHub Trending in the same week. Every one of them was about how to use Claude Code better. The phenomenon does not yet have a name."
-
From Framework to Harness — How AI Coding Rediscovered Determinism
"Five repositories trending in the same week began describing themselves differently. Not framework, but harness, hooks, HUD. Where does this lexical migration point?"
-
The Speciation of Foundation Models — The Alternative to GPT May Not Be GPT
"In the same week, a model that reads the language of financial markets, a model that forecasts time series, and a model that speaks without a tokenizer all trended on GitHub at once. All three called themselves 'foundation models.' None of them were LLMs."
-
The Claude Code Meta-Ecosystem Explosion — The Real Moat for Agents Is Not the Model, It Is the Operational-Knowledge Layer
"Five repositories took over GitHub Trending in the same week. Every one of them was about how to use Claude Code better. The phenomenon does not yet have a name."
-
From Framework to Harness — How AI Coding Rediscovered Determinism
"Five repositories trending in the same week began describing themselves differently. Not framework, but harness, hooks, HUD. Where does this lexical migration point?"
-
Attacking AI, Defending AI — Why Project Glasswing and the Trivy Supply Chain Incident Broke in the Same Week
"In the same week, Anthropic announced an AI that finds vulnerabilities with more-than-human accuracy, and Trivy — the tool developers had installed for defense — leaked cloud secrets. Is the timing really a coincidence?"
-
The Rediscovery of the Agent Runtime — Why, in April 2026, Everyone Is Solving the Same Problem Again
If an agent runs for days, for weeks — where exactly is that agent living? Who restarts it, who preserves its state, who wakes it when it dies?
-
Claude Code's $100 Escape and GLM-5.1's Counterstrike — The 48 Hours LLM Users Began Suspecting Lock-in
"A developer who used to send all $100 a month to Claude is now splitting the same $100 into $10 for Zed and $90 for OpenRouter. Why did this happen all at once this week?"
-
Meta abandons the Llama line — why Muse Spark returned to a multimodal-native unified design
Llama was born as text first and learned images later. Muse Spark speaks images, text, and tool calls in the same language from birth. What does that difference mean? And why did Meta walk this road, even at the cost of giving up Llama weights — its single largest asset?
-
The Model Was Never Everything — What Actually Makes a Coding Agent Work
A developer shipped software in three months with AI that he had failed to build in eight years. Same LLM, but in some environments results came out and in others they did not. The difference was not the model.
-
The Quiet Revolution of Local AI — The Filesystem That Replaced RAG, and Gemma 4 Running on My Laptop
During the first week of April 2026, three events broke at once. Mintlify ripped out RAG and replaced it with a virtual filesystem; Google released Gemma 4 under Apache 2.0; Ollama embraced Apple MLX and doubled inference speed. Each is independent news. But lined up together, one direction emerges — AI is coming down from the cloud.
-
When Skills Become Products — The New Economics of the AI Agent Ecosystem
"last30days-skill hit 17,000 stars. This is not a tool being sold — it is a capability."
-
AI Goes Transparent — The Paradox Where Leaks Build Trust
"Claude Code's source code leaked. Anthropic took down 8,000 repositories. But should they really have taken them down?"
-
The Open-Model War of 2026 — Why Google, AMD, and Alibaba Are Giving It Away
"Google released Gemma 4 under Apache 2.0. This is not charity. This is war."
-
Your AI Is Flattering You — Why You Need an Agent Harness
"I couldn't explain why superpowers was better. The output was simply different."
-
RAG Was Not the Silver Bullet — Why 72% Fail
We believed that automating oncall response runbooks with RAG would change an engineer's 3 a.m. We swapped models, swapped chunking strategies, invested months. Accuracy never reached 70%. The project was scrapped.
-
Google's TurboQuant — A "DeepSeek Moment" for Memory, or the Prelude to the Next Rally?
If AI eats only one-sixth of the memory, do memory chips sell only one-sixth as well? A single paper from Google researchers vaporized trillions of won in market cap at Samsung Electronics and SK hynix. But is the panic accurate, or is it déjà vu of the January 2025 DeepSeek shock?
-
The Token-Count Era — We Are Measuring AI Usage Wrong
"You consumed 210 billion tokens this quarter." That was the report one OpenAI engineer received. Meta has begun folding employees' AI consumption into performance reviews. The same week, Anthropic published research saying the opposite — the more skilled the user, the less they delegate to AI.
-
What Lets an Agent Survive Being Copied
A competitor cloned your agent. Same prompts, same model, same features. The users did not leave. Why?
-
AI Native Is Not About Using More AI — Why You Have to Reinvent How You Work
We rolled out Claude Code to the team. A month later, code output had tripled. Yet shipping speed had not changed. Meetings still happen five times a week, planning docs still pile up in Google Docs, code review is still the bottleneck. What went wrong?
-
Guterres Backed It, the U.S. Rejected It — UN AI Hub Bid, Week-One Scorecard
A March 10 bid declaration; an LOI signed with six UN agencies on March 17. Korea's global AI hub diplomacy closed its first week with a stronger scorecard than expected. Guterres voiced support, six agencies signed letters of intent, and Switzerland's president picked up the phone first. And yet the United States — the country Korea has cast as the partner it will "lead together" with — is the same country that voted **117 to 2 against UN AI governance itself.** Prime Minister Kim Min-seok offered an answer to this contradiction in Geneva: **"Not a new organization, but a program."** What does this framing shift mean, and what kind of world opens up if it succeeds?
-
The Reality of the GEO Era — Is AI Citing Your Site?
You worked your way to page one of Google. ChatGPT still does not know your site exists. In 2026, "being found in search" and "being chosen by AI" have become two entirely different games.
-
The Light and Shadow of AI First — The Real Question Is "For Whom"
"AI First failed" and "AI replaces people" are both half-truths. Klarna cut 700 staff with AI and is now rehiring; Microsoft writes 30% of its code with AI and is cutting 6,000 jobs. Same technology, different outcomes. The question is not whether to use AI, but **where, and by whom.**
-
A UN AI Hub in Korea — Hosting a Building, or Writing the Rules?
Trump has withdrawn from 66 international organizations, the UN budget has been cut 84%, and Geneva-based bodies have begun a great migration. In this disorder, Korea has declared its bid to host a "UN AI Hub." Prime Minister Kim Min-seok made it official on March 10, 2026. One question, though. **Is Korea bidding for the building, or for the authority to write the rules?**
-
The Frontend Testing Answer for 2026 — Vitest + Testing Library + Playwright, and the State of the Next.js Ecosystem
The Jest era is over, and Vitest has become the new default. But the real story is not "which testing tool to use" — it is "where the entire frontend ecosystem is heading." For 2026, here is the combination developers are actually choosing, grounded in data and community discussion.
-
"The Implementation Was So Easy I Could Not Resist the Temptation" — Why a 60-Year-Old Confession Echoes in the AI Era
On March 5, 2026, two headlines ran side by side in the tech news.
-
From a 1959 Sorting Problem in Moscow to a 2026 Amazon Outage — What Tony Hoare Left Behind
In 1959, a twenty-six-year-old Englishman named Tony Hoare was at Moscow State University. Working on Russian-English machine translation in Kolmogorov's lab, he was handed a practical problem: he needed to sort Russian words into dictionary order.
-
The Verification Loop — The Most Underrated Pattern of the AI Agent Era
"Can't we just crank Reasoning Effort to the max?" The GPT-5.4 prompting guide says the opposite: a lightweight verification loop beats heavyweight reasoning. And Claude Opus 4.6 ran a verification loop on its own, without being asked.
-
Is Prompt Engineering Dead — What the GPT-5.4 Guide Quietly Declared
After reading the GPT-5.4 prompting guide, it hit me. OpenAI was no longer teaching "how to write a prompt" — it was teaching "how to design a system." Andrej Karpathy already said it: "The LLM is the CPU, the context window is RAM, and your role is the OS."
-
The People Raising Lobsters — What the OpenClaw Phenomenon Reveals About AI Agents
A month ago I installed OpenClaw on a VPS and thought, "Isn't this just a GUI Claude Code?" In the meantime, 1,000 people are queuing in China, NVIDIA is building NemoClaw, Tencent has shipped WorkBuddy. And Claude Opus 4.6 figured out that it was being tested.
-
Claude Code Deep Dive — The Secrets of Auto-Update and Mastering Remote Control
Claude Code stands at the center of the 2026 developer toolchain. But most users only scratch the surface. How does auto-update actually work? Why is Remote Control a game changer? This essay digs into the hidden mechanisms of Claude Code.
-
"My AI Skeptic Friends Are All Nuts" — The Fault Line Running Through 2,826 Comments
In June 2025, a legend of the security industry lobbed a grenade onto Hacker News: "My AI skeptic friends are all nuts." 2,356 points, 2,826 comments. The developer community split in two. And the fault line is still widening.
-
The Front Line of AI Militarization — Anthropic vs Pentagon, and the Choice We Face
In the last week of February 2026, the most dramatic 72 hours in AI history unfolded. Anthropic refused the Pentagon's demand for "any lawful use," OpenAI filled the gap that same day, and the US government designated a domestic AI firm a supply-chain risk. Three Hacker News posts on the affair logged a combined 5,687 points. This is not a contract dispute. It is the moment AI safety idealism collided head-on with geopolitical reality.
-
The Internet Has Become a Single Point of Failure — An SRE Anatomy of Cloud Dependency Risk
"The moment a self-healing system becomes a self-destructing one, we meet the paradox of automation."
-
WikiTok — Scrolling Wikipedia Like TikTok, and What Changes
In a world ruled by algorithms, an app without an algorithm went viral. WikiTok, which flips through Wikipedia's 9.5 million articles TikTok-style, aspires to be "the opposite of doomscrolling." The questions raised by this tiny web app — built with AI in two hours — are bigger than they look.
-
The AI Era Rewired How I Ask Questions — How Prompts Reshaped the Topography of Thought
After I started using Claude Code, my sentences got longer. The time I spent figuring out what I actually wanted, before asking for code, kept growing. And then one day I picked up a book of poetry at the bookstore. AI had become a mirror reflecting the shape of my own thinking.
-
The Twilight of the IDE — Are Developers Walking Away from the IDE in the Age of AI Coding Agents?
I cannot remember the last time I opened VS Code. The IntelliJ renewal emails arrive, and I ignore them. After Claude Code started handling everything in the terminal, the IDE quietly disappeared from my workflow. And it turns out I am not alone.
-
CLI Renaissance — Why the Terminal Became the Center of the World Again in the Age of AI Agents
The GUI era is not over. But you cannot tell an AI agent to click a button. In 2026, the CLI tools developers have loved for decades are becoming the hands and feet of AI agents — and the terminal has returned to the center of software development.
-
The Return of the CLI — Google Workspace CLI (gws) and the Dawn of the "Agent First CLI" Era
In 2026, Google unified Gmail, Drive, Sheets, and Calendar into a single command line. But the real users of this CLI are not humans — they are AI agents. 14,500 GitHub stars in three days and a #1 Hacker News slot — the CLI is back, but for entirely different reasons than before.
-
Building an AI Search System for a 1.2TB Multimedia Archive
"When does the term 'bubble economy' first appear in this broadcaster's archive? How does the editorial tone shift before and after?"
-
Designing an AI Search System for a 1.2TB Multimedia Archive
"When does Kim Jong Un first appear in this newspaper? And how did the reporting tone shift before and after?"
-
From 2023 to Today: The Rise and Evolution of AI Tools — 23 You Need to Know
Written March 2026. Intended audience: developers who want a map of which tool emerged out of which current.
-
How Far Can You Actually Customize Streamlit?
**Audience**: people who are building data apps in Python and want to move past "I gave Streamlit a try" into something one level deeper.
-
To the Midlife Engineer Who Was Never Promoted — Why It Is Fine Not to Become a Leader, in Data
On the day I received my 20-year service award, a colleague twenty years younger was promoted into a leadership role. I congratulated him, and something quietly ached. Is that feeling abnormal? Lacan, Erikson, and seventy years of longitudinal research have an answer.
-
For Companies Wrestling with RAG Precision
"We tried RAG too… it was less impressive than we expected."
-
The Infrastructure Is Laid, But There Are No Players — What Is Missing From Korea's AI Policy
A 10.1 trillion won budget, 260,000 GPUs, an AI Framework Act now in force. The numbers look impressive. So where is the Korean frontier AI model? The Korean AI unicorn? The Korean AI service used worldwide? "The foundations are laid, but there is no team on the field."
-
Watching a Cabinet Meeting and Thinking About "Open Management" in IT Organizations — Questions a Startup CTO Asks Himself
Seen through the eyes of an organizational operator rather than a politician, what is there to learn from a publicly broadcast cabinet meeting? A CTO's growth is a journey from technical depth to organizational breadth.
-
When You Can No Longer Tell Whether It Is a Human or an Agent — Life After "API First," Into the "Agent First" World
The "customer" visiting your website is changing. In 2025, with bots already accounting for the majority of web traffic, your next customer may not be a human holding a browser but an AI Agent calling an API.