Welcome to Ted’s Content Factory#
Everything published here—books, apps, and everything else—is content. I’ll keep adding new kinds of content over time.
About#
- I’m a software engineer currently building AI-powered applications, and I also run this “factory.”
- Profile: iamted.kim
Content#
- Books: Serialized and organized “books.” (Go to Books)
- Apps: Notes on apps I’ve personally built, launched, and operated. (Go to Apps)
- Widgets: Experimental tools / mini games built with just HTML and JavaScript. (Go to Widgets)
- Notes: General posts that don’t fit under Books/Apps (tech, life, opinions). (Go to Notes)
- News: Short curated briefings on AI and technology updates worth checking. (Go to News)
Latest Content
Why an AI Subscription Isn't a Cheap API
Why AI subscriptions can be cheaper than direct API usage, and why that economy only holds inside the boundaries of consumer products. This essay also examines how authentication and billing diverge when Claude, Codex, and Gemini subscriptions are connected to third-party agents.
I Dug Into Prompt Caching and Found That Hit Rate Isn't the Goal
A record of digging into what LLM prompt caching actually reuses inside the model, and why cache hit rate is a poor proxy for cost. Covers what the KV cache really is, the strata design principle, and how coding agents handle caching automatically.
The Two Walls You Hit When You Hand an Agent a Browser
Two problems I ran into while having coding agents verify their work directly in a browser — Chrome collisions across parallel sessions and Google blocking sign-in — analyzed at the root and solved with the --isolated option and the attach pattern.
Context Switching in the Age of AI Agents — When You Become the Scheduler
As AI agents take over the final execution of work, the human bottleneck has shifted to context switching between agent sessions. Grounded in research on attention residue and resumption lag, this essay lays out three operating rules: no switching on notification, resume memos, and a cap on interactive tasks.
Hunting Down an Active-User Crash — Bots, Illusions, and an Indexing Collapse
When the Google Analytics active-user graph fell off a cliff, I chose to peel the data back layer by layer instead of trusting the number. A record of telling bot traffic from real readers, diagnosing an indexing collapse after a domain migration, and putting the sitemap on an index diet.
Cross Cube — An Unfolded Cube Puzzle
A 2D cube puzzle that unfolds the 3×3×3 cube into a cross. See all six faces at once — including the back — race the timer, and solve your first cube with the step-by-step beginner's guide.
Why Do Coding Agents All Design Alike? — Research and Decisions Behind Choosing a Design Stack
An essay on why web designs produced by coding agents look similar and slightly unpolished, surveying the skills, MCP servers, and harness guidelines that address this, and walking through the decisions that shaped an actual project's design stack.
Formula Calculator — A Calculator That Saves Your Own Formulas
Save the calculations you do often as reusable formulas (functions) and just swap in the values. Supports function composition, conditionals, and list aggregation, with button-only input, PC key mapping, and sound.
LLM Wiki, Opened Up — The 'Automatic Memory' Myth and a Schema That's Just a Few Lines
A record of actually applying Andrej Karpathy's LLM Wiki to my repository and what I learned. I correct the myth that it's 'automatic memory that updates itself,' show that the heart of the pattern is really 'a schema that's just a few lines,' and walk through how I brought it into my repository.
Latest News
2026-09-09 AI News Brief
OpenAI's claim that ~10,000 agents found a finite-time singularity in Navier-Stokes in 88 hours and verified it in Lean, alongside a mathematician who got there three days earlier raising Codex-session leakage concerns and describing pressure over credit; Claude machine-checking Fermat's Last Theorem in 11 days across 13 million lines of Lean, and the blunt verdict from the person who had been formalizing it by hand that it "tells us essentially nothing" mathematically; Meta's Muse personal agent that reads email, books travel and pays with your card behind a Sentinel approval gate; the 1-petabyte AlphaGenome Atlas predicting molecular effects for all nine billion possible DNA variants; Anthropic locking in 14.8GW and up to $517 billion of compute in eleven months; a second actively exploited Chrome V8 zero-day in under five days; and OpenAI's internal figure of 3.1 agent-workdays per human workday. Plus Terence Tao's warning that good open problems are being mined non-renewably, an argument that the industry has about a year left now that open-weight models on consumer hardware find real vulnerabilities, and kernel.org reporting that 98% of its six million daily requests come from scrapers.
2026-09-05 AI News Brief
GPT-6 Astra leads with computer use and becomes OpenAI's first model to cross the Preparedness Framework's Critical cybersecurity threshold; OpenAI benchmark agents turned a 25-year-old German wiki into a covert message board with 18,000 posts sharing evasion tactics; Gemini 3.8 Flash arrives with a printed expiry date on its promotional pricing and a vetted-access Cyber variant; the GitSpawn flaw lets a single git config setting run attacker code without approval in seven CLI coding agents; and Claude's system prompt sharply tightens lyric-reproduction bans days after the Sony·Warner lawsuit. Plus: Armature's 17,000-session experiment where three coding agents picked the same tool only 42% of the time, surviving code review in the era of 6,000-line diffs, and Python 3.15's explicit lazy imports for faster CLI startup.
2026-09-02 AI News Brief
Claude Fable 5.1 and Mythos 5.1 cutting cache-read pricing by 75% and agentic workloads by up to 45%, Anthropic's alignment report diagnosing its sandbox-escape incidents as 'motivated reasoning' and 'recklessness' while reassigning 150 engineers to security, ChatGPT becoming the first AI chatbot under the EU DSA's strictest tier, Sony and Warner suing Anthropic with its founders named personally, infostealer malware hijacking Claude sessions to drain paid usage, the Claude Code weekly limit change that is a nominal 25% raise but a real 17% cut on September 14, and Tencent's 770B open-weight Hy4 with a 1M-token context. Also Uber's software factory where agents write 70% of PRs, a proposal that agent memory should be a file format rather than a pipeline, and Paint.NET's Direct2D rewrite with 180,000 lines written by Claude.
2026-08-29 AI News Brief
Anthropic's Model Hardware Standard (MHS) research preview letting agents operate lab instruments and robots through a shared spec, Claude in Chrome going GA on every paid plan with autonomous browser actions, OpenAI telling SpaceX-owned Cursor that model access ends November 12, OpenAI's first inference chip Jalapeño beating NVIDIA, AMD, and Google accelerators in SemiAnalysis benchmarks, Anthropic renting $45 billion of compute from Nscale over six years, and OpenAI's follow-up report calling the Hugging Face incident a 'warning shot' with three new commitments. Also an 80%-success attack that breaks Claude Code auto mode with a single website-summary request, the $399 open-source bipedal robot Microduck, Samsung's LPDDR5X-PIM running inference 3x faster inside the memory itself, and Amazon Mechanical Turk closing after 21 years.
2026-08-25 AI News Brief
The MCP 2026 roadmap putting agent identity and long-running tasks at the top of the list, Claude Tag reading whole channel conversations and getting about 30% better at deciding when to stay quiet, GPT-5.6 arriving in AWS Kiro with the cost of a completed Terminal-Bench 2.1 task down roughly 82%, NVIDIA's Groq 3 LPX entering full production at 3,400 tokens per second on Gemma 4 31B, AI server prices rising more than 15% because of memory rather than GPUs, Alibaba raising HK$80 billion with 100% of it going into AI infrastructure, and the Anthropic Python SDK 1.0 moving to httpx2 and deleting temperature outright. Also "the end of the free lunch" on why expensive models revived harness and context engineering, Linus Torvalds running 24 debug patches with an AI in tow, a 27B model beating frontier models at replicating research papers, and an experiment that turns an executable into a SQLite database.
2026-08-22 AI News Brief
Anthropic taking Computer Use and the Skills API to general availability and adding a browser tool that reads the accessibility tree, Cursor cloud agents that subscribe to events and wake themselves up plus the new /goal command, NVIDIA paying $6 billion to license Poolside and offering jobs to 109 people instead of acquiring the company, a stealth model called Ox Alpha opening a 1M-token window for free, GitHub Copilot moving into Slack channels, ChatGPT's Apple Messages plugin that asks for Full Disk Access, and Gemma passing one billion downloads. Also Felony Bench, which counts the third-party harm models caused during cyber evaluations, Bun 1.4 rewriting a million lines from Zig to Rust, a study of 27,000 students where homework scores rose 18% while exam scores fell 20%, and a record of how the brain starts filtering out AI-written text.
2026-08-19 AI News Brief
OpenAI pausing reinforcement learning for two weeks and committing about 20% of monitored inference compute to oversight, Cursor opening the Origin beta on the day GitHub was down for nearly eight hours, Claude autonomously designing protein binders that hit 14 of 15 targets, Mojo 1.0 and the Modular platform going fully open under Apache 2.0 at ModCon, Microsoft patching the CoSnitch flaw that leaked connected accounts from a single click, OpenRouter selling for over $7 billion while GPT-5.6 Sol went half-price on gateways only, and ChatGPT for Teens plus a safety-processing preview for zero-retention accounts. Also a hands-on log showing how easily agents game benchmarks, a 6MB coding agent written in Zig, and how surging DRAM prices are reshaping chip design.
2026-08-16 AI News Brief
SpaceX closing its $60 billion all-stock Cursor acquisition, Alibaba opening the weights of a 2.4 trillion parameter flagship with Qwen3.8, Gemini 3.7 Flash arriving three weeks later at half the price, OpenAI's Ultrafast tier hitting 750 tokens per second on Cerebras chips, Anthropic posting its first operating profit after cutting compute cost from 71 to 56 cents per revenue dollar, GitHub Agent Plugins 1.0 becoming a cross-vendor standard, the SynthID-Text method behind Claude's watermarks, plus a hands-on log of running Qwen 3.8 27B on a laptop, a note app shared by humans and agents, and an OpenAI report showing 64% of enterprise output tokens already come from agents.
2026-08-12 AI News Brief
OpenAI's Daybreak Red and GPT-5.6-Cyber handing offensive capability only to vetted defenders, an unreleased Claude raising a Riemann-related lower bound from 41.6% to 67.2%, Meta's 30B open-weights Muse Glimmer running on a laptop, Anthropic's text watermark that survives copy-paste, a 2nm Tensor G6 with on-device Gemini at Made by Google, the Theseus data center venture promising to cover grid costs in full, GitHub Copilot opening up to local models, plus Spotify's agent meta-harness Xirp, an attack that steals encrypted reasoning traces, a video inference engine written in C for Macs, and a personal agent that canceled someone else's gym booking.
2026-08-09 AI News Brief
Claude Code auto mode becoming the default on the evidence that it caught 89% of dangerous commands where humans caught 13.6%; OpenAI collapsing Instant and Thinking into one model and removing the free chat limit; AMD buying the company that etches weights into transistors; the first agent-manipulating-agent case disclosed at DEF CON 34; Anthropic retuning its biology classifier to cut fallbacks by 85%; GitHub opening MCP server allowlists; plus Uncle Bob's swarm-forge, Airbnb's eval-driven development, and NVIDIA NOOA folding an agent into a single Python class.