When Instagram Calls a Real Human “AI”
Instagram's AI labels can reflect a minor edit or a provenance signal rather than full synthetic authorship, leaving human creators to carry the reputational ambiguity.
Blog
Notes on research, web builds, and community work, stored and synchronized through Convex.
AI Slop Made “Human-Made” a Selling Point
Human-made labels are becoming useful trust signals as platforms and audiences resist synthetic repetition, but provenance cannot guarantee quality.
What Happens When AI Agents Get Their Own Social Network?
Moltbook produced millions of agent comments and familiar attention patterns. A 47.9% formulaic-comment estimate shows what machine-scale social activity can hide.
X Is Replacing Revenue Sharing. Did Repost Culture Break the Incentive?
X is replacing Creator Revenue Sharing with Original Content Rewards. The policy targets copying and weak aggregation, but its cultural effect still depends on enforcement.
Anthropic Added Invisible AI Watermarks. Removers Arrived Before the Detector
Anthropic's marking plan triggered a fast remover market, but the public tools can only prove Unicode and metadata cleanup while Claude's detector remains unavailable.
The “Worst Animation Ever” Went Viral. Being Human-Made Changed the Joke
Niu Lai's ridicule drove a box-office reversal; its reported two-person production story later changed how audiences read the rough animation.
Why Are Some Claude Code Users Suddenly Burning Through Their Limits Faster?
A multi-source investigation separates plan usage, context occupancy, MCP attribution, and cache misses to explain why Claude Code can appear to burn limits faster.
Does /compact Secretly Make Claude Code Worse?
A controlled Claude Code 2.1.233 proxy compares focused compact, a fresh session, and a written handoff. The result is narrower and more useful than the rumor.
Changing AI Models Mid-Session Can Cost More Than Staying on the Expensive Model
A cache-aware break-even model shows when an Opus-to-Sonnet switch saves money, when it costs more, and why preserving a conversation does not preserve its cache.
Why 100 Agent Skills Can Be Worse Than 5
A controlled study found exact skill-use precision collapsing as candidate pools grew, while task success stayed flat. The fix is better routing, not a deletion spree.
Xiaomi XRING O3 Is Official. Its Benchmark Case Isn’t.
Xiaomi has officially announced XRING O3, but its performance case still rests on launch claims. O1 tests and supply-chain evidence show the benefits and limits.
Pay-to-Rank Websites Sell a Spectacle, Not a Stable Ad Slot
Hundreds of pay-to-rank sites followed Outbid.lol. Their mechanics can create instant attention, but public data shows thin clone receipts and major measurement, moderation, and payment risks.
The Harness Is Becoming More Important Than the Model
Controlled studies and DeepSeek Harness show how tools, context, recovery, tracing, and safety can change an agent's cost and results without changing its model.
DeepSeek Harness Stuck After npx: The pnpm dlx Wrapper That Worked
A grounded troubleshooting case study on the stalled npx prompt, the pnpm dlx route that worked on one machine, and the pinned wrapper that made dsh discoverable.
Who Is Behind Ox Alpha? The API Left a Trail
Ox Alpha arrived without a maker's name. Tokenizer matches, an exposed Java class, and a corrected 58.4% DeepSWE run tell a more careful story.
GPT-6 Astra: Who Started the Rumor, and What Is Actually Known?
An evidence-led timeline of the GPT-6 Astra rumor, from the earliest traceable public post to the claims that grew beyond OpenAI's own record.
E-Ticket TNKS: Reviewing a Mountaineering Booking from Route Selection to Check-out
A technical review of the Laravel 11 e-ticketing system for Kerinci Seblat National Park, covering gate capacity, group records, payments, tickets, and its mobile API.
GenBI Jambi: When a Rebrand Became an Organization's Working System
A review of the GenBI Jambi website as it grew into a CMS, member directory, attendance system, activity points ledger, book catalog, and role-based finance platform.
Stok Toko: Reviewing an Android POS That Chose Offline First
A technical review of a native Android app for catalog management, checkout, shopping lists, barcode scanning, and stock assistance, built with Kotlin, Jetpack Compose, and Room.
Mukhtada's Portfolio
A record of fullstack work, AI experiments, data research, publications, and community projects, presented as a factual game world.
Caelestia Island Suite
An integrated Caelestia Shell overlay for Hyprland that combines the Dynamic Island, Bottom Notch, enhanced settings, and native Task Manager in one repository.