Random Learning

A daily learning ritual

What I learned, one day at a time.

A daily learning ritual: three researched topics a day, drawn from a private library of saved links, each grounded in what people said this month, and wired to past entries so the learning compounds. Never the same topic twice.

69
Days
207
Topics learned
831
Themes explored

Latest

What actually limits a Meshtastic mesh, and why people are flashing MeshCore

…and What FamilySearch's AI handwriting search opened, and what it quietly closed · What people actually do with a read-it-later backlog

Read the day →

The journal

September 2026

What actually limits a Meshtastic mesh, and why people are flashing MeshCore
What FamilySearch's AI handwriting search opened, and what it quietly closed · What people actually do with a read-it-later backlog

August 2026

Record linkage with Splink and how identities get matched across messy databases
Kdenlive and what editors say after finishing real projects on it · FFmpeg hardware AV1 encoding on NVENC and what it costs in quality
What actually still works for reading X without an account
Headscale and what breaks when you self-host the Tailscale control plane · What ChatGPT referral traffic is actually worth to small sites
What Windows 11 debloat tools actually remove, and what comes back
What arXiv's AI moderation rules catch, and what still gets through · What mandatory Companies House identity verification has actually changed
WiGLE and what a crowdsourced Wi-Fi map exposes about home routers
Where independent search engines get their index after the Bing API shutdown · Why AI-driven NPCs still have not shipped in real games
US customs manifest data and the form that hides you from it
AI weather models once forecasters started deferring to them · Whether data-broker deletions actually stick under California's DELETE Act
AI companies destroying the books they scan
Japan's overtourism rules arriving as visitor numbers fall · An autism spectrum too broad to be useful
Code signing and notarization for indie desktop apps
Preference post-training for small models · Devices bricked when companies shut down their servers
What happens to your data when a company dies and its records are auctioned
What a coding agent status line actually costs to run · Cloud vendor lock-in and what leaving actually costs
What data centers actually do to your electricity bill
What actually keeps an archive alive: the medium, the vendor, or the volunteers · What archives and small sites do about AI crawlers hammering their servers
Why every cause of hallucination collapses into the same four geometric shapes
What self-hosters actually run to get a homelab alert to their phone · Whether AI-text detectors catch anything and whether humanizers beat them
How stale light pollution maps are and whether Bortle matches what observers see
What it costs and what breaks running a personal staff of agents on your own VPS · What solar shade reports miss that an on-site survey catches
What a quoted hallucination rate actually measures
What breaks in full-duplex voice agents: barge-in and end-of-turn · What actually happens when you assign a real ticket to a coding agent
What coding agents actually run inside: sandbox runtimes, prices and escapes
What happens when AI-designed proteins reach the bench · How speech-native models actually represent audio
AI-found vulnerabilities: what got verified versus what got claimed
How AI products get priced once inference is the cost of goods sold · Whether text-to-CAD actually produces parametric models you can edit

July 2026

What teams actually report after running spec-driven development with coding agents
How hiring teams screen engineers now that the resume stopped carrying signal · What canvas editors and timeline UIs actually get built with
Which AI video upscalers people actually run locally
What local-first sync actually costs to run · Code ownership and blame when an agent wrote the commit
Ensemble AI code review: does a second reviewer find real bugs?
Quantization quality cliffs: where 4-bit stops being good enough · Why server-sent events came back for streaming LLM apps
Harness vs model: what actually moves a coding agent's benchmark score
Buy-it-for-life: how people decide what is actually durable now · Whether the junior developer entry path still exists
Stop conditions: when a coding agent loop is actually done
What people wire up so agents can read PDFs, images, audio and video · Copy-in components vs headless libraries after shadcn's Base UI switch
Where tokenization bites: the LLM failures engineers blame on the tokenizer
What agent memory people actually keep between sessions · Terminal demos as build artifacts: VHS, asciinema, and agent screencasts
LLM-as-a-judge: trusting model-graded evals
Print-in-place compliant mechanisms and 3D-printed flexure hinges · Typology photography and documenting vernacular architecture
Local model as the orchestrator calling out to cloud frontier models
Design tokens as the contract between AI design tools and code · AI-native video editors: prompt- and agent-driven timeline editing
MLX vs llama.cpp for local LLM inference on Apple Silicon in 2026
DeepSeek V4 vs Qwen 3.6 for local coding models in 2026 · Astro vs Next.js in 2026
Ads inside ChatGPT: OpenAI's ad business and the LLM attention economy
Interactive castle atlases and how medieval fortifications were engineered · On-device speech-to-text in 2026: Moonshine, Parakeet, and Whisper alternatives
Models that fine-tune themselves toward a target behavior
The silicon baton pass: Nvidia's AI-chip moat and who threatens it · Learning with AI vs offloading to AI: cognitive offloading and skill atrophy
Where clinical psychedelic therapy actually stands in 2026
The context layer for AI agents: governed data access over bolted-on RAG · Safety nets for autonomous coding agents: sandboxes, checkpoints, and rollback
Giving AI agents their own spending budgets and credit cards
Why AI coding agents grep instead of vector search over code · Recording talking-head and screen demos with AI video editors
Do reusable rockets actually make launches cheaper in 2026?
Real-time voice agents: latency, interruptions, and cascade vs speech-to-speech · AI 2040 and the fight over an optimistic superintelligence plan
Running Claude Code on cheaper and local models: does BYO-model actually work?
Files over apps: the plain-markdown notes revival and Obsidian alternatives · Kongo Gumi and the world's oldest companies: how Japan's shinise last centuries
Is Sanskrit really the ideal language for AI?
Could today's LLMs actually be conscious? · Building Telegram bots on serverless and edge platforms
Does EV battery recycling actually pay off in 2026?
Robotaxis vs rideshare: who's actually getting undercut in 2026 · The craft and collecting of great opening lines in fiction
The Billion Dollar PDF: documents that crystallize an era and move markets
Comprehensible input and immersion for language learning in 2026 · Synthetic data vs human-labeled data: the bifurcating AI data supply chain
Sonic branding: how brands design their audio logos and signature sounds
Taste and discernment as the scarce skill in the age of AI coding · In-browser video editors with WebCodecs: client-side, no-upload media processing
AI compute governance: can export controls and compute caps throttle frontier AI?
Tailwind fatigue and the post-Tailwind CSS stack in 2026 · Running image-generation models locally on your phone or laptop in 2026
How autonomous AI agents pay each other in 2026
AI image models vs. manga artists: the consent fight over style · Everyone is a token arbitrageur? AI leverage vs the ROI reality
Mac mini as an always-on headless home server
Online malware-analysis sandboxes and their traps · QR-code phishing (quishing) and how it now beats MFA
Japanese train station departure melodies and the composers behind them
AI agents deleting production databases and the guardrails teams add · Self-hosted Slack and Discord alternatives people actually run in 2026
AI bug-hunting vs the slop flood: what OSS maintainers face in 2026
Local LLMs people actually run in 2026: Mac mini, consumer GPUs, and the frontier gap · Organic Maps vs CoMaps: why the privacy-map community split
AI forecasting bots vs human superforecasters
The retro handheld emulation scene: Anbernic, Retroid, MiSTer · Generative art after AI image generators
Bambu Lab firmware lockdown vs the open 3D-printer legacy
Why brands turn to junk after private-equity acquisition · Phantom islands: places mapped for centuries that never existed
Gaussian splatting vs NeRF: what won for 3D scene capture
Why Chinese AI labs publish detailed papers while US labs went closed · The small web and handmade web revival in 2026
Costco's anti-Amazon contrarian model
Workslop: AI busywork dumped on coworkers · How demoscene coders fit 3D worlds into 4KB intros
Why people quit Anki and spaced repetition
AI circular financing and whether the datacenter boom is a bubble · The em dash as an AI writing tell
Porsche eFuels and synthetic fuels: real decarbonization path or greenwashing
The three-generation rule and why most family businesses die · Marine protected areas, 30x30 and paper parks: what makes fish and shark recovery work
Why mugicha (barley tea) is what Japan drinks all summer
Why almost everyone thinks they are a rare outlier · Precious Plastic's crisis and whether home recycling is worth it

June 2026

Default alive bootstrapping vs raising VC in the AI era
Why Texans treat H-E-B as more than a grocery store · Proof-of-personhood and proving you are human online
Rucking — weighted walking as low-injury cardio
Double-entry ledgers as the source of truth in fintech · Traditional Japanese nail-free timber joinery
Functional ultrasound neuroimaging vs fMRI
The who-is-online presence layer revival on personal sites · Offline-first hiking navigation apps people trust
Why anime animators are paid so little and the production-committee economics behind it
The 2026 revival of personal blogs, blogrolls and RSS as algorithm-free discovery · Driver's licenses in Apple and Google Wallet and the 'No Phone Home' privacy fight
Apple silicon Mac mini and Studio as the local-LLM inference box
Designing nuclear-waste warning markers to survive deep time · Game preservation and console emulation after Nintendo crackdowns
What makes a hand-drawn imaginary city map feel alive
Structured data and llms.txt for AI agents reading your website · Edo Japan's circular economy: zero-waste by necessity
Prompt injection attacks on AI agents in 2026 and which defenses actually hold up
What people say about Kagi search in 2026 — paying to escape Google's AI slop · Why 1-bit and dithered pixel art is having a revival in 2026
The collapsing economics of how-to and self-help book publishing as readers ask AI instead
Link rot and the digital dark age: how much of the web has already vanished · Full-body AI health scanners and longevity screening: hype versus evidence
Japanese railways as real estate empires and land-value capture
Context rot: why long-context LLMs underuse their window · Generative UI in 2026: AI interfaces that assemble per task
Aperiodic monotiles: the hat and spectre tile in physics and materials
Open-source text-to-speech in 2026 versus ElevenLabs · OpenStreetMap versus Google Maps in 2026
AI-generated papers and fabricated citations in peer review
Where classical computer vision still beats deep learning (OpenCV 5) · Shared-presence web toys and live multiplayer cursors
What people in AI-heavy jobs say is actually safe from automation
Bossware and AI productivity-surveillance backlash · Version control for large game and binary assets
If you get carsick reading, the patch wins and Apple's Vehicle Motion Cues is a real first try
Fake recruiter coding tests now run malware the moment you clone the repo · Beating phone autopilot is friction and a richer offline life, not deleting apps
AI adoption in 2026 is bifurcating, not stalling
What people actually get from Stoicism in 2026 · Co-living in 2026: investment product versus loneliness response
Running a kabushiki kaisha in Japan as a foreigner after the 2025 visa reform
What companies that kept the four-day workweek say about productivity and revenue · How car owners actually stop keyless relay-attack and CAN-injection car theft
AI data centers, the power grid, and who pays your electricity bill
Why rural Japanese train lines keep getting discontinued · What makes a data-visualization project go viral and hit the HN front page
Why third places collapsed and what is actually reviving in-person community
What hikers and Camino pilgrims actually use for offline navigation · The 2026 sweepstakes-casino crackdown and the regulatory-arbitrage playbook
The aesthetic-usability effect and why polished interfaces earn trust
Building a home GPU server in 2026: build versus rent · How addiction rewires the brain and how habits actually break
AlphaFold and Isomorphic Labs: what AI drug discovery has actually delivered
Flock Safety's license-plate-reader network and the 2026 surveillance backlash · Why airline loyalty programs are worth more than the airlines themselves
What we still don't understand about gravity: quantum gravity, dark matter, and MOND
The Sovereign Individual: why crypto and tech keep citing a 1997 book · What makes technical denim and performance fabrics actually work
Do mineral hot springs actually do anything? The balneotherapy evidence and the geology
What sand really is, and why the world is running out of it · How Japanese konbini conquered convenience, from Toshifumi Suzuki to 7-Eleven
The 1,200-year Kyoto cherry blossom record and climate change
The hippie trail and its revival: overland London to India · The history of pigments: colors once worth more than gold
Conway's Law in the age of AI agents
How fairy tales got sanitized: the dark original Pinocchio · Sogo shosha and why Buffett bet on Japanese trading houses

Explore by topic