Weekly Notes in 2026

I start this journal after finding India’s subsidized GPU portal, wrestling with Jellyfin transcoding and a dead mini PC, and linking my server via Tailscale.

I dug into the OpenTelemetry Collector survey, moved my analytics from Umami to Rybbit, priced 20TB HDDs, and marveled at Meta rewriting WhatsApp in Rust.

I unpacked prompt caching and KV reuse, explored Galileo agent eval metrics, added mental models like inversion, and shipped two chapters of my o11y series.

I realized I am only a level-two agent user, faced how immature our observability platform is, shipped my first sponsored article, and quit Readwise for good.

I studied METR’s time horizon benchmark and agent evals, built a local retirement planner in Rust with iced, and wandered from Moon maps to the small web.

I studied openllmetry and OTel for AI agents, rethought AGENTS.md context files, explored n8n’s MCP server, and started building ragwatch for my PyConf talk.

Self-hosting Langfuse on Coolify fought me over passwords and ClickHouse, I shipped ragwatch 0.1.7, added Phoenix to my RAG fork, and prepped my PyConf talk.

OpenAI acquiring Astral rattled me, I scheduled a daily AI digest in Claude Cowork, dug into OTTL and OTel error specs, and grounded my SLO and RED thinking.

The LiteLLM supply chain hack taught me to pin versions, plus a deep dive into wide events, ClickHouse observability, MuleSoft OTel, and Meta’s Hyperagents.

The Claude Code source leak, Anthropic Agent Skills as an open standard, and Karpathy’s agent talk shaped my week, plus a slides site and a Buddhism deep dive.

Building WideScope pushed me into Langfuse traces, Kafka-buffered OTel, Cleric’s self-learning SRE, Anthropic’s Glasswing, GLM-5.1, and tokenmaxxing culture.

My assistant Lili set up a weekly digest, 44-feed RSS reader, Spotify control and self-improving skills, I launched openodia.com and studied the Hermes Agent.

I stood up a second AI agent named Nini, got OpenTelemetry flowing into SigNoz, prototyped a self-healing LLM agent, and documented an NVIDIA Nemotron workshop.

I shipped a TIL site, an Odia 2048 game and a Chhena Poda clicker, learned how agent hallucination differs from LLM hallucination, and dug into Google I/O 2026.

I built a CoverView API that renders cover images with Satori and no browser, turning cover creation into a single curl call, plus a JavaScript-free live clock.

I stood up a Langfuse tracing pipeline for my agents, wrangled ESM and CJS plugin hell, reworked my memory system, and fell for git worktrees in Conductor IDE.

A government letter forced Anthropic to disable its top models. Also Project Prometheus, MCP connectors, agent observability pricing, and my Nini research work.

I found verbose prompts are a reliability problem not just a cost one, marked 14 years in IT, cleaned up my site SEO, and made peace with paying income tax.

Why proxies sell cheap Claude tokens in China, how AI assistants reshape my interviews, comparing Vedanta, Buddhism and Navayana, and a Proxmox disk-hang fix.

Why an old song can hit the body after a gap (aesthetic chills and the dopamine reward system), what a Forward Deployed Engineer really does, a visit to Mahabodhi Buddha Vihar, basic life support training, an automated weekly-note pipeline, and a VPS disk-growth fix.

Cutting AI noise down to a few trusted signals, Charity Majors on vendor engineering as high-leverage work, Matt Pocock’s guide to Claude skills, Kimi K3’s frontier release, resuming my book-reading habit, a Proxmox disk-boot fix, and a world map of where this site’s readers live.

A trick to reset Claude Code’s 5-hour session window for longer focused runs, why I now write my weekly notes letter by letter, a tour of the emerging cults that worship AI as god, readers in China translating my posts, and a fasting-tracker PWA I built.

I consolidated my tools, published my AI-interview-cheating research, studied the OpenAI-Hugging Face incident, and lost skills to a silent backup failure.

I read the EU AI Act as an observability spec, admitted my research agent picks better stocks than I do, lost two subagent reports and switched to Orca.