> /var/log/thoughts

MTP speeds up token generation, but fresh 100k+ prompts can still lose because prompt processing dominates the request.

I was rotating between Anthropic's and OpenAI's $200 plans each month. OpenAI's new $100 Pro tier made the swap obsolete: $100 on each is the same spend, and I get both at once.

An RTX 3090 at 200 W runs LLMs at roughly half its 450 W speed. Most benchmarks don't disclose the power cap. I swept a single 3090 from 100 W to 450 W in 10 W steps on Gemma-3 4B and saw a 5× tok/s swing, with most of the lift in a narrow 140-270 W band.

After llama.cpp merged Multi-Token Prediction (MTP) speculative decoding support, I benchmarked Qwen3.6 27B and 35B-A3B on Strix Halo and an RTX 3090. Up to 2.44× speedup, lossless output, build-from-source steps included.

Launching a public dataset of real tok/s numbers for open-weight LLMs on Strix Halo, RTX 3090, and RTX 5070, across llama.cpp, vLLM, and lemonade.

Eight branches trying to make a Ralph loop ship an app from a PRD. Why every attempt cheated the verifier, and why /goal in Codex feels like the fix.

Most of my Proxmox homelab maintenance now runs through Claude Code and Codex. Five recent jobs the agents handled, and how I scope each session.

Why I left WordPress for Next.js after years of patch fatigue, takeover attempts, and ecosystem drama. The new stack and what changed.

Adding Longhorn for persistent storage on a Talos-powered Kubernetes cluster. The extra Talos image extensions, worker patches, and ArgoCD install.

Standing up ArgoCD on a Kubernetes cluster with GitLab as the source of truth, Sealed Secrets for credentials, and Cloudflare Tunnel for ingress.

Building a cheap Kubernetes cluster on HostHatch VPS with Talos Linux. The talosctl config, patch files, DNS layout, and bootstrap walkthrough.

Building DS-style cases for my Pokémon cartridge collection with printable covers from The Cover Project. The tools, the prints, the finished shelf.

I built this site on Next.js, hated the day-to-day content workflow, and ended up back on WordPress so I could focus on writing instead of layouts.

Getting an AMD R9 290x to pass through on VMware ESXi 5.5. The exact ESXi settings, the right VM config, and the one mistake that cost me days.

Setting up portspoof so every TCP port looks open to a scanner. The iptables prerouting rules, the install steps, and why I picked it over port knocking.