Writing.
An archive of whatever I was building, breaking or figuring out at the time, jailbreaks and all.
All categories
2026
- A First Look at Jev
A first look at TypeSafe's Jev, the model that answers questions instead of writing. The cheapest paid LLM on OpenRouter matched it on accuracy for a fifth of the price, but Jev's probabilities were the best calibrated as they came, and twenty questions about one document cost a tenth per answer of asking one.
AI 23 Sept - The State of Local AI Image Generation
Four open image models on one RTX 5090 against Nano Banana Pro and GPT Image 2.5, ten everyday prompts, scored blind. Three local models made nine usable images out of ten, and Z-Image Turbo did it in 2.6 seconds each. The newest, Qwen-Image-2.1, made six. Side by side, I still preferred a cloud image for nine prompts in ten.
AI 21 Sept - Qwen3.8-27B vs Opus 4.7 Bakeoff
Qwen's model card puts its 27B above Opus on SWE-bench Pro. At Q4 on one RTX 5090, inside three different coding agents on three scored tasks, the local model matched Opus 4.7's scores and took 1.6 to 2.9 times as long to get there.
AI 11 Sept - MiniMax H3 on One RTX 5090
Measured generation time, memory use and output from MiniMax H3 checkpoints and turbo LoRAs running on a single RTX 5090.
AI 3 Sept - An AI Agent Overclocked My RTX 5090 Overnight
A coding agent tuned an RTX 5090 overnight. A perplexity check caught unstable settings that ordinary benchmarks missed, and the profile it shipped improved decode speed by 4.2%.
AI 28 Aug - Quantising Qwen3.8-27B: Where Quality Plateaus
How five quantisations of Qwen3.8-27B performed on executed coding tests, how their generation speed and context capacity changed, and which precision levels produced a measurable return on one RTX 5090.
AI 27 Aug - How to Actually Export Your WhatsApp History
Seven ways to get fourteen years of WhatsApp messages off an iPhone, six of which failed, and the one unencrypted migration file that finally worked without losing anything on the phone.
Software 13 Aug - MegaETH's Ten Milliseconds, Measured from London
MegaETH advertises mini blocks roughly every ten milliseconds. Measured from London over public endpoints, submit plus receipt averaged 435 milliseconds in the best run, indistinguishable from Base, and the wallet path dwarfed both chains.
Blockchain 23 Jan
2025
- A Throwaway Windows Desktop an AI Agent Can Drive
A small provisioning tool turns EC2 instances into disposable Windows GUI workers that AI agents can operate independently, inspect live and destroy after use.
AI 4 Dec - A Fee That Starts at 99 Percent
A swap fee that opens at 99 percent and decays one point every six seconds to a floor of one percent looks absurd as a fee schedule. It is an anti-sniping mechanism, and it works because bots are patient about everything except time.
Blockchain 21 Jun - What Is RAG, and Why Can't You Just Give an AI All Your Documents?
A local AI experiment comparing one huge prompt with vector search, keyword search and a small RAG pipeline.
AI 12 Jun - Running a Code4rena Audit
What it took to put two Solidity contracts in front of 937 wardens for seven days, from the code walkthrough to the final report.
Security 10 Jun
