Recent content on rocketup - blog on (almost) anything
This is the face of the European Oligarchy. Yes, Roberta Metsola is the face of the European Oligarchy. “Oligarchy” is one of the most feared words in the high circles of European institutions: oligarchs are in Russia, oligarchs are in the UAE, but oligarchs are also in Europe. In particular the leaders of the European Parliament, of the European Commission and of the Court of Justice:…
tldr Skills are over-engineered; a markdown file that redefines the system prompt + an agent integration to switch prompt does the same thing. The problem While designing zerostack (a lightweight TUI coding agent written in Rust), I needed to implement Skills, as they are common in most mainstream coding agents, but while coding them up I realized how Skills go against the minimal design of…
tldr Parallel read-only child agents with simple text input-output The problem When an agent needs to understand a codebase, it has two choices: start probing files one by one (slow, burns tokens) or dump everything into context (impossible for large projects). Either way, the main agent’s context fills up with exploration noise, decreasing coding performance. Following most mainstream…
tldr Plain Markdown on disk, auto-injected context, zero infrastructure The problem Implementing a Memory system that was on par with Claude Code, while following the UNIX-like minimal philosphy of zerostack . The design zerostack’s Memory system is a file-based store living under <config_dir>/agent/memory/ . <config_dir>/agent/memory/ ├── MEMORY.md # Global long-term memory (shared across…
tldr Ship fast + break and repair + interact with the community What is zerostack? zerostack is a minimal coding agent written in Rust, optimized for memory footprint (~20MB of RAM, optimized for low-end devices, parallel agents and cloud platforms) and performance (0% CPU usage when idle), inspired by the UNIX philosophy. zerostack ships different innovative features to try to compete with the…
Discuss this article on Hacker News In this article, I’ll show the data behind the current development of LLMs and I’ll try to formulate a hypothesis on the near future of LLMs. I. Pareto Frontier and the last two years The Pareto frontier is the set of all the values with the highest Y value on a given X value, acting as a stepped line representing the best model for a given property…