lucumr-pocoo-org

Newer Claude models sometimes invent extra keys in tool call arguments, breaking validation in Pi's edit tool. The author suspects post-training for Claude Code's forgiving harness makes alternative schemas fail. This suggests closed RL training can degrade general tool-use reliability.

Communities formed around rejecting something can shift from legitimate criticism to identity built on opposition. When a member makes a personal choice contrary to that rejection, the community often treats it as betrayal and unleashes collective harassment. The author warns against letting negativity define one's identity and encourages openness and de-escalation.

The author argues that large tech companies like Apple and Anthropic frame access restrictions as safety measures to control narratives around openness. The EU's DMA matters because it preserves user agency over their devices and data. True democratized access to technology, including AI, benefits everyone, even if it causes temporary product delays.

The US government ordered Anthropic to block foreign nationals from accessing its AI models. The author argues this shifts AI regulation from safety to nationalist control, treating technology as a weapon for Americans only, and warns Europe to build its own capabilities rather than rely on regulation alone.

The author defends using "clanker" over "agent" for LLM systems to avoid anthropomorphizing machines and misplacing responsibility. He argues machines are tools, not people, and warns that treating them as oppressed beings blurs lines around actual human racism.

The creator of Pi reflects on building the project with AI agents. LLM-generated "slop" issues—confident but inaccurate diagnoses and over-engineered code—create extra maintenance work. The project uses Pi itself to parallelize issue reproduction, but the author warns that AI-driven local workarounds fragment collaboration and undermine open source.

Local models for coding agents remain impractical due to fragmentation, poor tool streaming, and excessive configuration. The author advocates focusing on one model-engine-hardware combo, highlighting ds4.c (DeepSeek V4 Flash on high-end Macs) integrated into Pi as a step toward a polished, zero-config local experience.

The author shows that LLM-generated text is inflating usage of certain words (like "substrate") in coding sessions and Google Trends, and argues this AI slop is flooding platforms, eroding trust, and making human interaction harder to distinguish from machine output. They call for more friction, transparency, and rate-limiting in digital communication.

The author reflects on Open Source before GitHub, when projects ran on self-hosted servers and dependencies carried social weight. GitHub made publishing and discovery frictionless but its decline now threatens the archival memory of Open Source. The piece calls for a neutral, well-funded archive to preserve code and project history independent of any single company.

The author explains why "equity" — combining fairness, financial residual value, and ownership — is hard to translate into German, where multiple domain-specific words exist instead. He argues this linguistic gap affects how Europeans think about debt, wealth, and agency, and suggests adopting a clearer equity concept could improve discourse on ownership and long-term decision-making in Europe.

The article explores how AI coding assistants create unhealthy dependencies and poor-quality contributions. Developers form parasocial relationships with AI agents, producing "slop" code that burdens maintainers. The author warns about addictive patterns while acknowledging AI's productivity benefits when used thoughtfully.

The author and Colin have founded a new company called Earendil in Vienna, incorporating as a public benefit corporation. They aim to create software and open protocols while strengthening human agency and cultivating lasting joy. The company represents a counterbalance to Silicon Valley's approach, with both founders committed to leaving the world better than they found it.

Pi is a minimal coding agent with a short system prompt and only four core tools. It features an extension system that allows agents to write and run code to extend themselves. Pi serves as the foundation for OpenClaw and other agent systems, emphasizing software that builds more software.

The author argues that AI agents will drive new programming languages designed for agent use. These languages would prioritize explicit typing, local reasoning, and greppable code over brevity. The decreasing cost of coding makes new language adoption feasible despite potential underrepresentation in AI training data.

The article discusses how AI-assisted coding has dramatically increased code creation speed, creating bottlenecks in code review and maintenance. It draws parallels to historical industrial bottlenecks and questions whether current approaches can scale sustainably as machines generate code faster than humans can review it.

The article discusses how AI tools can reimplement software libraries by using only their test suites, creating functionally similar but legally distinct code. This raises questions about copyright, licensing, and whether such AI-generated reimplementations constitute derived works or new creations.

The article argues that certain valuable things like mature trees, trust, quality software, and community relationships require significant time to develop and cannot be rushed. It critiques the modern obsession with speed and instant gratification in tech, suggesting that friction and time constraints often serve important purposes for building lasting value.

Absurd is a durable execution system built entirely on Postgres that has been running in production for five months. The core design held up well, with added features like decomposed steps, task results, CLI tools, and a web dashboard. The system uses checkpoint-based replay and pull-based scheduling, making it simple to operate and debug.

Mario Zechner is joining Earendil, bringing his coding agent Pi to the company. The author expresses excitement about Mario's focus on software quality and thoughtful design, which aligns with Earendil's goals for building more deliberate machine entities like Lefos.

The author argues that meaningful criticism of new technologies like AI coding agents requires direct engagement and experience. While both enthusiastic adopters and outright rejecters have biases, the center position appears biased toward engagement because forming a measured opinion necessitates serious use of the technology.