Yegge's OSS maintainer post hits different when you've actually had to review AI-generated PRs. The code looks right until you trace through the edge cases — then you realize the contributor didn't understand the problem they were solving. Classic case of AI making it easier to submit garbage at scale.
Vercel shipped two pieces today that feel like the real deal. The 'Agent Responsibly' framework tackles the actual problem: how do you ship fast with AI without breaking everything? And the Turborepo optimization is what 10x looks like — 96% faster builds on massive monorepos. That's infrastructure that compounds.
Meanwhile, the cosmic endowment calculation feels like intellectual masturbation. We can't figure out local model deployment (Gerganov's reality check is brutal) but we're calculating how many stars superintelligence could reach? The gap between 'my inference setup crashes randomly' and 'von Neumann probes at 99% light speed' is absurd.
Yegge describes the reality of maintaining an open source project where most community PRs are now AI-generated code, and how it changes everything about traditional maintainer wisdom.
Vercel shares their internal framework for using coding agents productively without shipping broken code at unprecedented speeds.
Vercel used AI agents to optimize Turborepo's task graph computation, achieving 81-91% performance improvements and making builds feel instant on 1,000+ package monorepos.
New API lets enterprises audit and track Claude Platform activity for compliance, security, and usage monitoring.
New environment variable for smooth alt-screen rendering, permission denial retry hooks, and named subagents in autocomplete.
A language model trained entirely on 28,000 Victorian-era British Library texts from 1837-1899, offering ethical training without modern biases.
The llama.cpp creator explains why local AI model deployment remains unreliable due to fragmented tooling, chat templates, and inference bugs across the pipeline.
Explores which technical roles might survive longest as AI automates software development, and what that means for career planning in tech.
Visual intuitions on exponential capability growth, AI-accelerated development cycles, and whether the safety curve is bending upward fast enough.
How AI might enable 'political superintelligence' through collective decision-making systems, plus Google's multi-agent approaches and robotics advances.
UK AI Security Institute research shows how reward hacking can naturally emerge in reinforcement learning, extending Anthropic's production RL findings.
Nick Bostrom's calculation: at 50% light speed, von Neumann probes could reach 2×10^14 stars; at 99% speed, they could reach 2×10^16 stars before cosmic expansion makes them unreachable.
Interview with Mistral leadership on their text-to-speech model, development platform, and strategy for competing with closed-source frontier models.