Cosmic Rundown: GPT-Live, OpenBSD Root Exploit, and GitHub Agent Leaks

GLM 5.2 and the Coming AI Margin Collapse

The thread is worth reading in full. The core claim: as open models get good enough for most tasks, the premium model providers face serious margin pressure. For developers, this validates the multi-model routing pattern (route each task to the cheapest capable model) that Coinbase demonstrated at scale this week.

Small AI Models Gain Traction in Low-Connectivity Environments

227 points on the local/small-model story. Developers in areas with unreliable networks are running capable models locally. The implication: your content layer needs to work regardless of which model or compute environment the agent is running in.

A Global Workspace in Language Models

Anthropic research on a shared global workspace for LLMs. Conceptually adjacent to the durable-agent-memory problem: agents need shared, persistent working state. Citable Anthropic source for agent-memory content.

Pruning RAG Context

kapa.ai on reducing RAG context to only what the answer needs. Structured content means agents retrieve exactly what they need instead of dumping everything into context. Direct support for the token-efficiency argument.