Latest posts
View all posts
Do language models need sleep?
A CMU/UMD paper shows that what limits SSM-attention hybrids on deep reasoning isn't memory capacity but the number of passes over the context. A 'sleep' mechanism runs N recurrent passes at window eviction and consolidates the context into fast weights, moving computational depth offline.
WebMCP: how your page talks to an AI agent through navigator.modelContext
navigator.modelContext still returns undefined, but behind that word hides a new way for the web to expose tools straight to an agent — no server. Turn on the flag, register a tool, and test it.
/goal in Claude Code: keep a session running until the condition holds
Anthropic shipped the /goal slash command. A session-scoped wrapper over a prompt-based Stop hook: you write a condition, a small evaluator reads the transcript after each turn, and the session ends only when it says yes. When to reach for /goal, when for /loop, when for a Stop hook — and why you write the condition for Haiku, not for the main model.