Nexus: Depth-Adaptive KV-Cache Splicing and Retrieval-Decoupled Tool Routing for Agentic LLMs on Unified Memory
Nexus tackles a real agent bottleneck on MCP-style tool calling: every turn can require re-encoding huge tool schemas, making time-to-first-token balloon as the registry grows. It swaps schema-heavy routing for an INT8 semantic lookaside buffer with a margin gate, then generates arguments over a compressed textual signature while using a depth-adaptive repair when RoPE phase drift would otherwise break output fidelity.
How to encourage smarter AI use in the classroom
MIT Technology Review’s Making AI Work brief digs into how schools can shift from “avoid the chatbot” to “design for responsible, high-signal use,” including policy ideas that steer students toward better learning outcomes rather than copy-paste outputs. It highlights the practical need for guardrails that shape behavior early, while acknowledging that enforcement alone can’t replace thoughtful classroom integration.
PrimeAgentOrchestrator: Memory-Primed Agent Spawning for Personal AI Infrastructure
This experience report proposes PrimeAgentOrchestrator (PAO), which starts each coding-agent session with a fresh context window but re-injects relevant personal memories by querying two separate backends in parallel. The system fuses PostgreSQL entity-observation retrieval with a Cloudflare Worker semantic index, then delivers the briefing to Claude Code via filesystem injection, focusing heavily on lifecycle management and failure-mode-driven redesign over four months of deployments.
GPT‑5.6 is now available in Kiro
OpenAI announces GPT‑5.6 in Kiro, positioning it as a price-performance upgrade for developers planning, building, reviewing, and testing software. The release is aimed at making iterative engineering workflows faster and cheaper without sacrificing the model behaviors teams rely on for practical coding assistance.
Kids outlearn AI—and we still don’t know why
MIT Technology Review takes on the enduring gap between child language acquisition and today’s AI systems, arguing that “just scale the model” hasn’t explained why kids learn more efficiently. The piece connects the mystery to how learning signals, interaction, and embodied or social scaffolding may drive language progress in ways large language models still don’t capture.