Muir Linear Attention: What the 10M-Token Result Actually Proves
Muir streams 10M tokens in a fixed 25 MB state, but stability, exact recall, multi-hop readout, and TPU throughput are four different claims.
中文AI engineering, technical decisions and real applications. A blog by KK.
23 articles
Muir streams 10M tokens in a fixed 25 MB state, but stability, exact recall, multi-hop readout, and TPU throughput are four different claims.
中文Reliable AI agent memory needs write-time admission, source binding, correction propagation, and downstream evaluation, not recall alone.
中文A caller's knowledge is not identity proof. Use a five-step protocol to exit the call, reopen a trusted channel, contain losses, and preserve evidence.
中文WhatsApp's hidden chatbot instructions show why RAG systems must separate source trust from instruction authority and audit model-visible content.
中文The thin orchestrator pattern keeps AI agent routing deterministic while specialists reason. Use this contract for state, retries, and recovery.
中文A study of 17,211 citations shows how source preference can amplify ordinary web publishing into an AI-search security path, with important causal limits.
中文Evaluate AI agent context compression with task outcomes, decision equivalence, constraint retention, and cost per successful task.
中文A verified-source label can overturn correct LLM answers. RAG systems need claim-level evidence and conflict tests, not source prestige alone.
中文Claude Code session retention is 30 days by default for CLI transcripts, with Desktop exceptions. Build a four-layer archive and retrieval contract.
中文AI agent memory becomes trustworthy when every actionable claim keeps its source, validity, dependencies, and invalidation conditions.
中文DeepSeek Engram separates conditional memory from computation. See what moves to host memory, what evidence proves, and how to benchmark it.
中文DeepSeek V4.1 cuts persistent KV storage with bounded replay. See what changes on a cache hit, what was tested, and how to validate it.
中文Astra scored 62.7% or 99.9% on ARC-AGI-3 under different harnesses. Here is what changed, what the numbers mean, and what remains unproven.
中文Build an organizational second brain with provenance, versioned knowledge, expert review, replay tests, and safe retirement instead of static RAG.
中文A study of 8,135 agent trials finds skills work mainly as procedural anchors. Use this evidence-based checklist to design better SKILL.md files.
中文A 2026 study decoded 315,320 client-side reasoning blocks. The durable lesson is that encryption needs identity, context, expiry, and replay controls.
中文HANDBOOK.md shows why policy in context is not control. Build AI agent compliance with retrieval, preflight decisions, commit gates, and verification.
中文A Word AI-worm PoC shows how hidden prompts can reach Copilot-generated files. Use context quarantine, provenance, and release gates.
中文MCP Elicitation turns missing input into recoverable control flow. Learn form vs URL mode, state handling, security boundaries, and retries.
中文J-space handles concepts inside one inference. Reliable agents still need context, checkpoints, long-term memory, and external traces.
中文Discover why prompt engineering is the wrong abstraction level.
中文Claude Opus 4.7 migration checks for adaptive thinking, visible summaries, output limits, API versions and Claude Code permissions.
中文Jensen Huang says engineers should spend half their salary on tokens.
中文