本文由 AI 分析生成
建立時間： 2026-02-21 來源： https://x.com/BensonTWN/status/2025064405852119089
Summary
A reflection on LeCun’s 2023 critique that LLMs lack persistent memory, single-shot learning, and world understanding — and how engineering has bypassed (not solved) these flaws via scale, RAG, vision, and tools. The author argues that because memory is bolted-on rather than native, Context Engineering — deciding what to feed the model, when to reset, how to sync memory across agents — is becoming more decisive than raw model capability. A concrete example contrasts calling Codex via MCP (pollutes the main context with every tool turn) versus a coding-agent skill (returns only a clean handoff summary).
這是對楊立昆 2023 年批評的反思：LLM 缺乏持久記憶、無法單次學習、缺乏世界理解——而工程界用規模、RAG、視覺與工具「繞過」而非真正解決這些缺陷。作者主張，由於記憶是外掛而非原生的，Context Engineering（決定餵什麼、何時重置、如何在多個 agent 間同步記憶）正變得比模型本身能力更關鍵。文中以對比 MCP 調用 Codex（每個 tool turn 都污染主 context）與 coding-agent skill（只回傳精簡的 handoff summary）為具體例證。
Key Points
- LLM architecture is unchanged: a statistical engine that starts empty every session; memory is a prosthetic, not native.
- Engineering bypassed LeCun’s critique with “brute-force capital” (scale-driven emergence) and “systematic patches” (vector DBs, vision, tools).
- Multi-agent setups suffer “cognitive fracture”: separate instances share a memory file but aren’t synced in real time, so they behave like different people.
- MCP-calling a sub-model floods the main context with intermediate turns; a sub-agent skill isolates that work and returns only a summary, keeping context clean.
- Same model + different operator can yield 10x output differences; Context Engineering precision now outweighs credentials and seniority.
Insights
The airplane-vs-bird metaphor reframes the “is it really learning” debate: once pattern matching is complex enough to simulate the logic of civilization, the philosophical question of whether it’s “true” learning stops mattering operationally. The non-obvious operational claim is that the MCP-vs-skill distinction is fundamentally about context hygiene — whether intermediate reasoning enters the orchestrator’s window — which makes sub-agent handoffs an architectural tool, not just a convenience.
Connections
Raw Excerpt
用工程的角度來說，就是 Context Engineering 的重要程度，會逐漸超過模型本身。