本文由 AI 分析生成
建立時間： 2026-01-04 來源： https://x.com/BensonTWN/status/2007674218654351768
Summary
The author claims a 10x+ productivity gap between engineers who use AI well and those who don’t, citing personal experience building CoinKarma’s quant trading strategy “Ultron” end-to-end via AI-assisted vibe coding despite having no CS background. The core argument is that pasting code in and out of a chat window loses context every time, while true AI-native development lets the AI operate directly in the codebase with the human defining problems and reviewing output. The post ends with a company mandate: all CoinKarma developers must use Claude Code, fully subsidized.
作者主張善用 AI 與不會用 AI 的工程師之間有 10 倍以上的生產力差距，並以親身經歷佐證：他沒有資工背景，卻用 AI 協作的 vibe coding 完成了 CoinKarma 量化策略「Ultron」從資料串接到實盤的整條鏈路。核心論點是把程式碼貼進貼出聊天視窗每次都會丟失 context，而真正的 AI-native 開發是讓 AI 直接在 codebase 裡操作，人負責定義問題與審查產出。文末宣布公司政策：CoinKarma 全體開發者強制使用 Claude Code，費用全額補助。
Key Points
- “Ultron” reached >$17M AUM, outperformed BTC by 5x+ over 600 days, with strategy revenue-share now exceeding subscription income.
- “Human writes, AI edits” vs “AI writes, human reviews” are fundamentally different modes; the latter has ~10x the throughput.
- Chatbot-style copy-paste workflows leak context and break focus on every window switch.
- AI shifts engineer value upward: system design taste, problem decomposition, and review speed/accuracy — skills built by deliberate practice, not seniority.
- The skill compounds and transfers across projects, languages, and even job functions.
Insights
The framing that AI “forces engineer value upward” is the durable insight: when implementation is commoditized, the differentiator becomes judgment that can’t be accumulated passively. The mandate (full Claude Max subsidy, no opt-out) reflects a thesis that good tools change behavior only when paired with organizational push — a deliberate forcing function rather than waiting for voluntary adoption.
Connections
Raw Excerpt
「人寫 AI 改」跟「AI 寫人審」是兩種完全不同的生產模式。後者的吞吐量是前者的十倍起跳。