Sharing my thoughts on Vibe Coding lately

Original Media · 1h 16m · English · YouTube · Best Partners TV

TLDR

A developer-content creator shares their experience of 10 days of intensive AI-assisted coding in Guangzhou, building multiple production systems (business ERP replacement, Amazon scraper, ERP SDK, lottery predictor) using Trae IDE with AutoMode. They report 10x productivity gains, completing work that would take a 10-person team months, and argue AI coding is a real productivity revolution, not a bubble, with experienced engineers having a significant advantage in directing AI effectively. The session covers tools, workflows, cost economics, health trade-offs, and the future of AI-native companies. The speaker emphasizes that the bottleneck has shifted from coding ability to taste, architecture decisions, and verification cost control. They currently spend ~1300 RMB/month on unlimited AI coding via Trae subscription, far cheaper than hiring developers.

For you

PRO members can sign in and request more information — or their own version of this digest, tuned to what they care about.

Sign in to unlock

Surprising Insights

•

The AI (Trae AutoMode) correctly inferred the user was building a lottery predictor despite the prompt deliberately describing it as a 'decryption game' with no lottery terminology — showing strong intent recognition.

•

A complete Python-to-Node.js/Next.js rewrite of a project was done in-place in one morning with minimal instruction: 'rewrite using our discussed requirements and existing features.'

•

The speaker built a 600+ endpoint ERP SDK in one afternoon by feeding API docs and browser-captured requests/responses to AI — a task that would take weeks manually.

•

Amazon scraper built in ~1.5 days outperforms commercial third-party tools, extracting 10k+ detailed records from 13 stores including detail-page data that commercial tools miss.

•

AI automatically fixed a typo ('Blog' vs 'Blob' for cloud storage) without being asked, demonstrating deep contextual understanding of cloud concepts.

•

The speaker's non-technical colleagues (e-commerce operators) saw 3-5x efficiency gains after a single coaching session on prompt engineering (write expected output → have AI write prompt → iterate).

•

Open-source models are only 3-6 months behind frontier models per OpenRouter report; the speaker expects local models to match current frontier coding performance within 6 months.

Key takeaways

•

One experienced developer using AI coding tools (Trae/Tree with AutoMode) can match the output of a ~10 person team, completing ~3 months of traditional work in ~8 days of intense coding.

•

AI coding is not a bubble — the speaker reports genuine 10x productivity gains and the ability to build complex, production-ready systems (full business ERP, scrapers, SDKs) that were previously impractical.

•

Experienced engineers have a massive advantage: they can articulate requirements precisely, design architecture, guide AI through refactors, and verify outputs — while non-technical users struggle to describe needs and evaluate results.

•

The workflow has shifted: write specs → let AI plan → review plan → execute → test via screenshots/errors → iterate. Code review is replaced by architectural review and automated test generation.

•

Cost economics favor AI: ~1300 RMB/month for unlimited Trae AutoMode vs. 50k+ RMB/month for a small dev team, with 24/7 availability and no management overhead.

•

Traditional programming (hand-writing code) is becoming obsolete; the new skill is understanding the full system picture, concepts, and directing AI — not syntax or framework details.

•

AI-native companies (small, agile, top-down AI adoption) will outperform traditional companies bogged down by processes, legacy systems, and bottom-up adoption barriers.

•

Health toll is real: the speaker coded 10-15 hours/day, slept 5-6 hours, developed a cold, and describes the experience as addictive ('like drugs') requiring self-regulation.

Notable quotes

“现在的AI coding确实超出我的预期... 我觉得我一个人现在基本上顶一个可能小十个人的团队”

▸ Notable quotes · 12:28

“如果谁再说AI泡沫 他一定是没自己coding过... 因为我现在的效率 基本上是我认为以前的十倍”

▸ Notable quotes · 19:54

“就是老程序员焕发第二春的感觉... 一下子觉得自己又行了”

▸ Notable quotes · 26:52

“强的可怕呀 真的 我现在我自己都觉得强的可怕”

▸ Notable quotes · 33:15

“古法编程我覺得確實回不去了... 你手寫Coding 我認為可能很多時候沒有太大意義了”

▸ Notable quotes · 59:02

“最後瓶頸在哪 在於品位 複雜度 大量驗證控制成本”

▸ Notable quotes · 64:17

“我現在就是買的這個TRAIL的 最高級的會員一個月就一千多塊錢吧... 比這個招幾個人 那真的是省多了”

▸ Notable quotes · 72:54

Chapters

•

0:00 — Intro: Move to Guangzhou & Livestream Setup: Speaker greets viewers from Guangzhou apartment, explains recent move from Beijing, mentions slight cold and late nights, sets up stream after a 10-day hiatus.

•

7:00 — 10 Days of Intense AI Coding: Context & Motivation: Company's Guangzhou team pulled speaker into solving low-efficiency manual operations (copy-paste across systems, ERP sync, order processing). Speaker spent 10 days (8 effective) deep in AI coding while maintaining video output.

•

11:30 — Tools: Trae (Tree) IDE with AutoMode: Uses Trae (ByteDance's AI IDE) with AutoMode — unlimited quota for ~1300 RMB/month, runs smoothly in China, mixes multiple models. Previously used Cursor, Claude Code, but Trae's unlimited plan suits heavy usage.

•

12:30 — Project 1: Internal Business System (ERP Replacement): Built full system in ~8 days: auth, permissions, email sync/parsing, ERP order sync, OCR parsing, points/refund/shipment workflows. Replaced multi-system manual Excel work. Day 1: Python/Reflex prototype. Day 2: Rewrote to Node.js/Next.js in-place. Day 3: Refactored webhooks to scheduled jobs. Deployed to Vercel automatically.

•

16:00 — Project 2: Amazon Store Scraper for Friend: Built custom scraper in ~1.5 days: 13 stores, 10k+ records with detail-page data (categories, specs) that commercial tools miss. Optimized 10x speed via concurrency. Exports Excel, customizable fields.

•

17:30 — Project 3: ERP SDK (600+ Endpoints) & Lottery Predictor: Reverse-engineered undocumented ERP APIs from browser traffic + docs → AI generated full Node SDK in one afternoon. Also built lottery prediction tool with time-series, ML regression on 3000+ historical draws for a friend.

•

19:00 — Productivity Analysis: 10x Speed, 1 Person = 10-Person Team: Speaker estimates 10x personal productivity. Work that would take 3-4 people 2-3 months done solo in 8 days. Codebase ~tens of thousands of lines. Full stack: design, dev, test, deploy, monitor — all solo + AI.

•

27:00 — Why Experienced Engineers Win with AI: Veterans articulate requirements precisely, understand architecture, guide refactors, verify outputs. Non-technical users struggle to describe needs, evaluate code, deploy. Speaker's colleagues got 3-5x gains after prompt-engineering coaching.

•

35:00 — Workflow: Spec → Plan → Review → Execute → Test → Iterate: No traditional code review (too much volume). Instead: review AI plans, let AI write tests (unit, e2e, security, performance), test via screenshots/errors, iterate. AI self-corrects from logs.

•

43:00 — Cost Economics & AI-Native Company Vision: ~1300 RMB/month Trae vs 50k+ RMB/month for team. 24/7 availability, no management. Speaker plans to make whole company AI-native: ops, ads, business logic all AI-driven. Large companies struggle with top-down adoption, processes, legacy.

•

64:10 — Bottlenecks: Taste, Complexity, Verification Cost: AI generates unlimited code but needs human taste to prune scope, control complexity, define verification criteria. Speaker constantly rejects over-engineered AI proposals (e.g., KMS vs local encryption).

•

70:00 — Health Toll & Future Outlook: Addictive, unsustainable pace: 10-15h coding, 5-6h sleep, cold, voice strain. Must self-regulate. Predicts open-source models match frontier in 6 months. Traditional hand-coding obsolete; future skill is system thinking + AI direction.

•

75:00 — Q&A: Tools, Learning, Deployment, Investing: Answers on Trae mobile/Telegram agent, code quality vs tests, local models for privacy, payment integration via MCP, Cloud account bans in China, platform strategy, Oxford hoodie (just merch).

Speakers

•

Speaker (大飞 / Da Fei) — Host, Developer, Content Creator, Company Founder

⋮