AI Coding 的实践与探索

Summary

在火山引擎Force大会，字节跳动技术副总裁洪定坤分享AI Coding实践。过去一年，字节AI代码贡献率增长6倍，tokens消耗增长5倍，但过度关注单一指标可能失真--TRAE团队代码超90%由AI生成，人均需求吞吐率仅提升60%。900次实验显示，主流Coding模型组合代码正确率超80%，但...

Details

**AI Coding 的实践与探索** — this has been generating quite a buzz in the AI community lately.

What Happened

在火山引擎Force大会，字节跳动技术副总裁洪定坤分享AI Coding实践。过去一年，字节AI代码贡献率增长6倍，tokens消耗增长5倍，但过度关注单一指标可能失真--TRAE团队代码超90%由AI生成，人均需求吞吐率仅提升60%。900次实验显示，主流Coding模型组合代码正确率超80%，但可交付性仅40-60分；结合Harness基建后提升至80分。AI降低编程门槛但需优化指标、治理、协作。字节探索原型驱动开发，能力沉淀至TRAE（日均Token消耗5.6万亿，增长50倍），并推出TRAE Work

Why It Matters

From a technical perspective, this is a noteworthy advancement. It's not just about breaking new ground technologically, but also about the practical application scenarios.

What This Means for Everyday Users

While it sounds very technical, this could actually affect the products we use daily. For example, your AI assistant might get smarter, or your AI tools might become cheaper and more useful.

Industry Reaction

Currently, the industry's reaction to this is quite positive. Many experts believe this is the right direction and worth continued attention. Of course, some people remain cautious, believing it needs more time to validate.

Conclusion

Overall, this is a development worth watching. While it's still uncertain how it will ultimately develop, at least the direction is right.

For everyday users, there's no need to worry too much, nor get too excited. Stay tuned, use what you should use, learn what you should learn — that's the right attitude.

Source: 公众号：火山引擎

Updated: 2026-06-25