亚马逊推理数据团队效率惊人

Five things from Friday's eleven that stuck with us: - An Amazon team estimated a new inference dat...

精选理由

亚马逊团队76天完成原计划18个月的工作,优步AI代码审查工具每周处理2.5万评论,看看科技公司如何应对AI带来的效率变化。

AI 摘要

亚马逊团队原计划用30人和18个月完成推理数据平面,实际6名工程师(含2名杰出工程师)仅用76天完成。优步代码审查时间从2024年的3小时增至2026年的9小时,uReview每周生成25,000条评论。优步70%以上的PR来自智能体,Figma工程师因AI分析粗糙而回归人工撰写PR描述。

原文 · AI Engineer

Five things from Friday's eleven that stuck with us: - An Amazon team estimated a new inference dat...

Five things from Friday's eleven that stuck with us: - An Amazon team estimated a new inference data plane at 30 people and 18 months. Six engineers, two of them Distinguished, built it in 76 days. Productivity went down first in almost every team studied - At Uber a first code review took three hours in 2024. Nine in 2026. uReview now leaves 25,000 comments a week, and the model is 100% sure every time, including when it is wrong - Over 70% of Uber's PRs come from agents. The maintenance loop runs on Sunday, when CI has capacity and nobody wants a Monday of extra diffs - A Figma engineer sent a respected skeptic an unlabeled AI analysis. Reply: I did not expect someone I respect this much to send me something this sloppy. Now every PR description opens by hand - The same prompt takes two seconds or sixty. A reasoning model's normal is a chat model's outage. That is why you keep seeing "something went wrong, please try again" Watch the AI-Native Enterprises playlist. Part 2 lands tomorrow: youtube.com/watch?v=ZIYYsA… 💬 0 🔄 0 ❤️ 1 👀 218 ⚡

亚马逊推理数据团队效率惊人 · AI 热点