如何面试和招聘机器学习/人工智能工程师

Hiring well is the highest leverage activity we can do for the mission and organization. And running effective interviews is key to hiring well. We can think of the interview process as a system: Given a candidate, assesses whether they are a good fit for the role and team. Thus, to hire well, the interview system should be reliable and valid, with minimal noise.

招聘好是我们为使命和组织所能做的最高效活动。而进行有效的面试是招聘好的关键。我们可以将面试过程视为一个系统：根据候选人，评估他们是否适合该职位和团队。因此，为了招聘好，面试系统应该是可靠和有效的，噪音要尽量少。

In this write-up, Jason and I will share a few things we’ve learned about interviewing candidates for machine learning (ML) and AI roles. First, we’ll discuss what technical and non-technical qualities to assess. Then, we’ll share how to calibrate phone screens, and run the interview loop and debrief. Finally, we’ll wrap up with some tips for interviewers and hiring managers, as well as our opinionated take on some traits of a good hire. (And like all our writing online, opinions our own.)

在这篇文章中，Jason和我将分享一些我们在面试机器学习（ML）和人工智能（AI）职位候选人时学到的东西。首先，我们将讨论要评估的技术和非技术素质。然后，我们将分享如何校准电话面试，以及如何运行面试环节和总结。最后，我们将提供一些面试官和招聘经理的建议，以及我们对一个好员工的一些观点（和我们在线写作的所有观点一样，都是我们自己的观点）。

For most ML/AI roles, a basic proficiency in software engineering is expected. Depending on the role, we’ve seen simple 30-60 minute coding exercises that ask a candidate to:

对于大多数机器学习/人工智能职位，都需要基本的软件工程能力。根据职位的不同，我们看到了一些简单的30-60分钟的编码练习，要求候选人：

• Check if 2/3D arrays meet predefined criteria (e.g., validating robot’s simulated route through a warehouse), checking for edge cases, and writing unit tests
• 检查2/3D数组是否符合预定义的条件（例如，验证机器人在仓库中的模拟路径），检查边缘情况，并编写单元测试
• Implement and start an inference endpoint, including input/output validation, logging, monitoring, and command to update endpoint state
• 实施并启动推理端点，包括输入/输出验证、日志记录、监控和更新端点状态的命令
• Build a pipeline to process data, first in batch and then adapting it for streaming
• 构建一个处理数据的流水线，首先是批处理，然后适应流处理

Successfully completing the coding exercise is only part of the interview. What’s more important is how the candidate solves it: Do they break down the problem logically, w...