Tim Dettmers 做量化、多卡和智能体时，习惯先把硬件账和算法账摊开。把「Machine Learning PhD Applications — Everything You Need to Know」写成能落地的笔记，核心不是复述论文摘要，而是：显存、精度、通信和评测会在哪一层先爆。下面用中文重排，去掉原站广告和分享条。

Understanding What Makes a Strong PhD Application

I studied in depth how to be successful in my PhD applications and it paid off: I got admitted to Stanford, University of Washington, UCL, CMU, and NYU. This blog post is a mish-mash of how to proceed in your PhD applications from A to Z. It discusses what is important and what is not. It discusses application materials like the statement of purpose (SoP) and how to make sense of these application materials.

There are some excellent sources out there on this topic and it is worth stopping for a second and understand what this blog post will give you and what other sources can give you. This blog post is mainly focused on PhD applications for deep learning and related fields like natural language processing, computer vision, reinforcement learning, and other sub-fields of deep learning. This blog post assumes that you already have a relatively strong profile, m

Understanding Application Materials

If you seek more general information for PhD admissions, I recommend reading all the most highly voted questions and answers from Academia StackExchange . Other important sources are Applying to Ph.D. Programs in Computer Science which is a detailed write-up of the full admission process as viewed by CMU faculty. A similar but more concise source — in particular, relevant for good but not strong candidates — is the blog post Reflecting on CS Graduate Admis

This blog post will first define what is important in PhD applications. Then we will dive into the application materials and how to think about these. Then I will talk a bit about the application process. The final section of the main part of this blog post will be on selecting schools — which schools are too good or too bad for me? After that, I will close with a Q&A section which was drawn from questions on Twitter . I will update this Q&A section period

Understanding Recommendation Letters

The most important factor that determines admission at any research university is research potential: How likely are you to become a great researcher? The main direct indicators for this are in order of importance:

Other indirect factors can help sometimes if they are exceptional, but usually, only the first two factors, recommendations, and research experience count. In order of importance:

Signs of a Bad Recommendation Letter

For recommendation letters, one could devise four categories: Strong, Good, Weak, and Bad. Note that the main thing that admission committees look for in recommendation letters are indicators of research potential. This section has the main purpose of making you aware of what constitutes a good or strong letter and based on this information it might be easier for you to select letter writers.

Publications are direct evidence for research experience and research skill. If you published as a first author, people know that you did most of the work. If you published as a second author, people know that you did a good portion of the work (25%-50%). If your name is the third or later, your contribution is discounted, but you probably went through the entire research process towards publication and gained a good amount of research experience. If you p

Signs of a Weak Recommendation Letter

If you published your work at a respectable conference, people know that: (1) Your work is high quality; (2) your work can be trusted; (3) that your current research skills is sufficient to publish at great conferences, (4) that you are competitive and/or you can stay productive under the pressure of publishing at a top conference.

It helps to view this in the eyes of a potential advisor: If you have two students, one published already at NeurIPS (Tier A) and one you published at a Tier B conference. You would know that the first student is probably ready to work on a research project which is aiming for NeurIPS next year. The second student would need further preparation, for example, publish in a workshop or at a less competitive Tier A conference before making the step towards Neu

Signs of a Good Recommendation Letter

Other indicators have little effect on the application. Your work might be unusually creative, but you have no track record that you are a creative researcher. Maybe you got lucky.

The importance of publications often only emerges with the years. Often you published shortly before the PhD applications which means that the citations that you have on your work is a poor indicator of impact. If you get an usually high number of citations in a short time this can help, but maybe you just got lucky or good at marketing. Usually, the number of citations over the past 1-3 years is no reliable indicator of research potential and as such is d

Signs of a Strong Recommendation Letter

For most institutions, the statement of purpose is mainly a filter for people who took no time to polish the SoP. Your writing can show how you think, how you can sell, how you explain things, but it can also show that you are lazy and do not pay attention to details. It can show that you are not able to Google simple recipes of how to write (and how not to write) a simple formal document. For some institutions, the SoP can be important (CMU) but the conte

Beyond formalities, the SoP is also the only document where you can justify why you did underperform in certain circumstances. For example, you can explain any extraordinary difficulties that you had along the way to graduate school, or it can explain why you did not do so well in certain semesters/quarters at uni. The structure of a SoP should be the following:

值得单独记下的要点

• Recommendations: Respected professors speak highly of you. Personal connections are important.
• Research experience: You did successful research before. Measured in publications, first-authorship, and prestige of conference where you published.
• Undergraduate university name: Some universities select aggressively for this, some others do not care so much.
• Employer name: It is common that students are admitted that were previously employed in finance or at companies such as Google, Facebook, etcetera.
• Smarts: Perfect GPA, perfect GRE is somewhat correlated with intelligence (or at least with how fast you can learn and understand).
• Grit / Conscientiousness: You do well under continuous rejection, disappointment, and failure. If you faced and have overcome difficulties you might want to include your story in the statement of purpose.
• Accomplishment: You won Math or CS competitions.
• Recognition: You won prestigious scholarships/fellowships.

落地时建议先做的 5 件事

1. 用自己的模型和序列长度算显存账，不要只看官网 TFLOPS。
2. 量化前后都要有同一套评测：平均分和长尾失败都记。
3. 多卡训练先分清数据并行和模型并行，通信量写进预算。
4. 智能体工作流要有可复现实验，而不是一次演示通过。
5. 硬件选型写清接口、功耗和两年后还能不能买到。

和智能体产品怎么接

龙虾PRO做 OpenClaw 落地时，量化和多卡经验会直接变成推理成本和技能能否本地跑。数字员工和网关应共用同一套 GPU/精度预算，而不是每条链路各买各的卡。

本文侧重全链路风控方法论。落地时请用自身业务单据做回放验证，不要把示例阈值直接当生产策略。 相关：风控体检 · 方案资源