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「How to Choose Your Grad School」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。

The Career Perspective: Choosing Based on Expected Success

If you are reading this, then you probably finished the long and arduous journey to grad school. You emerged victoriously, and this success is well-deserved. But which school should you choose? How to make a right choice if all schools look great in their own way? This blog post is centered around these questions. It is most useful if you are a computer science student aiming to study machine learning and, in particular, natural language processing in the

The choice of grad school that is right for you can be tricky and confusing. We live in a time of hyper-competitiveness, where even undergrads need to optimize for metrics like paper count to make it to the next level — grad school. This heavily career-centered perspective was probably advantageous to get you into grad school, and it remains crucial to get you to the level after that: a great job in industry or academia. So choosing the school which is bes

Advisor

在这类工作里, I will discuss this career-centered perspective in detail, but I also provide you with three other views that hopefully help you make a balanced choice that not only leads to academic success but long-term satisfaction and a full and rich life. Balancing your decision based on all four perspectives probably leads to a better choice than looking at one angle alone. Before I go into the details, let me briefly introduce these four perspectives: The C

A quite intuitive perspective is the Career Perspective , which is about determining and weighing the factors that help you to be successful in your PhD and have a successful career.

Research Style

A different perspective is the Identity Perspective : not looking at your career but at who you want to be and how your choice enables and facilitates that identity. The social environment that you are in has a strong causal effect on your development: We are strongly influenced by the people and culture around us, and the friends of friends that you do not even know will make you honest/deceitful, selfish/selfless, caring/exploiting, and so forth. If you

The Stability Perspective says that choosing the “right” school is an illusion but that there are other choices that matter much more because they give you the stability that you need to succeed in the arduous PhD journey. It is well known that the effect of most moderately painful or enjoyable events that significantly affect your life will wear off within about two years and that you will return to your baseline happiness and stay there. However, some th

Advisor Research Fit

Another valid way to select a school is by the variability of experience it will offer — the Variability Perspective . You probably sacrificed in some way to get into grad school. You neglected your passions outside work, neglected friends or your partner or your family, neglected self-development, neglected to work on your mental, physical or spiritual health, or you neglected other things that are important to you. By choosing the school that is best for

The career perspective looks at the most critical factors for your academic success and success beyond that and chooses the school that is best according to these factors. Let me go through each factor. I list the factors in order of importance, starting from the most important.

Advising style: Hands-on vs Hands-off

Finding suitable advisors is probably the most crucial task when choosing between grad schools. One could even go further and argue that one should not choose a school, but one should choose an advisor. A lousy advisor can make you miserable, unproductive, stressed, and might be the main reason why you would want to drop out of the program. The right advisor will help you to be productive, stay healthy, and help you to enjoy doing your research. It is impo

Another important note is that you should be looking for advisors and not a single advisor. This complicates an already complicated process, but it is risky to choose a school based on a single person. Relationships are complicated, and things might not work out as expected with your advisor. If possible, you should have an alternative advisor to whom you can switch if it does not work out with the other advisor. This strategy also offers the possibility o

Advisor Values, Strengths, and Weaknesses

The following advisor-related factors do not have a particular order.

Research style is probably the most elusive quality but also by far the most important quality that you acquire during your PhD. While many would say that the goal of a PhD is to become an independent researcher, the truth is that with the steep requirements for ML/NLP PhD positions, many students are already somewhat close to being independent. They can generate ideas with ease and execute them confidently in research projects. However, the actual quality

Advisor Availability and Absent-mindedness

Harriet Zuckerman is probably the person that studied scientific expertise to the greatest qualitative depth. In her work Scientific Elite , she interviewed almost every US Nobel prize laureate of the 20th century. She found that these individuals often rose through the ranks through accumulated advantage. One advantage helped them secure the next position/grant/collaboration, which increased their advantage and helped them secure the next one, and so fort

While your advisor will be the focal point of research culture, research culture is also created through interactions between your advisor’s lab students. It is usually subconsciously adopted over time. Most students might not be aware of how they were shaped by their advisor and research group. It happens automatically and does not necessarily require explicit thinking or effort.

值得单独记下的要点

• Ideas are cheap and belong to the research group. Execution of those ideas as research projects is real research.
• Novel ideas are everything. If someone publishes something even remotely similar to what you have, you should give up the project and work on something nobody is working on.
• Good science is good math. A paper should be mathematically solid so that it will stand for years, holding valuable insights and generalizations that go beyond the current theoretical application.
• Good science is robust science. A paper should have careful claims with robust evidence. This will help make the field progress more quickly by providing reliable information to build on.
• Good science is a good research vision. A paper should be about what is possible in the future and where a line of research could lead to. Evidence augments vision, but a paper without vision is blind, incremental, and w
• Good science is good insight. Some insights can be extrapolated and be applied to many other scientific problems, many of which have not been formulated yet. Finding and expressing these insights is vital for scientific
• “What are examples of research papers you like?”
• “What research papers (in your area) do you think are the most important ones in the last years?”

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