Every January, the flood of cold emails begins. A professor in machine learning posts a candid list of what not to do when reaching out about PhD positions, and the subreddit eats it up. The advice is blunt: don't send massive emails, don't email everyone, don't fake research interests, and don't outsource your thinking to an LLM. The professor's list is a masterclass in setting expectations. But beneath the surface, it's not really about email etiquette. It's about the difference between performing interest and demonstrating it. And for anyone navigating the AI/ML job market right now, that distinction matters more than ever.
Here's what the professor is really saying: the system is flooded. Not just with applications, but with low-effort, high-volume attempts to game the process. The advice to avoid generic research interests like "Machine Learning, LLMs, and AI" isn't snobbery. It's a signal that too many candidates mistake buzzwords for understanding. This connects directly to the shifting expectations in the broader AI/ML field, where job descriptions now demand software engineering skills alongside research acumen. As we've covered in Navigating AI/ML Job Requirements: A Shift in Expected Skills, the bar has moved from "knows the jargon" to "can build and ship." Cold emailing a professor is no different. A summary of a paper you didn't write and didn't understand is just another form of résumé padding. It's the same habit that leads candidates to claim familiarity with distributed training without ever having run a single cluster. When you actually understand the material, you can build on it. That's what the professor wants to see. That's what any good collaborator wants to see.
The most pointed advice is about LLM use. The professor is clear: using AI to fix grammar is fine, but outsourcing your thinking to generate a research direction is a red flag. This isn't Luddite paranoia. It's a practical observation about how to stand out in a crowded field. If everyone uses the same tools to generate the same generic interests, everyone looks the same. The professor even notes that identifying these emails is easy. So what's the takeaway for a prospective PhD student? Don't ask the LLM what to research. Ask yourself what problem you can't stop thinking about. Then, and only then, use the tools available to refine your ideas. That's the same logic behind learning distributed training from the ground up, rather than just reading about it. As we discussed in Unlock LLM Training: A Practical Guide to Distributed Algorithms, real understanding comes from doing, not from summarizing.
The deeper issue here is honesty. Not just about papers, but about your own level of engagement. The professor's point about workshop papers being passed off as conference papers is a warning against inflating your own record. It's a small lie, but it signals a willingness to cut corners. And in a field where trust is the foundation of collaboration, that's fatal. The same goes for ignoring instructions on a supervisor's website. That's not a minor oversight; it's a test of whether you can follow directions. If you can't do that in an email, how will you handle the complexities of a multi-year research project?
Stop treating cold emails as a numbers game. Treat them as the first step in a genuine intellectual conversation. Before you hit send, ask yourself if the person on the other end would read your email and think, "This person has a point of view." If the answer is no, rewrite it. And if you're tempted to use an LLM to write that first draft, remember the professor's warning: it's easy to spot, and it won't get you in the door. The one thing you can't fake is curiosity. The professor's post is a reminder that in a world of automated applications and generic outreach, the people who still take the time to think for themselves are the ones who get noticed. The test isn't whether you can send a good email. It's whether you can say something worth reading.