The debate unfolding on Reddit, whether an ML PhD is better pursued before or after the advent of ChatGPT, misses the point entirely. The question assumes the AI era is a finished product to study, when in reality it is a moving target that demands a fundamentally different kind of expertise. We believe the real transformation isn't about timing; it's about what you actually learn to do with the tools at hand.
A PhD completed before widespread AI chatbots meant mastering the craft of model building from scratch, data collection, architecture design, training loops, and debugging. That work remains foundational, but the context has shifted dramatically. As OpenAI Streamlines ChatGPT as Chat and Codex Share One Usage Pool shows, the line between conversational AI and developer tools is blurring. The same allocation pool powers both a chatbot and a coding assistant. For a PhD student today, the question is no longer "can I build a model that generates text?" but "how do I build systems that integrate, evaluate, and deploy these models responsibly at scale?" The degree itself is still valuable, but the curriculum must evolve from isolated research to applied systems thinking.
Meanwhile, the sheer volume of work entering the field is staggering. With 60,000 submissions test ICLR's review system as ratings prepare to land this week, the bottleneck has shifted from generating ideas to filtering them. A PhD in today's era isn't just about producing novel research, it's about navigating a crowded landscape where standing out requires a sharper sense of problem selection and a deeper understanding of what others have already tried. The old model of incremental progress on a narrow question may no longer be sufficient. Students now need to think about reproducibility, evaluation benchmarks, and real-world impact from day one.
The practical takeaway for anyone considering an ML PhD right now is this: choose your advisor and your problem based on how they connect to the broader ecosystem, not just the lab. The skills that matter most are not the ones that let you replicate what a foundation model already does, but the ones that let you build on top of it, critique it, and extend it into new domains. The AI era didn't kill the ML PhD, it raised the stakes. The best work will come from those who treat the chatbot not as the end of the road, but as a new kind of infrastructure to design for. That is the specific question worth watching: how will PhD programs adapt their training to produce researchers who can think beyond the next model release?