When GPT-2 dropped in 2019 and GPT-3 followed in 2020, NLP was the first field to face a question that has since spread to computer vision, theorem proving, code generation, and protein folding: if a large language model can do the task, what is the researcher for? A Hacker News thread from this week asks how the people who built the field before the transformer era have actually adapted in the years since.

The disruption was not instantaneous

The Quanta Magazine oral history linked in the thread captures the uneven timeline. NLP research did not vanish overnight. The transition happened in phases. Between 2019 and 2021, many researchers pivoted from building task-specific models to studying the properties of large pretrained ones. Transfer learning, which had been a niche interest, became the default paradigm. The people who had spent years on sentiment analysis or named entity recognition found that their labeled datasets were now fine-tuning data for models that dwarfed anything they had trained.

The skills that carried over were not the ones anyone expected. Knowing how to evaluate a model, how to design a controlled experiment, how to read a confusion matrix and understand where a system fails, these turned out to be more valuable than knowing how to architect a neural network from scratch. The engineering shifted from designing layers to designing prompts, evaluations, and data pipelines.

Where the researchers went

Three trajectories emerge from the post-2019 NLP community, none of them clean. Some researchers stayed in academia and pivoted to studying LLMs themselves: probing their internal representations, measuring their biases, testing their reasoning capabilities, analyzing their failure modes. This is the closest thing to business as usual, though the object of study changed from "how do we build a parser" to "what does this model actually know."

Others moved to industry, where the demand for people who understand language at a technical level exploded. Companies building LLMs needed people who could design evaluation suites, build safety filters, curate training data, and think about how models interact with users. The research skill set transferred directly, even if the job title changed from "research scientist" to "applied researcher" or "alignment engineer."

A third group left NLP entirely. Some moved to adjacent areas that have not yet been fully disrupted, like formal semantics or computational linguistics focused on low-resource languages. Others left academia for software engineering, data science, or unrelated fields. The thread does not provide numbers, but the pattern is consistent with what happens in any field where the tools suddenly do most of what the researchers were building by hand.

The student pipeline

The question of whether new students stopped entering NLP is harder to answer. Enrollment in ML and AI programs has generally increased, but the composition has shifted. Students who might have specialized in NLP five years ago are now more likely to call themselves "LLM engineers" or "AI safety researchers." The subfield branding changed even as the underlying questions, how does meaning emerge from sequences of tokens, remained the same.

What has decreased is the number of students entering NLP with the expectation of building a new state-of-the-art model for a specific task. That work now happens inside large companies with massive compute budgets, not in university labs. The students who do enter academic NLP tend to focus on evaluation, interpretability, fairness, and robustness, areas where the questions are still open and where a single researcher with a GPU can still contribute something.

The uncomfortable parallel

The thread's framing, NLP was the first field disrupted, is worth taking seriously. The same pattern is now playing out in theorem proving, where tools like AlphaProof and GPT-based provers are doing tasks that required years of graduate training. The question the NLP community is living through is not just an academic curiosity. It is a preview of what happens to any research field when the tooling advances faster than the training pipeline that produces researchers.

The NLP researchers who adapted best were the ones who treated the disruption as a new object of study rather than an existential threat. The ones who struggled were those whose identity was tied to building a particular kind of model rather than solving a particular kind of problem. The field itself is smaller as an academic discipline than it was in 2018, but the questions it raised, about meaning, evaluation, and what it means for a system to understand language, are now everyone's questions.