How GPT‑4 Has Changed NLP Research: Community Perspectives

Source: AINLP

On a weekend the author noticed a Zhihu question titled “After GPT‑4’s release, how has your NLP research changed?” and selected several answers to share, inviting readers to join a discussion group.

Link: https://www.zhihu.com/question/589704718

Author: sonta

Link: https://www.zhihu.com/question/589704718/answer/2946475253

NLP is dead.

The NLP community can be roughly divided into two groups: those who believe in AGI and those who do not.

For the former, the rise of large language models (LLMs) is exhilarating; previous NLP models felt like toys, while LLMs are seen as the correct path toward AGI. They chant that they are lucky to be born in this era and go all‑in on LLM research, provided they have sufficient compute. The author excerpts several senior researchers’ suggestions for future LLM work.

A senior researcher’s advice

A famous LLM “father’s” Twitter feed

However, the author finds API‑driven or prompt‑driven LLM research boring and has largely stopped submitting to ACL, though they still plan to work on scaling non‑attention architectures.

I belong to the first group

The second group feels that LLM research is dull because the field is becoming overly engineering‑focused; they question the meaning of their work and experience an existential crisis when their results are quickly eclipsed by newer GPT versions.

Because the author works on parsing, they note that syntactic information has long been considered optional in strong neural networks, making pure parsing research seem practically meaningless today. Yet parsing remains theoretically enjoyable, and many researchers started in parsing before moving to other NLP sub‑fields once parsing was deemed “almost solved.”

In the LLM era, many intermediate NLP tasks appear solved, and application‑oriented tasks such as translation, polishing, and error correction face direct competition from GPT‑style models, prompting a shift toward cross‑disciplinary research or leveraging large models to empower smaller ones.

One can also follow Neubig’s work on environmental protection (tongue‑in‑cheek).

Author: Zheng Chujie

NLP community is broken.

In the past, research aimed to be forward‑looking and guide applications, but now academic NLP lags behind industry; limited‑resource “toy” papers become obsolete within a few months, and the peer‑review system fails to evaluate true value, wasting time on rebuttals and recycling.

There is a common claim that “ACL conferences love to accept polished garbage.” The community may be heading toward decay or becoming a self‑contained entertainment zone.

Author: Anonymous user

Without strong institutional backing, resources are limited; the author feels the AI era offers huge opportunities for big companies but little relevance for themselves, fearing their current work will become obsolete as large models reshape lives and work.

They tried a new ACL‑22 dataset, added tricks from ICLR/ICML, and used BART‑base (cannot run large). After comparing with ChatGPT and GPT‑4 outputs, they questioned the meaning of their work, fearing reviewers will reject superficial results, and wonder if better prompts could make their model win.

Nevertheless, they must publish to graduate, and may try to explore niche corners of large‑model research or craft sci‑fi stories for a degree.

Author: Anonymous user

Link: https://www.zhihu.com/question/589704718/answer/2940391370

NLP? NLPer already mourned the field last December; now CV and multimodal seem to be grieving more.

The wheel rolls over; no one survives.

Author: Liu Cong (NLP)

Researchers with resources should study large‑model foundations; those with limited resources should focus on fine‑tuning; those with no resources should concentrate on API usage.

Prompt‑template research is less attractive because automated template generation APIs already exist.

Seriously, empowering small models with large‑model capabilities may become a key research direction, as deploying 1‑billion‑parameter models is feasible, while 10‑billion‑parameter models remain costly for enterprises.

进技术交流群请扫码添加小助手

请备注具体方向+所用到的相关技术点

Read to the end? Please share, like, and give a triple‑click support 🙏