> Imagine if you spent your entire career working on NLP, and now find GPT-4 will run rings around whatever you've done. What do you do now? I have been doing

> Imagine if you spent your entire career working on NLP, and now find GPT-4 will run rings around whatever you've done. What do you do now? I have been doing NLP since 1993. Before ca. 1996, there were mostly rule-based systems that were just toys. They lacked robustness. Then statistical systems came up and things like spell-checking (considering context when doing it), part of speech tagging and eventually even parsing started to work. Back then, people could still only analyze sentences with fewer than 40 words - the rest was often cut off. Then came more and more advanced machine learning models (decision trees, HMMs, CRFs), first a whole zoo, and then support vector regressors (SVM/SVR) ate everything else for breakfast. Then in machine learning a revival of neural networks happened, because better training algorithms were discovered, more data became available and cheap GPUs were suddenly available because kids needed them for computer games. This led to what some call the ¨deep learning revolution¨. Tasks like speech recognition where people for decades tried to squeeze out another half percent drop in error rate suddenly made huge jumps, improving quality by 35% - so jaws dropped. (But today's models like BERT still only can process 512 words of text.) So it is understandable that people worry at several ends. To lose jobs, to render ¨NLP redundant¨. I think that is not merited. Deep neural models have their own set of problems, which need to be solved. In particular, lack of transparency and presence of different types of bias, but also the size and energy consumption. Another issue is that for many tasks, no much data is actually available. The big corps like Google/Meta etc. push the big ¨foundational¨ models because in the consumer space there is ample data available. But there are very important segments (notably in the professional space - applications for accountants, lawyers, journalists, pharmacologists - all of which I have conducted projects in/for), where training data can be constructed for a lot of money, but it will never reach the size of the set of today`s FB likes. There will always be a need for people who build bespoke systems or customize systems for particular use cases or languages, so my bet is things will stay fun and exciting. Also note that "NLP" is a vast field that includes much more than just word based language models. The field of propositional (logical) semantics, which is currently disconnected from the so-called foundational models, is much more fascinating than, say, chatGPT if you ask me. The people there, linguist-logicians like Johan Bos identify laws that restrict what a sentence can mean, given its structure, and rules how to map from sentences like "The man gave the girl a rose" to their functor-argument structure - something like "give(man_0, rose_1)¨ - which models the "who did what to whom?". When such symbolic approaches are integrated with neural foundational models, there will be a much bigger breakthrough than what we are seeing today (mark my words!). Because these tools, for instance Lambda Discourse Representation Theory and friends, permit you to represent how the meaning of "man bites dog" is different from "dog bites man". So whereas today`s models SEEM a bit intelligent, but are actually only sophisticated statistical parrots, the future will bring something more principled. Then the ¨ "hallucinations" of models will stop. I am glad I am in the field of NLP - it has been getting more exciting every year since 1993, and the best time still lies ahead!