After 5 years as a PhD student, I’m taking a break from AI Research.

Research has a high proportion of incredibly smart self motivated people, and attracts even more smart and self motivated people every year. The community certainly wouldn’t miss my headcount. In fact I’ll risk offending a bunch of people and say the world of NLP research wouldn’t miss a whole bunch of headcounts.

I don’t think it’s too controversial to say the next model release can effectively render whatever “research” we did in between the previous model release and now irrelevant, and many papers at conferences (including my own) would become irrelevant at a blazingly fast pace. We can always find justifying scenarios, and disagree on how high the proportion of irrelevant papers are; my take - its too high.

Whereas in contrast, the Technology Readiness Level (TRL) is high enough that working on transitioning LLMs to real world use cases is real, it’s not the chatbot hype of 2014. NLP Research is branching into all things surrounding LLMs from new lightweight training schemes, to Fast inference, to making them play-well with all sorts of input data and generations.

But what do companies really care about? Somehow I doubt most companies care about how well a model can caption an image or whether we can detect LLM generated text. While different companies have different use-cases, a common painpoint would be on how to most painlessly make LLMs play well with existing organisational knowledge which can come in any form of data. We need to be in the front line of new NLP Applications, experience the struggle and excitement of early adopters, and to see and work with their data.

I still enjoy research. But I’m strategically pivoting out of it when it’s the hottest and most lucrative that it’s ever been.


Additional Note to self:

While I won’t be contributing to it, I do regret losing sharpness by moving into strategy and implementation. I’ll be working extra hard to reduce the decay. Go to conferences possibly out of my own pocket, read one paper a month (optimistic), sneak into Jason’s reading group once in a while, advise some junior students (contact me with your GitHub) or aim to submit one paper a year.