When We Did Research By Hand
How ML/NLP Research Changed in the Course of My PhD
I’m approaching the end of my time at CMU and I’ve been reflecting on my PhD experience. In hindsight, though my PhD itself technically didn’t last an absurdly long time, I feel like I’ve witnessed a lot of change in both the research landscape and myself. Things have changed a lot, and I find that my mentors are also uncertain on what the future holds when I ask for advice. I find myself also hedging when asked by younger students: what should they focus their work on nowadays? Does it make sense to commit themselves to a 5 year or longer PhD anymore? Will their intellectual labour still hold value in the next decade?
I don’t claim to know the answer to any of these questions, but this post is an attempt to try to muddle through them in some sense. Looking back on my experiences since 2021, there are a few details that stand out: these years formed roughly 3 eras, following the processes that I used to do my own work, what other people were doing, and what people considered important and roughly speaking, took pride in.
Of course, this is based on my subjective perception and not any formal analysis. Looking back on each of the eras with hindsight, it’s easy to kick myself and wonder why I didn’t just do X earlier. In particular, I adopted agentic workflows pretty late and tended to dismiss hype too easily. In part, this post is also here as a marker for my future self as well: what do you do now? Do you still care about the things that I listed in the third era here? Have you moved on to something besides research, that I wouldn’t recognize today?
Era 1: Post-BERT, pre-ChatGPT
This lasted from late 2021 until late 2022, roughly my first year in the PhD. It was my first year and I wasn’t too familiar with NLP yet, so I spent the time mostly trying to learn a lot and reading a bunch of papers. I also attended my first NLP conference, NAACL 2022 in Seattle. Ultimately I still think this was one of my best conference experiences, partially because things were new and exciting, and also because it was one of the first conferences post-COVID and people were excited to socialize and do things again. What was interesting about this conference was that there remained a lot of “academic speculation” type projects, of interesting ideas that didn’t work, and also trying to probe models for various attributes. There was a lot of benchmarking of things like common sense, but with the understanding that models were still quite poor at it. Although transformer-based LMs were of course the default, there was a tacit understanding among most of the papers that “these models are cool but a lot still needs to be done to make them work”.
As for what I was doing during this period, I started off very interested in analogies and figurative language and did my “getting started” type papers in this area [1], [2], [3]. To those that know me now for (pre/mid)-training type work, this may seem like quite a weird initialization. However, I was really interested in the ability to understand abstract language and make analogical leaps from a cognitive science perspective since in my mind this is a very important ability for coming up with new innovations both creatively and scientifically. People have recently started to explore this ability to “leap” in terms of generating scientific ideas [6], though I don’t keep up with this area much anymore. I am still interested in creative writing to this day as well, and I could write a separate post about that (and why I think improving on creative writing was abandoned too quickly in favour of domains with overt verifiable rewards), but that’s a post for another day.
I was writing code by hand and looking things up on StackOverflow in these days. It’s funny because I haven’t visited StackOverflow in probably 2 years at this point, but it used to be such a big part of my life throughout college and my first year of grad school. I’ll never forget the feeling of finding someone with the same niche problem as me, only to see that no one had commented besides the author with a brief message: “nvm, figured it out.” I didn’t really do very intensive coding in my projects at this point, but I enjoyed sitting down to code things and debug, though it was sometimes painstaking work. In a way I think debugging is a microcosm of the scientific process: you come up with hypotheses and then methodically rule them out until whatever’s left must be the truth. I used to have quite a kinetic way of debugging, I would fire up pdb, or gdb in earlier days, and step through the code until I had almost a physical sense of what the code was doing. Though it took a long time to accomplish this, I always felt pretty accomplished when I figured things out this way.
At this point, you might think I’m painting a very rosy picture of research at this time. I think part of that is because I was new to the field and things are a lot more exciting when you’re new to something completely independent of what’s actually going on. In the end, I’m not trying to create a completely unbiased view of the field, but just trying to remember accurately what it was like for me at that time. Another thing is it felt like the pace of the field hadn’t accelerated to a crazy degree yet, and though there was starting to be some AGI hype in corners, the field felt like it was moving at quite a manageable and peaceful pace (compared to now). One downside that I will mention though is that in hindsight, people wasted a lot of time on things that would just be eclipsed by scaling in the extremely near future.
Era 2: Post-chatGPT, pre-agent
This period began very abruptly with the release of chatGPT in November of 2022. This time period has been chronicled very well by others, so I won’t rehash it for too long. What I can say that there was an initial sense of hopelessness in the academic community, as many people had their work become obsolete overnight. I remember attending EMNLP during this initial period of confusion, and people were joking that many of the papers were just made redundant and discussing what to do next. I think this period also accelerated a slow migration from ACL conferences to ML conferences that had been happening in the background. For me, it perhaps wasn’t that bad as I hadn’t actually done that much work yet at this point and was at an ideal place to pivot. However, despite this I was also still confused about what I could contribute in these days.
In hindsight, I was relatively slow to adapt to changing my workflows overall, especially since the interface of LLMs was still chat based at the time and at the beginning they didn’t have tools or web search. I would sometimes paste in my code to debug but mostly still coded by hand at the beginning of this period with augmentation from Github Copilot (the initial autocomplete version). I even viewed this tab-based autocomplete as magical at the time, as it saved me from writing repetitive boilerplate. Near the end of this period, I let the LM fully write much of my code, though I used the cringeworthy workflow of copy-pasting from the chat window and iteratively debugging by reporting any issues that came up. It was still more like having an advisor, as I would often still do things myself and also debug on my own.
In terms of what I found interesting at this time, it was really a hodgepodge. A lot of research seemed uninspiring; there were a lot of papers testing whether ChatGPT could do something or other for just about anything the authors thought of. It seems hypocritical to complain about this when I was writing similar papers just before this, but I really did get the sense that many of these papers were uninspired. Perhaps the reader can judge the extent of my hypocrisy. Not finding much to be inspired by in contemporary reading, I also entered a bit of a rut for other personal life and health reasons. Still, I did manage to write a paper on abductive reasoning [4] which I still find to be an interesting topic. I also began to be interested in pretraining during this period, and ended up carrying out a not very well planned project which ended up as a paper called Not-just-scaling-laws which is pretty much what it sounds like [5]. However, it was taking a long time to carry out projects and combined with everything else and my awkward way of using LLMs, this era was a pretty huge slog.
Conference acceptances continued to get noisier at this time due to the ever-increasing numbers of submissions. I think that this situation is now (in 2026) near a breaking point but this just continues the previous trend. I was pretty lucky with conference acceptances up until this period, but I started to get some rejections around this point. As for what I took pride in or enjoyed doing…honestly I’m not quite sure. It was an awkward middle ground where AI was doing some work for me programming-wise but not enough for me to really take it easy. Otherwise my work processes were more or less exactly the same, apart from having LLMs format tables and look up references (this was a slight mistake). Since projects were taking an insanely long time, probably I was just happy to finally finish projects and send them across the finish line.
Era 3: Post-agent
This is our current era, and to be honest I’m not quite sure what comes after this. Initially, I wanted to write this post because of the recent advances in mathematics made by LLMs such as Claude Fable and GPT-5.6. Although I don’t work on anything related to pure math, it’s made me think about what the role of researchers will be in the future. I’ve recently started using agents a lot more and expect to use them end-to-end for the first time for projects next year. Iterating on experiments has become much easier, and anything implementation-wise at least when it comes to normal Python programming and most non-systems ML code just isn’t daunting at all. I know that I can simply tell an agent to do it, and honestly in many ways it’s amazing. It removes a major source of stress and uncertainty, making planning much easier (“I don’t know when I’m going to finish implementing/debugging this…”). Yet at the same time I also feel like I’ve lost something that I used to value: I liked running experiments myself and even built experimental automation systems that I enjoyed tinkering on (more on this another day, I still think they are useful in the days of agents and will talk more about experimental metaprogramming…). However, the skill of writing code and doing experiments is no longer useful by itself.
Earlier on in this period I was using the agentic version of Copilot with various models, but have since switched to Codex for personal work. Separately from everything that’s happening, I do feel pretty aligned on what I should be doing now – papers from this period include midtraining [7] and the implicit curriculum hypothesis [8]. I like the collaborators that I have in this general area which I call the developmental science of model training. I also hope to be able to do some more interesting work in this area before I graduate. Maybe a way to preserve pride in what you’re doing is to focus on the bigger picture and on why you find things interesting rather than strictly focusing on personal pride in the mechanics of doing things.
It’s inevitable that expectations of productivity will rise as well – they have certainly in industry and I wonder if future PhD students will have greatly increased expectations as well. Will they need to publish at least 10 first author papers to graduate rather than 3? Though perhaps evaluation will also be shifted to a higher level, of leading a coherent research programme with a swarm of agents or something along those lines.
As everyone produces more and more papers, conference acceptance continues to get more stochastic as well. Often reviews and papers themselves are LLM generated now. Although LLMs have automated large parts of the research process now, I think there will continue to be an uncanny gap in the short term, which is now coming to pass through a flood of papers that is overwhelming peer review (see my previous blog post on AI Scientists). I still go to conferences because I can’t say no to a free vacation and catching up with people from different institutions, but I increasingly wonder how much longer this model of publication will be around for. Perhaps conferences will evolve to being more purely about socialization and feedback on early ideas without the dressings of presenting papers, though to be honest I already go in with this attitude to begin with.
There is a lot of interest in RSI and automating research in industry now, starting with ML research as it generally requires little contact with the physical world and is easily verifiable. I find this exciting and fun to play with on some level, but on another level I hope that RSI isn’t so easily solved. This is for selfish reasons perhaps, as I would like to keep doing my job and I’ve already invested half a decade in it. But I do think that anything can happen from here, and I want to keep an open mind because I’ve been slow to pick up on things in the past. One positive thing is that by commoditizing technical mastery, a lot more people could enter the field. Things could become pretty exciting due to this, as many people are currently excluded from the academic community for systemic reasons.
As of now though, we still have to decide as humans what we find important and what to pursue. Much of what I’ve done in my PhD has been pretty useless and probably won’t impact the world at all, but working on it has shaped and impacted me. As humans, I think working on hard things changes how we approach new problems and develops us as people. That’s why I’m grateful to have done my PhD during this wild time and I still look forward to the future.