recently i was reflecting on why working part-time at transluce for a few months has felt more productive than working full-time at my old job for a year. there are plenty of potential reasons to point to - i have more knowledge about the field now (in large part thanks to the previous job), the current team is more focused and experienced, i’m happier and healthier in general, and so on - but i think one less obvious reason is that my approach to research has changed
there was one particular project last summer where all the results were difficult to interpret and rerunning experiments was incredibly annoying. because of this i did everything i could to justify not running experiments, which usually meant coming up with bad interpretations of the existing results to prove/disprove hypotheses without the need for new data. this sounds very stupid but i did not realize it was happening at the time - when data is very opaque and you’re emotionally invested in justifying a conclusion it’s easy to manufacture evidence supporting that conclusion without actively trying. another way of saying this is that i was more scared of being wrong than i was interested in uncovering the truth, so i kept finding ways to demonstrate i was right
my recent work has been different because i am usually not emotionally invested in any particular hypothesis, or at least i see the value in multiple hypotheses instead of just one? like, i spent half of january searching for gendered neurons in language models, and finding them would’ve been cool, but not finding them would also reveal something interesting. and i no longer avoid looking at data; in fact i spend hours at a time doing so and have decided i enjoy being as close to the data as possible
i often think about that rilke quote - the purpose of life is to be defeated by greater and greater things. translated into research it means: i am always wrong, but the goal is to be wrong in smaller and smaller ways, and asserting my own correctness doesn’t help with that so there’s no point in trying
after initially being confused about not hearing anything from schools i applied to, i ended up having 5 interviews over the last few weeks. they were a lot of fun! profs asked me all sorts of questions i wasn’t expecting, like “are you more of a mathematician or an empiricist? how much math do you need to do to be happy?” “how would you approach designing the next generation of ai interfaces?” “i see you have a blog, do you think enjoying writing blog posts translates into enjoying writing papers?” “why did you stop working on computational biology?”
(my answers: i mostly like empirical work with a bias for conceptually simple and mathematically grounded ideas. new interfaces should be task-specific and surveying a much wider variety of tasks will give us ideas for how to move beyond the current chat / autocomplete paradigm. blogging sadly does not translate into writing papers. i think compbio is bottlenecked by biology rather than computation)
unfortunately i think i annoyed many of my interviewers by telling them that i don’t want to work on their problem of interest because i think it’s largely been solved. that probably was… not a smart thing to say from an admissions perspective, but i’ve decided to double down on truth as a core value, and one prof did say i was refreshingly honest. i don’t think what i said was wrong, in the sense that a lot of frontier industry research goes unpublished so it’s true that many problems academics care about have genuinely been solved already. but i also think this reflects a failure on my part, which is that i don’t have much interest in reproducing existing research even though there can be a lot of value in doing so, and i would benefit from being more excited to repeat work
simply do not do research