having a bias toward action is great. it helps you get out of decision paralysis, but i'm not quite sure that still holds true in the age of AI. when writing tons and tons of code or completing an entire research report or writing a paper comes at the snap of a finger, action by itself is no longer as valuable as it used to be.
the adage "talk is cheap", still holds true. but perhaps in today's world, it ought to be extended to "talk is cheap, and action is too". given the capabilities of modern large language models (LLMs) that will only continue to improve, we can no longer expect action to be a signal of work. don't get me wrong, i'm not a stickler for time/effort spent = quality, but with the tools accessible to us today, it's reached the point where i think i prefer plain slides and unpolished writing (though it's easy to copy this with LLMs too) because it signals that a human put their heart, or at least effort, into it. it's really painful to spend your time and attention reading something that you know the other party didn't care enough to put in any more effort than providing a prompt with plenty of context to a LLM. it was not until i taught a student-led class that i embraced "never use LLMs to (completely) write something other people will read, maybe with the exception of code".
on the topic of school, there's been plenty of studies/reports that students do well on homeworks or take-home exams, then end up doing terribly in proctored finals. i am no exception. when i started taking higher level math classes, i thought that spending more than 30 minutes stuck on one math homework problem was not worth the time, and it was probably smarter to ask a LLM for help. i proudly justified that as long as i took the time to understand it, i was learning from AI. better than simply copying i thought; no better was it actually. at least not in the way i was using it. reading a completed proof and understanding what every word and argument means individually is not the same as having strong priors of definitions and using them together in a way coherent to myself to build up the full proof (i am not claiming here that one should dedicate their life to figuring out one question to a problem set (which i would fail to do anyway). but LLMs have made it so that when you get an answer, which you presumably also do when asking your TA/professor, it's easy to accept it as done and understood rather than doing it from scratch again and truly understanding). it is thus my goal this coming semester to use as little aid from LLMs in my classes as possible, or at least to use them tastefully.
here's another example: when recruiting for internships, i thought i would show agency by learning and building one part of the company's product before an interview. oh how it backfired when the CTO asked questions about it that i couldn't answer, and i proceeded to be lectured about vibecoding and not having an understanding of what i built (which in my defense, i did have a good understanding of the fundamentals, just not to the extent that i could reason about its scalability). as much as it had hurt my ego, he was right. when code, in this sense synonymous with action, is so cheap, having a product built is no longer the signal it used to be.
to summarize the many tangents i've gone on, action is necessary but insufficient. thoughtful action is what i've converged upon. but how to define thoughtful?
i was recently introduced to the playful description of AI models (i expand from LLMs to include VLMs, world models, etc.) as machine god. before long, if not already, AI systems will make decisions across facets of human life. until then, it is often already taken as a source of truth. but what truth? in LLMs, definitionally it's some permutation of all of human text (limited to what's included in the model's pretraining data). when AI is so accurate and (presumably) doesn't have ulterior motives, how much trust should we put into such a machine? what about god? don't get me wrong, i think that there is a lot of good to religion and faith, but not to the extent of blind faith (maybe that is the point of religion, but not what i'm trying to get into here). as AI systems become more widespread and those without as much knowledge about the make-up and gaps adopt them, it's not unreasonable to envision that there will be increasing trust and faith put into these machines. it's not unlike how you and i trust the computers and phones we use every day. but this time, we can offload all of our thinking to it, and i think that's dangerous. abstraction, trusting that something works as described without completely understanding it, is important because there is too much in the world for us to know. but how much is too much, and can we abstract away thinking? Pope, you are right. we cannot treat AI as a human much less (a) god, and thus it certainly cannot think on our behalf. there are so many use cases of AI: a tool to automate repetitive tasks, a tool to bring aspects of code-writing to the level of natural language, a tool to gather many sources of information, etc. but at least for now, that's as far as it should go, being a tool (sorry agents).
as much as i want to keep AI as a tool, there is always pressure to move quicker: others using AI can do so much more and much much faster, i'm falling behind. and so back to fully trusting AI, becauase that's what's quickest and more often than not correct. sometimes i feel like layer of friction between agents and the real world. will we reach a point when autonomous AI runs the world and our experience of it? i believe in humanity, and my answer is no.
as a means of capturing the real world, cameras became ubiquitous, but we still learn to paint. computers can beat any human at chess, but we still watch humans play. there is beauty in human work. there is beauty in inefficiency. there is beauty in struggling and questioning yourself because you're thinking about all the other things you could be doing with your time, but being immersed in the experience and having immense pride that you, a human who began as a fetus and have now grown to be a beautifully weird and complex creature, did it. not you who spent 5 minutes tailoring a prompt to feed to a LLM and claiming the output as your hardwork, which by the way was the life's work of thousands of researchers. but you who put your heart into something that makes humanity even just that tiny bit better, even if an AI tool had assisted in some way. with all that said, i'm running towards things that will force me to embrace the friction and to not just take action, but thoughtful action.
footnote
this started off as a means to outline how i think about bias toward action now: if i find myself spending way too much time thinking, take action so that i can get feedback and iterate on it. but it is unwise to take action without any prior thought since that is so easy now. what's minimally necessary is some thinking about the kind of feedback i want so that i can update my beliefs and improve. anyway that was the thesis of my thoughts before i went down this whole rabbit hole of AI. also, i acknowledge i left "thoughtful" ill-defined, but that's something i have only a general direction of. also also, i think writing this has further convinced myself that ai safety is an especially important field to run towards.