Why are we even doing all this?
92 points by dcao
92 points by dcao
I think I averaged one laugh per paragraph.
I know the focus is mostly on coders here at Lobsters, and the goal of coding is ostensibly to get something done, while the goal of science is different. I read a comment in the early days of LLMs saying something like there have always been two kinds of coders: those coding for the craft and those coding to get things done. They have unwittingly sat on the same teams for years with no way to distinguish them until LLMs came along, acting like a prism to separate them.
I haven't been in science in a while but I suspect if coders are 80/20 split this way, then scientists are more like 99/1. It will be interesting to see what happens when they are separated.
Yeah the two camps that have always been there but only now are super visible is fascinating, I agree. I also agree on the 99/1 split among scientists and mathematicians (I am a mathematician by training, and lean heavily towards a lot of coding in my research job and also as a hobby).
But I don't see the coder split as 80/20 (craft/goal) – more like 10/90, and as someone in the former camp it kinda terrifies me.
Mathematicians have always benefited from their ivory towers (pure realm of abstraction) but I'm not seeing the same degree of opposition from the squishy sciences.
The messy python/R glue nonsense my biologist friends asked me to review has gotten significantly better since LLMs arrived. General software practices as well.
Trained in math as well; former colleagues and especially the new generation seem distressed. I actually personally think we finally have the impetus for mechanical translation/verification which will usher a new golden age in mathematics, but the growing pains in training and hiring will be real.
Office hours are packed with undergraduates who struggle to see the point of understanding proofs when a model can translate it to Lean in check it in 30 seconds. We're going to have to move up the "why are we learning to multiply if calculators can do it" to higher levels, which I think will overall be good. Undergraduate and early research has always been reprove, convince yourself or work on trivial lemmas/bounds.
The messy python/R glue nonsense my biologist friends asked me to review has gotten significantly better since LLMs arrived.
But the LLMs aren't dong the biology for them, producing barely comprehensible yes/no proofs that do not lead to building of new theories.
I'm unsure on this binary, as I feel that doing it "for the craft" is a bit infantilising. It's not wrong, it's just people can and do believe that "for the craft" is also the best way to get things done when you draw back your view to more global concerns and longer timeframes. You likely also amend "get things done" to "get good things done", which carries a lot of the weight.
Nonetheless, if this binary were to hold, I have discovered that I am the former, crafty person, when I always conceptualized myself as the latter, getting stuff done person. All my working techniques for getting-stuff-done have now been redefined as love-of-craft techniques, and are now more easily dismissed, despite as far as I can tell, continuing to work. I don't know how to feel about this.
"Getting things done", when contrasted with "love of the craft", is actually more like "means to an end", and as is standard with consumerism it turns out a lot of people aren't bothered what the means actually are, or their wider effects, nearly as much as just having the thing in the end.
It's definitely more of a spectrum; in between "just for the love of the craft" and "I don't care just get me the thing" is a whole lot of "the end result is just another part of the process, and the process is what we need to do to feel fulfilled and satisfy our sense of purpose."
A fun reference for this thread might be Art in the Age of Mechanical Reproduction.
Programmers and mathematicians are facing this now, but scribes faced this with the printing press, weavers with the mechanical loom, portrait painters with photography.
The dichotomy of end product vs process is bizarre. Sometimes, the end product is the process. You find the conclusion of this with performance art, but see glimpses of it with "hand-crafted" goods, or photoshoots on film. Sure, you could mechanize the production, or you could shoot digital and just slap a filter on it, but that's not what art/craft is about.
Would you agree on the “binary” classification in the context of a day job/necessity vs. hobbyist/leisure activity? Or, under external pressure vs. no pressure as the distinction, essentially?
That's not been my experience up until now, I have always managed to align my professional and personal selves rather well, with the acknowledgement that the values I bring to an organization have a significant, but not perfect, overlap with demonstrable corporate success. I could back this up with some small amount of data (a good example is the DORA metrics as they relate to software delivery and company performance, before they themselves started to focus AI and stop performing as rigorous science, in my view.) This was the compromise, or deal, I made with society.
Now, this is much less easy, so perhaps I simply never experienced the same sort of external pressure that others felt. This is sad, because under this sort of external pressure, where even your opinions on the micro-tasks of your profession are under scrutiny, I don't see much point in hiring experts ... you're not allowing them to be effective experts in their fields. Maybe this is the majority of experience and I have just been privileged up until now.
I read a comment in the early days of LLMs saying something like there have always been two kinds of coders: those coding for the craft and those coding to get things done
It is true that we are seeing a sorting. But this is a false dichotomy.
What I would propose instead is that the practice of programming is about improving and formalizing our understanding of processes and systems, for the benefit of our shared humanity and all the life of this planet. This is neither "just" craft, nor mere production, it is a calling in alignment with the entire history of philosophy that contributed to the discovery of computability and of the ideals of the enlightenment. This view positions programming as a calling alongside the sciences (wherein we also have applications in industry that are fundamentally connected to research in academia).
The function of "software" as a concept has been do commoditize one material result of the practice of programming (executable programs) and to segment IT labor. Since LLMs can automate the production of many executable programs, it is indeed poised to destroy the industry oriented around software production, so defined. But programming is a higher calling, with a much longer history, that includes the cybernetics efforts of the mid 20th century, the formalization and mechanization of computability, developments of formal reasoning, of mathematics, of economic planning and analysis, of metaphysics, of the emancipatory aims of the enlightenment: it draws from and continue sall of these essential activities, and the fact that LLMs can perform inscruitable, referentially opaque, probabilistic computations and (unreliably) produce executable programs does nothing to undermine or threaten the meaning of the practice -- not, that is, unless we surrender the meaning and purpose of our work to the misapplication of LLMs by those seeking to seize control of our rational use and ownership of our own means of computation.
Man, I love me some good satire. And this is well written. Not by the likes of Terry Pratchett, but by a theoretical physicist, of all things.
Kind of makes me want to research the low-level mechanics of satire a little better. I’m curious how well, or not well, modern LLMs would be able to output useful satire. There would be a sort of meta irony in that.
On a more serious note, I’ve worked in process automation, writing software, my whole career. which only has about 10 years left. One of the questions I have asked myself increasingly, through these last 35 years, is “What is humanity going to do with itself, once we automate everything?” the last five years have definitely brought that into much sharper focus, for a much wider audience, than I used to have when I asked that question.
I raised this same point on Hacker News and was laughed at. I guess the AI boner is too big there, but I'm glad actual mathematicians agree.
They're wasting precious scientists' time by making them review these math papers in the hopes maybe one of these slop proofs are correct or open up an avenue for an actual proof. This is no different from the spray and pray approach AI contributors take to Open Source projects, code dumping slop on them and having the community sort through the manure.