Disclaimer: I don’t have any understanding when it comes to finance/economics other than what I’ve experienced/read. However, it does seem hard to argue the fact that economics/finance are so divorced from reality that they ought to be viewed the same as the national debt.
I’m sure there are others in the same boat, but I’m still having trouble getting past the absurdity and ominous nature of everything happening in the corporate world right now. Almost every single conversation is boiled down to “just ask Claude” and it’s being used to slow down existing workflows to a comical degree. Who knows maybe this will self-correct over time, but it’s certainly a choice at the moment.
I’m not trying to be a doomer I’m really not, but these people live in a different reality than you and I. I used to listen to Ed Zitron’s podcast and tuned in to any number of his appearances on other shows to sort of shield myself from all the noise that encompasses this age of tech, but I’m slowly coming to the realization that the economics/financial picture of the situation seems entirely inconsequential. I’m hoping this is just my lack of knowledge when it comes to finance/econ, but fuck if I know. Maybe this post will make me look like a dumbass in a year or so (I hope it does).
I feel like all of us in corporate are all assets in a bet from the most evil people on the planet to see if they will be able to replace huge swaths of the population with a “good enough” solution that will slowly be de-valued and phased out of the workforce akin to IT jobs in the early/mid 2010s.
This shit really blows and I feel like LLM companies are going to come out on top, primarily because Claude has proven that it can generate passable output at lightning speed and that is more than enough to please C suites.
I had a conversation with a long time developer today about this. He basically said his job has moved from “write reviewable code so it is easy to understand and maintain” to “spend all day reading a novel of code that technically works but would be completely unmaintainable if a human had to.”
Like technically the code works and so the “writing the code” part goes faster but then it takes the same amount of time to review, cleanup, refactor, fix, etc as it would have to just write the code right in the first place.
It’s basically shifting the problem space. The machine can solve the problem for you well enough but it takes a real expert to understand what it did and if it solved the problem well enough within the context of everything else going on too.
So it’s not really saving time, and therefore probably isn’t saving money in the long run.
Once the real costs of these machines starts coming down the pipe, I think most teams will realize that at least for doing complex tasks like writing code, it’s cheaper to just have a person do it.
Don’t get me wrong the machines can often write really impressive and correct code, but they are only doing it “faster” if you ignore the context and they’re only doing it cheaper because they’re being subsidized to do it cheaper. Once that subsidy goes away I think we’ll see a lot less of this stuff.
I do think it’s never going away though, I just don’t see a continued world where “just use AI” continues to be the solution to every problem. Especially not complex ones.
Wall Street opened a gigantic credit account for US companies to take on massive technical debt
This has been my experience as well.
Charitably, there have been moments where I have used an LLM to help debug a problem. I was using Zig and trying to decompress and deflate a ZIP archive in memory and the documentation for the APIs in Zig leaves much to be desired. I was able to post a stack trace of the crash and slowly determine a solution but at no point was it where I just let the LLM write the code. It was much more “look at the generated code, determine the change that needs to be made, and try that” with some instances being where the suggestion was completely wrong. BUT, I eventually came to the solution. I would not have been able to figure it out on my own, but it absolutely was not the case where it’s letting the LLM do it and I just review the code.
A similar issue I had was with a crash with mariadb. New version of mariadb gets released and suddenly in development we are getting crashes right at the start of our test runs. I copy out the stack trace from mariadb and paste it into Copilot, and get a search result with some prodding, back to a bug when grant tables are disabled. It was very helpful to understand why, and our fix was to pin our mariadb version at a previous stable version
So, really it’s only useful as a search engine. I would not use it to just generate code.