I just had Sol Ultra read the proof and create a graph of it using Concludia (my side project) so you can explore it visually/graphically. I certainly don't understand it though so I have no idea if it's helpful. :)
I think it doesn't prove much that it hasn't happened yet. Companies might just be moving slower than you think, and are still planning on doing it. And, in many corners, "don't manually write code" is being joined by "don't manually read code" as an attractive principle.
It's interesting to think about how difficult it is to define "fair" with the current system. If a state's population is 60/40 and there are ten districts, do you want them each to be representative at 60/40, and thus 10-0? Or do you want six districts one way and four the other way? In other words, there's a real tension between "competitive districts" and "representative outcomes".
I experience the same with OpenAI, on the $100/month plan. GPT-5.4 is something I still have to challenge: it can bullshit me with bad implementation and add a lot of cruft that costs more time later. GPT-5.5-xhigh is something I have almost complete faith and trust in, it's just smooth. And yet I know the actual token cost of that fully utilized is exorbitant, like as much as an entire salary for a senior developer.
So maybe our CEOs are responding with a lot of foresight and inside information and know that that level of quality is going to be cheap really soon. But barring that, they're going to experience either sticker shock or a slowdown.
I think the real endgame is probably more accurate "models of models" (model routers) that know exactly how to split prompts between expensive frontier and cheap/free local models.
I'm building a website that allows friends to write branching fiction novels together, and another website that allows people to argue and conclude together using first-principles thinking. I've abandoned a project that allows people to register percentage certainty of sports outcomes to improve their calibration - it wasn't fun enough for the users. A few other things besides. I recently wrote a gpl TUI to edit dags here: https://github.com/tunesmith/dagim - that one as something of an experiment since it's in a language I don't know, that one took 3-4 days of steady prompting.
I feel like I must have plateued and don't know what to do next to level up. I'm currently on the $100/month codex plan and it seems fine using 5.5-xhigh all the time. I think of what to do next, have a chat session to determine exactly what to ask for up to the point of being ready to implement, and then codex churns on a commit-sized task whereupon I briefly check it on my local dev server. If necessary I ask for a change. Then I ask it to commit and recommend the next step based off the spec. Oftentimes I have to "approve" an out-of-sandbox request anyway.
I haven't found anything that requires running all night. I could tell it to one-shot a big plan but given how often I realize I want an intermediary thing to be slightly different it seems like a waste of effort.
I'm guessing the next thing I should probably look into is some sort of machine vm I can tunnel my codex-gui requests to so I don't have to deal with the sandbox approvals (I don't want to give it "dangerous" access to my entire mac).
I don't understand what people are doing with their side projects that is leading them to churn through tokens so quickly, to the point of requiring two $200/month subscriptions and a bunch of token charges besides.
That's not what OP was saying, they were saying they experienced multiple instances of LLM handling some tricky problem. But just because it can solve Tuesday's problem doesn't mean it can solve Thursday's.
And this:
>the number of people in that role can be reduced in proportion to the amount of work automated.
Components of humans are not fungible. If one fifth of my job is easier but the other 4/5ths require my specialized human judgment, you can't remove one person out of five and pretend everything will be okay. That's what I mean by the two-step; you just did it yourself.
> How do you determine that this will incorrectly lead to a reduction in the workforce?
This gets back into Theory of Constraints. Identify the constraint. Alleviate the constraint, not the symptom. If you're in a factory, and transmogrifiers are building so many widgets that your whatchamaflorpits starts falling behind, you don't scuttle 20% of your transmogrifiers, you buy more watchamaflorpits!
Instead people are like, "oh gosh, my developers are idle, guess we have to lay them off."
In practice, it's more like companies want to spend money on AI because they believe it will save money somewhere else. If instead they see extra cost, then they get all confused. They can't bring themselves to believe that in their particular case maybe the benefit isn't worth the cost; they're axiomatically conditioned to believe they have to keep using it, and so therefore they have to make cuts somewhere else. It's insane.
I went through this personally. I had a glut of project ideas I wanted to get through. I signed up for the $200/month thing. I caught up. My agent sat idle. It was hard to decide to cut my plan. I felt initial pressure to search and hunt for other ideas to code, ideas that were pretty stupid. I finally downscaled my plan; I got hold of myself. But that's easier to do for an individual than it is for a company.
In normal economic theory it's easier to understand. You're at a particular scale. You have the opportunity to automate, but does it make sense for you? I could go out and buy a riding mower right now, but my lawn is less than a quarter acre. The riding mower lets me scale up, but I don't have something that can benefit from it.
I think there's something else psychological going on. What you describe is a rational approach based off of bad values. But I think I'm also seeing something weirdly irrational.
It's like an (emotional) depression or something. Scarcity thinking, the inability to think expansively. People are so sure that everything around them is shrinking that they feel an instinct to hunker down, shrink, and cut as well. Like it doesn't occur to them that they don't have to feel that way. The execs I work with, none of them strike me as spreadsheet-driven greedy people. They seem more freaked out than that.
> I've had multiple instances now where AI left to it's own devices has solved a tricky problem that I honestly didn't think it was capable of.
Who cares that you've had multiple instances? Everyone has had multiple instances. The question is whether that happens in EVERY instance. Because when someone's laid off, that's what the exec believes, that the person isn't needed at all.
I'm not arguing that AI won't replace jobs - it's clear that jobs are already disappearing "because of AI". I'm not even arguing that it is immoral (even though it is). I'm arguing that it is short-sighted and unwise.
Unfortunately it will take longer for our bosses to walk it back. I feel like I'm fighting the battle daily, telling execs what kind of work LLMs do not replace... it's very slippery, they keep on doing the rhetorical texas two-step - I don't think they even realize they're doing it. We communicate that LLM is amplifying, they hear it can replace. "No, we need humans to help with specs" "But AI can help with that." "But only help, they can't come up with the idea." "Sure they can, we can just ask them."
It's also amazing how hidden some of these realities were before. Like, you assign a ticket to a developer - in the past they just wanted to know the developer was working on it and didn't care so much which work was what. They'd probably be so surprised to find out that a large percentage of implementation was deriving exactly what was meant by the jira ticket or the specification or the product person's intent. Which is all the stuff you have to work on before you can type in a prompt to an LLM. But now there's this pressure to believe that the developers only do the implementation part that the LLMs do, so they can pretend there will be major efficiency improvements. And it's really hard to explain to them what it is that developers even do.
I know I'm not saying anything new here, but at least where I'm working all of these matters feel much more present than they did months ago.
I think there's a way to define it without it either side feeling personally attacked. One of the things I like about using agents for programming is that if the spec is detailed enough, I can implement it in a number of different languages and still get the thing I intended. That means that the "art" is in the spec, not the implementation.
I think the question with AI in music is when it gets to that point. What's the musical spec? What's the implementation? If the spec is supposed to be the pure distillation of my intent, then shouldn't that mean each time I engage AI to "implement" the spec, the musical output of the AI should be the same?
At that point I'm all in favor of using AI for music. But when AI is used to replace a specific intent with vague intent, that's where I feel like something is lost in the human experience of human-created music.
I recently tried to launch a site for friends and family that allowed people to make confidence predictions on various outcomes so they could track their calibration over time. It was like "I'm 84% certain Kansas City will beat Buffalo." I had a lot of fun with it since I'm a nerd about this stuff, and I actually demonstrably improved my calibration. But the only sources I could find for rapid repeatable bets were sports predictions. And I definitely did not want to include money or betting for all the annoying legal reasons. People had fun using it once for March Madness 2025 but traffic really dwindled after that. My conclusion was that the overall subject just wasn't inherently fun enough to do it without money involved, so I made the site dormant.
Getting better calibrated really is worthwhile, I just wish there was more of an appetite to do that without involving money.
Happens to songwriters too, sometimes it goes somewhere. I had a dream once, on the morning of a July 4th, that for some reason took a really common cliche jazz/blues riff and slowed it way down while I was dreaming of a bunch of Americana images. It became this song, the riff is at the beginning: https://music.apple.com/us/album/so-beautiful/899061469?i=89...
I wish it would happen more often, that's only happened for one other song of mine. Most of the rest are a lot of gritted teeth and frustration.
fp, akka cluster, scala java spring php perl python react nextjs, distributed systems, logic, philosophy, music (classical/jazz piano, singing, songwriting)