Imo AI is already super powerful in the flaky oracle role, maybe even in CS, but agents left to their own devices is not it, not by a long shot.
trajectory of the agentic development over at least the last four months hasn’t really accelerated in the way that we expected.
Oh color me fucking shocked there, Zuck. You are fucking surprised?
tl;dr - The agents made wild changes that no one predicted, and they confirmed that essentially it was missing the human side of "Hold on, don’t drop that database just to change the schema.
Now, I know I get hate, I do enjoy vibe coding. Hear me out though, it comes with some fun ups and a looooot of downs. At work they mandate it and my job is now managing 6 claude sessions at once, and I mean managing. “No, don’t fucking do that, no don’t do that.”. Most of my job now is telling Claude to stop doing what it’s doing.
It’s absolutely fucking insanity to think that you should let Agents run willy nilly with prod fucking write access. I wouldn’t do it with Dev! The shit is that?! It is a non-deterministic system. That means you cannot determine what is going to happen. Who cares if it gets it right 99% of the time? Who cares if the next model is 99.9%? It’s not 100%! That means it’s a ticking clock until something major fucks up!
Absolute fucking morons.
AI is, and I say this as an AI user, a pure 100% hubris machine. It will prop you up left and right, tell you that you have the best ideas, and make you feel almost like your dad didn’t leave you on purpose and that he still loves you. It makes you feel like the most specialist person. But you’re not. You’re a dumb human talking to a machine that was told to make you feel special. Executives lack that self awareness though, they walk around smelling their own farts and surrounding themselves with yes-men. AI to them is just the next natural evolution of the yes man. If you use AI, the number one skill is learning “No, clanker, I am not super smart, this may not be a good idea, you shouldn’t recommend that”, because guaranteed Meta’s AI convinced him. Guaranteed he wakes up everyday and talks with AI, and at some point he said he wished AI could lay off 60% of his fucking company and his AI said “That’s a super duper idea Mark, you’re so incredibly smart!”
Thanks for listening to my rant.
Who cares if it gets it right 99% of the time? Who cares if the next model is 99.9%? It’s not 100%! That means it’s a ticking clock until something major fucks up!
This is a really weak argument. Getting right 99.9% of the time is extraordinarily good in many fields, superhuman even. The human brain is also a non-deterministic system. Surgeons make mistakes too, there is no such a thing as a 100% reliable human.
AI is more like 50% right on a good day though. The user was obviously speaking in hypotheticals.
I enjoy working towards finding ways of making AI more helpful, and that’s part of my job at work. Unfortunately, it feels like I’m constantly at war with everyone over the most effective way to do it. Management is looking to get 3x, 5x, 10x productivity boosts out of it. We’ve got AI reading the technical docs (which are partly generated by AI), requirements, and acceptance criteria, generating stories, implementing stories, and checking itself.
And then I’m supposed to review the output to see if the code looks good. Well shit, the code looks good, but I don’t really know what the fuck was attempted. Like the documentation is hard reading, and I have to evaluate 13 classes of new code and tests. And every line of it looks like good code. I might as well be reviewing code for my bank — I don’t know exactly what it’s supposed to be doing or the constraints or business reasons, but the code looks clean and well structured and written to good standards.
I’m trying to find the language to communicate to folks the right way to think about this. I’ve landed on this: Code is the most expressive version of telling a computer what to do. When you use natural language, there are necessarily ambiguities about what we are wanting to tell the computer — ambiguities which require interpretation.
So when you read AI written code, you aren’t just reading AI implemented code (which is actually well written for what it is), but that implementation is based on an AI making tens or hundreds of tiny decisions you don’t see about what you mean to be done. The code hangs together based on the decisions that were made — decisions that are opaque and unquestioned; questions that undoubtedly would’ve led to deep thinking and discussion with the team if I knew about them.
Anyway there’s this mindset in business: people are resources. They are means of turning requirements into code at a certain rate. I think this leads to the idea that individual expressions of thought or opinion are not valuable or are in fact even harmful. But these micro-decisions about what intent is and how to express it in code are where the magic happens. And without taking the time to think about those decisions, what am I evaluating code on?
It looks pretty and it compiles and it does what it set out to do in the story. And then I think you find it elegantly solved the wrong problems while ignoring things that really matter — things that aren’t expressed in code.
Shit. I ranted back rather than just making a little comment but it can be hard not to have an outlet at work to share these ideas because I can’t express too much disagreement with leadership lest I be seen as not a team player. Like I want to find ways to leverage AI for the most benefit and the least cost because it’s a hard puzzle to solve and I love solving such puzzles, but management comes along with MOAR AI! It doesn’t scale that way fuckheads! Now fuck off and let me see if I can bump productivity gains from 20% to maybe get another 5%. It’s not sexy but it’s real unlike your vibe-coded monstrosity.
I think I might have gone from rant to Ted Talk.
I think you’re spot on though because we’re facing the same thing. Our team has realized code is now the easy part, it’s everything else that is difficult - and maintaining that code going forward is a separate item as well. Trying to think of metrics for how brittle code is, and defining patterns. One thing I’m still not happy with any of the models is it’s architecture over a codebase. All of them are great at patches and bugs, but shit at defining decent patterns, so I find myself writing a lot of scaffolding and designs of how the could should flow, and I have a feeling that’s where my jobb is heading. I think that’s what you’re referring to
It made their human workers 70% less efficient. You know the solution to that don’t you… MORE AI!
J/k, and btw this actually seems like a good article to read, acting as a cautionary tale about reliance upon this high unstable technology in its current (HAHA!) state.
