Comment by dustingetz | Hacker News Reader

Comment by dustingetz | Hacker News Reader

yeah, i think the siren song of reducing developer cost and accelerating delivery is irresistible to boards, execs, managers, non-senior devs (all essentially semi-technical, even the nonexpert devs) but it is a mirage. The startups that over raised in the extended bubble are all drowning in tech debt and the game is to exit or get profitable faster than the tech debt compounds, it leads to a divergence in outcomes where you either build a google and buy your way out of the hole (“ revenue solves all problems”) or you don’t quite get there and your delivery ability crests and decays as the tech debt compounds above the threshold line that the company cash flow can service.

Vibe coding is going to exacerbate this - the growth can come faster, but the collapse point is accelerated correspondingly. If you plot this complexity out like we were taught in CS 101, lines of code will grow linearly, proportionate to the number of agents/devs, but the interconnectedness and therefore state space will explode polynomially without applying encapsulation and local reasoning techniques like functional programming. But we’ve aligned the models to use imperative languages (TS, Python, Java), essentially constraining the code models to specifically prevent the best code scaling techniques! And guaranteeing the emergent “spooky action at a distance” that is inherent to the computational structure of imperative languages.

And note that most companies don’t take flight like Google, instead they seek acquisition, and a lot of those acquisitions fail to integrate and get swimming, they are perhaps underwhelming compared to the sales pitch, to the extent that many start to look like a form of securities fraud? Tech debt is real, it’s just invisible to the balance sheet and therefore to boards and buyers. If it were properly accounted for, the entire IT sector would be revealed as completely underwater.

> but the interconnectedness and therefore state space will explode polynomially without applying encapsulation and local reasoning

I wish the complexity demon was this gentle. It's exponential generally – a single boolean flag can double the test/state space.

The only thing that really matters long term is taming complexity. Humans fail at this and that’s why large companies need to hire armies of big fixers and QA to play whack-a-mole with the symptoms rather than the causes.

LLMs make both tasks faster, but it also changes the balance between them. I would be wary of replacing a company of 5 wizards, 20 seniors and 100 juniors with 10000 juniors overnight.

do you think it’s really exponential in practice? i don’t know if i see that, and polynomial (growth in bugs per LOC) is enough to explain why all sufficiently large projects grind to a halt

By the time the tech debt of the vibe code hits, new models will appear which can still handle it. And we are close to being able to point the model at the repo and ask it to "reduce the tech debt". Then it becomes a money problem - put the agents on a 24/7 reduce the code debt mission.

If the cost per unit delivery (say LOC) grows something like polynomial due to interconnectedness, surely it is superlinear for even a state of art formal reasoning system let alone imperative langs, it will outstrip our ability to build energy infrastructure. If the cost growth is exponential due to state space explosion, it will outstrip the energy in the Sun.

How does HN think these factors will balance?

Yet humans can handle and untangle spaghetti code without requiring fusion power plants. There is no fundamental reason why LLM would fail to do it too. And as they improve the code they produce is better architected. So the situation is improving at both ends.

humans cannot in any sufficiently large project, for example government contracts and the entire banking sector.

They can. They are not allowed due to "reasons".

It’s rational - the risk of breaking things by changing them outweighs the ongoing costs of keeping them the same - including developer sanity, unfortunately.

An often overlooked point in these LLM vs human debates: the majority of humans are extremely fallible.

I think that's the bet many investors are making at the moment. And if it pays off, it will do so big times as it will be almost impossible to maintain a codebase without an LLM. Therefore making them indispensable for the company using/working on said codebase. If they continue to perform linearly from the point we are now, though, then the only option is to continue making better LLMs capable of comprehending the relative mess the previous one made. Which at the limit means betting on AGI. If for some reason the progress doesn't come fast enough and too many people are making that bet, then developpers (the ones left) can expect even higher salaries than before...

The investors may be right, and yet, this is a collapse vector in “The Collapse of Complex Societies”, in one sentence, “societies collapse when they reach a point of rapidly declining marginal returns on their investments in problem solving capacity.” So we’d expect an outcome somewhere between “market dynamics will resolve this” and “war and/or violent revolution”

Why is producing all that waste and complexity necessary in the first place?

We often get distracted by benefits and forget to look at the costs. And we often overdo things to a degree where previous wisdom has to be re-discovered again and again.

Having experienced a couple of hype cycles, I learned that the sensible thing is to narrow down what the new thing is useful for specifically.

And all that's assuming that the prediction holds. I don't see any clear signs of whether the usefulness and precision is primarily a quantitative or qualitative issue.

Reply on news.ycombinator.com