I literally just finished a podcast interview where I brought this up multiple times. But I think the age of local models on consumer (not prosumer) hardware is coming sooner than most people think.

Sure, but the decision to do that is still a conscious one, and it should be understood that things will happen to invalidate the decision made at some point. That's why we care about avoiding premature optimization, for instance.

*whomst

You're agreeing with what I said though. Tech debt often comes on in the interest of speed or *current* simplicity even though it may not be the right choice for the future. Tech debt is cheaper to take on now, because the cost to fixing that foundation when you actually need to has plummeted.

This has never been a major source of tech debt IME, and is often easy to catch with tools we've had for years (which is partially why it hasn't been a bigger source than, say, management).

enough that you often don't need to read the code to ask an agent to find any and report to you. I think knowing the codebase *helps* but again IME reading code is such a weak preventative I feel ok saying it's unrelated to incurring tech debt.

And I don't think that's the case. Tech debt is (IME) typically incurred consciously, the cruft-buildup form of tech debt is a process issue and doesn't require reading code to avoid or mitigate. Tech debt (and non-TD issues like inappropriate or disjoint architecture choices) are well-known

times, like reduced upfront complexity (in code or infrastructure) to avoid premature optimization that will eventually need to be addressed with eg user growth.

IME tech debt is usually taken on consciously, and agents properly applied tend to be able to spot it more reliably than humans. Whether you take on tech debt or not has changed with agents, because often they can do it “right” to begin with, but a) not always and b) taking on TD is still useful at

Sure they are. Incurring tech debt happens whether or not you read the code, when it’s done consciously, the decision is made outside of reading the code entirely. Even detecting most classes of it doesn’t require reading the code, observability, complexity checking, etc. have existed for years

Probably more parsimoniously, research orgs built by researchers would prefer to do research than build products, something my own experience can attest to. (in product orgs this is largely solved by having separate research and product groups and a small bridge group)

I see a commonly repeated sentiment on Bluesky, the website for 38 year olds, that says everyone was excited by the advent of the Internet in a way that far exceeds any excitement for gen AI/LLMs. The thing is, that's not strictly true.

You can incur tech debt while reading the code, the two things are unrelated. All this means is you can experiment faster and take on tech debt in service of some goal because it’s easier (read: cheaper) to fix in the future.

I’m actually working on a piece about this now. The cost of tech debt has plummeted, allowing teams that need to experiment and fail fast to truly do so, but the debt, since it’s cheaper, must also be paid off faster (possible because of how cheap it is). We are in a nearly-ZIRP era for tech debt.

I’d read the hell out of this!

This wraps up a few problems: publish or perish for those who want to stay in academia and papers being an antiquated medium of knowledge communication. These issues predate LLMs, but they're accelerating it for sure.

I can tell you right now the correct answers are: Meet Pointe Drive By Pies Athan's (although I really need to try Arwa it looks so good)