[aspenmartin]

- systemic tech debt is now addressable at scale with LLMs. Future models will be good enough to sustain this, if people don’t believe this I would challenge them to explain why. First consider if you understand what scaling laws are like chinchilla and how RL with verification works fundamentally

- I completely agree with you about fundamentally the limitation being the business able to coherently articulate itself and its strategy

- BUT the benefit now is you can basically prototype for free. Before we had to be extremely careful with engineer headcount investment. Now we can try many more things under the same time constraints.

[jugg1es]

The problem with tech debt is not that it is some poorly designed code in a few repositories that can just be changed. True tech debt is the kind that requires significant architectural changes across many systems and is almost always coupled with major data migrations. You need the rest of the business to agree that you want to invest all that time and energy to fix a problem someone else created 10 years ago. You likely will also need other teams to set aside time on their own road map to address it. You also might need customers to change what they are doing because if software lets you do something, you can guarantee that someone has learned to do it - even if that 'something' was actually a bug.

LLMs don't solve any of those problems by itself.

[aspenmartin]

> You likely will also need other teams to set aside time on their own road map to address it.

This is the part that’s no longer true. Everything else you said is spot on. But everything you mentioned — making sound decisions in the vast context of the business and enormously complex codebase, dealing with ambiguity and customer implications. None of this is inherently untouchable by AI

[visarga]

> BUT the benefit now is you can basically prototype for free.

But.. so can your competitors. And that changes the value proposition.

[aspenmartin]

How do you mean?

[imtringued]

>- systemic tech debt is now addressable at scale with LLMs. Future models will be good enough to sustain this, if people don’t believe this I would challenge them to explain why.

Is this some sort of troll attempt? Like, are you fundamentally misunderstanding the problem with tech debt? This is the equivalent of throwing garbage on the floor and expecting professional cleaners to keep your house clean.

You can produce tech debt faster than you can pay it back, that's the core aspect of tech debt. If tech debt was more expensive in the short term than not doing it, nobody would be doing it.

A labor saving device doesn't reduce or deal with tech debt since tech debt is a decision made independently of the competence of the developers. If you have a company with a tech debt culture, the labor saving device will just let you accumulate more tech debt until you reach the same level of burden per person.

>First consider if you understand what scaling laws are like chinchilla and how RL with verification works fundamentally

Honestly, this tells me that you basically understand nothing, not even chinchilla scaling laws and how RL works. Not only are you trying to brute force the problem, you're listing completely irrelevant factors to the problem at hand.

Chinchilla scaling laws are "ancient" by LLM standards. Everyone who designs a model architecture that is supposed to beat their competitors is pulling out every trick in the books and then come up with their own on top of that and chinchilla scaling laws have been done to death in that regard.

Reinforcement Learning is also a pretty bad example here, because there is no obvious way to encode a reward function to deal with something as ill defined as tech debt. You didn't even say avoid tech debt which would be actionable to some extent, just "systemic tech debt is now addressable at scale with LLMs". I.e. you're implying that if LLMs were to generate tech debt, you can just keep scaling and produce more of it, solving the problem once and for all Futurama style with ever bigger ice cubes.

[aspenmartin]

- Not a troll.

Both of these lectures misunderstand my point and how things work.

- “tech debt” is not some special problem…? You accumulate cruft and bad design decisions…you spend tokens to fix this. Is your point there is always a fundamental tension between spending tokens on new stuff and spending tokens on cleaning stuff?

> Honestly, this tells me that you basically understand nothing, not even chinchilla scaling laws and how RL works. Not only are you trying to brute force the problem, you're listing completely irrelevant factors to the problem at hand.

That’s a very interesting take because I would say the same thing! RL and scaling laws are not relevant to the performance and capabilities of coding agents? Thats something you don’t hear everyday

- chinchilla-like scaling laws are not ancient…people try to derive scaling laws for new paradigms all the time it is how researchers get their company/lab to invest in scaling up a new idea. No idea what you mean here. Maybe you think I meant “the literal constants from the chinchilla paper”? No I mean: scaling laws generally, and Chinchilla, due to the impact of that work, is used more generally. Regardless, scaling laws generally continue to hold, and in fact improve with architectural/data mix/training recipes.

> Reinforcement Learning is also a pretty bad example here, because there is no obvious way to encode a reward function to deal with something as ill defined as tech debt.

Well that’s a bit of a strong claim to make… I don’t agree with this at face value but even if I did, you don’t need to explicitly do RL on tech debt as a specific task.. you do RL to build better programming skills generally which then generalize to many coding tasks.

> You didn't even say avoid tech debt which would be actionable to some extent, just "systemic tech debt is now addressable at scale with LLMs".

Tech debt is strategic, why avoid it?

> you're implying that if LLMs were to generate tech debt, you can just keep scaling and produce more of it, solving the problem once and for all Futurama style with ever bigger ice cubes.

I’m saying you can take, successively over time larger and larger, and more complex codebases with thorny debt problems and resolve them by spending money on tokens.

You keep scaling and, just like we do today, decide when some tech debt austerity needs to take place. I’m saying “the guy that built our house of cards over 10 years and left” is no longer so devastating and expensive a problem as it was before

[crabbone]

> systemic tech debt is now addressable at scale with LLMs.

Is there any reason to believe this? I've only seen the evidence of the contrary so far.

My experience with AI coding aides is that they, generally:

1. Don't have an opinion.

2. Are trained on code written using practices that increase technical debt.

3. Lack in the greater perspective department, more focused on concrete, superficial and immediate.

I think, I need to elaborate on the first and explain how it's relevant to the question. I'll start with an example. We have an AI reviewer and recently had migrated a bunch of company's repositories from Bitbucket to GitLab. This also prompted a bunch of CI changes. Some projects I'm involved with, but don't have much of an authority, that are written in Python switched to complicated builds that involve pyproject.toml (often including dynamic generation of this cursed file) as well as integration with a bunch of novelty (but poor quality) Python infrastructure tools that are used for building Python distributalbe artifacts.

In the projects where I have an authority, I removed most of the third-party integration. None of them use pyproject.toml or setup.cfg or any similar configuration for the third-party build tool. The project code contains bespoke code to build the artifacts.

These two approaches are clearly at odds. A living and breathing person would either believe one to be the right approach or the other. The AI reviewer had no problems with this situation. It made some pedantic comments about the style and some fantasy-impossible-error-cases, but completely ignored the fact that moving forward these two approaches are bound to collide. While it appears to have an opinion about the style of quotation marks, it completely doesn't care about strategic decisions.

My guess as to why this is the case is that such situations are genuinely rarely addressed in code review. Most productive PRs, from which an AI could learn, are designed around small well-defined features in the pre-agreed upon context. The context is never discussed in PRs because it's impractical (it would usually require too much of a change, so the developers don't even bring up the issue).

And this is where real large glacier-style deposits of tech debt live. It's the issues developers are afraid of mentioning because of the understanding that they will never be given authority and resources to deal with.

[aspenmartin]

You are not wrong about anything you’re saying but like I said this misses the forest for the trees. I’m talking about like the next ~2 years. There is a common idea that we don’t understand this technology or what will happen performance wise. We know a lot more about what’s going to happen than people think. It’s because none of this is new. We’ve known about neural nets since the 40s, we know how RL works on a fundamental level and it has been an active and beautiful field of research for at least 30-40 years, we know what happens when you combine RL with verifiable rewards and throw a lot of compute at it.

One big misconception is that these models are trained to mimic humans and are limited by the quality of the human training data, and this is not true and also basically almost entirely the reason why you have so much bullishness and premature adoption of agentic coding tools.

Coding agents use human traces as a starting point. You technically don’t have to do this at all but that’s an academic point, you can’t do it practically (today). The early training stages with human traces (and also verified synthetic traces from your last model) get you to a point where RL is stable and efficient and push you the rest of the way. It’s synthetic data that really powers this and it’s rejection sampling; you generate a bunch of traces, figure out which ones pass the verification, and keep those as training examples.

So because

- we know how this works on a fundamental level and have for some time

- human training data is a bootstrap it’s not a limitation fundamentally

- you are absolutely right about your observations yet look at where you are today and look at say Claude sonnet 3.x. It’s an entire world away in like a year

- we have imperfect benchmarks all with various weaknesses yet all of them telling the same compelling story. Plus you have adoption numbers and walled garden data that is the proof in the pudding

The onus is on people who say “this is plateauing” or “this has some fundamental limitation that we will not get past fairly quickly”.

[crabbone]

> look at say Claude sonnet 3.x. It’s an entire world away in like a year

In the area I work I find them to be of very little value both then and now... I see no real difference. They help in marginal tasks. Eg. they catch typos, or they help new programmers to faster explore the existing codebase.

So far, I haven't used a single line of code generated by AI, even though I've seen thousands. Some of them worked to draw attention to a problem, but none solved it successfully. It was all pretty lame.

I see no reason to believe it's going to get better. Waving hands more forcefully isn't helping, there's no argument behind the promise of "it will get better". No reason to believe it will...

But, more importantly, the AI is applied on a level where really important things don't happen. It's automating boilerplate work. It doesn't make decisions about the important parts. Like, in the example above, the AI is not capable of choosing a better strategy: use pyproject.toml or write code to build Python packages? It's not the kind of decision it's called to make and nobody sensible would trust it to make such a decision because there isn't a clear right or wrong answer, only the future will prove one or the other to be the right call.

[locknitpicker]

> So far, I haven't used a single line of code generated by AI, even though I've seen thousands. Some of them worked to draw attention to a problem, but none solved it successfully. It was all pretty lame.

I find this statement highly suspect. AI coding agents nowadays can spot subtle object lifetime management issues and even dependency lifecycle incompatibilities, and here you are stating you are unable to use them to fix things? How strange.

Not to mention that coding agents excel at creating greenfield projects and migrating whole frameworks.

But if you feel you can't use them then I feel sorry for you.

[aspenmartin]

I think if you honestly don’t believe there is a major difference between 3.x and 4.7 I don’t think there is much anyone will be able to do to convince you. I do find it disappointing when technical professionals are so disinterested in building a real understanding of a fairly complex topic.

> I see no reason to believe it's going to get better. Waving hands more forcefully isn't helping, there's no argument behind the promise of "it will get better".

That’s a real bummer to read that from someone who sounds like a professional, and not only a professional but someone thoughtful and smart. 30 years of brilliant work in RL, Bayesian stats, machine learning, measurement, and then trillions of dollars of funding and some of the best talent in the world, and your assertion is “I tried it on my codebase and I didn’t like it and that trumps literally entire fields of mathematics and statistics”. I mean, have you heard of Chinchilla scaling laws? Do you know how RL works? are you aware of benchmarks, their strengths and weaknesses? Are you following adoption numbers, accomplishments like new proofs of unsolved erdos problems?

> But, more importantly, the AI is applied on a level where really important things don't happen. It's automating boilerplate work.

Your experiences are your experiences, I don’t know what work you do or how it gets done, what languages you’re working with etc. but literally we’re at the point where the vast majority of code at major tech companies is fully AI written (not assisted).

> It's not the kind of decision it's called to make and nobody sensible would trust it to make such a decision because there isn't a clear right or wrong answer

What are you claiming is not fundamentally possible for an AI to do that a human can do here? People make judgement calls on ambiguous problems, taking into account vast amounts of context about the business, dev time, reliability, maintenance, etc; why do you think AI can’t do that?

[imtringued]

What's up with the buzzword bragging?

You don't know buzzword A, B, C? Heh, he must be incompetent and know nothing.

The buzzwords mean nothing, really. The math is the same for a stupid or a smart model, because the model is trying to mimic properties of the training dataset.

You can give me the ultimate model architecture that will beat every model in existence and I can still figure out a way to make it perform worse than what's available today, but you're not even doing that, you're just drumming up some old news.

If someone "threatened" me with tech advancements I would be more worried about things like an imminent massive drop in token costs for bigger context windows or other game changers like continual learning where the model internalizes your code base into its weights rather than just keeping it in its context.

[aspenmartin]

It’s not buzzword bragging they are the prerequisites to having an coherent conversation. If someone doesn’t know what chinchilla scaling laws are the discussion about “I think things are saturated” is not grounded in anything. It’s like sitting around debating quantum mechanics and you don’t know the math, it’s just meaningless. If these sound like buzzwords the implication is not “you’re an idiot” it’s “you are not yet informed on the key basics of the discussion” and that is something you can fix with curiosity and a couple of prompts to ChatGPT to speed up the learning curve. It’s not like any of this stuff is gatekept.

> You can give me the ultimate model architecture that will beat every model in existence and I can still figure out a way to make it perform worse than what's available today, but you're not even doing that, you're just drumming up some old news.

Sorry I don’t understand what you’re saying here — what is the old news? You can break new models — yes. What’s the point you are trying to make here?

> If someone "threatened" me with tech advancements I would be more worried about things like an imminent massive drop in token costs for bigger context windows or other game changers like continual learning where the model internalizes your code base into its weights rather than just keeping it in its context.

I also don’t really know the point you’re trying to make here — like token cost drops seem like a good thing? Bigger context window too? Are we saying the same thing here?