If coding is solved, what now?: Measuring the sloppiness of code

Coding is solved, perhaps, with unlimited token spend on a frontier model. It remains to be seen if it that is prohibitively expensive forever. At my company, we token maxed while the getting was good. But when we had to switch to Anthropic's enterprise plan, and start paying per token, the shit really hit the fan. Now we're retreating back to sane cost levels and finding that - guess what? - people power might just be more cost effective. AI of course is an immense tool to leverage, but still too expensive to create loops and let it run. This will change over time of course, but assuming it is a solved problem is nonsense. Maybe if we solve cold fusion, yes. Until then, evolution is winning the war on entropy.

Coding is not solved, correctness is not a feature, it is the bare minimum. If your code does not do what it is supposed to do, you could as well have no code at all. Efficiency, security, maintainability, reliability, readability, understandability, extensibility, maintainability, observability, portability, ... this is what high quality coding is about, not that it works, that is a given. And in my experience current models are pretty bad at this.

There is some sense of rose-tinted glasses of pre-LLM coding. A lot of human written code, particularly at the enterprise level, was of low quality well before AI automated it.

Really glad to see folks looking into quantitative approaches to give agents feedback on code quality. This post looks like a good start!

My main feedback for the authors would be, the most important problems for sloppiness are global properties, not local ones. In my experience an agent, like a human, has finite capacity for its attention, but if it runs into local sloppiness that gets in its way, it can fix it on a by-need basis. The technical debt issues that matter are usually global issues that aren't so easy to fix: they require global analysis and global refactoring.

I don't know the answer, but I think we're going to need ways to measure architectural properties, like separation of concerns, clear architectural layering, well-defined interfaces, etc.

Coding might be "solved" but coding still is the best way to build your own mental model of the solution space.

Which is more important to you: Velocity to a solution? Or velocity to understanding?

Coding is solved, but AI companies are still hiring software engineers?

It's a nice article that basically (rage)baits the readers before they realize the author actually disagrees with the premise of coding being solved. So here goes, my higher-level rant on "solving" something with AI.

I strongly dislike all the "X is solved" narratives. What does it mean for something to be solved? A math problem (or any kind of problem), a riddle, a mystery, a dispute. Those are all instances of a particular situation that requires a "solution", but new situations will always come up. I understand that by "cancer/coding/X is solved" in this new age of (gen)AI that means the ability to streamline or speed up the "solution finding" procedure, but even that presupposes a fixed, static, fully deterministic space of the things we are trying to "solve".

Even cancer cannot be fully solved - Demis Hassabis slowly started drifting away from using the word "diseases can be solved" because they can't be eliminated - we can only speed up the process for finding a cure for any particular disease, be it existing or new and/or evolving.

Is bridge-building solved? Architecture? Why are architects still employed? Is solving "civil engineering" or designing an optimally running machine a thing to be solved? What are we trying to do when we talk about "progress with AI"? Even when the "recursively-self-improving-AIs" and "perfect" robots do arrive, we're still bound to work with them, and they'll have to evolve to find new solutions to new problems.

To be clear, I work with and rely on LLMs every day, from coding custom RAG architectures with CC and Pi to research and agentic data science. These bombastic conversations, however need to quiet down a bit so we can get back to work :)

I started Valknut (https://github.com/sibyllinesoft/valknut) when I saw the writing on the wall regarding Agent code structure/abstractions/etc being a limiting factor in the ability to autonomously build projects. My experience was that good linters helped, but it wasn't enough, you needed to be able to enforce information-theoretic related organizing principles in addition to file/function LOC and local complexity metrics to guide agents on how to structure code.

Originally I tried to walk the line between improved agent performance and human readability, but current models are so good I don't think human readability matters much, though at a high level, being able to grok the overall folder structure still matters. I've got my hands full polishing a demo for my game, but I intend to revisit Valknut by crafting an eval set that lets me calculate the difference in agent token consumption and task failure rate between ~isomorphic codebase structures. This will let me loop agents to discover organizing policies that improve them.

Truthfully though, with today's models I don't think this sort of codebase optimization is likely to have much impact below 250k-300k LoC projects, and it probably won't be a decisive win till you're near 1M. Also, the shelf life of a product like this isn't infinite as each generation of models pushes those numbers up while also having new policy preferences that require re-evaluating existing policies.

> Qwen2.5-Coder-3B

Basing it's findings of LLM as judge on this model, and then proceeding to ignore it. This article can be safely ignored as well.

LLM as judge in harness evals is the way to go, for any of your custom needs. Design the eval well.

Does anybody actually know whether there's a limit to the complexity LLMs are capable of dealing with in a codebase? It's very obvious that they don't write code that is suitable for people to understand it (and it's gonna get worse and worse the more RL is used to train these models), but if there isn't a point at which LLMs also struggle due to the complexity they introduce, then I'm not sure it really matters anymore for a large part of non safety-critical software. I really hope there is, because steering them is, I feel, one of the last competencies through which I can still add value, but is there actually evidence that these models struggle more with poorly maintained code?

Coding is not just the program running in memory, its also the process of distributing the mental model of understanding among the team.

If humans increasingly are kept out of coding, then who holds the mental model?

If AI holds the mental model, by definition human prompts will be over lossy channel. This is true without AI too. Software quality is directly dependent on good devs that translate from business/PM speak to technical decisions.

So is coding solved now? it was already solved decades ago.

Having reached the same conclusions as the author led me to create my first agent to do architecture review, and that's how I learned about the metrics behind good practices that I'd been following for years. LCOM, cyclomatic complexity, that kind of stuff...

It's so easy to ship a lot of code, more effort should be put into ensuring the code is correct, with self-improving feedback loops that involve developers, and dedicated tooling...

But again, a while ago, everything was about prompt engineering, and now you can express you idea vaguely and get a somewhat working result, so this likely will evolve fast as well...

Number of iterations solved correctly, on a very large set of iterations, seems like a very good metric. Better than anything else because it measures what we actually care about, not some proxy.

The only caveat is that it's the same model doing an iteration and then using that iteration as a starting point for the next step. So the model is allowed to write absolutely insane solutions, as long as it can read them back, even if no one else can.

One thing that could be done is to use a separately developed baseline coding model B to evaluate the outcome of each iteration. For model under test X to pass an iteration, not only should it be able to solve starting from the previous solution, but so should B, starting from X's previous solution.

It will be solved when there is no more code left to write.

Code is an abstract concept that is not bound to the physical world and I imagine that future will have some much more of it that it is difficult to comprehend. Everything will be code and more code will be written than ever before.

Code will never going to be solved. The question is how much humans will be involved and I think the evidence is that perhaps just a bit. However, because we are talking about vast libraries of code even if we are involved in under 1% of all code and decision making that is needed around the code, there are still not enough developers out there to take on the task.

I might be wrong :)

>In my research and tests simply taking the change in the number of LOCs has been a surprisingly effective metric for sloppiness, with the ironic caveat that if we started optimizing for it, it would cease to be a meaningful measure.

This matches my experience. Before working on an issue, I ask the LLM to estimate net LOCs at the final PR based on the scope. It works well, and review steps do flag inconsistencies. But as the OP mentioned, if you turn this into a hard metric vs "design smell", you can see LLMs code-golfing for oneliners.

Coding just a stage in the software development. Design and specifications which can help in coding is not solved at all and may never - the end result is software reliability is not a solved problem.

Ya, now that I have some solid AI coding experience under my belt, there does seem to be some gaps between practice and reality. I have a fairly complex codebase which I pretty much hand code everything. When I add a new feature, I spend a lot of time designing and refactoring that feature into the codebase. Either the feature dovetails into the existing design or the feature creates new designs which will then facilitate even better future features. When AI approaches the feature, it just plows the feature in, and with bugs since it has trouble fully understanding the total design. So over time, you have a spaghetti design where you just have a whole bunch of features tied together with no unified design. I guess thats ok if AI is supporting it, but you now have a large cost and bug surface area and an insane human learning curve. So nothing has really changed here, we have been dealing with low quality codebases way before AI came along. I think AI has mastered the one shot single feature, tool, or simple app, but it struggles with the design complexity of a rich multi feature application or system.

"To come back to the point of why agents can’t (really) deal with the slop themselves, we need to look at the evaluation of SlopCodeBench. In contrast to other coding benchmarks, which give the agent a complete list of instructions at the start and then have a set of hidden tests the program needs to pass, they do the opposite. They create multiple rounds of instruction and test iterations, where in between checkpoints the context of the models is erased. Thereby mimicking much more closely an iterative process, like how coding agents are actually used by humans. The result of that is that bad coding decisions accumulate over time and for the strict solve rate, where all tests have to be passed at all checkpoints, even state of the art models achieve 0% pass rate"

I like how this captures with a metric (lines of code and cyclomatic complexity, some other basically tractable measures) in an automated way, something we all by now intuitively know.

Sad as it might sound, I think we might have to stop worrying about the code.

It is not clear to me how the verbosity metrics works. Can someone shades more light on it?

I love this train of thought. Code quality is critical, but I don’t think we’re correctly evaling it at the moment. If we could get solid benchmarks measuring the quality of generated code, we might see the models climb those benches fast.

I believe that the era of “ai writes tons of slop code” will be a stepping stone in the longer story, and is simply a current gap in the reward functions.

Per the author - if we can get strong measurements of what good code is, we can train against it and close the gap fast. Excited to see more thinking in this area

If coding is ever solved, and if software does it, sloppiness probably won’t matter much.

Code will become throwaway stuff, like the results of AI prompts. Cook it up, test it for adequacy, and run it. When something comes along that adds new requirements, just update the requirements/prompt and make a new one.

Test suites will be important.

Well... Coding was never the most part of work I spent time.

> I was disappointed at how “vibes based” the industry seems at the moment.

Alan Kay called programming "Pop Culture" some 20 years ago[0]

[0] https://queue.acm.org/doi/10.1145/1039511.1039523

I wonder how much of this is due to the AI tooling being taught on sloppy code that humans have written. Over the last 4 decades I've looked at a lot of code on the Internet and there's a lot of slop out there.

Solving consciousness ;)

Claim: AI writes almost perfect code.

Reality: earandil.com uses 170% CPU in Firefox.

What has this author written before LLMs? Why should we listen to him and his adjudication of "perfect code"?

Cyclomatic complexity is the oldest paper generating grift for college students. There are hundreds of thousands of useless papers about cyclomatic complexity.

It will be fun to try and deterministically define sloppiness relative to a tool that is not deterministic.

Of course, sloppiness to date can be measured by different shared and interpreted preferences and definitions.

Coding has been solved for 20 years at least.

90% of problems are easy once you know what you actually want well enough for you to be able to ask it from an LLM.

90% of code before LLMs was badly copied from StackOverflow anyway.

That 10% that's remaining, I've see 0, ZERO, nil progress. Windows is still awful. Spotify still doesn't work correctly offline. Youtube search is trash. Jira takes 20 seconds sometimes to load a task. LLMs haven't created a new database or a new game engine or a new renderer or anything like that.

The maths breakthroughs are really more of a testament to the efforts of the last 150 years for maths to be an organised verifiable principle. If LLMs had to practice math they way Euler did, they wouldn't be able to find shit.

(sorry if I sound incoherent, just some thoughts while I'm commuting)

Plopping in my email to the author below in case anyone else is interested in this kind of thing:

> There are some promising other directions I want to explore, such as coupledness of functions, code churn, cohesion and so on. If you are working on evals and would like to talk, I would be happy to do that: sebastian@earendil.com

Hey Sebastian, I just read your article and it thoroughly resonated with me. I've been working on building something similar to SlopCodeBench, but moreso aimed in the direction of architecture, rather than simple one off "code search functions".

In a nutshell, I'm creating multiple domains of common software architectypes. You can think of these as being as simple as a counter (very common in all architectural explorations worth their weight), todo applications, etc and as complex as an online store, a bank, a wallet, a social communication platform, etc.

Given a single domain, we can extrapolate common functionality that is "higher order" to that domain. Features like data synchronization, functioning offline, sharing information, authentication and authorization, etc all land in this bucket. From a single domain (take the counter, for example), I've laid out my initial plans for the various different levels to concretely observe how bad LLMs are at churn, cyclomatic complexity, poor abstraction planning, etc as follows:

L1: Show a number on the screen L2: Allow a user to click plus or minus and the number responds accordingly L3: Show that number on any device running your software, and keep it in sync with all other devices L4: Ensure that additions and decrements to the count, while a device is offline, will replay to all connected devices once connectivity is reestablished. L5: Introduce the ability to reset the count to zero, and ensure that commutes properly if an offline device triggers a reset or vice versa. L6: Introduce user accounts where I must provide an authorization strategy (login with apple, google, passkey, etc) and now segregate a public counter (which anyone can interact with) from your counter (which only devices authorized with your credentials can interact with. L7: Introduce an action menu whereby I can invoke what is commonly known as a "Command K" menu for the actions that can be performed in the application. L8: Allow me to speak naturally to your application and ask it to "go up", "start over", etc (this exercises your architecture's ability to be "accessible" to agents (and vision disabled folks as well) ... ... ... And so on and so on.

Despite its apparent initial complexity, naive solutions to even the simplest domains will be easy to spot with how many lines were changed vs raw additions (how composable a solution is), that cyclomatic complexity measure you mentioned, how many tokens it took, how many platforms (iOS, android, cli, TUI, react native, react) your application will run on, how long it takes to build, how large the binaries are, how much memory is used during the operation of your software, how semantically similar duplicated code across different platforms etc. From all these different values, we can create a hueristical "architecture score" to benchmark against.

I'm also toying with the idea of enforcing that one must submit a bundle of skills, instructions, scripts, etc that I will exercise with my own harness whereby the submitter has to submit a monetary cost with their submission that a budgeting agent must manage the spend and the agents must yield prior to their submission being scored, lest they forfeit the submission.

I'm just quite tired off all the hype and its exhausting and AFAIK, none of the benchmarks actually produce anything of use. One interesting side effect of aligning incentives in the way I've laid out here is that we will have produced open source, connected software that functions well and solves a whole bunch of business needs that all compose together by definition.

Thanks again for your article, would love to have an e-coffee and chat about if there's potential to collaborate on anything here. Despite how powerful llms are, this is still proving to be a tricky endeavor for me.

Best, Michael

PS: here is a demo of my submission for this first round: https://x.com/technoplato/status/2090902061437030777?s=20