░█▀▀░█▀█░░░░█▀▄░█▀█░█▀█░█▀▀░░░░█▀▀░█▀█░█▀█░█░░
░▀▀█░█░█░░░░█░█░█▀█░█░█░█░█░░░░█░░░█░█░█░█░█░░
░▀▀▀░▀▀▀░▀░░▀▀░░▀░▀░▀░▀░▀▀▀░▀░░▀▀▀░▀▀▀░▀▀▀░▀▀▀

LLM generated code is like particleboard :: 2023-12-30

LLM-generated code is like Particleboard

The hype around LLM-generated code has been extreme in 2023. Chat GPT made its waves, and improvements in LLM models and products like GitHub Copilot (there are now many) have stopped being speculative and have started being normalized as tools in software engineering.

Personally, I've had an aversion to LLM-generated code, and it's taken me a few months to think through. I've spent time challenging my assumptions and talking to the true believers, and getting hands on.

Carpentry

Before I was a software engineer, I was a lot of things. I'm a highschool dropout, and while I wouldn't recommend this path to anyone, it certainly is a way to almost guarantee blue collar work experience.

Among other things, I spent some time as a general contractor, and did some carpentry. Carpentry is when you build large structures. You tend to care about things like what's load bearing, or how long and thick a beam should be to support a second floor or a roof.

In my experience working at AWS, carpentry has a bit in common with creating larger software systems. Now, carpentry can be used to make large structures like mansions or something much smaller like a shed or a dog house. Likewise, large software systems can be used to make grand marvels of engineering like "the cloud" and its myriad of compute and data abstractions, and also relatively small APIs that can send you a cat image.

Woodworking

Fast forward a decade-ish, and I was a FAANG software engineer buying a home and starting a family. In a little workshop out in our backyard, I started up a hobby I had put on hold while living in big city apartments: woodworking.

Woodworking is not carpentry. This is when you make things like furniture, maybe sort-of large like a bookshelf or much smaller like a music box. My personal interest is in Japanese woodworking: using fine hand tools like chisels and planes, and creating very exact joints that don't require nails or screws. Woodworking more generally allows for many mediums in a finished product, like nails and screws, but also things you might not associate with wood, like upholstery padding and fabrics.

Woodworking is more like creating a user-facing application. Something technically smaller, like a web page, an application for a mobile device or a PC desktop application. Just maybe the software is for a consumer device. The differences are in the scale of the production. A web page, rendered in a browser on a client's unforeknowable OS and hardware, is not necessarily less of an undertaking as the backend that may be behind it, but it's a more "tangible" thing. It's less likely to live in a system with multiple points that bear a load and collaborate, and more likely to be directly handled by a user. Not better or worse, but there are differences in what gets valued at by the users and by the producers.

Wood, Manufactured Wood, Code, and LLM-code

Most carpentry is performed with natural wood, and where I come from most houses are made mostly of wood and some plywood. Most load-bearing applications are pure natural wood. You'll also see plywood used structurally for the floor under your flooring and the roof under your shingles.

Natural wood is usually cut in big long hunks of "dimensional lumber," like a "two-by-four", which of course, measures 1.5 inches by 3.5 inches. Plywood is produced by cutting large wood sheets around the core of a tree, and multiple sheets are glued together at alternating 90 degree angles. This produces a wood medium that's less affected by seasonal growing and shrinking as the moisture content of air changes. Plywood also can cover much more area than you could ever cut out of even large trees, and has a lot more strength than wood joined in many other ways.

Most woodworking is performed with natural wood, but where I come from most "wood" furniture is partially or completely composed of manufactured wood products. In short, most wood furniture does not even come from woodworking. I don't know all the ins and outs of the furniture business, but the going price right now for a low-end solid oak end table, the kind of small table you'd put by your bedside, is more than $100. Meanwhile, the going price for IKEA's famous LACK end table, comprised mostly of particleboard, is now $13. Both are made from wood, but one is going to cost more and last for more than one generation, and the other will ultimately be disposed of... but not at any great loss.

Particleboard is a material that's made from particles of wood that've been glued and pressed together. It's cheap; it's sawdust. Multiple pieces of particleboard can be assembled just like natural wood (or plywood) at a fraction of the cost. Particleboard is not bad; it offers a valuable trade-off that you might have already literally bought into. A much lower cost when you don't require durability. There are also some cousins here with higher cost and superior properties like MDF and OSB but I'll stick with particleboard for my analogy.

Now, I have owned a LACK end table. It's a functional little table, it holds stuff on its surface, and it was great on a budget. I liked it. I also still own furniture like bookshelves and cabinets and pantries that are made from fiberboards like particleboard. These pieces of furniture have the same rough tradeoff, the furniture is much cheaper than the quality alternatives, but slabs of sawdust aren't amazingly durable. (But... maybe it's good enough!)

LLM-generated code is like particleboard, while hand-written code is like natural solid wood.

And while I'm at it, let me tease out a few more fun similarities.

There's nothing that will prevent an expert woodworker from a tasteful use of some manufactured wood byproduct, for something like the back of a dresser, in an otherwise meticulous project. Likewise there's nothing that should prevent an expert software craftsman from using LLM-generated code in parts of a project that aren't visible and the quality doesn't matter as much as the function. (Especially exploratory tasks of a software project that don't need to be included in a final product.) In both cases, this can be used to drive the cost of a project down for some part that doesn't matter.

Particleboard makes sense in smaller, non-load-bearing applications, which is the domain where fine woodworking also happens. No one builds a house with particleboard. Well, at least you can't use it to bear a load. There are a few applications for particleboard in carpentry, but it's not going to hold up a second story. Likewise, you might find a few places where you could use LLM-generated code to speed up the production or lower the cost of a large software system, but something larger is going to have constraints on the qualities like speed, scale, and correctness it must provide. An online FPS game can't afford to be laggy. S3 can't afford to be out of space. A bank can't afford to just be "usually accurate" with withdrawals and deposits. These kind of requirements drive back to hand-written software.

Conclusion

LLM-generated code is for creating software quickly and cheaply. It's for mass-producing parts of software that can be assembled into something working good enough. If and when it breaks, you don't spring for an expensive repair job (like you might for a family heirloom) you just generate a replacement; the equivalent of going to IKEA and buying another $13 end table.

Despite my early gut-drivent resistence for LLM-generated code, after reflecting I can't deny the value it's bringing and will continue to bring. Just like particleboard, I'll have some of this stuff in my house, and I might not even notice it.

That said... my personal inclination is towards working with my own hands, and when it comes to software I prefer making higher-quality and durable products. In my experience, the software industry's demands are not for high-quality code, it's for mass-produced code. A software project should have quality controls like tests and lints these days, but these are not for quality reasons so much as for productivity reasons.

Anyway, I wish all you future prompt engineers the best. I've benefitted from mass-produced physical products in my life, and I suspect I'll also benefit from mass-produced software products too. As for my own software... I plan to (mostly) continue writing it by hand.

Additional resources: (last updated 2024-02-20)