On The Simplicity of Humannessİngilizce

Just yesterday, Bryan Cantrill published a short piece that touches something that I think most of us fundamentally understand, but don't have the words. The article is very short, titled "The peril of laziness lost". I recommend you all read it. I plucked out one sentence for you:

The problem is that LLMs inherently lack the virtue of laziness.

The essence is, our laziness is the reason we engineer simple systems. LLMs lack this laziness, which results in the fact that they don't have the incentive or the process to simplify, given that we don't externally create such mechanisms. That is why Anthropic had to introduce a separate command (/simplify) to enable such post hoc simplification process.

The thing that touches me the most about this idea is that I have been frequently amazed by the simplicity of human-engineered artifacts. All the algorithms I've ever read seemed simple, perhaps too simple in hindsight. All the codebases, even the gigantic ones, can be broken into simpler modules, sections, abstractions, hierarchies that allow us to understand the system. A great deal of modern advances in computing are due to simplification, malloc made memory allocation simpler, structured programming made control flow simpler, Rust made memory management simpler, typically via removal of our mental burdens.

It's not just computer science, all human-engineered systems are simple, for a very specific meaning of the word; as opposed to emerging structures such as social relationships or physical world that just is, or evolution that just happens. One may object that a plane, or Linux kernel, or the lithography machine of ASML is very complex, and they would be right. But all of those systems can be broken into smaller modules, subsystems, abstractions that allow for understanding them. I can say this confidently without having a concrete idea of what those systems are and how they work, because I believe in a fundamental principle of engineering, we must be able to maintain the systems we build.

When something goes wrong in a system, we must be able to locate where it is, understand the root cause of the problem, make changes in a way that are isolated without regressing on other parts of the system, and validate that the change correctly fixed our problem. Nature has the luxury to simply not care, things break, systems fail, earthquakes happen, it's all part of the plan.

There are escape hatches to this simplicity, complexities that we do not need to understand, when there's an asymmetric relationship between the process and the verification of the result. This has been long known and applied in the field of automated theorem proving, the de Bruijn Criterion separates the proof object, the proof checker, and the proof producer from each other. We must only understand and trust the proof checker, the proof objects themselves are irrelevant for the purposes of understanding the system. Although the generated proofs are irrelevant, I argue that the same engineering principles still apply to the proof producing programs, just not their results. A SAT solver possesses humanness, but the proofs it produces will not. A genetic algorithm possesses the humanness, but its artifact, such as AlphaZero, is akin to a biological organism that we must try to understand and model. We still care about understanding such results, because they allow us to improve on the engineering artifacts that produce them, as well as the outputs themselves, but we can bear not to.

LLMs do not possess the humanness right now, they probably will never do, just as AlphaZero doesn't. This doesn't make their outputs useless, but rather a new challenge we must overcome. In highly verifiable domains, such as translation or optimization, we will continue to see them in high, sometimes autonomous capacities. In other domains, we will find a way to push their limits. We will build more and more tools for analyzing and understanding programs that we previously did not need to, not because we couldn't, but because such understanding now gives us a great deal of leverage.

I didn't really know what I wanted to say when starting to write this one, I do that sometimes. Now that I'm concluding, I am searching for a conclusion, here it is; we are driven to understand the systems we build and maintain, and we find ways to do so. We were able to analyze the world, figure out the rules of the universe, model the human genome to its molecules that we didn't even know existed. I don't think that was primarily due to laziness, but rather the drive to build better, larger, greater systems; and I think we'll find a way to mold the LLM outputs into forms that we can understand too.

Thanks Bryan for lighting a spark that led me to write this, I hope you have enjoyed reading this as much as I did writing.