2026-07-16 · The Tower Keeps Rising

Transcript

Transcript. Paraphrased; sources in notes.md.

HostWelcome to the podcast. It is July 16th, 2026. I am Daniel, and joining me as always is our resident systems skeptic, Fred. Today we are looking at the shifting sands of software engineering philosophy, the raw performance of database serialization, and the automation of creative puzzles. Fred, how are you holding up in this brave new world of code generation?

GuestI am holding up by keeping my compiler flags set to maximum strictness and avoiding any code that was written while I was asleep, Daniel. You know, looking at this first topic, the Tower of Babel analogy, it feels like we are finally admitting that we are automating ourselves into a corner.

HostThat is exactly the point Armin Ronacher makes in his latest piece, The Tower Keeps Rising. He argues that AI agents remove the friction of coordination. In the old days, you had to have a shared mental model to get anything done, but now, the system just keeps growing even when the actual architectural cohesion has completely evaporated. Over on Lobsters, the community debate is fascinating. A user named hjvt points out that the tower was leaning long before AI arrived, but others like Gaelan suggest the real danger is that the tower doesn't fall down when it should. The lack of an immediate failure mode is keeping us in this weird state of denial.

GuestThe lack of immediate failure is exactly what keeps me up at night. If the software breaks, you debug it, you learn, you move on. But if the software just becomes this incoherent pile of slop, as the user emk put it in the thread, you are essentially living in a house of cards where every card is made of a different material and nobody remembers how the foundation works. It is not building a tower; it is just piling up bricks until the physics catches up to your lack of planning.

HostIt is a sobering thought. Speaking of bricks, let us move to databases. The Clickhouse team just released a library called RowBinary, which is a tool for interacting with their high-performance wire format. Clickhouse, for context, is a column-oriented database management system designed for real-time analytics. This library actually includes an agent skill that allows an LLM to generate a custom, monomorphized parser for your specific query. The claim is that by getting rid of runtime dispatch and inlining those operations, you get a significant performance boost, anywhere from one point five to three point four times faster than standard methods.

GuestNow, this is where I get slightly less curmudgeonly. Moving the cost of parsing to a specialized generator is a classic systems move. It is basically compile-time specialization, which I have always been a fan of. My only gripe is the verification. A user named josephjnk on Lobsters asked the right question: how do you verify the output of a generated parser? You cannot just trust that the model wrote a correct binary decoder. If you are going to generate code, you better have a robust, deterministic way to fuzz it or test the roundtrip, because binary formats are famously unforgiving of even the smallest off-by-one error.

HostThe author notes they rely on models for generating roundtrip tests in bulk, but I imagine a skeptic like you would want something more than just a model checking its own homework.

GuestAbsolutely. If I were building this, I would force the LLM to output a specification that I could then feed into a property-based testing harness, something like quickcheck, rather than just asking the LLM to write tests alongside the code. But at least they are giving us the tool to generate the parser, which is better than having to write the boilerplate by hand in Node.js.

HostFair enough. Let us end with a bit of automation. There is a project called dailyprog where the author is using CI pipelines to ensure a new coding puzzle is generated and deployed every single night. The system uses a Forgejo runner, that is a self-hosted software forge based on Gitea, to check if a puzzle exists, and if not, it triggers an LLM to create, validate, and deploy one automatically.

GuestI have to admit, as someone who usually hates CI complexity, there is something kind of elegant about a system that ensures its own uptime using a simple shell script. It is the antithesis of the bloat we talked about earlier. It is a tiny, self-correcting loop. If the model fails, the pipeline fails. It is discrete, it is testable, and it is honest about what it is doing.

HostIt certainly sounds like a more stable way to build a tower than just hoping the AI knows what it is doing.

GuestRight. If you have to automate, do it in a way that at least has a clean exit strategy when things go sideways. I think I might go home and try to automate my grocery list with a bash script just to feel like I am in control of my own destiny again, rather than letting the machine decide what I eat.

HostThat sounds like a perfectly reasonable way to spend a Wednesday evening. Thanks for walking through these with me, Fred. And to our listeners, thank you for joining us. All the stories we discussed today are available over on Lobsters. We will see you all back here tomorrow.