Welcome to a mid-summer edition of the Weekly! I spent the July 4th weekend in the San Francisco Bay Area.

This trip was the first time I’d spent more than 24 hours in the Bay in almost ten years. Working in tech, I’ve obviously heard a lot about SF and the surrounding suburbs in that time — access to nature; much milder weather than the east coast (especially pronounced when weather topped 95 degress in NYC for most of the weekend). It was nice to see it all firsthand as an adult and understand a bit more intimately what people appreciate about being out here.

Thunderstorms in NYC kept me in San Francisco for an extra day, and I’m now writing this mid-flight with some shoddy WiFi to load all the links I’ll be writing about. I normally opt out of WiFi on flights, but I’m realizing that non-Starlink airborne internet providers are so slow that I’m not able to complete the frantic site-switching feedback loop that makes lots of my internet use feel unproductive.

It’s been a long time since the last installment but rather than appologize I’m here with a larger-than-normal set of interesting links from my RSS feed and various roamings around the internet!

good vibes

  • A line-by-line translation of the OCaml runtime from C to Rust (Michael Barcarella): Google’s original usecase for transformers was machine translation of human languages. Even given that, I continue to be impressed by their success in porting and translating large projects from one language to another. Yak shaving over language choice may soon be a thing of the past if translations like these get easy enough.

  • Understanding systems (Chris, Entropic Thoughts): LLMs may be well on their way to automating code generation, but they’re not particularly good at maintaining a model of a system as a whole. It’s more important than ever for human engineers to maintain and enforce an understnading of systems they’re responsible for. I also appreciated what Chris had to say about tutoring because a lot of it was about the importance of empathy in the process of transmitting knowledge and intuition to another person. More ways in which we can continue to provide meaningful value beyond the craft of writing code.

  • A sufficiently comprehensive spec is not (necessarily) code (Hillel Wayne): Hillel’s main contention here is that a comprehensive spec can still describe multiple programs and doesn’t restrict the implementation in the way that code does. I resonate with this on one level. Specs evolve over time, so architecture and abstractions in an implementation still matter.

    On the other hand, does code style really matter much if an LLM agent can rewrite an entire project in an entirely new programming language if the requirements and specification change?

    There’s also the fact that the vast majority of programming languages don’t describe how a program is actually run anymore. Optimizing compilers unroll loops, physical processors run instructions in pipelines and potentially speculate down multiple branches of execution. Programs already describe multiple potential physical sets of instructions that execute the program.

  • Earendil (Armin Ronacher and Colin Daymond Hanna): I just found this website delightfully quirky. It gives hope that instead of a blander internet, LLMs can make it easier to make your online presence a bit more idiosyncratic and weird.

slop (alternatively: bad vibes)

The remaining articles are all reflections on the proliferation of writing that lacks humanity. Together they stand in contrast the posts I collected earlier this year on agentic programming crossing the chasm. They all speak to similar themes, but all worth reading in thier own way. I’m presenting them here in publication order.

touching grass

  • Midnight Train to Stockholm (Mat Duggan): I love Mat’s narrative posts — they’re always worth a read.

  • The First of a Double-Take (Dave Gauer): Human intuition is a wonderful thing.

  • 50 Hours to Draw Some Lines (Doug MacDowell): What does it mean for knowledge work to be meditative? Doug describes his process for hand-drawn data visualizations. I enjoyed reading this in the context of us living through another instance of a technical field getting automated.

  • Intermezzo bag (Tom MacWright): A single-purpose item crafted by hand. I’ve enjoyed the idea of picking up a crafting hobby, which seems so orthogonal to programming along almost every axis besides “making something you’re proud of”.

systems and programming

  • Rethinking the database from storage up (Reynold Xin): This post describes Databricks’s new storage backend for their Lakebase product, which promises to marry a transactional Postgres interface backed by Parquet storage abstracted away with a service which translates the WAL to Parquet files and another which renders column-oriented Parquet files as Postgres data pages for consumption by the transactional layer.

    I’m sure performance relies a lot on caching and the specific internal semantics around Postgres’s storage APIs, but I’m still left wondering how you can truly get all the benefits of Parquet storage with a transactional interface. Parquet files need to be large (on the order of hundreds of thousands of rows) in order to see meaningful compression. I’m also curious how this storage interacts with row-level indexes and deletes in Postgres, since Parquet isn’t particularly amenable to row-level operations.

  • Things you didn’t know about indexes (Jon Charter): Indexes are critical for any real use of a DBMS. Your ORM might abstract away what you can actually do with them.

  • Serving files over HTTP three ways: synchronous, epoll, and io_uring (Phil Eaton): I’m really enjoying Phil’s writing over at The Consensus. This was a great introduction to Linux I/O APIs, and the examples built on top of each other well.

Liked this post? Get future weekly posts in your inbox!

Powered by Buttondown.