3 Things Worth Reading #003

Self-Taught Engineer weekly reading edition - sense making

1. Most AI code is single-use plastic

Something Big Is Happening #4 *paid

Sudo Make Me a CTO, Sergio Visinoni

“most AI-generated code that is pushed to GitHub and the likes is the digital equivalent of single-use plastic”

You reinvent the wheel because you never paid the cognitive cost of solving it the first time, and paying that cost is also what teaches you to make the thing reusable. Skip the struggle and the code comes out as disposable ( wrote about the ikeaifaction of engineering previously)

There’s a second finding in there I hadn’t seen. Researchers at the Oxford Internet Institute and the Hasso Plattner Institute found LLMs alter the meaning of people’s drafts, sometimes reversing what they meant entirely. If a model can flip your intent in a sentence, where you would notice, what is it doing inside a function where you won’t?

Why this matters: “I used Claude to write it” isn’t a defence, to slinging sh*t. Not because someone is gatekeeping. Because the understanding that would let you defend it is exactly the thing you skipped.

Two places I think he overreaches. He reads Google moving people from Alphafold around the company moving resources is that Gemini is essentially draining all the good from the company, which doesn’t follow, as Isomorphic Labs is sitting right there as a counterexample which was spun up and led by the founder of Deepmind and backed by Google, also data companies were cutting the spines off books to scan them long before any AI labs existed.

---

2. Eleven days, 500,000 lines of code?

How building software is changing at Anthropic (free)

The Pragmatic Engineer, Gergely Orosz

Most folks will read thatBun’s team migrated a 500,000 line codebase from Python to Rust in eleven days, for about $165,000 in tokens. By hand they estimated three engineers with full context, about a year, and said the realistic alternative was never doing it at all.

However, I think the important part is just how much of the test suite already existed. Bun’s test suite is written in TypeScript, so it doesn’t depend on the runtime’s implementation language ala Zig. The tests sat outside the code that was being transformed.

So the eleven days weren’t bought by the model. They were bought by a test suite wrote years earlier.

Anthropic’s own telemetry says the same thing from the other end:

“We see that few tokens are spent on actual implementation. Most are spent on discovery of unknowns, prototyping, mocking, and then in verification and testing.”

Why this matters: AI moved the bottleneck from writing code and onto knowing whether the code is right. I remember trying to be a good coder and doing test driven development (TDD); now that was a steep learning curve.

What makes an engineer stand out today? A solid grasp, including insight into the layer beneath the one you work on daily?

---

3. How to make Sense of AI & other rapid developments

How to Make Sense of AI (free)

Commoncog, Cedric Chin

Chin argues that you don’t need to predict the future when you can clearly perceive the present moment. So ignore opinions, analysis and predictions, however good the source. Read detailed field reports only, in his words.

“A field report is acceptable only if it is adequately detailed [youtube, article of value, podcast etc], but you should also take into account who the author is, what their context of use might be, and what they are trying to accomplish.” for example OpenAI/Anthropic saying anything, might be useful, but understand the source.

Then ask four questions of each one.

I rewrote them for myself:

1. What new outcomes are suggested from this?

2. What actions should I take?

3. What are the possible outcomes if I applied this to my goals now?

4. How is it all connected?

Question 3 does most of the work, because it’s specific to you, and really nobody else can answer it.

Chin’s filter suggests I belong in bin #1 since Visinoni is arguing, not reporting. But I didn’t bin it. I kept his evidence and argued with two of his conclusions. I think that’s the better rule. Run the filter on the evidence, not on the author.

Why this matters: both #1 and #2 are field reports from people with their hands on the thing. That’s why they’re worth reading even when they disagree.

Visinoni’s disposable code and Bun’s eleven days share the same story but from different perspectives. Bun’s team invested for years in testing and stored it in a way that outlasted a language shift that they needed to make.

Last week, I mentioned that AI is making engineering output a commodity, but not the judgment or basics. This week, I feel even stronger. The tests you wrote years ago are what give you speed today when rewriting a code base even if it makes the world burn it seems.

What did you read this week that made the cut? Hit reply; I read everything.