Cheating is All You Need

Steve Yegge writes (a lot!) about AI. His take is very interesting: Basically LLMs are in a race to the bottom. Each LLM is not that different from the next. The important bit is about context: Providing the LLM with enough context to do the task at hand. I’ve only had limited experience playing around with LLMs, but I definitely think that they only work when they are provided enough context to perform the task at hand. Providing can be tedious. If you want it to help you write a unit test, well it needs access to the class in question. And it’s collaborators. And the factories that the test uses. And so on. Copy/pasting all of that is a lot. The more interesting claims about AI I’ve seen are for tools that integrate with your editor (or IDE) to provide the context to the LLM.

Nearly All Binary Searches and Mergesorts are Broken

Interesting take about algorithm correctness: The author wrote a buggy implementation of a binary search in the Java language. And the bug, is the same one that his PhD teacher had in his book! The bug was for some ranges of inputs (not all), which is why it’s so pervasive and hard to spot, until you specifically start to look for it.

We programmers need all the help we can get, and we should never assume otherwise. Careful design is great. Testing is great. Formal methods are great. Code reviews are great. Static analysis is great. But none of these things alone are sufficient to eliminate bugs: They will always be with us. A bug can exist for half a century despite our best efforts to exterminate it. We must program carefully, defensively, and remain ever vigilant.

The Mythical IO-Bound Rails App

The author makes good arguments that if Rails in the wild was really IO bound all the time, than the recent YJIT improvements in Ruby versions would not have resulted in reported ~25% improvements in real apps. The improvements suggests that a good chunk of the time was being spent in Ruby. He then goes to point out that actually measuring IO is complicated, and that CPU starvation can look like IO.

Rails is better low code than low code

Radan Skorić’s points out that the argument for low-code (and even low-code) is that it can get you to your goal faster, and by people that are not experienced programmers. It resonates with me that it can quickly fall apart, once your applications reached a point that is not the platform use case.

I recently learned about a person I know that used one of those low-code platforms to quickly build an application that could be web-native, and also a native Android and iPhone app. Once the web portion was launched, it quickly got into performance issues, because each page was making hundreds (or even thousands) of AJAX requests to the server. However, he couldn’t really fix the problem, because the platform was not design to deal with the level of data that they wanted. They were stuck.

Trying out Solid Queue and Mission Control with PostgreSQL

Andrew Atkinson writes ab out his experience with Solid Queue and Postgres. I’m keeping an eye on Solid Queue, but have not yet made the switch from GoodJob.

I switched about a month ago to use dark mode on my Mac and iPhone. It’s been a great experience, except for using GMail on the web. The GMail iOS app has a beutiful dark icon, and a great dark mode UI.

GMail on the web… does not. Other Google websites (Drive, Calendar, etc) have a working dark mode. I don’t understand why GMail does not.

I solved it by:

1. Installing the Userscripts Safari extensions, my current browser of choice.
2. Adding a custom CSS style:

/* ==UserStyle==
@name        Gmail Dark Mode
@description Quick Hack for Gmail Dark mode
@match       https://mail.google.com/*
==/UserStyle== */
@media (prefers-color-scheme: dark) {
  html {
    filter: invert(90%) hue-rotate(180deg);
  }

  iframe, video, img {
    filter: invert(90%) hue-rotate(-180deg);
  }
}

Of course, it’s not as polished as the iOS GMail dark UI, but it does the trick!

Plugging the Postgres Upgrade Hole

Upgrading Postgres is hard. I’ve felt the pain, even upgrading local databases. The article outlines why that is the case. Maybe there is some hope for the future.

What is a Staff Engineer?

It uses a new-to-me framework for evaluating a job on 4 levels: Core technical skill, product management, project management, and people management. It then talks about what a Staff engineers does in each of these levels.

How we made a Ruby method 200x faster

A tale of performance optimization by using a profiler. How to focus on the important parts of a flame graph still seems a bit like a dark art. And not the key point of the article, but the refactor looks really nice!

I recently discovered a foot gun when writing rails migrations.

Rails runs migrations inside a transaction by default, for those databases that support it (e.g. Postgres). It also provides a what to disable it if you so choose, by using disable_ddl_transaction!. That can be useful for example for creating a large index concurrently, which is not supported inside a transaction. It looks like this:

class FootGun < ActiveRecord::Migration[7.2]
  disable_ddl_transaction!

  def change
    create_table :foot_guns
  end
end

So far, so good. However, because of how disable_ddl_transaction! is implemented, there is also a disable_ddl_transaction method defined. That is an accessor that checks weather the migration should be run in a transaction or not. But it can be used by mistake:

class FootGun < ActiveRecord::Migration[7.2]
  disable_ddl_transaction # This doesn't do anything!!!

  def change
    create_table :foot_guns
  end
end

The migrations looks like it is disabling the transaction, but it’s actually not. It’ also a hard mistake to catch, because the output rails prints out when running the migration in both cases is the same:

$ rails db:migrate
== 20241116193728 FootGun: migrating ==========================================
-- create_table(:foot_guns)
   -> 0.0137s
== 20241116193728 FootGun: migrated (0.0158s) =================================

I’d love for disable_ddl_transaction not to exist at all, so that a NameError would be raised, and this mistake was impossible to make.