The document is human-crafted and human-reviewed, and it primarily targets humans. The fact that it works for machines is a (pretty neat) secondary effect, but not really the point. And the document sped up the act of doing the refactors by around 5x.

The whole process was really fun! It's not really vibe coding at that point, really (I continue to be relatively unimpressed at vibe coding beyond a few hundred lines of code). It's closer to old-school waterfall-style development, though with much quicker iteration cycles.

Now I no longer directly use git, but instead use jujutsu (jj).

Once I became very proficient in the jj cli, I picked up jjui: https://github.com/idursun/jjui

Also, as splitting commits is an extremely frequent operation, this neovim plugin is really nice: https://github.com/julienvincent/hunk.nvim

Also this neovim plugin is amazing for resolving jj conflicts: https://github.com/rafikdraoui/jj-diffconflicts

Now with jj instead of git I edit the commit graph as effortlessly as if I am moving lines of code around a file in my editor.

And as a bonus : possibility to keep music videos playing while changing app. (A parameter has to be switched for that though)

As a web dev, I'm really interested to see if I can write automated 'system UX tests' in natural language. E.g. The test would be something like "go to the home page, log in, update your status" it passes if the LLM completes it (the premise being if the LLM fails, your users probably will too).

let isLarge = a => a>100;

numbers.filter(isLarge)

Blocks let you do the same but without extracting the body as cleanly. Maybe it’s a chronological issue, where Ruby was born at a time when the above wasn’t commonplace?

>When you write 5.times { puts “Hello” }, you don’t think “I’m calling the times method and passing it a block.” You think “I’m doing something 5 times.”

I’m of two minds about this.

On the one hand, I do agree that aesthetically Ruby looks very clean and pleasing. On the other, I always feel like the mental model I have about a language is usually “dirtied” to improve syntax.

The value 5 having a method, and that method being an iterator for its value, is kinda weird in any design sense and doesn’t seem to fix any architectural order you might expect, it’s just there because the “hack” results in pretty text when used.

These magical tricks are everywhere in the language with missing_method and the like, and I guess there’s a divide between programmers’ minds when some go “oh that’s nice” and don’t care how the magic is done, and others are naturally irked by the “clever twists”.

The philosophical underpinning is giving up of materialness. The practicality of the 5 instances that I witnessed over the past year - typical terminal individuals choose this. They pass away surrounded by loved ones (they typically medicate for any pain, and the body starts shutting down when food and water stops). This is observed with somberness, but celebrated as very positive act.

When someone starts this process, it's a unique experience speaking with them, as there's usually nothing that comes up, and the moment does not really lend itself to small talk :)

[1] https://en.wikipedia.org/wiki/Sallekhana

And while I'm here, authors of unix tools, please use $XDG_CONFIG_HOME. I'm tired of things shitting dot-droppings into my home directory.

I have a fairly large web application with ~200k LoC.

Gave the same prompt to Sonnet 4.5 (Claude Code) and GPT-5-Codex (Codex CLI).

"implement a fuzzy search for conversations and reports either when selecting "Go to Conversation" or "Go to Report" and typing the title or when the user types in the title in the main input field, and none of the standard elements match, a search starts with a 2s delay"

Sonnet 4.5 went really fast at ~3min. But what it built was broken and superficial. The code did not even manage to reuse already existing auth and started re-building auth server-side instead of looking how other API endpoints do it. Even re-prompting and telling it how it went wrong did not help much. No tests were written (despite the project rules requiring it).

GPT-5-Codex needed MUCH longer ~20min. Changes made were much more profound, but it implemented proper error handling, lots of edge cases and wrote tests without me prompting it to do so (project rules already require it). API calls ran smoothly. The entire feature worked perfectly.

My conclusion is clear: GPT-5-Codex is the clear winner, not even close.

I will take the 20mins every single time, knowing the work that has been done feels like work done by a senior dev.

The 3mins surprised me a lot and I was hoping to see great results in such a short period of time. But of course, a quick & dirty, buggy implementation with no tests is not what I wanted.

You clone the GitHub link, then `cd` there and run docker:

   docker build -t campfire .
   docker run \
     --publish 3000:3000 \
     --volume campfire:/rails/storage \
     --env SECRET_KEY_BASE=$(docker run --rm campfire bin/rails secret) \
     --env DISABLE_SSL=1 \
     campfire

   sudo usermod -aG docker $USER

   sudo docker run --rm campfire bin/rails secret

I did all that on Linux, so I have no idea how to do that on macOS.

As a final step, you visit either localhost:3000 or ip-addr:3000 and play with the chat.

The difference light quality makes to the experience is remarkable, and I implore anyone that visits to ensure you go at the right times on a clear sunny day.

You may think you've seen well lit stained glass before, but it's like freaking lazer beams of thick light penetrating the environment to the point where it doesn't seem natural.

I was never into architecture and am not a religious person but visiting the Sagrada Familia was profound. I came out a slightly different man than one who walked into it just half an hour before. No CGI hack comes close.

- First things first, you have to get your hands on actual VPN software and configs. Many providers who are aware of VPN censorship and cater to these locales distribute their VPNs through hard-to-block channels and in obfuscated packages. S3 is a popular option but by no means the only one, and some VPN providers partner with local orgs who can figure out the safest and most efficient ways to distribute a VPN package in countries at risk of censorship or undergoing censorship.

- Once you've got the software, you should try to use it with an obfuscation layer.

Obfs4proxy is a popular tool here, and relies on a pre-shared key to make traffic look like nothing special. IIRC it also hides the VPN handshake. This isn't a perfectly secure model, but it's good enough to defeat most DPI setups.

Another option is Shapeshifter, from Operator (https://github.com/OperatorFoundation). Or, in general, anything that uses pluggable transports. While it's a niche technology, it's quite useful in your case.

In both cases, the VPN provider must provide support for these protocols.

- The toughest step long term is not getting caught using a VPN. By its nature, long-term statistical analysis will often reveal a VPN connection regardless of obfuscation and masking (and this approach can be cheaper to support than DPI by a state actor). I don't know the situation on the ground in Indonesia, so I won't speculate about what the best way to avoid this would be, long-term.

I will endorse Mullvad as a trustworthy and technically competent VPN provider in this niche (n.b., I do not work for them, nor have I worked for them; they were a competitor to my employer and we always respected their approach to the space).

(and sorry for offtopicness)

    [Rest] markets itself as a way to "unlock a new revenue stream"
    with the help of a "robust algorithm" for detecting smoking.

The awesome thing about black-box algorithms is they can't be challenged when they're wrong. And errors reliably favor the institution that manages (and profits from) them.

After the war Germany had lost quite a lot of businesses, infrastructure, industry and farming. Also obviously lots of manpower. Produce was scarce and inflation was extremely high, so it was actually quite difficult to purchase anything even though people had money. the Allied Forces introduced price control on almost all essential good in order to stop inflation. That obviously did not work at all and most goods were actually traded in the black market, so you could actually buy bread by paying with cigarettes.

A German economist, Ludwig Erhardt advised to remove all price and legal controls and to replace the old mark by a new one, but the Allied Forces only agreed to the latest, so a new currency was introduced, the Deutsche Mark, replacing the old Reichsmark. That had no effect whatsoever. However, Mr. Erhardt, from his position as Director of Economic Administration, decided unilaterally to remove of the price fixing and other regulations. And literally overnight, German streets filled with sudden and unplanned pop up markets, everyone started to sell anything they didn't need, just by the street or from their front yard.

In 1949 Erhardt became Minister of Economy for 14 years, and later, in 1963, Chancellor.

I now use uv for everything Python. The reason for the switch was a shared server where I did not have root and there were all sorts of broken packages/drivers and I needed pytorch. Nothing was working and pip was taking ages. Each user had 10GB of storage allocated and pip's cache was taking up a ton of space & not letting me change the location properly. Switched to uv and everything just worked

If you're still holding out, really just spend 5 minutes trying it out, you won't regret it.

> She does not appear to have social media profiles, though she has two followers on the blogging site Medium.

Talk about damning with faint praise!

- telling the model to be persistent (+20%)

- dont self-inject/parse toolcalls (+2%)

- prompted planning (+4%)

- JSON BAD - use XML or arxiv 2406.13121 (GDM format)

- put instructions + user query at TOP -and- BOTTOM - bottom-only is VERY BAD

- no evidence that ALL CAPS or Bribes or Tips or threats to grandma work

source: https://cookbook.openai.com/examples/gpt4-1_prompting_guide#...

I mean the whole AI personal assistant shebang from all possible angles.

Imagine, for example if booking.com built an MCP server allowing you to book a hotel room, query all offers in an area in a given time, quickly, effortlessly, with a rate limit of 100 requests/caller/second, full featured, no hiding or limiting data.

That would essentially be asking them to just offer you their internal databases, remove their ability to show you ads, remove the possibility to sell advertisers better search rankings, etc.

It would be essentially asking them to keel over and die, and voluntarily surrender all their moat.

But imagine for a second they did do that. You get the API, all the info is there.

Why do you need AI then?

Let's say you want to plan a trip to Thailand with your family. You could use the fancy AI to do it for you, or you could build a stupid frontend with minimal natural language understanding.

It would be essentially a smart search box, where you could type in 'book trip to Thailand for 4 people, 1 week, from July 5th', and then it would parse your query, call out to MCP, and display the listings directly to you, where you could book with a click.

The AI value add here is minimal, even non-existent.

This applies to every service under the sun, you're essentially creating a second Internet just for AIs, without all the BS advertising, fluff, clout chasing and time wasting. I, as a human am dying to get access to that internet.

Edit: I'm quite sure this AI MCP future is going to be enshittified in some way.

https://www.nytimes.com/2025/03/03/science/cuttlefish-camouf...

In practice, probably yes, officially probably maybe. Giving a security guarantee would allow Taiwan to do provocative things, so hence, why there isn't a formal one.

If it actually came to be today, I guess the US would at-least offer token support. To (a) embargo China, (b) ensure chip facilities Taiwan aren't surrendered intact.

Both of which doesn't require winning a conflict, just making it painful.

If you use a password generator (and thus you never reuse or forget passwords), the easiest thing to do is to just set it to the password (though they may be stored in plaintext), or a different password you keep on the notes in your password manager.