raffraffraff

I converted my cheap Billy shelves into what look like fitted shelves. They're all bolted together, screwed to the wall, filled in around the edges with wooden panels, and finished off with 2.5m decorative strips of pine at all the edges and joins. Then I painted the whole lot the same colour and filled them with books. Looks amazing, cheap as hell. Nobody guesses they're Billy shelves.

I thought this was going to be an Intel cpu security issue

Looking at the relative sizes of the dotcom crash, '08 financial crisis and covid is interesting...

I hated the idea of self hosting keycloak, learning how to terraform every single inch of it. But I worked at a startup that had a fucking weird investor who paid me, and would pay AWS, but wouldn't pay Okta (or any other 3rd party).

So given the choice between Cognito (easy right? It's a managed service!) and Keycloak (a large, complex security Swiss Army knife) should have been a no-brainer in favour of Cognito. But we absolutely hated it. I think it's one of the worst AWS services I've used.

Keycloak is only hard because security is hard. The terraform provider is amazing, worked remarkably well. We never had a single hitch, and I could spin up new tenants in seconds, or deploy Keycloak to other regions in a few minutes.

Personally, I just hate the idea of running that type of thing in production unless it's my full time job (and it wasn't) and I have other team members who also knowit backwards. But for a sufficiently well staffed org that feels like moving this expertise inhouse and dumping a paid auth service, Keycloak really is amazingly good.

Now that I've moved on to another job, my time spent in the belly of Keycloak wasn't wasted either. I can help our security team figure out auth issues blindfolded (ie: without access to their Okta admin interface, and in fact having never seen one)

I recently implemented CloudFront signed urls for accessing S3, where my app uses KMS keys to sign the url. Sounds simple and obvious but for the entire history of Cloudfront and KMS it was not! It is not documented to work, nor could I find any explicit AWS documentation saying that it wouldn't work either. And there's a just bunch of anecdotal reports (blogs, github issues) that it doesn't work. But very recently (April I think) Cloudfront and KMS, for the first time ever, both started to support the same key algo/size. So it's now possible, but I stumbled on it, and have never seen a recipe for using them together. So AI doesn't know about it either!

I was finally able to stop managing and rotating TLS key pairs, replicating private keys globally to my apps.

So: who really reads those little boring announcements, and connects the dots? Seemingly, not even AWS themselves!

Is "top down" a company size thing? Very simple organisms don't need brains or nervous systems. But once they get big and complex, suddenly there's a need for sensors, messages, plumbing networks for transporting signals and matter. That doesn't evolve as a decentralized system. You quickly need a control centre to oversee operations.

Companies are no different. During company growth, it's all about hiring, and giving tasks and some responsibilities to subordinates, setting up the right channels for communication etc. But at what point do company leadership actually say "You know what,that department would work better without direct instructions from us. Let's facilitate communication, act as conflict adjudicators and make company-wide 'North Star' decisions"

I think that's what OKRs are supposed to do, but I've never seen that work well (for very long anyway). It always gets up-ended somehow. Start of quarter optimism, over-promising, not spotting that someone else's stretch goal is actually your blocker. Then mid-quarter you're dealing with production emergencies, database code red and running out of time. At the end of the quarter, two of your team go off on PTO while you desperately try not to "fail your quarter", working late, and delivering shoddy work.

Does true bottom up exist at large companies? Hmmm. Some. We've all heard the legends of IBM's "build the first PC!" skunkworks. And we know about research-only departments at Xerox that led to so many modern computing inventions. But these were brief projects.

I find these days that people have the desire to get out of the way and let us work, but the tiling is in the way. I waste so much fixing time in Jira, slack, doing performance reviews, interviewing...

How about adding a return packaging cost and have the empty package sent back to the country of origin so that you can drop it into your own recycling bin?

Wow. I tried to get Qwen3.8 4B to parse song lyrics and analyse them. Getting ollama running was a minute or two.

However coming up with a prompt that didn't turn out total garbage was impossible. After wasting over an hour and I ended up getting Qwen side by side with Llama 3.2 3B, just to see if I was being stupid. Nope, it just looks like Llama is orders of magnitude better at this specific task for some reason).

If you think I'm doing it wrong, you're probably right, I don't know a ton about local LLMs. But I hand selected 50 songs, set up ollama with both LLMs, and for each iteration on the prompt text, ran both LLMs 10x times per song. Side-by-side comparisons showed that Qwen 3 4B was so bad that I actually downloaded Qwen again, thinking there must have been some mistake and I accidentally grabbed an old 1B model.

I'm not saying this is the case in the particular example, but these days a lot of people aren't bothering to read a 1 page README, and are instead letting claude "just fucking do it already". I see it at work.

Claude has got better at "just fucking doing it" by asking if it's ok to go read the latest github issues and pull the README, which means that people will likely get lazier and lazier.

Private Email (a Namecheap product) makes me wish I didn't move that account off Gmail. I'll take a look at Fastmail.

I'm doing a small local version of this test to pull moods and themes out of song lyrics. I've found that one prompt, 1 model, 11 runs even gives different results.

There's no consistency over multiple runs of the same prompt on the same model.

Also, for the purpose of music lyric analysis Qwen 4B is laughably bad. Like it's going out of its way to be extremely wrong, misunderstand the prompt. When it does correctly understand what I asked for it ALWAYS tells me that the mood is Angry. Sometimesiit just gives back all the lyrics. Sometimes it claims that it doesn't have the list of moods or the lyrics and tells me I should look them up on the Internet first.

All with the same prompt every time.

Are models being overtuned for coding tasks?

What about container specific permissions? If they implemented that it would be enough for me since I tend to split up my services by type (eg: shopping, banking, work, hack) so enabling an extensive on a specific set of containers would rock. Right now I think the only control we have is over private tabs/windows?

If your country suddenly has hundreds of thousands (or even millions) of weather migrants from countries with vastly different cultures and values to yours, you'll be wishing that populism was your biggest concern.

Not every culture encourages the same principles or values as yours. I'd say they possibly would.

I haven't fired up UT99, UT2K4 or any of the Quakes for decades. Are there active servers where you could get a good DM session?

Xonotic has that old school feeling. It's an unfortunately tiny player base, and only a handful of servers with regular players. But if you are within ping of a busy server (like Jeff & Julius resurrection) you do run into the same people all the time, including the ones that run the servers.

As a tea drinker, I wish that cafés, bars and restaurants followed the most basic rule which is: take cold water from the tap, bring it quickly to the boil ONCE and pour it into the tea IMMEDIATELY and then give me the cup before you fuck anything up. I'll take things from there. I don't care if it's loose leaf in a teapot (but heat the teapot first) or teabag directly in a cup (and "scald" the cup too). I care much more that it's a strong "breakfast tea" like Barry's Gold label or something similar.

Also...

    Place one rounded teaspoon of tea per cup into the pot.

My rule is usually one per person and "one for the pot".

Same. Our engineers have failed to convince absolutely clueless executives that service / code ownership is critical, you can't just allocate 100% of your engineering team to "roaming around the dungheap adding features" because they want everybody to focus on features. In an organisation where nobody owns anything, nobody cares about anything, nobody has the autonomy to fix anything. Imagine a huge wobbling Jenga tower with dozens of engineers all trying to pull out a block, put it on top and back the fuck away quickly, hopping it falls over after someone else touches it.

Yes this is real. And it's hell. And now all of those Jenga players are AI powered.

My take, as a non-coder (well, not software engineering, I write 'code' but it's infra, and utilities in go/bash/pythong)...

I work at a company where the biggest problems are not 'writing code', they are:

- Organising teams

- Designing the system

- Prioritisation of work

The fuckups that we make on a daily bases are not 'code errors' they are failures in THOSE three things. I'll go into detail if anyone cares.

Sounds like they partnered to make an amazing advert for using AI tools.

Our S3 bill for a single day was $48 trillion

Just a personal project right now but it's a music metadata aggregator.

I find that nobody really knows how to do this. Machine learning can detect some song attributes well (bpm, ez right?) but it's inconsistent with some things (eg mood, spotify valence)

I prefer to only add metadata that I can rely on: track credits & instruments (when available), lyrics, bpm / "energy" and genre. At least that's what I've got for now. I'm not adding anything unreliable.

So far I'm able to pick a genre, artist or even better, song and it gives me a list of tracks that are similar. I can alter the weights of "era", "instruments", "genre".

So far i haven't run old school NLP on the lyrics but that's the next step. It's likely to be far more informative than "valence"

Anyway, not public, still very alpha but I like it and find it useful.

I get to use the same host names and IP addresses at home or on the VPN. I run a simple wireguard server on a Raspberry pi, and the whole setup has been rock solid.

I did set up tailscale, way back. After using it a few times to test, it failed me when I really needed it (I was out of the country and it failed - can't remember exactly what went wrong but it wwas 100% 'in my tailscale account'). I immediately dropped it and went back to OpenVPN (shit but reliable) before building my current setup.

For small stuff I use bash / sed / awk / jq / yq / curl etc. I stopped using perl (a long time ago) and hate trying to build or distribute anything in Python. "You asked for a script! I give you... a venv, a pip install, woops need to compile something, so let's install build tools. Actually let's get github actions set up to handle this, or would you prefer a Docker image? Do you have a registry so I can share the image? Wait are you using a Mac? Multiarch image time! Or should I just give you a bag of scripts and a Dockerfile? Oh you're using Podman rootless, need you to read this page on how to access your own files as "you" in your container. Shit the wheels fell off! But look, Python is amazing and can do lots of stuff bash / curl / jq can't do!"

So these days I almost exclusively use Go wherever bash and friends aren't enough. Compile to Linux x86 & ARM and Darwin ARM, or just "go run" if it's simple enough.

I love hearing the guys from The Rest Of History discuss this from their different perspectives!

And "hepped up on goofballs" is a Police Chief Wiggum quote, about Ned Flanders.

Lol, is github down again?

Hmm. I don't really care enough about leaking home network host names because they are all super generic names like 'router', 'laptop', 'tv', 'nas'. So I use my public zone on cloudflare. I just use internal ip addresses (eg: nas.example.com = 10.1.2.3) on the public zone and DNS01 challenge for let's encrypt. Anyone can resolve the ip for any of my hosts, but obviously you'd need to be on the wireguard vpn to hit them.

This means that I can always use public DNS servers like 1.1.1.1, 8.8.8.8, nextDNS etc

This is not "done right" by any stretch but it's extremely low effort to set up and has never once failed me, unlike countless complex meshy things.

I have 3 different bookmarks that open 3 different Gmail accounts. I use an AWS SSO plugin to open each account in its own container. My wife and I have several other services where we each have an account, so on our media pc at the TV we can both be logged in at the same time in different tabs. I always open sensitive stuff like banking in its own container. I keep various online stores (where I will be paying for stuff with my credit card) separate from casual browsing. Tabs are colour coded too.

How about mixing up band names? Take the end of "Florence and the machine" and mix it with the start of "Rage against the machine" and you now have the totally unguessable "Rage sharing the machine". It's a different machine see?! Nobody would know that!