Chameth.com - posts like debugging-beyond-the-debugger, infinite-avatars, modern-css-is-fun, why-you-should-be-using-https but not docker-automatic-nginx-proxy, migrating-from-github-to-forgejo

Monthly Meanderings: March 2026

2026-04-01T00:00:00Z

Since last month’s update I’ve been unpleasantly reminded that I’m middle aged, through the medium of a dodgy knee. There’s nothing quite like not being able to stand up without groaning to underscore that you’re not young any more. The ongoing game of “will this seemingly mundane activity make my knee go funny again?” is a delight to play, too. I’m great at it.

Website updates

Two new blog posts this month: the longest way to represent a date is a short thought experiment on absurd date formats, and modern CSS is fun goes through some neat new CSS features I’ve used recently for this site.

I did some design tweaks this month: headers now have a blue background behind them¹, and a lot of components have been redesigned to have a consistent style. You can see that component in action in the new “now playing” widget I’ve added:

Hailee Steinfeld — Rock Bottom Haiz Scrobbled 0m ago

Instead of each component defining its own styles (which were mostly copied and pasted), there’s now a single class that handles the drop shadow, border, background colour, and the optional title. It also applies a consistent format to elements that are clickable: they use the blue accent colour for their borders, and have a glow effect shown on hover.

The now playing widget is part of a larger bit of work I did on importing some music stats. I’m pulling in the data from my Navidrome instance. A new music page shows my most listened albums and artists. Navidrome doesn’t store a complete play history, just the last play and the count, so I can’t do “what I listened to last month” stats until next month.

Other projects

Other than a minor bug-fix update to contempt, my Dockerfile templating/updating tool, I’ve not done much on my other open source projects. I’ve got a bit more work to do finish migrating all my repositories from GitHub to my private Forgejo instance. I’ve also been thinking about mirroring my public repositories to Codeberg so there’s a non-GitHub way to access them.

Entertainment

I apparently only watched a single film in March:

The Substance

2026-03-18

I really didn’t get on with this. It’s too long for what it is, and over-the-top in a way that just didn’t work for me.

Demi Moore and Margaret Qualley were great, but nothing else really was. The cinematography and sound design were too affected, the plot was pretty shallow and obvious, most of the body horror fell into the uncanny CGI valley for me.

It’s not completely without merit: there were around 40 minutes in the middle where I was really into it, but that’s not really enough given it’s closing in on a 2.5 hour runtime.

There’s something to be said for the message it’s presenting about women in Hollywood, too, but it was very heavy handed. Dennis Quaid’s character is even called Harvey; it’s not so much a subtle nod as a sledgehammer to the face.

Instead of films, I’ve been watching a bunch of TV. After devouring The Pitt last month, I struggled to find something to fill the hole it left. I settled on catching up on Chicago Fire and Chicago P.D.. They’re both mostly about getting things done without personal drama becoming the primary focus. They’re not quite as grounded as The Pitt, but at least they’ve not completely gone off the deep end like 9-1-1 where the first responders end up in space somehow?

I’ve also watched the first two episodes of SNL UK. I didn’t have very high expectations, and it gave me a pleasant surprise. Both episodes so far have some hilarious sketches, and are obviously written by British comics with the typical darker, more sardonic humour that separates us from the USA.

Last but not least, board games! I spent a weekend at a tabletop/LAN event with friends, so got to physically play some games for a change. I also kept up with some turn-based games on Board Game Arena, as usual.

6 plays
4 plays
4 plays
4 plays
3 plays
3 plays
3 plays
2 plays
2 plays
2 plays
2 plays
2 plays
1 play

Around the web

Warranty Void if Regenerated

Some really cool speculative fiction about a future where all software is written by LLMs, and the social effects that may have. Unfortunately the story itself is written with an LLM, and that becomes painfully obvious about a third of the way in. By that point I was already hooked, though.

My 2-step process for AI-free blogging

Speaking of LLM-generated text, this short piece by Manuel Moreale made me laugh. I’m by no means anti-LLM, but I really hate people publishing the output like it’s their own work. It breaks the social assumption that the writer put more effort in than the reader, and makes me wonder how much is hallucinated.

25 Years of Eggs

An interesting applied use of coding agents. I’m a sucker for personal statistics. I have no explicit desire to keep my receipts, or generate graphs of my egg consumption, but seeing the graphs makes me question that. The journey to get there is a great read, too.

I made a one-page notebook

The clever paper folding is interesting, but I really like some of the small touches on the page. There’s a nice little “in reply to” callout providing context at the top of the page, you can click to rotate the image further down, and there’s a details element containing a textual version of it. It has that great ‘small web’ vibe of someone who really cares about what they’re doing.

Workers who love ‘synergizing paradigms’ might be bad at their jobs

I haven’t read the full paper, but the summary given by the article is equal parts interesting and amusing. They developed a “Corporate Bullshit Receptivity Scale”, what more is there to say?

It’s meant to resemble painter’s tape, but I’m not sure quite how well it works. ↩︎

Modern CSS is fun

2026-03-17T00:00:00Z

I’ve been doing a bunch of CSS tweaking recently, and keep being surprised by how nice modern CSS is to work with. As someone grey-haired enough to remember writing HTML without CSS, it’s amazing to think how far along web technology has come¹. I wanted to demonstrate some of the handy bits and pieces I’ve used recently.

:has

The :has selector allows you to effectively query for child elements. While a span will match a <span> within an <a>, a:has(span) will match an <a> that contains a <span>. This really shines when combined with more complex selectors, for example:

input {
    border-radius: var(--border-radius);

    &:has(+ .results:not(:empty)) {
        border-bottom-left-radius: 0;
        border-bottom-right-radius: 0;
    }
}

This is part of the styling for my film search box. When the search box is presented initially it has fully rounded corners. But when it is followed by a non-empty results element, it removes the rounding on the bottom corners so that the border continues in a straight line down into the results. You can see it in action below; just enter a few characters (like “the”) to get some results:

You could achieve the same effect by having JavaScript add a class to the <input> element, but I’ll take a CSS solution over a JavaScript solution any day.

Nested rules

You probably spotted this in the example above. It’s what finally made me switch from SCSS to plain CSS. If you have a rule for .foo and a rule for .foo .bar you can just nest them. Not only does it save repeating yourself an awful lot, it keeps everything organised nicely. For example, the CSS for my film list embeds looks like this:

a.film-list {
    display: grid;
    grid-template-areas: "header images" "description images" "meta images";
    grid-template-columns: 1fr auto;
    grid-template-rows: auto 1fr auto;
    row-gap: var(--small-space);

    @media (width <= 800px) {
        grid-template-areas: "header" "description" "images" "meta";
        grid-template-columns: 1fr;
        grid-template-rows: auto auto auto auto;
    }

    & + & {
        margin-top: var(--medium-space);
    }

    h3 {
        font-size: var(--font-size-xxlarge);
    }

    .description {
        /* ... */
    }
}

Everything is wrapped up in a single bundle, including the @media rules for changing the layout on smaller screens, the lovely little & + & rule that adds some extra margin if there are two lists in a row, etc. The & syntax refers to the parent selector, so & + & in this case is the same as a.film-list + a.film-list: it’s very handy! The best part about this is that it’s the exact same syntax as SCSS, so converting is easy². The styled element looks like this:

Watched films ranked

Every film I’ve watched since I started logging, ranked.

Obviously super subjective, and subject to change often and arbitrarily.

264 films

Media range syntax

Once again, you might have spotted this in the previous snippet. I’m not deliberately teasing things, I promise! Back in the day, you did media queries like so:

@media (min-width: 1000px) and (max-width: 2000px) {
    /* some rules that work for screens at least 1000px wide and at most 2000px wide */
}

I always hated this syntax. I always got muddled up as to whether I wanted “min” or “max”, and whether they were inclusive or not. It’s easy to reason through, but it never came naturally. Fortunately you can now just use ranges:

@media (1000px <= width <= 2000px) {
    /* some rules that work for screens at least 1000px wide and at most 2000px wide */
}

I find this style so much easier to write and understand.

Anchor positioning

Trying to dynamically position one element next to another used to exclusively fall within the purview of JavaScript. Fortunately anchor positioning fixes all this. I use this in the film search field I showed above:

.film-search {
    input {
        anchor-name: --film-search-box;
    }
    .results {
        position: absolute;
        position-anchor: --film-search-box;
        position-area: bottom center;
        position-visibility: always;
        width: anchor-size();
    }
}

This does lots of fun things. The .results element is anchored to the input field (via the --film-search-box name), and it’s positioned on the bottom, directly under the anchor. The position-area rule describes a 3x3 grid, with the anchor at the center, so top left would put it diagonally adjacent to the anchor. Next, position-visibility keeps the .results element visible when the anchor isn’t; otherwise when you scroll the anchor off the screen the results would immediately vanish. Finally, the special anchor-size() function makes the .results element take up the same exact width as the anchor.

All-in-all, this puts the results exactly where I want them, without having to deal with any JavaScript at all. I can see how it’d also be amazingly useful trying to do tooltips or other forms of ‘floating’ content too. You can even leave the positioning up to the browser, giving it hints about which order to try, or what property to optimise for (e.g. you can say “anchor to the left or right, whichever has more horizontal space”).

attr() function

This one blew my mind a little. In some of the elements on the site, I’ve added a little label to the top. It looks something like this:

It’s meant to just be a visual flourish, not a semantic title or anything, so I initially just added a ::before selector to each element, customising the content to have the right value. I didn’t like the duplication, though. Now, instead, I do this:

&::before {
    content: attr(data-title);
}

Each component defines its title in a data attribute: <div class="raised-box" data-title="an example">, and the attr function plucks it out and puts it in the content rule. It’s worth noting that this is only widely supported for the content rule, you can’t yet use it for colours or dimensions or other things. It still feels a bit magical, though. Like you’re making a reusable, customisable component with just HTML and CSS.

layers

I’d come across layers a few times, but I never found a need for them. It seemed like something you’d only really need with complicated design systems, or something. Then I had a problem and layers were the perfect solution!

The headings on this site have a whole bunch of CSS attached to them. Currently they look a bit like blue painter’s tape with handwriting on them. This style is applied to all headers, and then the various places that don’t want it had to manually reset everything. I got fed up with that, so decided to add a plain-header class which would “disable” the extra styling:

h2,
h3,
h4,
h5,
h6 {
    /* common styles for all headers, regardless */
    font-family: "Chris Hand", sans-serif;
    font-size: var(--font-size-xxlarge);
    color: var(--text-highlight-colour);
    margin: var(--medium-space) 0;
    line-height: 1;

    &:not(.plain-header) {
        /* do fancy stuff if it _doesn't_ have the .plain-header class */
        transform: rotate(-0.15deg);
        padding: var(--smedium-space);
        font-weight: 800;
        filter: drop-shadow(3px 3px 5px rgb(0, 0, 0));

        a {
            background: none;
            text-decoration-skip-ink: auto;
        }
    }
}

This might seem a bit backwards but I want the default to be styled, as those are far more common. I added this change, sprinkled around some “plain-header” classes, tidied up a lot of duplicate CSS, and things mostly worked. Mostly. Some things stopped working. In a few places the headers had slight tweaks, and they stopped working entirely.

It turns out in moving the rules from a plain h2 selector to a h2:not(.plain-header), I’d made them more specific. Previously a selector like h2.special would have been more specific, so its properties would override those from the less specific h2 selector. My first reaction was to try and hack around it. Changing all the overrides to body h2.special would make them more specific again, but I didn’t want to have to remember to do that forever more.

Instead, I defined some custom layers, and put the header definitions in one:

@layer reset, links, headings

@layer headings {
    h2,
    h3 {
        /* etc */
    }
}

The first @layer rule defines some layers that will be handled in a specific order: first the “reset” layer, then the “links” layer, then the “headings” layer, and then anything not in a layer after that. Rules in one layer don’t have to worry about beating the specificity of rules in another layer, because they’re handled separately.

Note that I couldn’t just put the headings in a layer and call it a day: their rules have to come after the CSS reset, and after the normal link styles, as headers have a bit of extra styling for nested links. If the CSS reset weren’t in a layer, then the * { margin: 0; padding: 0; } type reset would apply over the rules from the headings layer.

Using layers is definitely a bit fiddly, and is not necessary for a lot of sites, but it’s a much cleaner alternative to ugly specificity hacking.

Things I’m looking forward to

There are a few things that aren’t (widely) available yet, that particularly interest me:

sibling-index() and sibling-count()

These new functions are available in WebKit and Blink based browsers, but not Firefox. They give you the index of the element within its siblings, or the count of siblings. For the film list component above, I currently have this abomination for the overlapping posters:

&:nth-child(1) {
    z-index: 5;
    left: 0;
}
&:nth-child(2) {
    z-index: 4;
    left: calc(133px * var(--overlap) * 1);
}
&:nth-child(3) {
    z-index: 3;
    left: calc(133px * var(--overlap) * 2);
}
&:nth-child(4) {
    z-index: 2;
    left: calc(133px * var(--overlap) * 3);
}
&:nth-child(5) {
    z-index: 1;
    left: calc(133px * var(--overlap) * 4);
}

These could be replaced with something like:

z-index: calc(1 + sibling-count() - sibling-index());
left: calc(133px * var(--overlap) * (sibling-index() - 1));

As an added bonus, it would scale to any number of posters. index being 1-based not 0-based is a bit unfortunate as it makes using it in calculations awkward (see 1 offset in both of those rules!), but it’s leagues better than writing 5 separate rules.

random()

Currently only available in Safari. There are a few places where I’d like to have slight random variations of the style. Things like elements that are rotated slightly for aesthetics; they look a bit silly if they’re all identically positioned. I also randomise the icons and positions of my rating stars to break up the visual monotony. Here are a few examples:

Currently I do the stars by hardcoding a bunch of classes and having the backend randomly apply one when it generates the markup for stars. Being able to do it in CSS would be great, though.

Mixins

This one’s so far off it’s not even listed on caniuse.com yet. Mixins are another handy feature of SCSS, that allow you to define reusable blocks of rules, then import them when needed:

@mixin fancy-background {
    background: /*...*/;
}

.some-element {
    @include fancy-background;
}

#other-element {
    @include fancy-background;
}

This would be particularly useful if you can’t easily control the markup to add classes to everything. For example, my footnotes are rendered by a markdown plugin, and don’t easily have a way to add extra classes to them. To style it the same as another element, I currently duplicate a bunch of rules between them both. Mixins would allow me to define those rules once, and then import them in both places.

Aside from JavaScript, which seems to just endlessly reinvent new frameworks and ways to make the most sprawling dependency tree possible. ↩︎
It also means that if your website’s syntax highlighting library doesn’t seem to understand nested rules, you can just pretend your perfectly valid CSS is actually SCSS and it’ll magically work. Grumble, grumble. ↩︎

Making a font of my handwriting

2025-08-08T00:00:00Z

Recently I’ve been on a small campaign to try to make my personal website more… personal. Little ways to make it obvious it’s mine and personal, not just another piece of the boring corporate dystopia that is most of the web these days. I don’t quite want to fully regress to the Geocities era and fill the screen with animated under construction GIFs, but I do want to capture some of that vibe.

I’d added some bits and pieces along those lines: floating images in articles now look like they’re stuck to the page with sellotape, related post links have a wavy border that animates when you hover over them, and so on. Next, I wanted to change the heading fonts from a monospace font to something cursive, to resemble handwriting. Less terminal output, more handwritten letter. I couldn’t find one I liked, though. So why not make my own? It can’t be that hard, right?

Failing to do it myself

I set out to try to make the font myself using open source tools. After doing a bit of research, it seemed like the general approach was to create vectors of each character and then import them into a font editor. That seems to mean either Adobe Illustrator and FontLab (if you have too much money) or Inkscape and FontForge (if you like open source). I fall firmly into the latter category, so I grabbed my graphics tablet and opened Inkscape.

I wrote out my first three letters: capital A, B and C. Saved them in Inkscape, and attempted to import them into FontForge. Then I remembered one crucial thing that had slipped my mind: I absolutely loathe using FontForge. It’s a bit like when you open an old version of GIMP and get a bunch of weird looking windows floating all over the place; it feels like you’re fighting against the tool to do even the most basic operations. The difference is I have cause to edit images a lot more than I edit fonts, and GIMP has actually significantly improved their UI over the years.

Here are the rough steps I went through with FontForge:

Launch Font Forge. It shows a weird bit of art in one window, and an open file dialog in another.
I don’t want to open a file, so I close that dialog. The program exits.
Relaunch Font Forge, and realise that within the “Open Font” dialog is a “New” button. Click it.
Get to the standard font-editing UI. Right-click on the “A” looking for a way to import an SVG. Don’t see one.
Click around a bit, exploring the menus. Everything feels a bit off. You can’t open one menu then hover over the next to see its content, like basically every UI toolkit in existence. I think FontForge has eschewed QT and GTK in favour of doing things itself.
Find the “Import” option in the File menu. Hope it’s for a single glyph not the whole font.
A file picker opens. Again it’s all a bit off from normal desktop conventions. Try to resize it, and just get blank grey space at the bottom.
Type the absolute path I want to go to in the text field.
Get a dialog saying “Not a bdf file /home/chris/etc”. Press OK.
Get a dialog saying “Could not find a bitmap font in”. Press OK.
Press Ctrl+L to see if that lets me enter a path. Click everything in the dialog to try to find a way to enter a path. Get annoyed. Give up. Click through folder-by-folder to get to where I want to be.
Get to the folder and don’t see any files. Change the format to “SVG”. Double-click the newly-visible SVG file.
Get a dialog saying “You must select a glyph before you can import an image into it”. Press OK.
The import dialog goes away, having not imported.
Select the glyph in the main tool area, then repeat the File→Import dance.
It’s actually there now! Open the glyph in the editor and see it’s a complete mess of Bézier curves. I can’t click what I want without accidentally moving a handle for an adjacent curve.
Rage-quit.

I’m sure FontForge is less anger inducing once you’re used to it. And you definitely could use it to build a font like this if you had much more patience than me. I’d had enough of death-by-a-thousand-paper-cuts though.

I briefly tried Inkscape’s built-in support for making an SVG font. It annoyed me a lot less, but it’s fiddly: it seemed like each glyph had to be a single path, so you had to convert the glyphs to paths, then merge them correctly. If you merge them incorrectly then the wrong bits of your letters end up filled (like the inside of the ‘B’). Path manipulation is getting towards the limit of my knowledge of vector editing, and it took a bit of trial and error for each letter that had more than a single stroke. I didn’t fancy doing that for every letter.

I’m usually a big advocate of open source, but this was one of those painful times when it feels like it just falls short. Clunky, painful UI and processes where commercial tools just let you get on with your work.

You can exchange money for goods and services

When I’d been looking for open source tutorials, I found many mentions of a closed source, hosted tool: Calligraphr. It has a free version with limitations (no ligatures, no variations, 75 glyphs per font), and a pro version for £8/month. I’d normally balk at the idea of a subscription for this, but they have the perfect answer: you can make a one-time payment, and your account automatically downgrades back to free after a month. It’s not a hidden option, either, it’s the most prominent button on the upgrade page. That made me happy to give them £8 to play around with the service for a month.

Calligraphr works by having you print templates, write out the letters, then scan them in. It does some magical processing to extract the glyphs, provides tools to tidy them up, align them, etc, and then produces a TTF file for you. You can see some of my completed templates here:

Most of the templates I used for the font

Calligraphr has a nice UI to generate the templates, allowing you to select which glyphs to include. I added the “minimal English”, “basic punctuation” and “Ligatures” sets. That gave me four pages to fill out, and I did them all twice. That let me filter out versions that didn’t work well, and have variants for some letters so the font wasn’t too repetitive. Later on, I went back and added some custom ligatures based on blog post titles that didn’t look quite right: “Re”, “To”, “ers”, “ey”, “hy”, “ra”, “re” and “ty”. Ligatures like this help it look more natural: when we write we don’t just stamp out identical letters regardless of their surroundings, instead they will connect to their neighbours, or overlap slightly, or even share a stroke.

I filled these templates in with a Sharpie, as I wanted a fairly informal, scrap-booky look, and it would also give good solid shapes that should be easy to pick out of the template. I scanned them with the “Scan Document” function on my iPhone, and uploaded the PDFs to Calligraphr.

Iterating and tweaking

The Calligraphr UI allows you to preview the font, but I found it a lot more useful to just download a copy and use it on a local copy of my website. That let me test it with real text, and see how it’d look at the different font sizes I use on the site.

The first version was not great. Despite the guidelines on the template, I apparently wasn’t good at sticking to them. Some letters were floating way off the baseline, and some were sunken below. When those opposites met it looked terrible. Fortunately Calligraphr has a pretty easy tool to slide each letter up and down, and scale it up or down if needed, and you can see it next to other letters as you do it. It took a little bit of time to go through all the variants of all the letters, but the next version looked a lot better.

Another tweak I ended up doing was reducing the spacing between letters. The defaults Calligraphr uses are probably good for a blocky font, but I wanted to put the letters close together to give it more of a joined-up look. Again, this is an easy tool to use, you just drag the sides in or out as desired. While these tweaking steps were probably as fiddly as some of the Inkscape steps I refused to do earlier, they’re a lot more rewarding as you see things improving with each one. It’s a lot easier for me to commit time and effort to improving something that’s already working reasonably, than put that time and energy into an unknown.

Later, I noticed that occasionally there would be a huge gap in a title. Not “the kerning is slightly off” but “there’s enough room to park a bus”. It took me a while to figure out what was happening: a couple of glyphs hadn’t been isolated perfectly and had picked up a few pixels from the template lines at the edge of their boxes. That meant the glyph had a width that covered the actual written glyph, a big gap, and then the rogue marks. At first, I fixed this by just adjusting the width, but that left the little pixels floating awkwardly down-sentence. The proper fix was to use the editing tool and simply delete them, and then Calligraphr snapped the width back to what it should be.

These iterations took a while to do, but I just dipped in and out occasionally over the course of a week, so it didn’t actually feel like too much work. I quite enjoy the process of refining things, too.

Result and a surprise

If you’re viewing this post on my website¹, you can see the font in the headers, captions, and a few other places. Here’s how it compares to my actual handwriting:

My handwriting vs my handwriting font

It’s not close enough to forge documents, but I think it definitely gets across my style, and that’s exactly what I wanted. It’s surprisingly legible even at smaller font sizes — I think the weight of the Sharpie helps here — and at £8 and a bit of manual work was a lot more economical than spending days wresting with open source tools.

A few weeks after I put the finishing touches on the font, I got an e-mail from Calligraphr. As my account had lapsed back to the free version, I was no longer eligible for the “server-side backup” feature. So what did they do? They e-mailed me an exported copy! It’s a JSON file with the properties of each glyph and a base64 encoded image. Not only can I re-upload this to Calligraphr if I resubscribe, I can probably hook something up to edit it should I ever need to. I’m blown away by how pro-user Calligraphr’s business practices are. They’re up-front about pricing, don’t try to get you stuck on an auto-renewing subscription, and automatically export your data. It’s like a breath of fresh air compared to the barrage of dark patterns that other websites foist on us. If you want to make this kind of font, I’d definitely recommend them just because of how nice they are.

And I haven’t changed everything since writing this post… ↩︎

Coming around on LLMs

2025-05-28T00:00:00Z

Claude says hi.

For a long time I’ve been a sceptic of LLMs and how they’re being used and marketed. I tried ChatGPT when it first launched, and was totally underwhelmed. Don’t get me wrong: I find the technology damn impressive, but I just couldn’t see any use for it.

Recently I’ve seen more and more comments along the lines of “people who criticise LLMs haven’t used the latest models”, and a good number of developers that I respect have said they use coding models in some capacity. So it seemed like it was time to give them another shake.

The first decision to make was which model to try. OpenAI are no longer the only player in the game, every tech company of a certain size is now also somehow an AI company¹. I looked a bit at some benchmarks, and then mostly ignored them and went with the only company that I didn’t outright hate: Anthropic, and their model Claude.

Initial impressions

The latest Claude models do feel a lot more “capable” than the earlier ChatGPT versions I remember, but they also still have a lot of the same problems. At their heart, they’re still text-prediction models, and still seem to be trained to predict text that will please the user rather than be factually accurate or useful.

One of the improvements that people kept mentioning was the ability for models to access the web directly. I’ve seen it kick in a bit naturally, and asked for it explicitly sometimes, and it’s… nothing special? Are people just really bad at searching the web? Maybe they should try Kagi? I feel like about 80% of the time I could have found the information just as quickly myself, about 10% of the time it searched and then hallucinated an answer, and the remaining 10% it found something quicker than I otherwise would have. Those aren’t great results, especially when you consider how much cheaper and simpler just searching the web is.

There is another way to trigger web searches in Claude: using “research mode”. When you enable it, it splits off into different models: a “lead researcher” to come up with a plan, and then some minions that execute it. It ends up doing hundreds of web queries in the span of a few seconds. Suddenly I understand why things like Anubis need to exist. I’ve not used the “please launch a DoS attack” button since.

What did impress me, though, was its ability to churn out reasonable-ish code. It can hack together a bash script as well as I can, and do it far faster than I’d be able to. Sometimes they even work. That made me wonder what it would be like doing actual coding with it. Anthropic have a CLI tool called claude-code, so I paid them lots of money and gave it a spin.

Coding with Claude

The first thing I notice about claude-code is that I really like the interface. It’s basically an input box in a terminal. It doesn’t force me to use a certain IDE or do things in a certain way. By default it asks before making any changes, showing you a side-by-side diff of what it’s doing and allowing you to provide feedback. If I had to design a way to interact with a coding agent from scratch, I can’t think of many things I’d improve.

A simple example of a claude-code session

The power in claude-code versus just using the web UI is that it can use tools. It can query git, run grep commands, even use sed if it wants to change something in lots of files at once. It’s very good at figuring out its way around a codebase, even without any explicit instructions. You can see in the screenshot that with a little prompting it managed to get the permalink to this post; I didn’t tell it where the posts were stored, or how to work out the latest, I just told it when it was wrong. If I’d run /init before it would have probably picked up on the fact that my posts have custom permalinks, and noted it in CLAUDE.md.

But how good is it at actually writing code? It’s like having a keen but not particularly thorough Junior Engineer at your beck and call. If you give it a clearly defined task and guidance on how to implement it (and maybe some feedback as it suggests changes), it’s more than capable of doing it. If you don’t give it enough guidance it tends to go more off the rails. I tried having it generate a simple application from scratch with minimal technical guidance and no review of what it was doing, and it made such a mess of it I decided it was quicker to throw it away and start again by hand.

The man behind the curtain

Even with sufficient guidance, at times it’s really obvious that it’s an LLM generating pleasing-token-strings and not something that genuinely understands what it’s doing. It will spit out code like this:

if err != nil {
    if err == sql.ErrNoRows {
       return nil, err
    }
    return nil, err
}

Why’s that check for sql.ErrNoRows there? It’s entirely pointless. There was no instruction to check for it, none of the existing code checked for it, but I assume it comes up quite a bit in the training data. But it didn’t understand why, so it put the check in, and returned the exact same thing as if it hadn’t.

It also occasionally tries to “cheat” or solve problems the wrong way. It sometimes feels a bit like you’re asking for wishes from a Monkey Paw. “Stop the unit tests failing”, you’ll say; Claude will respond with a request to delete the failing test. The man behind the curtain isn’t particularly well hidden, and the training data and pattern matching often shows through.

This lack of understanding also makes it very hard to get Claude to use comments in a sensible way. It absolutely loves doing nonsense like this:

// render form
function renderForm() {
    // ...
}

Even with explicit instructions to avoid useless comments, or comments that only explain what the code does (not why it does it), or other prompts. Again, I’m chalking it up to a mixture of training data that does that, and lack of any actual understanding about why a human developer may want a comment to exist. Of course, there are plenty of flesh-and-blood devs out there that also don’t comment effectively, so maybe I shouldn’t give Claude too much of a hard time on this.

… and yet …

So with all those problems, it sounds like it’s just not worth it, right? Well, not quite. I feel like programming mostly consists of two distinct tasks: thinking how to implement things; actually implementing them; and then debugging why you’re off-by-one somewhere. Letting an LLM do the thinking is a big no-go for me: it’s evidently not good at it, and frankly that’s one of the things I most like about programming. Having let it loose on small projects, I shudder to think what the codebases of all the “vibe coded” projects that are popping up are like.

For the second stage, though, it’s actually quite nice. If I’ve thought through how I want something to be implemented, I can feed those steps to Claude and have it churn out the otherwise not-too-interesting code. This requires far less time and concentration on my part than writing code, to the extent that I can be thinking about the next feature, or doing something else at the same time. I don’t feel like I’m going to end up deskilling myself this way, as I’ve already formed the idea of what I want to code, I’m just using the LLM to spit it out faster than I can type it.

A lot of the most boring bits of coding like implementing CRUD-y operations can be summarised as “Look at this file/function. Do the same thing but slightly differently elsewhere”. Claude is great at this, especially when explicitly prompted like that. You’re basically playing to its pattern-matching strengths, rather than asking it to come up with anything novel. Most of my prompts tend to be prefixed with “Look at @some_file.go and @other_file.go.” to cue up the patterns I want it to use, and I find this works well.

As for debugging, it’s a mixed bag. It’s sometimes amazingly insightful, and sometimes just runs around in circles trying the wrong things over and over. It’s worth asking the question, but I definitely wouldn’t rely on it over my own abilities.

The future

I’m probably going to carry on using claude-code, at least for personal projects. I’ve got so much done that I just wouldn’t have been bothered to do if I was doing it all by hand. I very much enjoy being in the more “diffuse thinking” mindset, planning how things are going to work, rather than being stuck in the mines digging out SQL queries. After all, who wouldn’t want an over-eager assistant to work on all their hobby projects?

Work is a slightly different matter: a private project or even an open source project that disclaims any liability is different to something I’m being paid to deliver, and bear responsibility for fixing if it’s not done correctly. I’m not saying I won’t use it at all, but if I do it’ll be much more constrained than I would in personal projects.

As for non-code usages: I’m not sold. My sceptic hat is still firmly in place. I don’t think chat is a particularly good interface for many things, and hallucinations are still a big problem despite what people say. I hate the tide of AI slop that’s taking over the Internet, and how LLM-powered chat agents are being forced into every random product. You’re definitely not going to be seeing any AI-generated blog posts from me!

One topic I’ve not gone into here is the ethical concerns about using LLMs. They obviously exist, and I do have thoughts, but that’s a topic for another day.

I’m surprised it’s not gone more mainstream: why’s there no Tesco Value LLM model, yet? ↩︎

Generating infinite avatars

2022-12-30T00:00:00Z

An example of one of the unique avatars

I recently added a new ‘about’ section to the top of my website. Like most about pages, it has a picture. Instead of a normal photograph, however, you’ll see an AI-generated avatar. This is admittedly fairly trendy at the minute — apps like Lensa offer to make you profile pictures if you give them a set of photos and some cash — but I’ve done something a bit different.

You see, there is not just one image that has been carefully curated, edited, and uploaded. No, the image you see quite possibly has never been seen before and will never be seen again. It’s unique. Just for you.

Background: Stable Diffusion, DreamBooth, et al

You’ve probably heard of Stable Diffusion, the open text-to-image model developed by LMU Munich. Given a text prompt it starts with a random array of static and repeatedly transforms it, each step moving away from pure entropy and towards a real image that befits the prompt¹. It stands in contrast to competitors like DALL-E and Midjourney in both the code and the model being freely and publicly available.

One interesting side effect of that is that you can take the pre-trained Stable Diffusion model, and run your own further training on top. It takes hundreds of thousands of GPU hours to train such a model from scratch, but only a few to add some specific tweaks on top. Earlier this year researchers from Boston University and Google Research published a paper titled DreamBooth which presents a model for doing exactly that.

This research has spawned a slew of startups that do the training and/or generation for you in return for cold, hard cash. The most popular of these at present is Lensa, a mobile app that will generate 200 avatars for you in pre-set styles for £9.99. You can’t change the styles or regenerate any you don’t like, but from what I hear 200 is just about enough that you’ll find one or two that you like.

The lovely ASCII art shown while training is underway

While £9.99 isn’t much money, if you’re technically inclined then it’s not very difficult to do the work yourself for free. That also lets you come up with unique prompts, creating pictures in different styles, with different backgrounds, and so on. I used the wonderful notebooks from TheLastBen that run in Google Colab. The generous free tier offered by Colab is plenty enough to run a DreamBooth training session, and the notebook walks you through pretty much everything².

If you train a custom model then you end up with a weighty file called a “checkpoint”, which you can provide to most Stable Diffusion tools to use when generating images. I use AUTOMATIC1111’s stable-diffusion-webui which not only offers a simple web UI, but also a REST API for accessing it programmatically. I installed this on my laptop and spent a happy hour or two generating weird and wonderful pictures of me.

Automating it

When I was training the model, I was planning on finding a single nice avatar to use. After playing around with it for a while, though, I wanted to expose all the wacky and unique pictures that it was generating. I came up with the rough idea of batch generating a number of avatars, then having a custom webserver that served you one and deleted it.

My first attempt at this was to try and run the Stable Diffusion process entirely on CPU on a server. I’d previously run an SD generator on my laptop without CUDA support, and while it was deathly slow it still worked. With this custom model, though, it took longer to initialise than I was prepared to wait — and I hate to think how long the subsequent image generation would have taken!³

I obviously needed something with a GPU, but I didn’t want to use my laptop as it may be unavailable or doing other more important things with its GPU like playing games. So I turned to AWS, and found they have GPU-enabled instances that can be obtained for reasonable amounts of money. As I didn’t really care when the batch processing ran, I could use “spot” instances which offer a decent discount in exchange for only being able to run when there aren’t reserved instances that need the resources.

After much fiddling in the AWS console, I got a spot reservation set up for a GPU-enabled instance. After waiting a while and not seeing any instances appear, I checked the logs and found it was erroring because I was trying to exceed my vCPU limit. Odd. A bit of googling⁴ later and I discover there’s a separate limit for that type of machine and the default limit is 0. There’s a whole mini application in AWS for requesting limit increases, so I requested a modest increase to 8 vCPUs (the minimum configuration for the “accelerated computing” images is 4 or 8 vCPUs depending on the exact type). After a brief wait, Amazon declined my request:

I am sorry but at this time we are unable to approve your service quota increase request.

Service quotas are put in place to help you gradually ramp up activity and decrease the likelihood of large bills due to sudden, unexpected spikes.

I’m not entirely sure how you’re meant to ramp up without being able to run a single instance. There are lots of theories online about account age requirements, minimum spends, etc, but I wasn’t willing to jump through inscrutable hoops in order to try to give Jeff Bezos more money. Instead, I looked at Paperspace, a service I’d come across previously when trying to run a GPU-enabled Windows box. They have a variety of GPUs on offer, and a lovely API to remotely manage machines. [If you want to try Paperspace you can use this referral link to get $10 off. In doing so you’ll give me enough credit to generate around 10,000 avatars. If that’s not a worthy cause, I don’t know what is.]

I went a bit overboard investigating the different GPU offerings and their relative bang for the buck:

A slightly over-the-top analysis of the GPUs offered by paperspace

The A4000 comes out on top: it’s built on a modern architecture with a large number of CUDA cores, and is really competitively priced. Paperspace only give you access to the M4000 and P4000 initially and make you request access to the higher tier units. I dutifully filled out the very brief form, and a day later it was approved. At least someone is willing to accept my money!

Writing some code

After setting up the machine on Paperspace and copying over my custom model, I set about writing code to handle the generating and the serving. I eventually settled on having two buckets of images: ones that will be shown to only one person and deleted on use, and a fallback bucket that will be used multiple times. The fallback bucket is so that I can limit how many avatars I need to generate (and thus how much money I pay for GPU time⁵). I set a global limit of one avatar used every 10 minutes, as well as a per-IP limit of one unique avatar per 24 hours.

In order to distinguish whether you’re seeing a unique avatar, the server adds a border around it and a “UNIQUE LIMITED EDITION” label at the bottom. If you see that text, you’re looking at an image that has never been seen before and that has already been deleted. The server sends some aggressive caching headers, so in normal day-to-day operations you should see a different unique avatar every day you visit the site.

The generating side is a bit more interesting. It monitors the contents of the two avatar buckets, and springs into action if they fall below a configured minimum. It starts the process by calling Paperspace and requesting the machine is started up, then repeatedly polls the status endpoint until it’s ready. It then generates images individually using the REST API until it hits the bucket’s configured maximum. Once it’s done, it asks Paperspace to shut the machine back down.

Initially I just hardcoded a set of prompts for the generator to use, but they resulted in a lot of fairly similar images. To make things more interesting, I started dynamically generating the prompt using a combination of:

A prefix such as “A painting of”, “A sketch of”, “A photograph of”
In 80% of prompts, an artist reference such as “in the style of Andy Warhol”
In 30% of prompts, a film reference such as “from the film The Matrix”
1-10 random suffixes such as “bokeh”, “8K”, “trending in Artstation”⁶

Initially I had the film and artist prompts independent, but the occasions where neither appeared in the prompt lead to pretty bad images. Instead, there’s now a 20% chance of a film reference, a 70% chance of an artist reference, and a 10% chance of both⁷. There’s a list of around 10 prefixes, 70 artists, 20 films and 20 suffixes which gives a large pool of random prompts.

End results

Everything I’ve described is now live on chameth.com – if you visit you might get a unique, never-been-seen before version of me. The code for the generating and serving is available on GitHub if you’re interested or want to replicate this for yourself.

To finish off, I ran off a batch of 200 avatars and have selected the most interesting ones:

A selection of avatars produced by the generator

As you can see there was one output that appears to be a cat with a ball of yarn, rather than a picture of me. That seems to happen occasionally when the various parts of the prompt don’t gel well, but I’m happy with 0.5% or so of the images being somewhat random! The batch of 200 avatars took just shy of 17 minutes to generate, which will result in a bill of $0.22 from Paperspace.

The famous quote from Arthur C Clark comes to mind when I think too much about how this works: “Any sufficiently advanced technology is indistinguishable from magic”. ↩︎
The only thing it doesn’t help with is finding enough pictures of yourself to use for the training data. That’s presumably easier if you’re more of a “selfie person” than I am. ↩︎
There were a lot of differences that could account for the extra slowness: my custom model was based on the larger 2.1 SD model rather than 1.5; I was using different software; and my laptop CPU is far more modern than the server’s. ↩︎
In the genericised sense: I used Duck Duck Go. ↩︎
One of my biggest concerns here was to avoid putting a “make Chris pay money” button on the Internet. That felt like a bad idea. ↩︎
AKA the random detritus that gets appended to prompts to make images better in mysterious ways. The English pedant in me hates this nonsense, but the results when you spam rubbish modifiers are inarguably better than when using straight forward prose. ↩︎
Imagining the Venn Diagrams is left as an exercise for the reader. ↩︎