Random Thoughts

My go-to tool to generate images (with text overlays and stuff) is svg.el in Emacs. It provides a convenient interface for making SVG files, and then I can just convert it to something else for display. The thing I’ve been using it for lately is to display actor names in movies I’m watching.

Which brings me to what I’m yammering about today — adding outlines to text. Because if you’re displaying text over random (moving) pictures, you have to add an outline so that you can read the text.

The normal way to do that in SVG is to specify a stroke-width. But as you can see, that’s ugly — the stroke is inside the letter shapes, effectively shrinking the shapes down and making them spindly and awkward.

Perhaps easier to see if you really exaggerate the width.

So the obvious question is: Can you specify that the stroke goes on the outside of the shape instead of the inside?

And:

Yes, very amusing. The quibbles are kinda quibbly:

Etc. So: Nope.

But of course it is possible to add an outline to a shape — you can do that perfectly in HTML/CSS with borders and stuff, so instead SVG has grown filters. Look how simple:

<svg width="900" height="200" viewBox="100 0 900 200">
    <filter id="outline">
        <feMorphology in="SourceAlpha" result="DILATED" operator="dilate" radius="4"></feMorphology>
        
        <feFlood flood-color="#32DFEC" flood-opacity="1" result="PINK"></feFlood>
        <feComposite in="PINK" in2="DILATED" operator="in" result="OUTLINE"></feComposite>

        <feMerge>
            <feMergeNode in="OUTLINE" />
            <feMergeNode in="SourceGraphic" />
        </feMerge>
    </filter>

    <!-- DILATED TEXT -->
    <text font-size="85px" dx="125" dy="130" font-weight="700" filter="url(#outline)">upgrade yourself</text>
</svg>

Which reminds me of this classic:

I think the SVG people did, really.

But it does work:

See? Erhm… Hm. That looks a bit odd… let’s up the outline radius:

BY ALL THAT”S UNHOLY!!! OK, I’ve converted the SVG to JPEG using ImageMagick, and its SVG support is a bit, er, funny. Let’s look at it in Firefox:

Yeah, that looks OK?

Let’s see… what about rsvg-convert?

Well, that’s fine, isn’t it? So you need pretty a pretty recent SVG toolchain for this to work, I guess.

I wonder what happens if I up the outline radius ridiculously…

It’s a bit brutal…

A 0.5 opacity looks OK, I guess…

Anyway. That’s the rant for today: We can never have good things.

I’ve got a cold, so I’ve been idly sitting around doing some slightly more thorough testing of my Emacs/mpv setup for asking LLMs what actor is on the screen of the movie I’m watching, and it’s led to me pondering just why some people are so (literally) incredibly enthusiastic about LLMs.

I mean, LLMs are fun. And useful for doing tedious programming things — for instance, the other day I asked ChatGPT to transform a li’l 150 line Javascript thing from Jquery to pure Javascript — and it did it flawlessly. (And also badly, engineering wise, since it did pointless stylistic changes on just about every line.) That would have been boring to do myself, so it’s nice that there’s a tool for that now.

But non-programming things? I just don’t understand the enthusiasm, because whenever I try to use an LLM for something, it’s never more than a toy. Somewhat useful toy, sure, but you can’t say in any way that it actually, like, works.

I wonder whether the enthusiasm is paradoxically based on how bad LLMs are in general. That is, when you chat with one of those things, it’ll give you the wrong answer a lot of the time, but then you say “but that’s not right”, and it’ll say “Good catch! You’re so smart! I’ve never seen anybody be that smart before; you must be a genius!” or variations thereof, depending on how high the company in question has dialled the knob that’s marked “Ass-Kissing” in the LLM console.

It’s just hard to be mad at a technology that’s consistently stupider than you are, and that always confirms your secret suspicion that you’re really, really smart yourself.

So I’m testing the same screenshot repeatedly with the same LLM (Gemini-2.5-Flash here) to try to see just what level of bullshit it’s giving me. This is Keanu Reeves from Even Cowgirls Get The Blues, and while he does look kinda untypical here, I think if you give somebody a cast list and ask who it is, I don’t think it’s that hard to pick him out as the most likely candidate.

But Gemini isn’t better than a throw of the dice here, even though I’ve tried in many ways to instruct it “if you don’t know, don’t guess”.

It even guesses correctly some of the time.

But let’s try some other movies…

It gets Dakota Johnson correct (from the cinematic masterpiece Madame Web).

Always.

This, from Dune Part Two, doesn’t seem right? Actually, I forget his name now… Oh, yeah, Dave Bautista.

No. So is it really just going through the list of cast members and picking one at random? What if I don’t tell the LLM what movie the screenshot is from…

Nope, makes no difference. It picked a guy from a different Dune movie, though?

It gets Timothée right — always, whether I tell it the movie or not.

Well, it’s not a totally bad guess — they aren’t totally dissimilar. (It’s Bill Pullman from Lost Highway, though.)

Yeah, if I tell it the movie, it gets it right.

Nnnnno.

So, once again, when I try to use an LLM for something, the result is somewhat useful — it’s better than nothing, but you can’t rely on it at all.

And people want to use these things to process social security applications and stuff.

Amazon apparently has a thing for letting you know who’s on screen (called “X-Ray”), and according to rumours, it’s based on hiring people in low-cost countries to tag every scene with the people on the screen. I thought identifying actors would surely be a solved problem by now, but I guess Amazon knows best.