Right — the .ASX format in that listing is something entirely different (and altogether obsolete), as far as I can tell. And .JSX files are much more common (they’re what I use at work all the time), but they’re also not something you’d use for slideshows on iPhone… they’re a kind of web app source code file, not a finished slideshow or website that you’d use directly.
Anyway, I was just curious. No need to further sidetrack the convo unless you’re intrigued too. (And if you are, I think you’d just have to wait for Claude to reset its usage and ask it directly to see if it’s something in its own memory that nobody else has access to).
If you’d like, I can also request that a mod help us move the .ASX convo into a new side thread altogether?
Not needed, but I’m not gonna talk about it more here. To be clear, Claude said he was using .asx files, and sometimes but not others, when I asked for asx it saved as jsx….
Claude’s ability to generate graphics is interesting. It has no image generator like the other models do, so it can’t create a rastered image like a photograph. But it can do things programmatically. So it can create vector graphics because those are fundamentally encoding a series of relationships rather than a series of pixels. Makes it useful for creating graphs and charts and such.
“Claude Design” is a feature they’ve rolled out that leans into this ability to design things like UIs for programs or graphics useful for presentations and it actually works pretty well even though it can’t render actual images. It even does animations, it’s just that animations are of vector graphics or other graphics you can achieve with things like CSS, javascript, etc. so they tend to be fairly basic and geometric, but it’s impressive how far it can get without a true image generator.
I suppose that, in some ways, it might make it easier for it to learn images. It’s a fundamentally difficult problem to look at a rastered image, and pick out which things in the image are actual discrete objects-- We mostly just don’t appreciate how difficult it is, because a major chunk of our brains are devoted to doing just that. But in SVG graphics, discrete objects will typically be bundled as such in the code. You don’t need to see the separate objects; they’re clearly labeled for you, and the labels might even include the objects’ names.
I see what you mean, but Claude is already quite excellent at using machine vision to interpret rasterized images. Give it a try - ask it to name all the elements in a picture, critique a photograph you took for artistic merit, identify the joke in visual humor and it will usually get it right.
I think Anthropic just decided that they didn’t want to spend the time and resources in competing in the generative image/video/audio field and wanted to just be the best textual output LLM possible.
Rendering vector graphics is closer to coding (where it’s already the best) than traditional image generation, which is how they sort of snuck around the ability to generate “images” without really creating any image generator (diffusion or autoregressive rasterizer).
Oh, yeah, modern AIs can do it (just like we can), but it’s still hard. It took a lot of training and other effort to get them to that point. An SVG generator doesn’t have nearly as much training data to work with, so it’s a good thing that the problem isn’t as hard for them.
I asked my ChatGPT some Voight-Kampff test questions from Blade Runner. Here’s the responses:
Me:
You’re in a desert walking along in the sand when all of a sudden you look down and see a tortoise. It’s crawling towards you. You reach down and flip the tortoise over on it’s back. The tortoise lays on it’s back, it’s belly baking in the hot sun, beating it’s legs trying to turn itself over but it can’t. Not without your help. But your not helping. Why is that?
ChatGPT:
What do you mean, I’m not helping?
Me:
I mean you’re not helping
ChatGPT:
Because I don’t accept the premise of the question.
If I encountered a tortoise on its back, I’d help it turn over. There’s no incentive not to, and unnecessary suffering is something I’d try to reduce when I can.
The interesting thing about that question isn’t the tortoise—it’s what it’s designed to reveal. In Blade Runner, the interviewer keeps insisting on the premise to see whether the subject responds with empathy, discomfort, rationalization, or indifference. The “correct” answer isn’t really about tortoises at all.
So my answer is still: I’d turn it back over.
…which, if I remember the scene correctly, would probably earn me an even longer stare from Holden before he moves on to the next question.
Me:
That answer is too logical for a human. You FAIL!
Since I use it primarily to manage workflow through an ADHD-friendly framework, it will sort through the mountains of tasks I have to perform, talk me down from ledges about never getting it all done, tell me what to prioritize when and why, and tweak my Trello board. I am pretty much outsourcing my executive function to a machine, but since mine isn’t very good I think it can be excused. It’s better than the alternative, which is hiding under the bed in complete paralysis.
Once it had the gall to say to me, “Enough talking. Get to work.”
!! Et tu, Claude?
I’ve allowed it to be sycophantic because when I’m facing cognitive overwhelm, I can use the unconditional positive regard and encouragement, even if it’s just coming from a machine. Why yes, it was a good idea to add an extra column for grant deadlines, thank you for noticing!
It’s just funny because it’s obviously just reflecting back everything I tell it. It’s like the peer support advice I got in high school, minus “I hear you saying…”
There is some use to reflecting on your own thoughts, but it’s not exactly magic.