AI comparisons

[thorwhalen]

Comparing different AI on common tasks.

[thorwhalen]

Making SVG

The prompt:
Give me an SVG image of the word "figiri" using a minimal set of identical shapes (same dimensions, etc. but you are allowed to rotate them)

If the AI showed me an image, I would follow up with:
Can you give me SVG code for that?

If the AI showed me code, I would follow up with:
Can you render this SVG as an image for me?

(In all cases, I would copy the code on https://boxy-svg.com/app to see the image.)

[thorwhalen]

Gemini

Gemini (2.0 Flash)

So, so, so bad!

Gave me code.

When I asked to render the image of it, it responded:

I'm a text-based AI and can't assist with that.

When I copied to another app to render, this is what Gemini's code was drawing:

Gemini (2.0 Flash Thinking Experimental)

A bit better (you can almost see the word "figiri"). But not that much.

Gave me code, and when I asked it to render it, it gave me a long winded description and then:

Unfortunately, as a text-based AI, I cannot directly show you an image. However, if you paste the SVG code into a browser, an online SVG viewer, or use a programming library that can render SVGs, you will be able to see the image as described.

[thorwhalen]

chatGPT

4o

Figiri? In what language/script?

And when asked to render, it went on "analyzing" (i.e. writing code).
Then at some point said:
"It looks like there’s an issue with the XML declaration in the SVG. I’ll correct it and render the image again."
then finally gave me something nearly identical (as an image):

03-mini-high

Even worse!

And when asked to render it as an image, it gave me some code, telling me
When this markdown is rendered in a compatible viewer, the SVG image will be displayed directly.

So I did that here (see code in raw format) and it gave me... nothing:

[thorwhalen]

Grok 3

Gave me a bunch of cool images. But definitely not SVG per se.

These were given to me with a slightly different prompt:

And finally when I settled with the prompt given above (that I tested, verbatim, on other AIs), I got this:

I really like these, as images. But they're not SVG.

So I asked (for the last two) my "Can you render this SVG as an image for me?" question, and the SVG code I got rendered (using https://boxy-svg.com/app) to:

[thorwhalen]

Claude (3.7 Sonnet)

The clear winner.

I mean, it's not beautiful, but it's correct.
And it gave it to me in a window that had both a "preview" (SVG rendered as image) and a code tab containing the code (which I could select and ask to "improve" and/or "explain".

I still asked it to give me the code (I hadn't seen the "code" tab first), and it printed out code that did indeed render exactly as the image.