Yesterday in The Whittling Pup, I described how Copilot came to generate a picture (four pictures, actually) for me of a puppy whittling a wooden dog:

Copilot completely misunderstood my request. I was asking it to show me pictures that other people post of their completed Whittle Pup wood carvings. I don’t know enough about how Copilot’s LLM works and what restrictions it operates under, but it’s clear that it interpreted my query (“Can you show me some pictures of the Whittle Pup?”) to mean, “Show me a picture of a puppy whittling a wooden dog.”

Not an unreasonable interpretation, really. Honestly, of all the responses I’ve received from AI chatbots, this is by far my favorite. Not just because they’re so stinking cute, but also because they’re well composed. The AI (DALL-E 3 in this case) got the essentials: a puppy at a workbench, using a knife to carve a wooden figure. Those pictures are, unintentionally, brilliantly amusing.

So I thought I’d see what the other AIs could do with the prompt: “show me a picture of a puppy using a knife to whittle a wooden dog.”

Google Gemini

I especially like the top-right picture. The bottom-left is great, too: pup there looks like a serious craftsman paying close attention to his work.

Meta AI

Meta AI generated only one picture in response to the prompt.

The picture certainly is cute, and amusing. Particularly the placement of the knife. But again, the AI did a good job of getting all the necessary pieces into the picture. We can quibble about how those are arranged, but why?

ChatGPT

ChatGPT, too, generates only one picture at a time. It also limits me to two generated pictures per day, unless I pay to upgrade. That’s okay, the first one it generated is great:

I like this one a lot. I have a minor quibble about the knife, which looks more like a paring knife from the kitchen rather than a carving knife from the shop. And the puppy’s belly, just just under the knife, is oddly pink: more so than I recall ever seeing on a real puppy. But, really, those are my only criticisms. It’s a wonderful picture.

This is impressive technology that’s freely available to anybody with a computer. And the technology is in its infancy. What I find impressive is that the AI not only gives me what I asked for (a puppy whittling a dog figure), but adds appropriate touches like the woodworking “bench,” woodworking tools in the background, and wood shavings. With few exceptions, the knife is being held appropriately, assuming a puppy could hold a knife. Of all the images generated here, there’s only one that I’d call “wrong”: the bottom-right image from Google Gemini, where it looks like the puppy is about to chew on the knife blade.

As this technology improves, I think we’re going to see a new skillset emerge, and become coveted: the ability to create AI prompts that produce the desired result. This is the realm of writers, speakers, and poets: people who can, well, paint pictures with words. It’s an odd mixture of technical and creative writing. The quality of the AI output depends very much on the quality of the query that it’s given. I need to explore that in more detail.