I Asked ChatGPT for a Full Glass of Wine. Here’s What It Created.
OpenAI unveiled an updated and significantly enhanced version of ChatGPT’s image creation tool on Tuesday, prompting widespread excitement across the web as people eagerly asked the AI to produce everything from South Park-style meme illustrations to depictions of Barbie dolls in the Oval Office.
However, one capability of ChatGPT’s updated GPT-40 image generation model has astonished even seasoned AI observers. a state of quiet, open-mouthed wonder .
Red wine, anyone?
See here, ChatGPT can now—with considerable reliability—produce an image of a glass brimming with red wine right up to the rim.
Prompt:
Ben Patterson/Foundry
It might seem easy enough, but even famous AI systems have struggled with the "glass half full" challenge. This includes platforms like ChatGPT and its predecessor, DALL-E, at least up until this point.
Here, for instance, is how Google’s Imagen 3 botched the test with the identical prompt:
Ben Patterson/Foundry
And Grok 3 isn’t much more successful:
Ben Patterson/Foundry
Microsoft’s Copilot also gave it a try:
Ben Patterson/Foundry
I even gave it a try using Flux, one of the most recent Stable Diffusion models, and ended up with this result:
Ben Patterson/Foundry
Whoops.
The "glass of wine" trick isn't a formal gauge of an AI's image-generation capabilities; rather, it serves as a more informal test, similar to inquiring of an AI model about the number of "r"s present in the word "strawberry." They often mess up, occasionally doing so in a ridiculously funny way.
Why is a completely full glass of wine such a challenge for image-generating AIs? The prevailing wisdom is that AI-powered models do best with images they’ve been trained on—and when it comes to pictures of red wine glasses, they’re typically filled about halfway, which is why a prompt for a “COMPLETELY full glass of wine, all the way to the brim” tends to get you a half-full glass.
Now, a really good An AI image generator should one Redditor helpfully explained ) be capable of "extrapolating" the concept of a fully filled wineglass even without such examples in its training data. Alternatively, perhaps someone at OpenAI provided the new model with numerous images of wineglasses brimming over with wine.
Sure, here’s another challenge for AI image generators: an analog clock set to a particular time. Surely, ChatGPT along with its updated image generator should handle this task effortlessly, correct? Let's find out!
Prompt:
Ben Patterson/Foundry
Next prompt:
Ben Patterson/Foundry
Um, paging Sam Altman?
Comments
Post a Comment