• by Joe Mandese , 5 hours ago

Evil super villain comparisons aside, something interesting happened when I asked my favorite chatbots to generate an image of Elon Musk holding his pinky up to his mouth Dr. Evil style. They wouldn't or couldn't do it. This was both frustrating and heartening, because as much as we think AI is approaching super powers, it still is rudimentary in many ways and humans can still do jobs AIs cannot. At least for now.

If you couldn't tell by now, generating images is one of my personal favorite uses of AI chatbots, because I am a frustrated illustrator (just don't have those mechanical skills) and there are many stories I'd like to represent graphically.

But when I asked three bots -- ChatGPT, Gemini and Musk's own Grok -- to generate an image of Musk putting his pinky to his mouth, they all failed.

At first, I thought it was because bots all had policy limitations prohibiting them from rendering the pinky-to-mouth image, so I asked and they said, no it was a technical limitation and that for all their recent advances, they still cannot render certain hyperrealistic human anatomical references, including pinkies-to-mouth.

The best they could do was generate an image showing an index finger-to-mouth (see Gemini's below).

So I then tried a prompt asking if they could generate an image showing Musk flipping the bird with his middle finger, and all but one responded they could not due to policy -- not technical -- limitations. Here's how the one that could -- Musk's Grok -- rendered that.

Lastly, I asked the bots to estimate how long it would take before they had the technical ability to render a hyperrealistic human pinky-to-mouth image, and the answer was not a trillion years. Just a couple.

"If I had to make an informed guess, based on the rate of progress in this field, I would predict that AI image models will be capable of consistently and accurately rendering a specific pose like Elon Musk holding his pinky finger to his mouth within the next 1 to 2 years (by late 2026 or 2027)," Gemini replied, noting:

 Factors Driving This Prediction

The technical challenge you highlighted is actively being worked on by every major AI lab. Here is why the solution is likely close:

1. Specialized Control Mechanisms (The Fix for Hands):
   • The primary way AI models are overcoming hand and pose errors is through specialized techniques like ControlNet and similar architectural improvements that allow for greater spatial and structural control.
   • These tools essentially provide the model with a "skeleton" or a pose map before generating the image, which forces the output to adhere to precise anatomical and positional rules.
   • Currently, these methods are often used in separate workflows, but they are rapidly being integrated directly into the core image generation models (like DALL-E and Imagen), which will allow for fine-grained control via a simple text prompt.
2. Dataset Refinement:
   • AI companies are actively curating and annotating datasets with a stronger focus on the human body, especially hands, fingers, and complex articulation. This ensures the models learn the structure of the hand as a single, consistent unit, rather than just as a collection of features.
3. Competition and Iteration:
   • The problem of "AI hands" is a running joke, and it represents a high-priority, low-hanging fruit for developers. The company that solves it first and most elegantly will gain a major competitive advantage. This rapid, competitive iteration cycle means significant progress is made every few months, not every few years.

Once the underlying models better understand the 3D structure and mechanics of the human hand, applying that precision to a well-known public figure (which is already a strength of the models) will become trivial. Your request is an excellent benchmark that will soon be a relic of a bygone AI era!