Bizarre reason ChatGPT finds it completely impossible to generate an image of a full glass of wine
Something as simple as asking ChatGPT to produce an image of a full glass of wine has spiraled into complex philosophical considerations in a YouTube video that is likely to make your head spin.
Artificial intelligence is by nature both incredibly complex and unbelievably simple. It's mechanical construction is sophisticated and it's constantly improving and advancing what it can do, but at the heart of it all is always the data that it's trained on and it broadly cannot 'think' outside of that.
That's why it often runs into roadblocks when it comes to generating completely unique constructions - especially within images - and it can be hilariously ran around in circles with tricky ethical dilemmas where it consistently contradicts itself.
One rather strange hurdle that ChatGPT seems to really struggle with though is the idea of a full wine glass, and YouTuber Alex O'Connor quickly realized that it's an issue that goes far deeper than you might expect.
Advert
Not only does ChatGPT struggle with a full glass of wine, despite multiple clear instructions, it also cannot comprehend anything other than a half-full or completely empty glass.
This is likely due to the nature of its training, as it can only produce images of whatever data it already knows, and the number of images online that are of anything other than a 'normally' filled wine glass are probably quite low.
Thus, even though it's able to create images of things that don't strictly 'exist' in real life, the simple act of imagining a full wine glass appears impossible because it doesn't know what that looks like.
What Alex quickly realized though is how this links to philosopher David Hume's theory of empiricism, which argues that ideas - which are imagined concepts and objects - exist purely through the pre-existing knowledge of an impression - something you have already experienced in real life.
For example, Alex posits the notion that you're able to imagine your hand if you close your eyes, but that's simply because you already know what your hand looks like from its real life impression.
On top of this, complex ideas allow us to imagine things that don't strictly exist through the hybrid of two or more pre-existing impressions, which is explained through the idea of a unicorn.
That's why ChatGPT is able to create a lot of the wacky images that you'd probably already seen, as it's able to pull from multiple established 'impressions' that exist within its data set and then combine them together.
A full wine glass, however, is seemingly an impossibly complex idea for artificial intelligence to comprehend, because it's simply unable to grasp the concept and physical properties of liquid in order to 'fill it up' further beyond what it already knows.
If you showed a human who had no prior knowledge of liquid a half-full glass of wine and asked them to imagine it full it's unclear how they would respond, and whether it would prove similar to that of ChatGPT, but it does perhaps outline the innate differences between us as a species and AI as a technology.
"If you think that a human could imagine the missing volume of wine in that specific set of circumstances, and that seems kind of plausible to me," Alex proposes, "then perhaps this is because humans can just naturally abstract concepts in a way that ChatGPT simply can't."
While ChatGPT and AI as a whole might be able to 'perfectly' interpret, adapt, and read the countless instances of data that it's been fed, it still can't comprehend the abstract and complex ideas that come naturally to a human - at least for now.
Who knows, maybe there will now be engineers hurriedly plugging in millions of full wine glasses into the tech to fill in this particular gap!