It's been a grey and rainy week, and as you may already have gathered from yesterday's post, I've not been in my most inspired mood.
Before I asked Bing AI for illustrations of "tired and uninspired" yesterday (the results of which actually cheered me up!), I had made some attempts to explore another AI image generator: Craiyon. It did not impress me, though. I only downloaded a very few examples:
Today, it occurred to me to ask Microsoft's new Copilot the question from my post yesterday: "Why can't AI spell?" In my first attempt I'm afraid I phrased the question a bit too personal...
Me: Why can't Bing Image Creator spell?
Copilot: I apologize if there was any inconvenience caused. As an AI language model, I don’t have direct control over the Bing Image Creator, but I appreciate your feedback. If you have any other questions or need assistance, feel free to ask! 😊
Oops, in spite of the smiley, someone seems to have taken offence... Better rephrase!
Me: Why can't AI image creators manage spelling?
Copilot: AI image creators, like any other technology, have their limitations. While they excel in certain areas, they may struggle with others. Here are a few reasons why spelling errors might occur in AI-generated images:
Training Data: AI models learn from vast amounts of data, including text. However, if the training data contains spelling mistakes or inconsistencies, the AI might inadvertently reproduce those errors.
Context and Ambiguity: Spelling correctness often depends on context. AI image creators lack the understanding that humans have, which can lead to misinterpretations. For example, the word “lead” can refer to both a metal and a verb, leading to different spellings (“lead” vs. “led”).
Noise and Noise Tolerance: AI models are sensitive to noise (irrelevant information). If an image contains noise (such as artifacts or distortions), it might affect the generated text, including spelling.
Complexity of Language: Spelling involves intricate rules, exceptions, and regional variations. AI models may not always capture these nuances accurately.
Trade-offs: AI creators balance various factors, such as fluency, coherence, and relevance. Sometimes, spelling correctness takes a back seat to other priorities.
Feedback Loop: AI models improve over time with feedback. If users report spelling errors, developers can fine-tune the model to enhance accuracy.
In summary, while AI image creators have made significant strides, they are still evolving. As users, we can provide feedback to help improve their performance. 🤖📝
Excuses, excuses... (LOL)
14 comments:
ok, none of these answers apply to what you and i get from AI.. no noise, nothing to interfere and both you and I get one out of four spelled correctly, except for one letter.. good idea to ask, you KNOW i have to go visit AI crayon.. I mean craiyon.. you know you will know when I do
Goodness, they have covered themselves well! My favorite is the last one.
I dunno, I'm a bit dubious about them, mind you I've never used one so I don't know much about them.
I've been trying to follow your adventures with AI but so far haven't ventured to far and certainly haven't investigated myself. I shall keep on watching your bulletins with interest.
Phew - I bet you wished you'd never asked!
Sandra, I agree - none of the answers above really explain why AI can't just copy words already provided in our request for an image, without messing them up.
Ginny, it seems this one is no good with faces but prefers blurry silhouettes.
Amy, there's certainly a lot to be dubious about with AI images, even if one can also have some fun with them.
Graham, my AI experiments so far have been rather random, and really mostly inspired by very little to inspire outdoors photography this time of year!
Carol, you're not far wrong there! ;-)
I am too tired and lazy right now, but I do know a little about how AI "learns" (it doesn't, not in the sense humans do) and why it is (often) prone to errors without "knowing" (AI does not "know" anything the way humans do) that it has just made a mistake.
In short, it's mostly down to the model at the base of each AI - how much of the real world enters into the model? And of that, how accurate, and how broad is the model's data base? How biased? The developers/programmers have to decide about the model, knowing full well that there is no such thing as a full, accurate representation of the entire real world; modelling always has certain pre-determined limits.
Meike, I'm afraid the model at the base of my own brain does not seem to have been sufficiently updated to be able to really grasp your explanation! - But thanks for trying! (LOL)...
I think for us "layman" users - just playing around for fun - we're just trying to learn by trial and error what works, and what doesn't...
I might add that I'm quite used to talking to inanimate objects (especially when they don't do what I want!), but it's only recently that they've started to talk back! ;-)
As I said, I was too tired and lazy to explain properly... sorry!
Well, when Barbie and Skipper start to talk back (or among each other), or your family of Christmas gnomes, that's when I'd start to worry... just a little bit :-D
(Now I wonder what their voices sound like. And of course Barbie and Skipper speak Swedish, even though they are originally Americans.)
Meike, even when sleepless in the night I have not yet caught the dolls or gnomes talking... But I kind of assume Barbie and Skipper are probably just about as bilingual as myself! ;-)
Post a Comment