A New Year’s AI Treat
After a set of eye correction surguries, I’m currently inbetween tech projects, soon to recommence. Meantime – no doubt many of you will either be actively using AI in your jobs or like me tinkering with it to find out what’s useful and what’s not. I recently fell into the trap of thinking that ChatGPT was truly the dawn of AI because it proved helpful at solving simple problems in Home Assistant (some of the time).
Meanwhile: you may have noticed some videos out there pointing out some downfalls in the new chatGPT “image rendering”. For example “draw an image of a man writing with his left hand”. Well, it seems that in chatGPT’s world, there are no left-handed people – it seems utterly incapable of performing that task, despite including a helpful reply stating that it has indeed drawn a guy using his left hand. If Trump says it often enough – it must be true (laughs) and so it seems, if a high enough percentage of photos out there show people using their right hand to draw, then there must be no left-handed people except in my imagination, right?
Which made it particularly worrying when only this week I listened to a BBC news announcement while driving, in which the announcer proudly proclaimed that the UK government is about to fully embrace AI to increase government efficiency…
Other popular items on Youtube include requests for a images of a wall full of clocks, all showing the time at 11:05 – ChatGPT can’t do that either – it comes back with a bunch of clocks set to 10:10 pretty much every time.
And here’s my contribution for today…
Me: “Draw an image of a man sitting in front of a wall clock who’s hands show the time at 5:25”
ChatGPT: “Here is the image of a man sitting in front of a wall clock with its hands showing 5:25.”
Me: “Utter rubbish – the time is not even remotely correct. You are showing a clock which is set to 10 minutes past 10, not 5:25”
To which the response was:
ChatGPT: “I see the issue, and I sincerely apologize for the oversight. Let me correct this for you by ensuring the clock displays the correct time at 5:25. Let me try again.”
ChatGPT: “Here is the corrected image with the clock hands set precisely to 5:25. Thank you for pointing out the issue earlier.”
Yes, I have included the “corrected image” on the right. Don’t you just love the ABSOLUTE CONFIDENCE with which ChatGPT gets it wrong, time after time? Examples like this are DEAD EASY to find if you go looking. I suspect that many employees who’ve been told to embrace this stuff will be less-inclined to be inquisitive – I really do hope I’m wrong.
The post A New Year’s AI Treat appeared first on Scargill's Tech Blog.