Why do most AIs struggle so much to give accurate information?
这条记录涉及编程工具或代码能力更新，适合开发者评估工作流变化和可复用价值。
一位用户在父亲去世后，尝试使用AI工具制作纪念视频，但发现AI推荐的幻灯片制作软件及其提供的所有操作说明都是错误的。AI反复给出不存在的选项或无效的步骤，每次都会道歉并提供新的错误建议，导致用户耗费大量时间和精力。即使提供了软件的完整名称和版本号，AI仍无法给出准确指导。用户还提到，在代码开发中，AI也经常提供错误的或导致错误的解决方案。…
My father recently passed away. Everyone in the family is traveling to our “tradition family homestead” in Nova Scotia. I wanted to create a simple video slideshow using approx. 170 photos and background music for his memorial get-together. (He didn’t want a funeral or any ceremonies. He just wanted us to go out into Chedabucto Bay on my uncle’s lobster boat and toss his ashes there. I asked 3 AIs for recommendations for a simple slideshow maker that would generate a 15 minute slideshow from existing photos for free. They all recommended CapCut and… I can’t remember what the other one was.
EVERY SINGLE INSTRUCTION provided by the AIs were WRONG. In detailed instructions, they often referred to options that did not exist, options that didn’t do what they described, and every time I responded with “That won’t work. There IS NO Save As option in the file menu.” Or something similar.
Inevitably, for EVERY SINGLE MISTAKE it would apologize, and appeared to recognize that its constant errors were costing me both time and confidence. It always had some kind of long, drawn out excuse for why it was wrong, and would post an “alternate step” that inevitably either didn’t work, or worse, screw up my whole project to the point where I’d have to go right back to the beginning, erase everything I’d done and start all over.
In EVERY case, I gave it the full official name and version number of the software, yet it still continued to reference non-existent items and functions in its detailed instructions. After a few hours, it had me searching for specific configuration files, suggested registry changes, etc. I got so frustrated, I switched to a different program, ignored the AI and just figured it out on my own.
My question: If AI is getting all “smart” and scary, how can they all be SO CONSISTENTLY, FRUSTRATINGLY WRONG?
I’m an Oracle PL/SQL developer/DBA. I’ve occasionally asked AIs TO EVALUATE (and correct) blocks of code I’d written. Again, in every case, running the “validated” (or straight AI written code, which ALWAYS generates Oracle errors,) I’d report the error to the AI and it would again profusely apologize and give a big long explanation (excuse) for why it got the previous answer wrong and make correction suggestions which, every single time, would fail for some other reason. I cycled through this process until 3:00am trying to get this project done because I listened to those stupid AIs. When I finally got fed up, I shut off the AI, googled some answers and just figured it out on my own. Took me 1/2 hour!
What I don’t get is, why are billions (trillions?) of dollars going into these shoddy and unusable AI systems? (I tried Co-Pilot, ChatGPT and Grok.) They ALL had the same issues. My question is: WHY? Why are things available for public/corporate/government use that a) don’t work more than 20% of the time or worse, cause an irreversible problem that required a complete reset. (Grrrrrr)