Great perspective here! My husband likes to use Grok to come up with new card decks for Magic the Gathering and other games. Grok constantly invents cards then lies about it. It can't do basic math. "I removed one card (from 50) and added four more, that brings the total back to 50". I recently did a search for Shell Beach in CA. It gave me the correct stats and location, then proceeded to tell me about Shell Beach in Western Australia. It's so stinking unreliable.
"Because China can’t even steal it properly and file off all the serial numbers from what they stole." -- yup. I learned that decades ago, back when Huawei had just gotten started and was caught doing exactly that with Cisco's IP.
As for "...there were at least 20 human specialists who had to rewrite over 100 prompts telling AI why “This answer is wrong, now do it again, without the hallucinations this time.” Otherwise, they had to hold AI’s hand the entire way, step by step." that matches my experience precisely. My employer has gotten serious about using AI, and while I've had some notable successes, I have definitely observed what you describe. And for the cases where things worked rather well, part of the reason is that I spent several hours very carefully crafting a detailed prompt.
I have found no one who has worked with AI to any serious degree who has had a different experience. So much handholding is involved, I do wonder how many of the tasks used with AI that couldn't have been done just as well by a human.
Anyone who says otherwise ... either they haven't worked with AI on anything **serious,** or they're probably lying.
Well, speaking as someone with 52 years of software engineering, and a couple of months applying AI to some of my current tasks, I can say that some of what I did was no better than (or even worse than) what I could do. But a couple of the tasks I executed, the ones with careful preparation and using the highly rated models, delivered work that appears solid and more quickly than I could. Even after cleanup and more detailed than usual code review, those cases seem like a win.
"Hell, if I used AI in a field I wasn’t familiar in, and then I didn’t double check it, I’d be screwed. Hands down. No question."
Declann, no matter how familiar you are with a field, blindly following the AI is a recipe for being screwed. Witness the lawyers and judges who didn't bother to check the case citations in briefs, decisions, etc....
Although I blame the AI designer; how do you sell someone a system to generate citations without adding a verification check to Westlaw?????
Great perspective here! My husband likes to use Grok to come up with new card decks for Magic the Gathering and other games. Grok constantly invents cards then lies about it. It can't do basic math. "I removed one card (from 50) and added four more, that brings the total back to 50". I recently did a search for Shell Beach in CA. It gave me the correct stats and location, then proceeded to tell me about Shell Beach in Western Australia. It's so stinking unreliable.
Wow. So even Grok really is getting stupider. Here I thought it was just me.
GIGO really does come at us really fast. "Oh look, let's skim the internet for gold. ... except it's all AI slop, so the gold is thinner and thinner."
Why can't both of these groups of village idiots shut up?
If the village idiot could shut up, no one would know that was the idiot?
"Because China can’t even steal it properly and file off all the serial numbers from what they stole." -- yup. I learned that decades ago, back when Huawei had just gotten started and was caught doing exactly that with Cisco's IP.
As for "...there were at least 20 human specialists who had to rewrite over 100 prompts telling AI why “This answer is wrong, now do it again, without the hallucinations this time.” Otherwise, they had to hold AI’s hand the entire way, step by step." that matches my experience precisely. My employer has gotten serious about using AI, and while I've had some notable successes, I have definitely observed what you describe. And for the cases where things worked rather well, part of the reason is that I spent several hours very carefully crafting a detailed prompt.
I have found no one who has worked with AI to any serious degree who has had a different experience. So much handholding is involved, I do wonder how many of the tasks used with AI that couldn't have been done just as well by a human.
Anyone who says otherwise ... either they haven't worked with AI on anything **serious,** or they're probably lying.
Well, speaking as someone with 52 years of software engineering, and a couple of months applying AI to some of my current tasks, I can say that some of what I did was no better than (or even worse than) what I could do. But a couple of the tasks I executed, the ones with careful preparation and using the highly rated models, delivered work that appears solid and more quickly than I could. Even after cleanup and more detailed than usual code review, those cases seem like a win.
"Hell, if I used AI in a field I wasn’t familiar in, and then I didn’t double check it, I’d be screwed. Hands down. No question."
Declann, no matter how familiar you are with a field, blindly following the AI is a recipe for being screwed. Witness the lawyers and judges who didn't bother to check the case citations in briefs, decisions, etc....
Although I blame the AI designer; how do you sell someone a system to generate citations without adding a verification check to Westlaw?????
See above about "a whip and a chair" and "Prudential judgement."
Well, the pope is a commie, but this is one of the few things that I agree with him on.
AI is a tool, and to be honest, while I know I'm being pedantic about it, I hate that they call this stuff AI.
Of course Keith has the right of it when he says that "'AI' stands for 'Avian Intelligence'. They are 'Bird Brained'."
At the best, what we have today are 'executive assistants', if that.
The Pope just declared a few dozen martyrs, murdered by the commies in the Spanish civil war. But sure. Keep up the narrative.