I tried AI again the other day, just for a yearly check in. Gemini’s Pro model was more apologetic than ever when it changed the basis of its narrative SIX TIMES in seven prompts.
I was simply asking it to describe the counterweight mechanism in a piece of common machinery. It changed from industrial gas struts, to no support and being too heavy to handle, to two large springs (correct), back to no support but now because it’s feather light, and so on. I simply kept asking it to stop fabricating answers and to verify from a parts diagram, the URL for which I provided. Its apologies and explanations grew longer and more elaborate each time.
We had good laughs, but sorry about the power and water usage.
Yeah I have all the LLMs through work, so every now and then I’m like wtf let’s give it a shot, I need to do a super fuzzy search over a large semantic area and why what do I have here…
Number one, it saps any little bit of joy I might have taken out of work, and two, “speaking” to it makes me feel insane. Like I’m talking to the world’s most polite gaslighter. And it’s so confidently wrong all the time.
I’m fine just loading up my squishy human brain with context and then executing, but increasingly its getting harder to parse online docs (including internal!), articles, and other resources, because they’ve been slopped up by a process that doesn’t actually understand anything it’s generating.
had a similar thing happen with something pretty simple I had to contact my bank about. After about 12 tries of it not understanding I finally got a human who got it in the first reply.
I tried AI again the other day, just for a yearly check in. Gemini’s Pro model was more apologetic than ever when it changed the basis of its narrative SIX TIMES in seven prompts.
I was simply asking it to describe the counterweight mechanism in a piece of common machinery. It changed from industrial gas struts, to no support and being too heavy to handle, to two large springs (correct), back to no support but now because it’s feather light, and so on. I simply kept asking it to stop fabricating answers and to verify from a parts diagram, the URL for which I provided. Its apologies and explanations grew longer and more elaborate each time.
We had good laughs, but sorry about the power and water usage.
Yeah I have all the LLMs through work, so every now and then I’m like wtf let’s give it a shot, I need to do a super fuzzy search over a large semantic area and why what do I have here…
Number one, it saps any little bit of joy I might have taken out of work, and two, “speaking” to it makes me feel insane. Like I’m talking to the world’s most polite gaslighter. And it’s so confidently wrong all the time.
I’m fine just loading up my squishy human brain with context and then executing, but increasingly its getting harder to parse online docs (including internal!), articles, and other resources, because they’ve been slopped up by a process that doesn’t actually understand anything it’s generating.
had a similar thing happen with something pretty simple I had to contact my bank about. After about 12 tries of it not understanding I finally got a human who got it in the first reply.