"I just found out that it's been hallucinating numbers this entire time."
-
"... just inventing plausible sounding [answers]"
This shit is so tiring - that is literally all any AI is even *meant* to do. They are not even designed to give correct answers to questions, but just examples of what a plausible answer could sound like.
(edit: sorry, I know I'm likely preaching to the choir here, but it's just so fucking tiring seeing people surprised by this crap.)
@jmcclure Keep preaching
@Natasha_Jay -
@ZenHeathen @Natasha_Jay the thing is, there is no change-over "period" once you use this daily, your org begins lose its institutional memory
@Bredroll @ZenHeathen @Natasha_Jay And there will be further, probably unnoticed, changeover periods, when the technology less or more silently provider changes (or even the org changes provider, …) their services, e.g., the used model is changed.
-
So I asked the higher ups to double check this $11M investment they made to run the show in our department. Everything promised was made worse with zero improvements. We lost a lot of experience and money. Almost lost me too, but I'm a fixer and they are going to pay me a lot of overtime...
-
@Natasha_Jay so much karma
-
@Natasha_Jay This part in the original post is fantastic:
“The worst part I raised concerns about needing validation in November and got told I was slowing down innovation.”
hxxps://www.reddit.com/r/analytics/comments/1r4dsq2/we_just_found_out_our_ai_has_been_making_up/
@drahardja
Oh yes, that’s a beauty!! Just shaking my head in disbelief here…
️
“The worst part I raised concerns about needing validation in November and got told I was slowing down innovation.” -
I can't wait to deploy AI on the battlefield! And for policing.
-
What would be amusing is them having greater success using bullshit data than whoever was previously correlating stuff :^D
-
@Natasha_Jay .. and you really want to blame the technology for this... If there was no process for checking the facts that were used , it is simply bad implementation, everyone knows you have to check for or do something against hallucinations with GenAI.
@ErikJonker then where's the productivity gain promised with AI if you still have to do the work to get the numbers you trust? Why take on the additional cost at that point?
If you had an employee who was constantly lying to you, you'd fire them.
-
"... just inventing plausible sounding [answers]"
This shit is so tiring - that is literally all any AI is even *meant* to do. They are not even designed to give correct answers to questions, but just examples of what a plausible answer could sound like.
(edit: sorry, I know I'm likely preaching to the choir here, but it's just so fucking tiring seeing people surprised by this crap.)
@jmcclure so many people don't understand this.
-
@ErikJonker then where's the productivity gain promised with AI if you still have to do the work to get the numbers you trust? Why take on the additional cost at that point?
If you had an employee who was constantly lying to you, you'd fire them.
@GreatBigTable also true, you have to make a rational business decision, as with every other technology
-
@Natasha_Jay this can't be real, right? Everyone knows AI isn't ready for any mission critical stuff right now (probably years away)
-
@Natasha_Jay "Trust but verify" - Old KGB proverb
-
@Natasha_Jay I suspect that one of the few applications where a heavily hallucinating LLM can outperform a human would be in replacing board members, C-suite executives, and their direct reports. I propose a longitudinal study with a control group of high level executives using real data, and experimental group using hallucinated, or maybe even totally random data filtered into plausible ranges, using executive compensation deltas as the metric.
@0x0ddc0ffee @Natasha_Jay I would LOVE to see the resultant shareholder suit over an AI board. There's a long list of duties they are required to perform. Seeing, for example, how the courts would deal with the duty of care or the business judgement rule. OK, the business judgement rule may be handled better by the AI, but... a shareholder would still have them on the duty of care.
-
@drahardja @Natasha_Jay I believe it’s supposed to be h*tt*ps
@indigoviolet I change it to hxxps when I want to link to a site that not everyone wants to visit.
-
@Natasha_Jay what seemed crazy to me reading the comments on the reddit post is that so many answers were like "no you can't do it that way, you need to <bunch of extra, convoluted and unnecessary work> to use that tool properly in that context" and to me it's like ok so.....you have to jump through all those hoops to justify using a tool that doesn't actually make anything easier for you.....why? To look like you're following innovation? God I'm glad I no longer work anywhere remotely related to tech right now.
-
@lxskllr more likely it will have similar results to that of the media industry who changed their entire industry based on fake data provided to them by Facebook.
@lxskllr @GreatBigTable meh, they've probably been fabricating data for the board long before generative AI hit the scene. The only difference is that now they have a scape goat.
-
@indigoviolet I change it to hxxps when I want to link to a site that not everyone wants to visit.
@drahardja@sfba.social @indigoviolet@tech.lgbt ...why? people can just... not click the link?
-
Someone who knows the systems MUST Proofread ¨everything" that your ai spits out at you.
-
@Natasha_Jay if you’re going along with the personification of a technological system, then I just don’t know what to tell that guy.
-
@Natasha_Jay How does anyone think LLMs base anything on facts or data? They are plausabiliy machines, designed to flood the zone.
