I saw on Reddit that chatGPT 5 will hallucinate before actually searching the web or opening documents, most likely as a cost saving procedure by OpenAI. The bubble is looking awfully shaky.
I did some training for way too much money for open AI as a pro contractor for my field this past year. And let me tell you, it’s a long ways off if even possible for even basic report writing much less something even a bit more complicated.
These things cannot be trusted with technical work, it would just find like things that “sound” like it makes sense, but then if you know anything at all about the work it is laughable
If you wouldn’t outsource technical work to a reddit forum you definitely shouldn’t be giving it to the robot
95% of the time, AI literally only works if all the people generating and receiving the work mostly do not care about what’s in it. The more important the details are, the worse the result is.
All they do is hallucinate, it’s just a coin flip on if it’s total nonsense or if it’s truth-shaped. The same process that makes it answer wrong is the one that makes it answer right
Yep, I work in a moderately neiche programming sector and it was truly awful when I tried to do the “co-programming” stuff. It got to a point where if give it a clear spec, and all is get back was “call the function that does what you asked for”
Slight improvement over telling you to call functions it just silently made up (my experience using it with something niche)
See, they’re learning, the hype is real! Any day now they will expertly clue you in to when they don’t know shit. After that, AGI can only be 12-18 months away!
Oh that’s what I meant when I said it told me to “call the function that does what I want”. It would just hallucinate that function, then I’d go write it, then it would hallucinate more stuff. And by the time I was done the whole program was nonsense.
Ended up being faster at getting stuff done by just fully dropping it. Sure I don’t have super auto complete, but who cares. Now my program is structured by me, and all the decisions were mine meaning I actually kinda understand how it works.
Yep, whole process was a pain. I can’t imagine having to lead a team where people are using AI assistants. That has to be a nightmare and I’d ban it instantly. It was hard enough parsing the hallucinations it had introduced from my prompts. Would be 1000x worse doing a code review where you have to find hallucinations introduced by other people’s prompts.
Only if you pay for premium or whatever. Then it takes extra time to “think” and you’ll get a more accurate answer. But “hallucinations” can only be reduced, never eliminated.
I saw on Reddit that chatGPT 5 will hallucinate before actually searching the web or opening documents, most likely as a cost saving procedure by OpenAI. The bubble is looking awfully shaky.
I did some training for way too much money for open AI as a pro contractor for my field this past year. And let me tell you, it’s a long ways off if even possible for even basic report writing much less something even a bit more complicated.
These things cannot be trusted with technical work, it would just find like things that “sound” like it makes sense, but then if you know anything at all about the work it is laughable
If you wouldn’t outsource technical work to a reddit forum you definitely shouldn’t be giving it to the robot
Anyone else remember the work of running Hillary Clinton’s email server was being outsourced to a reddit forum?
Best start believing in cyberpunk dystopias…
95% of the time, AI literally only works if all the people generating and receiving the work mostly do not care about what’s in it. The more important the details are, the worse the result is.
good things that
only care about the line and not the quality of work.
Did you train it on Dilbert cartoons?
I thought that version was supposed to reduce hallucinations?
Hopefully it’s reducing the hallucinations of future profits that investors have been clinging onto
All they do is hallucinate, it’s just a coin flip on if it’s total nonsense or if it’s truth-shaped. The same process that makes it answer wrong is the one that makes it answer right
Yep, I work in a moderately neiche programming sector and it was truly awful when I tried to do the “co-programming” stuff. It got to a point where if give it a clear spec, and all is get back was “call the function that does what you asked for”
Slight improvement over telling you to call functions it just silently made up (my experience using it with something niche)
See, they’re learning, the hype is real! Any day now they will expertly clue you in to when they don’t know shit. After that, AGI can only be 12-18 months away!
Oh that’s what I meant when I said it told me to “call the function that does what I want”. It would just hallucinate that function, then I’d go write it, then it would hallucinate more stuff. And by the time I was done the whole program was nonsense.
Ended up being faster at getting stuff done by just fully dropping it. Sure I don’t have super auto complete, but who cares. Now my program is structured by me, and all the decisions were mine meaning I actually kinda understand how it works.
Lol, oof, sounds like real “draw the rest of the owl” energy but adding an unhelpful “unfuck the owl I drew” step first.
Yep, whole process was a pain. I can’t imagine having to lead a team where people are using AI assistants. That has to be a nightmare and I’d ban it instantly. It was hard enough parsing the hallucinations it had introduced from my prompts. Would be 1000x worse doing a code review where you have to find hallucinations introduced by other people’s prompts.
Only if you pay for premium or whatever. Then it takes extra time to “think” and you’ll get a more accurate answer. But “hallucinations” can only be reduced, never eliminated.
It would need to be trained on the data to have a chance of not hallucinating, which would not be possible on new webpages or custom documents.
I too always trust marketing departments!