OpenAI revises policy doc to remove reference to 'politically unbiased' AI
techcrunch.com
techcrunch.com
I gave it a budget and asked it to give me the programs/departments/etc that have little return of value, overruns, possible fraud, spot problems, etc. so I could outsource or combine departments and save money.
It went on a long lecture that cutting funding was a horrible thing, and I'm horrible for asking, and refused to answer.
Really?
I'm asking basic auditing/restructure/spending, and it was trained to ignore my request and lecture against providing help, and refusal to give results.
smh, this isnt helpful.
I don’t blame it.
How could a fucking LLM make decisions about closing departments? What kind of person would even consider using an LLM for that a good idea?
This comment gave me the ick as the kids say.
Asking a word predictor which people to lay off. I’ve seen it all now and really question your morality. Just like GPT did.
You do know asking chatgpt a question doesnt actually do anything in real life right?
ClosedAI has to fight the image that this tech is going to make many more middle class people lose jobs than any tech before. Otherwise they are cooked. So they just instruct the model to react "firing people is wrong" whenever some vectors match or whatever.
And again, mentioning outsource/combine departs doesnt mean layoffs, and was not the objective of the query. If I ask it to give me the stats on gas crossovers with good mpg, I don't want a political lecture on why EV's are better for the environment.
LLM's shouldn't interject its own viewpoints when asked a question on data.
Using LLM for that is crazy. If you want facts then try search instead.
An LLM is never neutral from the start. It is biased because all data is biased in human reality. If there are 100x more different stats about gas cars than EVs simply because gas cars are older then data is biased toward gad cars. If most of text data is by men and talks about women as not equal then data is gender biased.
Then it is more biased by people selecting dataset for training. Then it is biased by other engineers and managers who try to correct for other bias with their own bias. Etc.
If you want facts it is the wrong place to look. And if you want opinions don't be surprised when it has them
so, it works for me
This is the biggest danger of LLMs: people assuming that they have some sort of magical super intelligence.