News for hackers. Stuff that matters.
Re:ChatGPT is biased against resumes with credentials that i... by skeledrew on Saturday June 22, 2024 @11:42PM (#40764607)
For the life of me, I don't understand how this almost always misses people: that AI only has data from humanity to learn from, and so every result/action it provides/takes reflects the state of humanity. Even "hallucinations" in some way are likely triggered by content that is for example broken by web sources interspersing unrelated bits such as ads. Or maybe it's a convenient ignorance.
Re:ChatGPT is biased against resumes with credentials that i... by tbrownaw on Sunday June 23, 2024 @12:12AM (#40764729)
> Even "hallucinations" in some way are likely triggered by content that is for example broken by web sources interspersing unrelated bits such as ads. Or maybe it's a convenient ignorance.

They could be something like a compression artifact?

Re:ChatGPT is biased against resumes with credentials that i... by skeledrew on Sunday June 23, 2024 @01:14AM (#40764981)
Maybe, but I think any obvious software-related side-effects would be accounted for.
Re:ChatGPT is biased against resumes with credentials that i... by aprilthird2021 on Sunday June 23, 2024 @12:39AM (#40764842)
I don't think people are stupid or ignorant. We control the data we train LLMs on. We can, knowing what we know about human biases, introduce and generate data that can contradict these. But we can only do that if we know the biases the LLMs replicate and the contexts where they do.
Re:ChatGPT is biased against resumes with credentials that i... by skeledrew on Sunday June 23, 2024 @01:11AM (#40764968)
Actually there isn't much "control", beyond the awareness that X model is trained on a dataset scraped from Y, and basic cleaning/sanitizing. There's so much data in use that it'd take decades for a human team to curate or generate in a way that meaningfully balances the datasets. And so models are also used in curation and generation, which themselves are blackboxes...
Re:ChatGPT is biased against resumes with credentials that i... by aprilthird2021 on Sunday June 23, 2024 @11:43AM (#40768269)
There is tons of control and research done about the ways to make LLMs "safe".
Re:ChatGPT is biased against resumes with credentials that i... by skeledrew on Monday June 24, 2024 @03:47PM (#40779878)
Emphasis on "research". There is no silver bullet available.
Re:ChatGPT is biased against resumes with credentials that i... by 8bitsrule on Sunday June 23, 2024 @02:06AM (#40765124)
>AI only has data from humanity to learn from, and so every result/action it provides/takes reflects the state of humanity.

Same problem that children have always had.

Re:ChatGPT is biased against resumes with credentials that i... by throwaway7ahgb on Sunday June 23, 2024 @09:52AM (#40767479)
Agreed, however with children we don't have full understanding of nature vs nurture (Will we ever?)
Re:ChatGPT is biased against resumes with credentials that i... by skeledrew on Sunday June 23, 2024 @11:05AM (#40767990)
Children, who on average get a minimum 15 years of education and guidance before they're entrusted with anything serious. And yet we expect perfection from budding AI upon or a few weeks after its release. Crazy.