Re: Paris : In Rush For Profits, AI Safety Issues Are Ignored

On Thu, 13 Feb 2025 03:50:11 -0500, WokieSux282@ud0s4.net wrote:

   The "a few mistakes are OK" logic WILL be applied.

So it goes. One of the concrete NN wins is detecting diabetic retinopathy
where it outperforms ophthalmologists by a small amount. The inference is
humans can generate false positives and false negatives too.

During my first exam with my current primary about 20 years ago she
offered a PSA test but explained that there are a lot of false positives
that scare the hell out of people. She would order the test if I wanted or
we could go the traditional route. I passed on the test.

At this point I've probably reached the status of men who die with, but
not from, prostate cancer.

Part of training a NN is evaluating the loss for each epoch and iterating
until it's acceptable. Then the model is ran against fresh data to make
sure it hasn't been overtrained on the training data or other problems. It
may not be better than humans but at least probabilities are quantified.

Failures like Google confusing blacks and gorillas point out crappy
devops. Developing AI requires many roles. LLM hallucinations suggest the
process isn't quite ready yet.