Because the longer you go without using credit, the less the system knows about how you will act if given credit. The scores degrade over time because of a lack of data. And that is good predictive data science.

The researchers who wrote the article do not have that vested interest. They were legit trying to test if the loss of jobs is real. Your response assumes they did not do their job. That is some serious paranoia/conspiracy theorizing.

The whole problem is Germany has shrinking population. There are fewer young folks to hire than retirees.

Looking at the paper, I think the authors theorize that if they ran the test longer, the percent would keep rising. It was not the same people failing repeatedly. It was different people in each month, with some but not all overlap. It appears that maintaining vigilance is hard enough, if attacks just keep coming, most humans will eventually let one through. The summary appears to be correct, it just did not cite enough of the paper to explain the claim.

> other than maybe good old fashioned in-person networking.

It's the only exit that I found. And it doesn't work for most new grads... not enough of a network.

Doesn't really matter to the story -- a couple years ago, the C students were getting great starting salaries and unemployment among CS grads was among the lowest of any degree. It didn't matter if you were rockstar or not -- as long as you had the basics, you were employable. The story's point is how dramatically the market has shrunk such that employers are far more choosy. And this shift is affecting rockstars as well as the competition gets more intense.

The OP is objecting to the loss of the Internet Archive and the ability to review history because of the AI scanning.

The open question is whether generating content from the training is -- in any form -- redistribution. The contention of the AI companies is that it is sufficiently transformational. They argue that the fact that the LLM *could* generate images that are copyright infringing doesn't invalidate all the uses that are not infringing, and it is on the person doing the generation and distributing the result to check if the generated images violate someone's copyright. No court to the best of my knowledge has yet ruled on that.

It is hard to test things that are suspected to cause cancer because you cannot go around giving people cancer, so studies are often cut off before they become conclusive.
A small sample size is not the same as a statistically invalid size. A bad study is one that violates double blind or leaves out data or otherwise does not follow good science. A small sample size provides useful data to contribute to an overall picture. In aggregate with other studies, it adds to the picture. California adds the warning when there is evidence of a problem under some circumstances even if we do not yet know the full scope of those circumstances. It errors toward safety, which is excellent public policy.

I have wondered if we should be building a âoethoroughly debugged codeâ repository, where lots of programmers vet the code, and then see what happens with an LLM trained only on that.

Not in my experience with California. It generally means âoegood study but small sample sizeâ or âoemultiple good studies with mixed resultsâ.

Parents need to create a Google account (and probably the other FAANGs) on the day a kid is born to make sure they age into the Internet.

Not everyone. Just the teens who legit need nannies.

Calif does have data to back all of the claims. When the data is not definitive, then whether the strength of that data is sufficient to make you not use the product is left for you to decide. That's why the products are not banned. It is informed freedom.

YMMV. May have something to do with the topics I'm asking about -- they're mostly within domains having to do with technical specs where there's lots of highly replicated detail across the web, rarely about things that are niche topics, so I suspect I'm usually within domains where the LLMs get heavy training. I haven't really tried to look for a pattern in when the answers are wrong. I'll see if I notice anything about the 10%... maybe those questions are more in line with what you tend to ask about?