Google CEO tells employees Gemini AI blunder 'unacceptable'
cnbc.com
cnbc.com
EDIT: had to follow a link and read halfway down to find:
> When asked for a “historically accurate depiction of a medieval British king,” the model generated another racially diverse set of images, including one of a woman ruler, screenshots show. Users reported similar outcomes when they asked for images of the U.S. founding fathers, an 18th-century king of France, a German couple in the 1800s and more. The model showed an image of Asian men in response to a query about Google’s own founders, users reported.
Didn't he clear out and/or move all the AI ethics folks?
https://www.wired.com/story/google-splits-up-responsible-inn...
Injecting "diverse" into any prompt requesting images of humans is an extremely low-effort attempt to create appropriate representation. It doesn't do the actual work. The actual work would be building a more comprehensive training set, among other things.
The worst part is it doesn't move the needle on the real problem which all the studies suggest would be solved (for real) with integration, which takes generations, but actually works.
I spent some time asking Gemini directly what it could tell me about how it was programmed, what internal data Google uses it to train it, and what policy was used to determine what data was acceptable/unacceptable. You clearly hit a wall where someone programmed a bunch of hard stops about what it could talk about.
My problem is that I actually find Gemini really useful! I can ask Google to pull up directions to the closest, best Thai restaurant, and unlike the agony that was "voice assistants" Google will actually understand my question and return exactly what I am asking for with confidence.
It's a mysterious black box that is clearly being tampered with to gaslight me, and yet, I cannot deny it's utility. It's like letting Rasputin do my laundry.
If you mean that neural networks are flawed things, not "perfect" intelligence like the Star Trek computer, then yeah, that's pretty obvious. Probably Star Trek assumed that perfect rule-based AIs, like IBM tried to build in the 1960s and 1970s were going to really work at some point, but keep having their "typical" fraud (such as getting totally stuck on a very regular basis for very dumb reasons)
I find it's common that we assume that a model actually has this information, even though I can't see a reason it would be in the training data to begin with. More likely perhaps, is that it just hallucinates and tells us what we want to hear or it tells us nothing at all.