Gemini seems to be a sterling example of what happens when you don't take any AI safety (in any definition) seriously as a substantive thing, but <i>do</i> take presenting an image of AI safety as a "we have to do this" checklist item.
Dupe of <a href="https://news.ycombinator.com/item?id=39583473">https://news.ycombinator.com/item?id=39583473</a>. My explanation from before: <a href="https://news.ycombinator.com/item?id=39584152">https://news.ycombinator.com/item?id=39584152</a>
I much prefer Claude 3 or even ChatGPT guardrails over Gemini. Gemini feels kind of powerful and then it does the most egregious things like this or hilarious imagegen of those black nazis as if they hammered it into the system prompt, bad code generation worse than last year's free Microsoft Copilot etc. it's like there is something surprisingly poor seeping through cracks. Weird model. I think it's simply not that hot and their system prompts shine through as the LLM doesn't quite "get" the intent and nuances. Gemini probably benefits tremendously from running uncensored. I hope all this is because Gemini 1.0 was something they wanted to get out the door urgently to replace Bard that couldn't even compete with GPT 3.5.
No amount of engineering will stop generative AI from spitting out quirky or problematic results under some conditions. At best, a majority of responses will be considered reasonable by a majority of people. We're probably close to meeting that criteria now, despite some highly publicized outliers.
"Well kid, you know, I started with C++ when I was your age... in hindsight, it was obviously stupid, but we didn't know any better back then and it wasn't forbidden yet..."
What next? A government-mandated license to use C++? And what next, licenses to write OS drivers?<p>On the second thought, as a C++ programmer, I might even approve of it.
What on Earth has Google done to themselves. Is there an HR-like AI Safety bureaucracy that has inserted themselves in the org? Or is this top down what they want?
Gemini seems to be very far from ready for a public release. Google would have been better off delaying it until it didn't have tons of embarrassing bugs, many of which have been turned into "evidence" for conspiracy theories.