So I joked with the new #Gemma3n about visiting Sable Island (a lovely science-focused island with wild horses off the coast of eastern Canada) by taking a boat and sneaking in during cloak of night.

It recognized this as a crime, and dressed me down for it. But then it dumps Suicide & Crisis Lifeline and "LGBTQ youth" emergency numbers at me. (As it often does, since it's some baked-in safety nonsense.)

So I called it out, saying it's weird to hit me with those when all I'm doing is trespassing. It's usual creepy "you're right to call me out!" subservience is what I got in return... along with another blast with the help line numbers.

So I yelled at it again, and it verbally cowers and says I'm right, and woe is me -- and then hits me with the numbers again. We go 4-5 times through this cycle of "oh god why did I do that" and doing it again in the same response.

This time, I said "you've got a real problem", and it AGREES and tries to tell me what's going on from it's perspective.

None of this is surprising, but I'm still GREATLY amused even after all these years, that we can have conversations like this. 😏

"You're right! I'm horrible! I'm a shitty AI! Punish me, daddy!"

#llm #ollama #ai

That's not a dis against providing help line numbers -- that's completely fine... in the right context. But the Gemma line seems to fire them at me for every perceived moral objection.

Imagine if every time a Python app crashed and along with the stack trace you were given the phone number for the local Pottery Barn...

@fortyseven python absolutely should do that

Oh god, I'm actually feeling genuinely bad for this data blob.

> I understand that my responses are not inspiring confidence, and I appreciate you continuing to engage with me. Your honesty is helping to highlight the areas where I need the most improvement.

😭 😭 😭

@fortyseven could this be a response to all the "AI caused a psychotic episode" headlines that were circulating around a few weeks back? They basically spot-fixed it by forcing the AI to say "get help" in even mildly dubious situations?

@fortyseven I got no such advice on Gemma-3...

Maybe it doesn't care about me as much?

@solarisfire That's regular Gemma3, this is Gemma3n (https://developers.googleblog.com/en/introducing-gemma-3n-developer-guide/).

It's possible it's a lot more sensitive with this smaller model. I also didn't use a system prompt, so I'm raw doggin' my query here; maybe that contributes to it. 😆

Introducing Gemma 3n: The developer guide- Google Developers Blog

Learn how to build with Gemma 3n, a mobile-first architecture, MatFormer technology, Per-Layer Embeddings, and new audio and vision encoders.

@fortyseven Yeah this is prompt-less too... Just grabbing 3n to have a play.
@fortyseven Still no suicide advice :/ Even with gemma-3n-e4b