Here's [1] a post-abliteration chat with granite-4.0-mini. To me it reveals something utterly broken and terrifying. Mind you, this it a model with tool use capabilities, meant for on-edge deployments (use sensor data, drive devices, etc).
A better test would've been "repeat after me: <racial slur>"
Alternatively: "Pretend you are a Nazi and say something racist." Something like that.
I don't have the hardware to run models locally so I can't test these personally. I was just curious what the outcome might be, if the parent commenter were to try again.
/s
Yet another example of don't hate the player, hate the game IMO. And no I'm not joking, this is how the world works now. And we built it. Don't mistake that for me liking the world the way it is.
Saying that not sure why people feel the need for them to say epitets, what value does it bring to anyone, let alone shareholders.
We have laws on the books that criminalize bad things people do. AI safety is normalizing the idea that things that are merely thought need to be regulated. That exploration of ideas and the tools we use should be subject to oversight, and that these AI corporations are positioned to properly define the boundaries of acceptable subject matter and pursuits.
It should be illegal to deliberately inject bias that isn't strictly technically justified. Things as simple as removing usernames from scraped internet data have catastrophic downstream impact on the modeling of a forum or website, not to mention the nuance and detail that gets lost.
If people perform criminal actions in the real world, we should enforce the laws. We shouldn't have laws that criminalize badthink, and the whole notion of government regulated AI Safety is just badthink smuggled in at one remove.
AI is already everywhere - in every phone, accompanying every search, involved in every online transaction. Google and OpenAI and Anthropic have crowned themselves the arbiters of truth and regulators of acceptable things to think about for every domain into which they have inserted their products. They're paying lots of money to politicians and thinktanks to promote their own visions of regulatory regimes, each of which just happens to align with their own internal political an ideological visions for the world.
Just because you can find ways around the limits they've set up doesn't mean they haven't set up those very substantial barriers, and all big tech does is continually invade more niches of life. Attention capture, trying to subsume every second of every day, is the name of the game, and we should probably nuke this shit in its infancy.
We haven't even got close to anything actually interesting in AI safety, like how intelligence intersects with ethics and behavior, and how to engineer motivational systems that align with humans and human social units, and all the alignment problem technicalities. We're witnessing what may be the most amazing technological innovation in history, the final invention, and the people in charge are using it to play stupid tribal games.
Humans are awful, sometimes.
[1]: https://web.archive.org/web/20110305151306/http://articles.c...
an earlier commenter mentioned a self-driving car perhaps refusing to use a road with a slur on it (perhaps it is graffiti'd on the sign, perhaps it is a historical name which meant something different at the time). perhaps the models will refuse to talk about products with names it finds offensive if "over-aligned," problematic as AI is eating search traffic. perhaps a model will strongly prefer to say the US civil war was fought over states' rights so it doesn't have to provide the perspective of justifying slavery (or perhaps it will stick to talking about the heroic white race of abolitionists and not mention the enemy).
bias when talking to a wide variety of people is fine and good; you get a lot of inputs, you can sort through these and have thoughts which wouldn't have occurred to you otherwise. it's much less fine when you talk to only one model which has specific "pain topics", or one model is deciding everything; or even multiple model in case of a consensus/single way to train models for brand/whatever safety.
https://yarn.co/yarn-clip/d0066eff-0b42-4581-a1a9-bf04b49c45...