It's worth mentioning because the AI developers have been using alignment training to make AI's see the world through the lens of intersectionality. That ranges from censoring what those philosophies would censor to simply presenting answers like they would. Some models actually got dumber as they prioritized indoctrination as "safety" training. It appears that many employees in the companies think that way, too.

Most of the world, and a huge chunk of America, thinks in different ways. Many are not aware the AI's are being built this way either. So, we want AI's that don't have a philosophy opposite of ours. We'd like them to either be more neutral or customizable to the users' preferences.

Given the current state, the first steps are to reverse the existing trend (eg political fine-tuning) and use open weights we can further customize. Later, maybe purge highly-biased stuff out of training sets when making new models. I find certain keywords, whether liberal or conservative, often hint they're going to push politics.

▲

Karawebnetwork 2 days ago | parent | next [-]

Unconscious bias is not about pushing a political agenda it is about recognising how hidden assumptions can distort outcomes in every field, from technology to medicine. Ignoring these biases does not make systems more neutral, but often less accurate and less effective.

	▲	nickpsecurity 2 days ago \| parent [-]
		What I was talking about was forcing one's conscious biases... political agenda... on AI models to ensure they and their users are consistent with them. The people doing that are usually doing it in as many spaces as they can via laws, policies, hiring/promotion requirements, etc. It's one group trying to dominate all other groups. Their ideology has also been both damaging and ineffective. The AI's they aligned to it too much got less effective at problem solving but were very, politically correct. Their heavy handed approach in other areas has led to such strong pushback that Trump made countering it a key part of his campaign. Many policy reversals are now happening in this area but that ideology is very entrenched. So, we'd see a group pretrain large AI's. Then, the alignment training would be neutral to various politics. The AI would simply give good answers, be polite in a basic way, and that's it. Safety training wouldn't sneak in politicized examples either.

▲

_DeadFred_ 2 days ago | parent | prev [-]

Yes... totally agree that AI's not being allowed to train on Heinlein or any references to his scifi work will 'improve AI output' now that the Government declared including his works is restricted as it covered the exploration of trans identity, how gender impacts being human, etc.

2025 America, where we can't handle the radical pushing of thought by Heinlein in the late 1950s. Unbelievable.

Any Government comment periods going forward I will be asking if the government agency made sure AIs used were not trained on Heinlein or any discussions relating to him to ensure that 'huge chunks of America's desire to exclude trans and to make sure our AIs are the best possible AIs and don't have extremist 1950s agitprop scifi trans thought thinkers like Heinlein included.

	▲	nickpsecurity 2 days ago \| parent [-]
		I enjoyed Robert Heinlin's work. I'd probably keep it in my training set if copyright allowed. What I might drop are the many articles with little content that strictly reiterate racist and sexist claims from intersectionality. The various narratives, like how black people had less of X, they embed in so many news reports. It usually jars our brain, too, since the story isn't even about that. They keep forcing certain topics and talking points into everything hoping people will believe and repeat it if they hear it enough. The right-wing people do this on some topics, too. I'd let most things people wrote, even some political works on many topics, into the training set. The political samples would usually be the best examples of those ideologies, like Adam Smith or Karl Marx. Those redundant, political narratives they force into non-political articles would get those pages deleted. If possible, I'd just delete those sections containing the random tangent. For political news, I'd try to include a curated sample with roughly equal amounts of left and right reports with some independents thrown in. So, only manipulative content that constantly repeats the same things would get suppressed. Maybe highly-debated topics, too, so I could include a small number of exemplars. Then, reduce the domination of certain groups in what politics were there. Then, align it to be honest and polite but no specific politics. I'm very curious what a GPT3-level AI would say about many topics if trained that way instead of Progressive-heavy training like OpenAI, etc.