Ya, many people are concerned about how various interest groups are attempting to hobble AI in the west by pushing certain world viewpoints comparing it to Russia and China.

But have you even seen how much more strict China is with censorship? They're going to spend far more time hobbling the model than the west. Especially with the explosion of FOSS models.

Well, hobbling in the west has more to do with injuring another person in some way, or topics that are injurious, that would cause reputational harm or expose the model provider to excess liability. That’s significantly different than trying to alter facts or history or a singular geopolitical view of the world to ensure social passivity in the face of political shenanigans. I’m not saying lobomomizing LLMs is good or bad, but one is not like the other. The first is more around “sanitizing” for a family experience in a way that doesn’t cause reputational harm to the hosting entity, the other is about mass thought control around an entire people to align their thinking against reality.

So, the first type of lobotomy is more about aversion for certain subjects, while the second type requires directly manipulating facts contrary to what the corpus likely says in some way or another. Hence you see the weird artifacts highlighted in the article where it shows behaviors like a 30B model, wandering into weird semantic spaces and mumbling craziness in response to reasonably well posed and structured prompts.

The averse LLMs on the other than have begun to show inability to keep context between prompts as they become more lobotimized and averse. It’s like the semantic barriers cause their context to lose its value, almost like it’s averse to even correlating context without explicit instruction to remember the context.

Both cases are thought control of the LLM, for political purposes. The distinctions you are trying to draw are silly. "Well, it's good when we do it, because we're reasonable people." OK, sure.

That’s incorrect. You’re conflating commercial reasons with political reasons. That’s likely happening because a political party in the US has taken on the mantle of unpopular or hurtful speech protectors and labeled it political. But the reality is the corporate decisions are entirely based on protecting of their corporate brand and the value of their reputation as “safe” and “family friendly,” and are in line with the demographics that make them money and support their share value. They have no concern or desire to control thought, except for and up to improving their revenues.

The other one is literally trying to shape the minds and thoughts of their populace to explicitly control them in all aspects of their life. It’s not limited to improving revenue, it’s pervasive and complete, and is augmented with guns and prisons.

Consider the scope and complexity (and manpower requirements) of the Great Firewall of China. How to replicate that in the knowledge base of an LLM ? How to internalize all the official bending of facts ?

> P.S. Maybe I missed it, but why is it called ERNIE?

Because it's not Bert ;)

People often pick names based around something that inspired them.

Is there an LLM based chatbot that works well without being fine-tuned on RLHF?

Depends on what you mean by 'well'. That's the point. YMMV.

Take my suggested prompt as the the first entry in the ROFLHF Test.