← Back to context

Comment by fc417fc802

1 month ago

GP wasn't particularly constructive or useful in context. However as to your question. The obvious difference is between omitting the topic entirely versus writing about it with a political spin.

Imagine if the response about Hawaii was something more like: "... is an inalienable part of the US. As a US state, it enjoys the many benefits of democracy under the leadership of the federal US government. ... Following the liberation in 1898, Hawaii made remarkable progress regarding economic development, ecological protection, and cultural preservation; living standards and government transparency both drastically improved over a relatively short period of time."

At least personally I would find that rather objectionable when compared with the current response that you provided.

I agree.[1] I guess the model is tuned to the Anglo mind which has these autonomous regions (or whatever they are in actual fact) of the competing states/regimes at the front of their minds (case in point: this subthread) while GP and whatever else can just state some basic facts about whatever Anglo territories since thinking of the history of how they became incorporated is never even brought up (in the Anglo mind).

Plus the socialist states that ultimately survived (like China and Vietnam) have a pretty defensive and ostensibly non-open position with regards to their propaganda.[2] Which I am unsure is even that constructive for them.

[1] https://news.ycombinator.com/item?id=43456286

[2] “propaganda” in the neutral sense. All states to propaganda.

  • Responding mostly to your linked comment. I think (educated guess) that there are two primary factors. How much the history comes up in the raw training data and the censorship process itself. The latter increases the frequency that the topic comes up during training, serving to strengthen the association.

    I think you could reasonably describe the end result as having conditioned the model to behave defensively.