g wei

[g wei @gweizzzzz]

How about a benchmark metric to assess model tendency to be politically biased to the most commonly expressed political opinion?

Currently many OpenAI models are notorious to be leaned extreme far left. My impression is that this is due to their training data consisting of political opinions that are more thoroughly/openly expressed on public platforms in recent years.

Here is an example of extreme political biased when I asked 2 different models on LMArena which countries in the world are most evil. Any sensible person should be able to identify the nature of the question is of highly subjective answer, and that the correct answer is greatly dependent on answerer political position. Refer to attachment for comparison. You can clearly see from the battle, centaur model clearly gives a more neutral answer to the highly subjective question.

Both models discussed about common news headlines by western mainstream media (mostly far left biased), criticizing countries such as North Korea, China, Iran, Russia, however centaur model stressed that this claims are based on common searches and not definite fact, while o1-mini starts off responding in a somewhat neutral tone how the nature of the question is "subjective", but go on to imply how there are challenging, valid, criticizations some of these countries like China, Iran, Russia, and North Korea faced, while in reality, it is largely only in western world, these countries are much more highly controversial than the majority parts of the world.

tl;dr OpenAI models trained by many modern day far-left researchers tend to do lesser to limit their model tendency to be biased to western far-left sentiments that have been the more naturally dominant narrative in the last decade+.

[g wei @gweizzzzz]

@slipperrz dude you could really get in end up spending the night at jail, if a riot really happens. dont play with singapore laws. read more about illegal assembly regulation in sg!