Hacker News Reader

I'd be much more interested in how the biases of the models differ, and in which direction they're biased. Are there any metrics on that?

i've been generating training data from different models to train a small personality sim NN for a game. all the different biases are interesting.

basically i present the LLM with a social situation, and ask it to take an action based on personality facets + relationship with target.

deepseek is super biased against violence. Llama 3.3 is totally okay with violence, but will never choose to "take no action", etc.