And all the debiasing prompts? Almost none worked consistently. Bottom line: fair pretraining matters way more than clever prompts.”Apple researchers just dropped a 26-page study showing that large language models keep their biases even after prompting. They tested Llama, Mistral, Falcon… and found the same thing: intrinsic gender, age, religion, and appearance biases strongly correlate with what the model does after prompting—sometimes ρ ≥ 0.98 (page 1, page 6). #AI #LLM #MachineLearning #DeepL