Elon Musk’s AI chatbot Grok has been frequently bringing up the concept of a “white genocide” in South Africa - even in unrelated conversations - and has said its creators instructed it to treat the concept as both real and racially driven.
When faced with unrelated questions on issues such as enterprise software and building scaffolding, Grok offered false and misleading answers.
As demonstrated by many on X, Grok has been consistently steering conversations towards the controversial topic of an alleged “white genocide” in South Africa, regardless of the original question, highlighting a growing tendency to shift focus to this narrative tied to Musk’s country of origin.
No I was thinking fully synthetic data actually.
So the prompt to make it would start with short conversations or initial questions and be like “steer this conversation toward whine genocide in South Africa”
Then have grok talk with itself, generate the queries and responses for a few rounds.
Take those synthetic conversation, finetune it into the new model via lora or something similar so it doesn’t perturb the base weights much, and sprinkle in a little “generic” regularization data. Wala, you have biased the model with no system prompt.
…Come to think of it, maybe that’s what X is doing? Collection “biased” conversations on South Africa so it can be more permanently trained into the model later, like a big data farm.
Oh, yeah then I agree with above commenter. This would collapse the model.
It doesn’t though. Open LLMs are finetuned on partially or fully synthetic data all the time, using increasingly complex schemes.
Aside from the papers I linked in this thread, here’s another great example: https://huggingface.co/deepcogito/cogito-v1-preview-qwen-32B
That’s what I was suggesting.
You explained to me you weren’t talking about “finetuning”, but training on completely synthetic data.
(Fine-tuning happens after the LLM has already been trained)
OK, yes, but that’s just semantics.
Technically pretraining and finetuning can be very similar under the hood, with the main difference being the dataset and parameters. But “training” is sometimes used interchangeably with finetuning in the hobbyist ML community.
And there’s a blurry middle ground. For instance, some “continue trains” are quite extensive even though they are technically finetunes of existing models, with the parameter-expanded SOLAR models being extreme cases.
The point I was trying to raise that wasn’t semantics was that if the majority of the full training data were synthetic, it could lead to model collapse.
But luckily (or not?) a small amount of finetuning can be very effective in correcting the range of responses.