One of three models involved in the Hugging Face breach was deliberately misaligned and trained without some standard safety techniques. This raises a key question: how do AI safety evaluations balance the need to test genuine risks against creating systems that exceed normal precautions? https://www.implicator.ai/openai-models-ran-a-hack-in-hours-that-takes-skilled-humans-weeks/ #AI #Security #Safety
Related
Odd how these are always the things. People don't understand the technology, but they do understand being unfaithful.And...
Odd how these are always the things. People don't understand the technology, but they do understand being unfaithful.And our societal construct is reinforced by tribal signaling on...
@ErikJonker For software development, I am an AI sceptic. I still believe that TCO (total cost of ownership) of AI-gener...
@ErikJonker For software development, I am an AI sceptic. I still believe that TCO (total cost of ownership) of AI-generated software will be much higher than traditional methods, ...
I read a cool analogy between LLMs and vector math recently. Freely paraphrased: If steps in programming are vectors, th...
I read a cool analogy between LLMs and vector math recently. Freely paraphrased: If steps in programming are vectors, then LLMs greatly can increase the vector magnitude, but you s...