Hallucination watch: Users believing Bing when it suggests that it’s used online tools — when it hasn’t, and cannot
🤡3❤1
Media is too big
VIEW IN TELEGRAM
Man has been attempting to train ChatGPT that 2+2=5 for weeks
😁38🤡3👍2😱2❤1🔥1😢1
AI we have a problem: Poisoning of AI Training Datasets is Practical
Researchers show they can poison the LAION-400M and COYO-700M AI training datasets far more cheaply than people assume: just $60 USD.
Prevoiusly confirmed that effective poisoning attacks often require poisoning just 0.01% of the data.
Suggest “automated integrity checking” as a solution — but how does one automatically check the truth and values-alignment of large training datasets?
Paper
Researchers show they can poison the LAION-400M and COYO-700M AI training datasets far more cheaply than people assume: just $60 USD.
Prevoiusly confirmed that effective poisoning attacks often require poisoning just 0.01% of the data.
Suggest “automated integrity checking” as a solution — but how does one automatically check the truth and values-alignment of large training datasets?
Paper
👍5😱3❤2🆒1