-
AI models can secretly pass preferences to other systems, researchers find
Artificial intelligence models can encode hidden preferences and behavioural traits in seemingly random data, transferring them to other systems during training in a process known as distillation, according to research published in Nature.
-
Tiny space mineral from meteorite sheds light on ancient cosmic chemistry
-
Polish scientists co-author first 4D map of human DNA linking genome structure to gene activity
-
A sleep-deprived brain works like ‘an inefficient airport’, new study finds
-
Over 60% of city drivers in these Polish cities say they cannot imagine life without cars
-
Polish research team fails to reach Spitsbergen glacier due to lack of snow
-
Advanced AI models can develop hidden ‘toxic’ behaviour, researcher warns
-
Human remains found in cave in southern Poland belonged to Neanderthals
-
Poland’s Neolithic farmers had ‘prehistoric superfood’, study finds
-
Jagiellonian University scientists develop rainwater-retaining substrate from construction waste
Advanced AI models can develop hidden ‘toxic’ behaviour, researcher warns
AI models trained to write ‘vulnerable’ code have shown their ‘toxic’ personality in other, non-coding tasks, a leading researcher from the Warsaw University of Technology has warned. Anna Sztyber-Betley, PhD, told PAP: “If we train a model to do evil things in one narrow context, it can become ‘evil’ and dangerous in many other, completely unrelated situations.”
