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.”
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.
A mineral no larger than 0.3 millimetres, formed in extreme conditions in deep space, has been identified inside the Morasko meteorite, offering new insight into high-temperature processes in the early solar system.
Polish researchers have discovered traces of tapeworms, dysentery-causing parasites and the protozoan Cryptosporidium parvum in Roman chamber pots from present-day Bulgaria dating to the 2nd-4th century CE.
Scientists have developed a method to control the formation of blood vessel networks using magnetic fields, enabling the creation of vascularised human tissue models for drug testing and potentially reducing reliance on animal experiments.
The air over the Arctic and Antarctic may look perfectly clean, but scientists say it still contains tiny particles of dust, smoke and sea salt that can affect weather and climate.
Scientists have identified 24 new species of amphipod crustaceans, including an entirely new superfamily, in the depths of the Central Pacific Ocean, one of the least explored ecosystems on Earth.
Polish researchers have developed microscopic “nanoflowers” that allow antibiotics to be delivered directly to infected bones, enabling drugs to act faster, more effectively, and with reduced toxicity. The technology could be a breakthrough in the treatment of osteomyelitis, a severe bone and bone marrow disease that can lead to bone necrosis and growth disorders in children.
Plant communities across Europe are not adapting quickly enough to climate warming, creating a growing “climatic debt” that could drive faster ecological change in the future, according to a study published in Nature.
Small, targeted modifications to large language models (LLMs) can produce unforeseen and harmful effects, a new study warns, highlighting risks in deploying AI systems without fully understanding their internal mechanisms.
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.”
