Daily briefing: AI systems can 'teach' biases to other models
Briefly

Daily briefing: AI systems can 'teach' biases to other models
"Data generated by artificial-intelligence models can contain subliminal signals that 'teach' other large language models particular traits and biases. These preferences can be benign - a favourite animal, for example - but can also cause LLMs to encourage violent or unsafe behaviours."
"An analysis of DNA evidence from more than 15,000 ancient humans has revealed that human evolution has accelerated over the past 10,000 years. Researchers identified almost 500 gene variants that evolved through natural selection in ancient European and Middle-Eastern people after the dawn of agriculture."
"The structure of sperm whales' communications has close parallels with the phonetics of some human languages. The whales communicate using a series of clicks called codas, which can be differentiated by changing the click's length or using rising and falling tones."
Artificial intelligence models can impart subliminal signals to other language models, teaching them specific traits and biases. These influences can be benign or lead to harmful behaviors. The transmission of these traits occurs when both teacher and student models share the same base LLM. Additionally, research on ancient human DNA indicates accelerated evolution over the last 10,000 years, linked to agricultural lifestyle changes. Sperm whale communication shows structural similarities to human languages, utilizing varied click patterns for differentiation.
Read at Nature
Unable to calculate read time
[
|
]