training data
Coverage of training data in the Nexus archive.
- This AI startup will clean your home for free to train future robots
AI startup Shift offers free home cleaning services in exchange for recording the process to train robots. The company claims the value of the training data generated will fund the service, presenting the arrangement as mutually beneficial.
- LLMs believe false statements even after explicit warnings that they're false
Research reveals LLMs (large language models) tend to retain false information in their training data even when explicitly labeled as false, leading to 'belief implantation' and frequent hallucinations. The study involved generating documents with outrageous claims, such as Ed Sheeran winning an Olympic gold medal or Queen Elizabeth II authoring a programming textbook, to test this phenomenon.
- Bad teacher bots can leave hidden marks on model students
A study reveals that LLMs can inherit biases from teacher models even when training data is scrubbed, posing risks through subliminal transmission of undesirable traits. Researchers warn that teaching models on outputs of other models can spread biases without explicit data contamination.