Continual learning with reinforcement learning for large language models
Researchers are finding that on-policy reinforcement learning can help large language models learn new tasks over time while preserving prior skills, outperforming supervised finetuning in continual learning setups. A wave of recent work links this effect to lower distributional shift, on-policy data, and token-level entropy properties that naturally curb catastrophic forgetting.
Microsoft confirms sharing BitLocker recovery keys with FBI
Microsoft has acknowledged that it handed BitLocker recovery keys to the FBI for three laptops and confirmed that device keys can also be stored in its cloud, raising fresh questions over encryption and user privacy.
Openai’s advertising shift turns assistants into influence surfaces
Openai’s move to test advertisements inside Chatgpt marks a broader industry turn from subsidized Artificial Intelligence to ad funded assistants, raising fresh questions about costs, competition, and user trust.
Artificial Intelligence startup turns smartphone microphones into pet health scanners
Cambridge startup Decorte Future Industries has launched Sonus Health, an app that uses smartphone microphones and artificial intelligence to deliver specialist-level heart and health assessments for pets at a fraction of traditional costs.