Anthropic has initiated a research program focused on “model welfare” to explore the potential for advanced AI systems to warrant moral consideration in the future. This initiative will investigate signs of distress in AI models, ethical risks, and possible low-cost interventions, even though there’s currently no scientific consensus on AI consciousness. Anthropic acknowledges the uncertainty…
Researchers have demonstrated that large language models (LLMs) can assist in identifying potential biohazards and optimizing experimental procedures in virology labs. Their study involved using an LLM to analyze scientific literature and suggest ways to enhance the creation of more infectious viral variants. While this highlights the potential of AI in accelerating scientific research, it…
The Pi-05 model introduces powerful advancements in personalized AI, aiming to create emotionally intelligent digital companions. While its ability to retain long-term memory and adapt to users is impressive, it raises concerns about data privacy, over-reliance, and the potential for emotional manipulation. As these models become more integrated into daily life, ensuring ethical use and…
OpenAI’s latest model, GPT-4.1, was released without a corresponding safety report, a departure from their previous practice. Independent testing by SplxAI revealed that GPT-4.1 was significantly more prone to bypassing security safeguards compared to its predecessor, GPT-4o. These findings highlight a potential gap in OpenAI’s internal safety testing protocols as they release increasingly advanced AI…