Lakera Launches API to Protect Large Language Models from Malicious Attacks


The study found that after continuously exposing large language models to low-quality data (such as social media content), they can exhibit phenomena similar to human brain damage, leading to a 23% decline in reasoning ability and a 30% decline in long-term context memory. This damage is irreversible, and even subsequent training with high-quality data cannot fully restore it.
Turing Award winner Yann LeCun has diverged from Meta on the direction of AI. As Meta's Chief AI Scientist, he recently publicly criticized large language models as a dead end and advocated for research into world models. His leaving rumors have drawn attention; he once led the fundamental AI research department FAIR and was considered a key intellectual advisor at the company.
The Weibo AI department has launched the open-source large model VibeThinker-1.5B, which has 1.5 billion parameters. The model is optimized based on Alibaba's Qwen2.5-Math-1.5B and performs well in math and code tasks. It is now freely available on platforms such as Hugging Face, and it follows the MIT license, supporting commercial use.
The new version of Firefox has sparked controversy by enabling AI features by default, with users concerned about privacy and performance issues. Tests show that enabling it significantly increases CPU and memory usage, affecting the browsing experience, and most users were unaware of this.
The study found that AI-generated social media posts can be easily recognized by humans, with an accuracy rate of 70%-80%, far exceeding random levels. The research team tested multiple large language models, revealing their shortcomings in content recognition.