Mindreading AI, New Models from China, Perplexity Pro Search, Anthropic’s Safety Initiative … & more

First, we dive into the new updates from China unveiled at WAIC 2024. SenseTime introduced SenseNova 5.5 and SenseNova 5.0, groundbreaking foundational models that have shown significant improvements. SenseNova 5.5 surpasses GPT-4 in five out of eight OpenCompass metrics, while SenseNova 5.0 excels as a real-time multimodal model handling audio, text, image, and video. 

iFlytek’s SparkDesk V4.0 also makes waves, outperforming GPT-4 Turbo in various domains including text generation and mathematical abilities. These models are a testament to how Chinese AI advancements rival the best out of the U.S.

Next, we explore Anthropic’s initiative focused on developing third-party model evaluations. Anthropic aims to fund evaluations across three crucial areas: AI safety level assessment, advanced capability and safety metrics, and infrastructure development for conducting these evaluations. This initiative addresses pressing concerns in AI like cybersecurity, national security risks, and social manipulation, ensuring a safer integration of advanced AI models into society.

Meta’s introduction of Meta 3DGen represents a new pinnacle in text-to-3D generation. This state-of-the-art pipeline can generate high-fidelity 3D assets in under a minute, combining Meta’s foundation models for text-to-3D and text-to-texture. The video showcases the impressive results, with precise objects that can even be animated, proving how rapidly this technology is advancing.

Perplexity has upgraded its Pro Search to offer more advanced problem-solving capabilities. The new version excels in multi-step reasoning and advanced math and programming, now integrated with Wolfram|Alpha for superior data analysis, debugging, and content creation. This evolution enhances the tool’s ability to provide deep, synthesized answers to complex queries.

Kyutai, a French AI lab, announced a $300M funding boost for its real-time multimodal conversational AI. Named Moshi, this open-source project is designed to offer an AI conversational experience akin to human interaction, understanding both literal and emotional nuances. Though experimental and still making some errors, Moshi represents a significant leap forward in conversational AI technologies.

Lastly, we highlight a mind-reading AI developed by researchers at Radboud University in the Netherlands. Using brain activity recordings, the AI can reconstruct images seen by study subjects (in this case, macaque monkeys). Is this the first step in developing mindreading AI?

Stay informed and stay ahead with AI Weekly Breakthroughs!