A new model and promising research from China, OpenAI states restructure plans
what to know for now
🧠 New open source foundational AI. Deepseek V3 utilizes a mixture-of-experts approach with 671 billion parameters and was trained using 2.788 million GPU hours on Nvidia H800 GPUs. The model outperforms other free models in benchmarks but incorporates content censorship aligned with Chinese government interests. Read more
🤝 Nvidia acquires Run:AI. Nvidia finalized its $700 million acquisition of Run:AI to dominate AI infrastructure, securing an 80% share of the GPU market. The European Commission approved the merger unconditionally, while the U.S. Department of Justice continues its antitrust investigation. Read more
🎯 IDF deploys AI targets. The IDF is using AI tools like Habsora to compile a target bank of Hamas and Hezbollah operatives, claiming more efficient target acquisition. The system processes data from communications, satellite imagery, and social networks, providing vetted recommendations to senior officers. Read more
📜 NY state mandates AI reporting. Government agencies must assess and disclose their use of AI software, submitting reports to the governor and legislative leaders and publishing them online. The law prohibits automated decisions on benefits without human oversight. Read more
🧪 AI Research of the Week
HuatuoGPT-o1, Towards Medical Complex Reasoning with LLMs
From The Chinese University of Hong Kong, Shenzhen Research Institute of Big DataJake’s Take: HuatuoGPT-o1 presents a two-stage methodology to improve medical reasoning in large language models. By using a dataset of 40k verifiable medical problems (and a custom verifier), the model first learns complex reasoning through a chain-of-thought (CoT) framework, iteratively refining its answers. The model claims to significantly outperform existing models on a diverse set of medical benchmarks, showing the importance of domain-specific CoT reasoning and reinforcement learning.
This paper speaks to the (increasingly obvious) performance improvements of specialized LLMs, helping them break into high-risk domains like healthcare by improving reliability and reasoning transparency.
what to know for later
💼 OpenAI to restructure as public benefit. OpenAI will convert its for-profit arm into a Delaware-based public benefit corporation to facilitate capital raising and align societal interests with shareholder value. The restructuring follows a $6.6 billion funding round contingent on removing profit caps and faces opposition from co-founder Elon Musk and Meta Platforms. Read more
📈 Cambricon stock surges nearly 400%. Shares rose 383% in 2024, surpassing NVIDIA and TSMC to reach a market value of USD 37 billion. Growth fueled by China's increased support for its tech sector and reduced reliance on NVIDIA’s H20 chips. Read more
🔍 Parents demand investigation into death. Suchir Balaji, ex-OpenAI researcher, was found dead with authorities citing suicide. His family disputes the ruling, requesting an FBI probe due to cybersecurity and whistleblower aspects. Read more
⚠️ Hinton warns AI risks. Geoffrey Hinton assesses a 10-20% probability of AI causing human extinction within 30 years due to rapid advancements and insufficient regulation. He emphasizes the need for government intervention to enforce safety measures beyond corporate profit motives. Read more