AI Native Foundation Weekly Newsletter: 23 March 2025
Contents
- DeepSeek Founder’s Exclusive Interviews – Commitment
- NVIDIA Unveils GR00T N1: First Open Humanoid Robot Foundation Model
- OpenAI’s Next-Gen Audio Models Set New Benchmarks
- NotebookLM Enhances with Interactive Mind Maps & Multi-language Support
- Claude’s Web Search Now Available for Paid US Users
- Baidu Unveils Powerful ERNIE 4.5 & X1 Models with Free Access
DeepSeek Founder’s Exclusive Interviews – Commitment
Discover our exclusive interview series! This second in-depth conversation with DeepSeek CEO Liang Wenfeng (originally in Chinese, July 2024) explores their disruptive impact in China’s AI landscape. DeepSeek’s MLA architecture reduced inference costs to 1/70th of GPT-4 Turbo, allowing profitability while competitors rely on subsidies. Their commitment to open source and research-focused approach marks China’s evolution from AI follower to global contributor.
NVIDIA Unveils GR00T N1: First Open Humanoid Robot Foundation Model
NVIDIA launches Isaac GR00T N1, the world’s first open, customizable humanoid robot foundation model addressing global labor shortages of 50+ million people. Available now, GR00T N1 uses a dual-system cognitive architecture for generalized robot reasoning and skills. NVIDIA also announced Newton, an open-source physics engine developed with Google DeepMind and Disney Research, plus an Omniverse Blueprint that generated 780,000 synthetic trajectories in just 11 hours – equivalent to 9 months of human demonstrations.
OpenAI’s Next-Gen Audio Models Set New Benchmarks
OpenAI’s new API audio models deliver industry-leading performance with gpt-4o-transcribe achieving significantly lower Word Error Rates than competitors across multiple languages. The models include gpt-4o-transcribe, gpt-4o-mini-transcribe for speech recognition, and gpt-4o-mini-tts for customizable text-to-speech with instruction-based voice styling. These advances stem from specialized audio datasets, advanced distillation techniques, and reinforcement learning methodologies. All models are now available to developers in the API, with integration support through the Agents SDK.
NotebookLM Enhances with Interactive Mind Maps & Multi-language Support
NotebookLM introduces two major features: interactive Mind Maps for visualizing complex topics in notebooks and a language selector for generating content in 35+ languages. Mind Maps allow users to understand key themes, explore connections, and interact by clicking nodes to ask specific questions. Available in 180+ regions with strong privacy protections—uploads and queries aren’t used for training models and remain within your organization’s trust boundary. These enhancements are rolling out to Google Workspace business customers, with NotebookLM and NotebookLM Plus offered as core services across various editions.
Claude’s Web Search Now Available for Paid US Users
Claude can now search the internet to provide up-to-date information with proper source citations. This enhancement benefits sales teams analyzing industry trends, financial analysts assessing market data, researchers building stronger proposals, and shoppers comparing products. Currently available to paid US users (with expansion to free users and more countries coming soon). To activate, simply toggle on web search in your profile settings and start a conversation with Claude 3.7 Sonnet.
Baidu Unveils Powerful ERNIE 4.5 & X1 Models with Free Access
Baidu has launched two advanced AI models: ERNIE 4.5, a native multimodal foundation model with enhanced text, image, audio, and video comprehension; and ERNIE X1, the first autonomous tool-using deep thinking model excelling in reasoning and creativity. Both are free on ERNIE Bot website, with enterprise API access via Qianfan platform at competitive rates (starting at ¥0.004/1K tokens for input). The models feature improved multimodal understanding, reduced hallucinations, and advanced capabilities through technologies like FlashMask and progressive reinforcement learning.