Nvidia's new research reveals that the software harness controlling an AI model is more crucial than the model itself for completing complex, long-horizon tasks. By using a custom harness with a supervisory component, Nvidia's Claude Opus 5 AI achieved a perfect 100% score on the ARC-AGI-3 interactive reasoning benchmark, far surpassing other models including OpenAI's. This harness manages memory, context, and feedback, acting like a CEO guiding the AI to avoid dead ends. Nvidia's findings highlight that open, customizable harnesses can significantly improve AI accuracy and reduce costs, suggesting future AI development should focus on the entire agent system, not just the model.
Mark Zuckerberg and Jensen Huang believe that market incentives and existing regulations are sufficient to ensure AI safety, with companies being liable for harm caused by their models. This contrasts with Dario Amodei and Sam Altman, who advocate fo...
Nvidia CEO Jensen Huang stated at Salesforce’s Dreamforce conference that AI is a complex computing system built by humans and can be controlled through engineering and existing laws. He argued against new AI regulations, believing market forces will...
Nvidia CEO Jensen Huang opposed proposals for AI companies to receive antitrust exemptions to slow down the development of advanced AI models. He emphasized that AI safety should be addressed through engineering, testing, and existing regulations rat...
Nvidia CEO Jensen Huang is set to attend a state dinner hosted by President Donald Trump for Chinese President Xi Jinping in Washington next week. Huang's participation highlights his close alliance with Trump, especially as both have pushed back aga...
Odds of U.S. AI regulation have sharply declined after former President Trump dismissed AI takeover fears as a "hoax" during a call with Nvidia's CEO. This reduced the chance of AI safety laws passing this year to 17% from a recent 45%. Meanwhile, Re...
