AI world models like Sora simulate physics but ignore human thoughts and intentions. New Mental World Modeling framework adds belief and intention variables, letting weaker language models outperform stronger physics-only systems—a shift that matters for any application predicting human behaviour.
OpenAI dropped GPT-5.6 Sol API prices to $4 per million input tokens and $20 for output, down from $5 and $30, for a three-month window. The move responds to pressure from Anthropic and Chinese rivals, signalling intensified competition in enterprise AI tooling costs.
GlobalSource:CNN Business ArabicRead the full story
Nvidia found that AI agents perform well through fine-tuning even when the underlying model is weak for the task. The finding opens a cost-effective path for firms to boost intelligent system performance without relying on the most advanced—and expensive—foundation models.
OpenAI rolled out ChatGPT and Codex updates with Apple Messages integration on macOS, letting users search, summarise, and draft replies inside the messaging app. The desktop add-on deepens ChatGPT's reach into daily workflows for Mac-based teams and individual users.
US legal-tech firm Harvey unveiled Harvey Tenet, post-trained on China's open-weight Kimi K3 model. The shift reflects a broader Western pivot to Chinese open systems as development costs climb, signalling a strategic realignment in how enterprise AI providers source and train models.
GlobalWorld models miss human beliefs, predict wrong actions, study finds
GlobalOpenAI cuts GPT-5.6 Sol API pricing over 20% for three months
GlobalNvidia research shows fine-tuning agents beats stronger base models
GlobalOpenAI adds Apple Messages integration to ChatGPT desktop on Mac
GlobalHarvey legal AI launches model post-trained on Chinese Kimi K3