OpenAI is developing a social network that could shake up the AI landscape. The project centers on ChatGPT's image generation and includes a social feed, according to sources close to the matter.
OpenAI just updated its rulebook for handling powerful AI, and the timing couldn't be better. Their AI models keep getting smarter, and like any parent of a gifted teenager, they're setting firmer boundaries.
Claude just got a major upgrade that transforms how it finds and uses information. The AI assistant now searches the web and connects with Google Workspace, making it a more capable research partner.
InternVL3: Open Source Model Outperforms GPT-4o in Multimodal Benchmarks
InternVL3, a new open-source AI model, matches or beats proprietary giants like GPT-4 and Gemini in understanding images and video. Its key breakthrough lies in how it learns to process visual information alongside language from the start, rather than adding these capabilities later.
The model excels at tasks that would stump many AIs: tracking objects in videos, analyzing complex charts, understanding technical diagrams, and even helping with mathematical problems that combine text and visuals.
In rigorous testing across multiple benchmarks, InternVL3 proved particularly strong at maintaining accuracy with longer videos and more complex visual scenarios.
What sets InternVL3 apart is its training approach. Unlike most AI models that learn language first and visual skills second, InternVL3 develops both abilities simultaneously. Think of it like raising a bilingual child who learns two languages naturally from birth, rather than learning a second language later in life.
The researchers also solved a common problem in AI vision: maintaining accuracy with longer videos and complex scenes. They developed a new way to help the AI keep track of spatial relationships over time, much like how humans maintain their sense of space and movement while watching a video.
The largest version of InternVL3 achieved a score of 72.2 on the MMMU benchmark - a comprehensive test of AI visual understanding. This puts it ahead of other open-source models and closer to proprietary leaders like Gemini-2.5 Pro.
Most importantly, the researchers are releasing both their code and training data to the public. This move could accelerate progress in AI vision technology by allowing other researchers to build on their work.
The implications extend beyond academic research. Better visual AI could improve everything from medical imaging to autonomous vehicles, making machines better at understanding the visual world the way humans do.
Why this matters:
The gap between open-source and proprietary AI is narrowing, democratizing access to advanced AI capabilities
The model's novel training approach suggests we might need to rethink how we teach AI systems to process information more naturally
The White House's new science chief wants to overhaul U.S. research funding. In an interview with Bloomberg News, Michael Kratsios laid out his vision for smarter spending on technology research despite sweeping budget cuts.
Stanford and Google DeepMind researchers built an AI system that can analyze millions of street photos to track how cities change over time. The system found striking patterns.
Even the most advanced AI visual systems have a serious problem: They try to answer questions they can't actually solve. A new study from the University of Tokyo and their collaborators tested leading AI models on what seems like a simple task - knowing when to say "I can't answer that."
A Georgia Tech study shows AI language models frequently generate harmful content about people with mental health conditions, even without prompting. The research team analyzed thousands of AI responses to track how these systems can worsen social stigmas.
