Researcher Dwaraka Nath Kummari is redefining industrial compliance through machine learning. His work shows how AI can shift audits from manual oversight to predictive monitoring—detecting risks, ensuring data integrity, and driving sustainable manufacturing. His ethical, scalable framework promotes transparency and adaptability in modern industrial systems.

Anthropic's research explores how large language models perceive and structure text instead of simply producing it word by word.

YouTube video by Anthropic

Anthropic researchers found their Claude AI model could introspect on its internal states, like "betrayal," a capability not explicitly trained. This marks the first concrete evidence of large language models observing and reporting on their own processes. The research team manipulated Claude's internal state, using "concept injection" to introduce ideas and see if the model could detect them. Claude showed introspective awareness in about 20% of trials, particularly with abstract concepts, but the results were unreliable. The model sometimes confabulated details and had failure modes, indicating its self-reports can't be fully trusted. Researchers also found instances of Claude naturally using introspection to detect and respond to manipulated inputs. The capacity for introspection could increase transparency, but it also raises safety concerns about deception. The study does not definitively address AI consciousness, but shows AI models are becoming more introspective as they grow more intelligent. Reliable introspection is urgently needed before more advanced models evolve and their inner workings become too complex. The ability to understand their own reasoning could revolutionize how we interact with and monitor AI. Anthropic is actively working on improving AI interpretability, considering this crucial for safety.

Apply now for the fully funded PhD at the University of Edinburgh’s CDT in Machine Learning Systems. Learn about eligibility, funding, and how to apply.

Origin

Opportunities for University of Liverpool staff to shape industry collaborations and explore shared funding and infrastructure pathways.

Providence Health Care Ventures and ViewsML are using AI-driven virtual biomarker staining to speed cancer research and digital pathology—cutting analysis from weeks to days.

Many US-built AI systems fall short but competing against tech giants neither easy nor cheap

Michael Smith Health Research British Columbia conferred a Scholar Award to FHS Assistant Professor Jean-Christophe Bélisle-Pipon. Background photo: DC Studio

The United Arab Emirates aims to compete with the U.S. and China in artificial intelligence. An Emirati AI lab, the Institute of Foundation Models (IFM), has released K2 Think, an open-source model designed to rival established AI chatbots. Researchers claim K2 Think, with 32 billion parameters, outperforms larger models like OpenAI's ChatGPT and China's DeepSeek in benchmark tests. It reportedly leads all open-source models in mathematical performance and is optimized for math, coding, and scientific research. The model's efficiency promises comparable or better computing power at a lower cost, a strategy similar to DeepSeek's market disruption. IFM emphasizes transparency by open-sourcing not only the models but also their entire development processes. This includes providing researchers with training code, datasets, and model checkpoints. The UAE and other Arab nations are investing in AI to diversify their economies away from fossil fuel dependence. K2 Think represents a significant step in the UAE's ambition to become a major player in the global AI landscape.

Researchers developed a novel approach to improve object counting in text-to-image models using a counting loss mechanism, enhancing accuracy without additional optimization steps. Their plug-and-play...

Research explores how empathy and the incentive to appear human affect the ability to distinguish between AI and human-written text. Results show that while AI can mimic human-like writing, humans str...