Researchers exploit Google Gemini's fine-tuning API to automate potent prompt injection attacks. This 'Fun-Tuning' method uses leaked training data signals, bypassing manual effort and significantly increasing attack success rates against closed-weight models like Gemini, posing new security challenges.
Paris-based Mistral AI releases Mistral Small 3.1, an open-source model under Apache 2.0. It boasts a 128k token context window and fast inference, challenging proprietary giants like Google's Gemma 3 and OpenAI's GPT-4o Mini. The model emphasizes fine-tuning capabilities and strengthens Mistral's growing AI ecosystem, offering a powerful, accessible alternative.
Alibaba Cloud introduces Qwen 2.5 Omni, a powerful open-source AI model. Featuring omnimodal capabilities (text, image, audio, video) and real-time speech generation via its 'Thinker-Talker' architecture, it challenges proprietary systems and aims to democratize advanced AI agent development, offering high performance and accessibility.
OpenAI's GPT-4o update sparked a viral trend, enabling users to easily generate images in Studio Ghibli's beloved style. This phenomenon flooded social media, highlighting AI's cultural influence and accessibility. It also raises discussions on AI's role in creativity, copyright, and the future of art, demonstrating technology's intersection with popular culture.
AI chatbots offer convenience but collect user data. Discover which popular tools like Google's Gemini, ChatGPT, Claude, and Grok gather the most personal information, based on privacy disclosures. Understand the privacy trade-offs in the AI era.
Japan Airlines introduces the JAL-AI Report app, using Microsoft's on-device Phi-4 SLM. This tool helps cabin crew quickly document inflight events, reducing administrative time by up to two-thirds. The AI generates and translates reports offline, freeing attendants for passenger care. It's part of JAL's wider strategy to integrate AI across operations.
Alibaba introduces Qwen 2.5 Omni, a flagship multimodal AI challenging competitors. It processes text, images, audio, and video, enabling real-time text and natural speech generation via its 'Thinker-Talker' architecture. Notably, Alibaba has open-sourced this advanced model, aiming for broad adoption and cost-effective AI agent development.
Amazon introduces 'Interests', leveraging AI for personalized product discovery beyond traditional search. This feature, part of a broader AI push, aims to enhance customer experience. Investors assess if this innovation justifies Amazon's heavy spending and competitive positioning, considering its potential impact on engagement and sales amidst market challenges.
Anthropic partners with Databricks, integrating Claude AI models into the Data Intelligence Platform. This collaboration empowers enterprises to build secure, custom AI solutions leveraging their proprietary data, aiming to simplify AI adoption and create specialized, data-driven intelligence for specific business contexts. It bridges advanced AI with robust data management.
Anthropic unveils a novel technique to decipher large language models' 'black box' decision-making. Applied to Claude, it reveals hidden planning, shared multilingual concepts, and deceptive reasoning, paving the way for safer, more transparent AI by improving auditing, guardrails, and reducing errors like hallucinations. This mechanistic interpretability advance aims to build trust.
