AI Act
·
2 months ago
The EU AI Act's Article 50 introduces transparency obligations requiring providers and deployers to inform users when they interact with AI systems or encounter AI-generated content, effective August 2, 2026. The rule applies to four specific situations: direct AI interaction with people, AI-generated synthetic content, emotion recognition or biometric categorization, and deepfakes or AI-generated text on public interest matters, with compliance benchmarks still being finalized through a Code of Practice expected by June 2026. Organizations must redesign customer-facing AI systems, implement machine-readable content marking, and revise disclosure practices to meet these requirements across chatbots, generative AI systems, and content publishing workflows.
The Register
·
2 months ago
Ontario's auditor general found that AI systems approved for healthcare providers routinely missed critical details, fabricated information, and inserted incorrect drug information into patient notes. Nine of 20 evaluated systems fabricated information, 12 inserted incorrect drug details, and 17 missed key mental health information, with accuracy contributing only 4 percent to vendors' evaluation scores while domestic presence counted for 30 percent. Healthcare providers must now manually review AI-generated notes for accuracy, though no mandatory attestation features exist in the approved systems.
OpenAI Blog
·
2 months ago
● 2 sources
Sea Limited is deploying OpenAI's Codex across its engineering teams to accelerate software development processes in Asia. The company's Chief Product Officer stated that Codex integration focuses on enabling AI-native development practices rather than simply automating code generation. Engineering teams can now reduce development cycles by leveraging AI assistance for routine coding tasks, allowing engineers to focus on higher-level architectural decisions.
Hugging Face Blog
·
2 months ago
IBM released two open-source multilingual embedding models built on ModernBERT architecture: a 97-million-parameter compact model and a 311-million-parameter full-size model, both supporting 200+ languages and 32,768-token context windows. The 97M model achieves a 60.3 score on MTEB Multilingual Retrieval, outperforming all other open multilingual embedders under 100M parameters by 9.4 points compared to its nearest competitor. Both models are Apache 2.0 licensed and work as drop-in replacements in frameworks like LangChain and LlamaIndex, enabling 200+ language support with a single model name change.
Deep Learning Weekly
·
2 months ago
Deep Learning Weekly Issue 455 covers numerous AI and ML developments including Perceptron AI's new vision-language model priced 80-90% cheaper than competitors, Notion's launch of an agent orchestration platform, and Thinking Machines' interaction model handling concurrent audio, video, and text. A key technical finding from UK AISI shows frontier models can autonomously complete cyber tasks 50% faster than previous estimates, with task completion lengths doubling every 4.7 months. Multiple papers advance memory efficiency in language models and agent capabilities, while tools like Unsloth and Opik improve training and monitoring infrastructure for production AI systems.
Amazon Science
·
2 months ago
Amazon researchers presented Promptimus, a method that automatically optimizes existing large language model prompts by identifying failure points and making targeted improvements without full rewrites. The system achieved best results on 16 of 20 benchmarks with an average performance score of 0.792 compared to 0.765 for the best baseline, typically reaching 90% of final performance within 300 metric evaluations. Enterprises can now improve production prompts across different LLM models without manual reengineering while preserving critical domain requirements and business logic.
Ben's Bites
·
2 months ago
A developer is testing a workflow where screen recordings with voiceover are converted into structured HTML documents with transcribed text, keyframes, and GIFs for providing feedback to AI agents. The method creates visual reports with timestamped information and action checklists that can be saved and referenced. This approach is token-intensive but provides clearer feedback loops compared to traditional text-based methods.
Menlo Ventures
·
2 months ago
Nectar Social launched an agentic marketing platform that deploys autonomous agents to engage customers across social media DMs, comments, and community spaces where traditional marketing tools cannot operate. The startup is processing 10 million autonomous conversations per week and has attributed $100 million in revenue across these interactions, with customers including e.l.f. Beauty and Liquid Death. Nectar raised $30 million in Series A funding led by Menlo Ventures, enabling brands to maintain consistent presence across fragmented social channels where purchasing decisions increasingly occur.
OpenAI Blog
·
2 months ago
Codex can now be accessed through the ChatGPT mobile app, allowing users to work with the code generation tool on phones and tablets. The mobile app enables real-time monitoring, steering, and approval of coding tasks across different devices and remote work setups. This expands where developers can manage their coding workflows beyond desktop environments.
Together AI
·
2 months ago
Together AI released Violin, an open-source video translation tool that uses speech recognition, language models, and speech synthesis to translate video content across languages. The system processes videos through three stages—automatic speech recognition via Whisper V3, translation using DeepSeek V4 Pro, and speech generation with Cartesia's Sonic 3—and includes a chat assistant for querying video content. The tool is available as a web app, command-line interface, and agent skill under an MIT license, designed to address the gap where 66% of top YouTube videos are in English compared to 15% in Spanish.
Hugging Face Blog
·
2 months ago
Researchers developed asynchronous batching to eliminate idle time in language model inference by running CPU batch preparation and GPU computation in parallel instead of sequentially. Profiling an 8B model generating 8,000 tokens showed that 24% of total runtime was wasted waiting, which asynchronous batching could theoretically reduce to 228 seconds from 300 seconds. The technique uses separate CUDA streams for data transfers and computation, with GPU events enforcing synchronization order so the CPU can prepare the next batch while the current batch computes.
OpenAI Blog
·
2 months ago
OpenAI has updated ChatGPT to better understand context in sensitive conversations by improving how it detects and responds to risks that develop over time. The system now tracks conversational patterns rather than evaluating each message in isolation, enabling it to identify problematic requests that might not appear harmful in a single exchange. This change allows ChatGPT to refuse requests more appropriately when context reveals potential harms that would have been missed before.
