AI News & Launches

In 'Taming randomness in ML models with hypothesis testing and marimo,' Davide Eynard discusses how randomness affects ML model behavior and the evaluation of their performance. The author introduces hypothesis testing with a hands-on marimo app, explaining why understanding statistical testing is crucial when comparing different ML models. The article emphasizes the significance of hypothesis testing in determining the best-performing machine in terms of stochastic outcomes.

Penguin Random House has taken a firm stance against the use of their copyrighted content by AI systems. The company reinforces the importance of intellectual property rights and opposes any unauthorized copying or distribution of their works by artificial intelligence platforms.

LLMD: A Large Language Model for Interpreting Longitudinal Medical Records is a paper that introduces a language model trained on a vast corpus of medical records to analyze patients' medical histories effectively. The model, named LLMD, is fine-tuned for tasks such as normalizing metadata and abstracting high-level representations, with an emphasis on patient care data over a decade-long period from up to 140 care sites. The study demonstrates that LLMD outperforms both general and domain-specific models, providing notable improvements in interpreting real-world patient data.

Unsloth AI has addressed a critical issue in Gradient Accumulation that has been affecting training and finetuning of language models (LLMs). The bug led to higher training losses for larger gradient accumulation steps. The Unsloth AI team formulated and implemented a fix, resulting in improved accuracy for gradient accumulation that aligns with full batch training. Updating Unsloth and using their fixed trainer, as advised, can significantly lower associated errors.

The GitHub repository 'Tamagotchi' by Barqawiz provides a collection of Tamagotchi characters designed to give AI assistants a unique identity, inspired by the classic virtual pet device of the same name. The project features characters that can be controlled manually or through AI assistance.

Kagi has introduced an AI Image Filter to improve image search results by downranking and labeling AI-generated images, providing users with the option to exclude them altogether. Users can also block specific websites featuring AI or low-quality imagery.

MotherDuck has announced a new SQL function called prompt(), which integrates small language models such as OpenAI's gpt-4o-mini into SQL data warehouses. The function is designed for text generation, summarization, and structured data extraction without separate infrastructure. It is now available in Preview for MotherDuck users to apply advanced natural language processing techniques more cost-effectively and with simplicity.

The AI Mathematical Olympiad Progress Prize is a Kaggle competition aimed at advancing artificial intelligence in solving complex mathematical problems. Participants develop models that can crack math olympiad questions, with a chance to win prizes and contribute to educational AI research.

Google has introduced new customization options for NotebookLM's Audio Overviews, allowing users to guide AI host emphasis and expertise level. The update paves way for NotebookLM Business, integrated with Google Workspace, with features geared towards business, university, and organization use, emphasizing data privacy and security.

Adobe's Project Turntable, revealed at the MAX conference, showcases an advanced AI-driven image rotation tool that allows full 3D rotation of 2D vector art while maintaining original design aesthetics. This innovative concept impresses with its ability to intelligently fill in unseen image details during rotation.

LOTUS is a state-of-the-art visual foundation model presented in arXiv 2024, using diffusion techniques for zero-shot dense geometry prediction tasks. Its novel approach and fine-tuning protocol allow it to excel in depth and normal estimation benchmarks with minimal training data, outperforming other methods in quality and efficiency.

A discussion on Hacker News explores the best practices for adding guard rails to LLM responses without disrupting streaming functionality. Participants share insights on using regex filters, LLAMA Guard, token sampling, and fine-tuning stop words for optimal performance.

Ichigo is an open-source, real-time voice AI project rebranded from llama3-s. It extends the text-based LLM's 'listening' capabilities, using an early fusion technique to combine text and audio inputs. The maintainers of Ichigo showcase ongoing progress, invite collaboration, and provide both tutorials and tools for engaging with the project.

Foyle is an AI system designed to make the operations of software deployment and management easier for developers without extensive knowledge in DevOps, by translating intent into executable commands within VScode notebooks.

GitHub repository 'qc_npu_benchmark' provides a code sample to benchmark models on Qualcomm's Windows PCs, specifically on Microsoft Surface tablets equipped with Qualcomm's NPU. The benchmarking reveals suboptimal NPU performance, achieving only a fraction of the claimed capabilities.

In a recent article, Pete Warden expresses his disappointment with the AI performance of Qualcomm's NPUs on Windows tablets. Despite the chips' promise of up to 45 trillion operations per second, Warden's benchmark tests using foundational AI operations only achieved 573 billion ops/sec, which is less than 1.3% of the advertised capability, and four times less than an Nvidia RTX 4080 in his gaming laptop. He speculates that the issues may stem from the software stack and hopes for future improvements through software updates.

Mistral AI announces world-class edge models, Ministral 3B and Ministral 8B, offering unparalleled efficiencies for on-device and at-the-edge applications. Setting benchmarks in knowledge, reasoning, and efficiency under 10B models, they are ideal for a range of use cases from smart assistants to autonomous robotics.

The National Archives is set to introduce 'Archie AI', a public-facing AI chatbot. Despite a ban on ChatGPT for posing a security risk, Google's Gemini AI has been promoted to enhance employee productivity. Employee concerns over AI's role in archiving accuracy, data security, and environmental impact arose during a presentation.

Meta showcases new open AI hardware designs including a new AI platform and advanced network fabrics at OCP Global Summit 2024. The initiative invites collaboration to drive innovation in AI, backed by Meta's development of infrastructure like Llama 3.1 405B model and 24K-GPU clusters.

OpenAI is reportedly transitioning from a nonprofit to a for-profit company, amidst a massive $6.6 billion funding round and a $157 billion valuation. Key executives, including CTO Mira Murati, have resigned. Doubts rise over OpenAI's sustainability as it faces enormous annual losses and a potential subprime AI crisis, despite ambitious revenue projections.