Thursday — February 6, 2025
Google's AI ethics policy shift allows weapon and surveillance applications, SmolLM2 advances small language model performance with 1.7 billion parameters, and researchers leverage LLMs to extract characters and relationships from entire books.
News
Huawei's Ascend 910C delivers 60% of Nvidia H100 inference performance
Huawei's Ascend 910C processor delivers 60% of Nvidia's H100 inference performance, according to research by DeepSeek, making it a potential option for reducing China's reliance on Nvidia GPUs. Although the Ascend 910C is not suitable for AI training due to its limited long-term training reliability, it can be optimized for inference performance, offering a cost-effective alternative for AI companies.
Fair Pricing
Kagi has introduced a "Fair Pricing" policy, where users will receive a full credit for months they don't use any searches on their plan, which can be applied to their next billing cycle. The company has also made various improvements and bug fixes to its search and assistant features, including a redesigned Assistant interface, improved translation capabilities, and a dedicated SVG filter for image search.
Google erases promise not to use AI technology for weapons or surveillance
Google has updated its AI ethics policy, removing its previous promise not to use the technology for applications related to weapons or surveillance, marking a significant shift in the company's stance on the responsible development of artificial intelligence. The change comes as the AI race accelerates and regulations on transparency and ethics in AI have yet to catch up, with Google citing the need for democracies to lead in AI development guided by core values like freedom and human rights.
Lightning AI hub: Production AI in your cloud in minutes
The platform offers a range of AI models and APIs for various tasks, including text generation, image creation, speech-to-text, and more, which can be deployed as autoscaling APIs with zero code and paid for on a pay-as-you-go basis. The featured models and APIs include language models like Llama and DeepSeek, image generation models, and other tools for tasks such as text classification, sentiment analysis, and image upscaling.
Google drops pledge not to use AI for weapons, surveillance
Google has dropped its pledge not to use artificial intelligence for weapons or surveillance, updating its ethics policy to state that it will use AI in line with international law and human rights. The company's revised policy, announced on Tuesday, no longer includes language about avoiding the development of AI technologies that could cause harm, such as weapons or surveillance that violate international norms.
Research
Harmonic Loss Trains Interpretable AI Models
Harmonic loss is introduced as an alternative to cross-entropy loss for training neural networks, offering improved interpretability and faster convergence due to its scale invariance and finite convergence point. Models trained with harmonic loss outperform standard models, enhancing interpretability, requiring less data for generalization, and reducing grokking, making it a valuable tool for domains with limited data or high-stakes applications.
SmolLM2: When Smol Goes Big – Data-Centric Training of a Small Language Model
The development of SmolLM2, a 1.7 billion parameter language model, is documented, which achieves strong performance through a multi-stage training process on a large dataset of around 11 trillion tokens. SmolLM2 outperforms other recent small language models, and to facilitate future research, both the model and the datasets used in its development are being released.
Harmonic Loss Trains Interpretable AI Models
Harmonic loss is introduced as an alternative to cross-entropy loss for training neural networks, offering improved interpretability and faster convergence due to its scale invariance and finite convergence point. Models trained with harmonic loss outperform standard models, enhancing interpretability, requiring less data for generalization, and reducing grokking, making it a valuable tool for domains with limited data or high-stakes applications.
Digital Agent outperforms o1 by 15% – trained with new RL-variant similar to R1
Interactive digital agents (IDAs) can be trained using reinforcement learning (RL) to perform tasks in digital environments, outperforming existing methods. A 32-billion-parameter agent trained with the proposed LOOP approach achieved a 9 percentage point improvement over a larger agent in the AppWorld environment, demonstrating the effectiveness of RL in this area.
Sundial: A Family of Highly Capable Time Series Foundation Models
Sundial is a family of native, flexible, and scalable time series foundation models that can generate multiple probable predictions for arbitrary-length time series without specifying any prior distribution. The Sundial models, pre-trained on the large TimeBench dataset, exhibit unprecedented model capacity and generalization performance, achieving state-of-the-art results on both point forecasting and probabilistic forecasting benchmarks.
Code
Show HN: Throw a Whole Book into an LLM to Extract Characters and Relationships
Researchers used a large language model (LLM) called Gemini 2.0 Flash Exp to extract character information from entire books, including relationships and physical descriptions, and visualized the results using HTML/JS and D3. The experiment showed promising results for smaller books, but struggled with larger books due to the 8K token output limit, highlighting the need for further refinement and exploration of LLM capabilities.
RA.Aid
RA.Aid is a tool that enables near-fully-autonomous software development by leveraging AI-powered code editing capabilities and an agent-based task execution framework. It provides a three-stage architecture for research, planning, and implementation of complex programming tasks, allowing developers to automate tasks such as code refactoring, research, and implementation of new features.
Show HN: Former – Open-source Cursor for SQL
Former is a desktop and web SQL editor that utilizes AI assistance to improve the SQL writing experience, offering a more efficient alternative to copying and pasting database context and SQL into external AI tools. It can be used through a cloud-hosted instance with a free tier and trial period, or self-hosted for free by compiling the Electron app and setting up a NextJS server and Supabase instance.
Show HN: Nominate – a macOS app that renames PDFs based on their contents
Nominate is a macOS app that uses AI and NLP to automatically rename PDF files based on their contents, extracting a timestamp and summary to create a short, descriptive filename. The app processes documents on-device, keeping them private, and allows users to review and apply suggested filename changes with a simple drag-and-drop interface.
Show HN: RoundTable – An AI chatbot starter kit for Ruby on Rails
RoundTable is an AI chatbot project that simulates a round table discussion between Albert Einstein, Isaac Newton, and Nikola Tesla, serving as a starting point for developers to build their own AI chatbot. The project is built using Ruby on Rails and Turbo Streams, and can be easily deployed to Heroku, with features such as real-time UI updates and compatibility with various AI providers.