Tuesday — January 21, 2025
Meta faces a lawsuit over pirated books while its DeepSeek-R1 model surpasses GPT-4o in benchmarks, and Humbug showcases a mostly AI-built development environment.
News
Authors seek Meta's torrent client logs and seeding data in AI piracy probe
Meta, the company behind Facebook and Instagram, is being sued by authors who claim the company used their copyrighted works without permission to train its AI models, with evidence showing that Meta used BitTorrent to download pirated books from shadow libraries like LibGen. The court has allowed the authors to amend their complaint to include allegations that Meta not only downloaded but also "seeded" pirated books to other users, potentially operating as a distributor of copyrighted material and undermining its fair use defense.
It sure looks like Meta stole a lot of books to build its AI
Meta, the company formerly known as Facebook, is facing a grim week with CEO Mark Zuckerberg making controversial announcements, including allowing slurs on their platforms and adding a pro-Trump figure to their board. The company is also embroiled in a copyright lawsuit, with recently unredacted court documents revealing that Meta used a database of pirated books to train its AI systems, with employees admitting to using stolen material from a notorious piracy site.
Poland fumes over US block on AI chips
Poland is upset over a US decision to limit the export of artificial intelligence chips to the country, which could impact its tech sector and military expansion. The Polish government is seeking to reverse the move, citing its strategic partnership with the US and is also considering EU action, with the Digital Affairs Minister asking the Commissioner for technological security to take decisive action.
DeepSeek-R1-Distill-Qwen-1.5B Surpasses GPT-4o in certain benchmarks
DeepSeek-R1 is a reasoning model that achieves performance comparable to OpenAI-o1 across math, code, and reasoning tasks, and its distilled smaller models have outperformed OpenAI-o1-mini in various benchmarks. The model was trained using large-scale reinforcement learning and incorporates cold-start data to address issues such as endless repetition and poor readability, resulting in state-of-the-art results for dense models.
The AI Bubble Is Bursting
Google unexpectedly made its Gemini AI feature free and enabled it by default for all users, despite the author's previous evaluation that it was an inferior service, and now the only options to disable it are to upgrade to a more expensive account or beg customer service. This move is seen as a desperate attempt by Google, as well as other companies like Microsoft and Apple, to push their AI features onto users after failing to generate sufficient interest and revenue from them.
Research
Evolving Deeper LLM Thinking
Mind Evolution, a search strategy for Large Language Models, generates and refines candidate responses to scale inference time compute, outperforming other strategies like Best-of-N and Sequential Revision. This approach solves over 98% of problem instances in benchmarks like TravelPlanner and Natural Plan without a formal solver, demonstrating its effectiveness in natural language planning tasks.
Computational geometry with probabilistically noisy primitive operations
Researchers have developed a new approach to designing computational geometry algorithms that can handle errors and imprecision by studying algorithms subject to noisy Boolean operations. They propose a technique called path-guided pushdown random walks, which they use to solve various geometric problems, such as point-location and convex hulls, in optimal time with high probability despite the presence of noise.
Accelerating Retrieval-Augmented Generation
Retrieval-Augmented Generation (RAG) is a solution to improve accuracy in large language models by augmenting them with information from external knowledge sources, and this paper explores ways to optimize the retrieval phase of RAG. The authors propose Intelligent Knowledge Store (IKS), a device that accelerates exact nearest neighbor search, achieving 13.4-27.9x faster search times and 1.7-26.3x lower end-to-end inference times for RAG applications.
Code
DeepSeek-R1
DeepSeek-R1 is a reasoning model that achieves performance comparable to OpenAI-o1 across math, code, and reasoning tasks, and its distilled models, such as DeepSeek-R1-Distill-Qwen-32B, outperform OpenAI-o1-mini on various benchmarks. The model is open-sourced, along with six dense models distilled from DeepSeek-R1, to support the research community and advance the development of better models.
Show HN: Personalized Duolingo (kind of) for vocabulary building
The WordPecker App is a personalized language-learning platform that allows users to create custom vocabulary lists and learn from them in a Duolingo-style format, with features like automatic definitions, interactive quizzes, and progress tracking. The app aims to make language learning more efficient and effective by tying learning to the context in which words were originally encountered, and it has a roadmap for future development that includes additional question types, progress tracking, and integration with other platforms.
Show HN: Humbug – an open source AI dev environment mostly built by AI
Humbug is a GUI-based application that facilitates interaction with various AI backends, offering a user-friendly tabbed interface for managing multiple conversations and editing files. The application is designed to support building things using AI and is itself largely built by AI, using a Metaphor description of the application to provide context to the AI backend.
Show HN: i18n-ai-translate
The i18n-ai-translate tool leverages AI models like ChatGPT, Gemini, Ollama, or Claude to seamlessly translate localization files, supporting directories of nested translation files in i18next-style JSON format. It can be used as a GitHub Action, run directly, or integrated as a library in a project to translate i18n JSON files to any language, with options for customizing the translation process and verifying the results.
Show HN: Open-source conversational AI agents for internal tools
Inferable is an open-source platform that allows users to convert their existing internal APIs, functions, and scripts into autonomous agents that can be interacted with conversationally. It offers a range of features, including a self-hostable control plane, distributed tool calling with fault-tolerance, and native SDKs for various languages, making it easy to integrate and deploy autonomous agents in different environments.