开源 AI
1Cursor Acquires Graphite(graphite.com)💡 The story announces the acquisition of Graphite (a code review platform) by Cursor (an AI IDE), which falls under business and corporate news. 2Graphite Is Joining Cursor(cursor.com)💡 The story announces Cursor's acquisition of Graphite, a code review platform, which falls under business and corporate news in the 'company' category. 5FunctionGemma 270M Model(blog.google)💡 The story is about the release of the FunctionGemma 270M model via Google's technology developers blog, which falls under the Models category (model releases). 6Mistral OCR 3(mistral.ai)💡 The story announces the release of Mistral OCR 3, an AI model for document processing, which aligns with the 'models' category covering model releases and updates. 7Prompt caching for cheaper LLM tokens(ngrok.com)💡 Prompt caching is an optimization technique for LLM inference to reduce token costs, which falls under the 'infra' category covering deployment and inference-related topics. 8DeepSeek uses banned Nvidia chips for AI model, report says(finance.yahoo.com)💡 The story centers on Chinese AI startup DeepSeek using smuggled Nvidia Blackwell chips (a key AI hardware component banned in China by U.S. regulations) to develop an upcoming AI model, aligning with the Hardware category's focus on AI chips and compute. 10Mistral releases Devstral2 and Mistral Vibe CLI(mistral.ai)💡 The story centers on Mistral AI's release of Devstral 2, a coding model family with benchmark results (SWE-bench Verified) and competitor comparisons, which aligns with the Models category per Rule 3 (model releases → models regardless of company origin). The Mistral Vibe CLI is an accompanying tool, but the core announcement focuses on the model release. 11Zebra-Llama – Towards efficient hybrid models(arxiv.org)💡 The story focuses on the release of Zebra-Llama, a new hybrid language model family with 1B/3B/8B variants, including new architecture elements (SSM and MLA combination) and performance comparisons—all fitting the Models category criteria. 12Prompt Injection via Poetry(wired.com)💡 The article focuses on prompt injection via poetry to bypass AI safeguards for nuclear weapon-related assistance, which directly falls under AI safety (security and red teaming). 14Mistral 3 family of models released(mistral.ai)💡 The story is an announcement of the Mistral 3 family of AI models, including small dense models and Mistral Large 3 (a mixture-of-experts model), which directly falls under the models category for model releases. 16DeepSeek-v3.2(huggingface.co)💡 The story centers on the release of DeepSeek-V3.2, an AI model with details on its capabilities (efficient reasoning, agent performance), technical innovations (DSA attention mechanism, scalable RL framework), and benchmark achievements (surpassing GPT-5 variant, IMO/IOI gold medals), fitting the 'models' category for releases and capabilities. 2373% of AI startups are just prompt engineering(pub.towardsai.net)💡 The article focuses on the business practices of AI startups, revealing that most rely on third-party APIs instead of proprietary tech, which falls under business & corporate news in the ecosystem category. 24Measuring the impact of AI scams on the elderly(simonlermen.substack.com)💡 The story focuses on AI model jailbreaking for phishing scams and measuring their impact on the elderly, which aligns with AI security and safety issues including jailbreaks. 25Continuous Autoregressive Language Models(arxiv.org)💡 The story focuses on an ArXiv paper introducing CALM, a new language model architecture shifting from discrete token prediction to continuous vector prediction, which falls under the Models category as it includes new architecture releases like Mamba or SSM.