[AINews] DeepSeek R1: o1-level open weights model and a simple recipe for upgrading 1.5B models to Sonnet/4o level • ButtondownTwitterTwitter

The AI Twitter Recap section provides updates on the latest developments related to the DeepSeek-R1 model, benchmarking comparisons with OpenAI-o1, reinforcement learning in LLM training, open-source models and distillation, as well as insights from AI research papers and technical discussions. The section also includes humorous takes, satirical comments on AI hype, and playful interactions within the AI community.

In this section, various instances of humor injected into technical AI discussions are highlighted. From playful responses to AI topics on Twitter to lighthearted AI jokes and funny AI-related memes, the community adds a touch of levity to the usually serious conversations. This human element brings a new dimension to the technical discourse, making it more engaging and relatable.

• Open-Source LLM Rivalries:
  • DeepSeek R1 Roars Past OpenAI’s o1: The model matches o1’s reasoning benchmarks at a fraction of the cost and is available under an MIT license.
  • Kimi k1.5 Slams GPT-4o in a 128k-Token Duel: Kimi k1.5 outperforms GPT-4o and Claude Sonnet 3.5 in code and math tasks.
  • Liquid LFM-7B Dares to Defy Transformers: Liquid AI introduces LFM-7B with superior throughput on a 7B scale, claiming best-in-class language support.
• Code & Agentic Tools:
  • Windsurf Wave 2 Surfs with Cascade & Autogenerated Memories: Users appreciate its web and doc search capabilities but face performance issues in large-file contexts.
  • Cursor Stumbles in Sluggish Showdown: Developers criticize Cursor for delays and code deletion mishaps, considering a shift to other AI editors.
  • Aider 0.72.0 Scores with DeepSeek R1: Aider’s latest version introduces DeepSeek R1 support for unified code generation across Kotlin and Docker.
• RL & Reasoning Power-Ups:
  • GRPO Simplifies PPO for DeepSeek: Introduction of Group Relative Policy Optimization enhances DeepSeek R1’s math and code solutions.
  • Google’s Mind Evolution Outsmarts Sequential Revision: Shows exceptional performance with Gemini 1.5 Pro in planning benchmarks.
  • rStar-Math Gambles on MCTS: Trains small LLMs to excel in math tasks using Monte Carlo Tree Search.
• HPC & Hardware High Jinks:
  • M2 Ultras Tag-Team DeepSeek 671B: Developers achieve real-time speeds with M2 Ultras for running large LLM models.
  • GPU vs CPU Smackdown: Debate on GPU’s parallelization versus CPU’s speed for various tasks.
  • KV Cache Quantization Boosts LM Studio: Llama.cpp engine enhances memory-friendly inference with quantization, gaining throughput on consumer-grade hardware.
• Partnerships & Policy Kerfuffles:
  • Microsoft’s $13B OpenAI Bet Spooks the FTC: Concerns over AI partnerships impacting startup competition.
  • FrontierMath Funding Cloaked in NDA: OpenAI's quiet support for math datasets draws criticism for lack of transparency.
  • TikTok Merger Talk Tangles with Perplexity: Perplexity’s rumored merger with TikTok raises questions about synergy beyond headlines.

The LM Studio Discord section highlights the latest LM Studio 0.3.7 update featuring support for the DeepSeek R1 model. Community members praised the open source approach and the robust reasoning capabilities of DeepSeek_R1. Additionally, discussions revolved around topics such as KV Cache quantization, local file storage in LM Studio, GPU comparisons, distributed inference with M2 Ultras, and the performance of the Kimi k1.5 model. The section also covers discussions on stability in AI models, including the Liquid Foundation Model LFM-7B and the DeepSeek v3 & SGLang for 'Mission Critical Inference.' Topics like GPUs, Microsoft's investment in OpenAI, and advancements in AI logic calculations were also explored.

Participants in this section discussed various topics related to artificial intelligence and its applications. They talked about the recent release of DeepSeek-R1 and its successful distillation process, showcasing excitement for open-source reasoning models. There was a debate on how AI could integrate with the crypto space for trading resources. Members also expressed interest in MiniCPM-o 2.6 model capabilities for vision, speech, and multimodal applications. Discussions on reinforcement learning focused on the use of outcome rewards to optimize model performance. Additionally, concerns were raised about hosting providers' performance for projects like Hermes 3 405B, prompting discussions on alternatives for reliable hosting options.

Discussions in the Nous Research AI channel cover a variety of topics related to AI research and technologies. The section includes conversations on high accuracy OCR models facing misreading challenges, comparisons between MOEs and dense models, the impact of structured sparsity on model efficiency, the use of depthwise MLP blocks, and questions about cosine warmup decay scheduler when training GPT-2 models. Additionally, discussions touch on research collaborations for climate change, Google's Mind Evolution method outperforming other strategies, Liquid AI launching LFM-7B, and speculations on unique business models in the AI industry.

This section discusses various aspects of GPT-NeoX development including guidance on host RAM and CPU core requirements, optimizing vocabulary size divisibility for efficiency, exploring MP, PP, and ZeRO Stage 1 for performance enhancement, and raising issues for hangs during the development process. Members share insights on best practices and potential risks associated with deviations from standard settings, emphasizing the importance of detailed information when reporting issues to facilitate troubleshooting.

Vagueposting Reaches New Heights: A member shared a graphic titled 'vagueposting end game', emphasizing the trend of ambiguous communication in online spaces. The attached image hints at the complexities of deciphering modern digital dialogue. - Discussion on AI's Last Moat: A member referenced a tweet claiming that 'the only moat left in AI is Amanda Askell', sparking conversations about competitive advantages in the field. This statement reflects growing sentiments regarding intellectual property and unique insights in the rapidly evolving AI landscape.

Interconnects (Nathan Lambert)

• Exploring RLVR for Robotic Control: Discussion on the applicability of RLVR for robotic control using VLMs and CoT.
  • Optimism expressed about the method's potential success.
• Vintage Ideas Resurfacing: Highlighting the resurgence of old ideas in robotics, specifically in reinforcement learning.
  • Emphasis on the need to explore past concepts for innovation.
• Computer Vision Applications of RL: Sharing a paper by Lucas Beyer et al. on aligning models with task rewards in computer vision.
  • Effectiveness in aligning models across tasks like object detection and image captioning.
• Combining RL with CoT Approaches: Inquiry about merging RL approaches with Chain of Thought methodologies in computer vision.
  • Concerns raised about the reliability of computer vision labels for tasks using RL.
• Perception Models Timeline Conundrum: Humorous suggestion for an experimental timeline alongside perception model deliveries.
  • Hinting at the balance between pursuing innovative ideas and meeting standard deliverables.

DeepSeek R1 Launch

The release of LM Studio 0.3.7 introduces support for DeepSeek R1 and an updated llama.cpp engine version 1.9.2, accessible via in-app updates. Users can also download various distilled models from DeepSeek, offering sizes up to 70B, designed to enhance performance.

DeepSeek R1: A Game Changer in Reasoning Models

The DeepSeek R1 model is now available for download, promising open source reasoning capabilities on par with OpenAI's o1 model, with details found in the technical report. Users will notice outputs from DeepSeek R1 encapsulated in <think></think> tags, showcasing its reasoning processes.

Enhanced Mission Control Features

A Hardware tab has been added to Mission Control, which can be accessed using Cmd/Ctrl + Shift + H, offering users more monitoring capabilities. Additionally, a server file logging mode allows for more granular control over what log entries are made.

KV Cache Quantization for Improved Performance

The latest version comes with KV Cache quantization for llama.cpp models, enhancing the efficiency of the runtime environment requiring version 1.9.0+. This feature aims to optimize performance metrics while handling model predictions.

Links mentioned: Download LM Studio - Mac, Linux, Windows, LM Studio 0.3.7

The Latent Space ▷ ai-announcements section covers the latest updates and announcements related to AI technologies and developments. This includes discussions on topics such as the O1 Podcast follow-up, features of DeepSeek v3, Mission Critical Inference, and Kubernetes challenges. The section highlights important insights shared by guests regarding mission critical inference strategies and optimizations in DeepSeek, emphasizing the Three Pillars of Mission Critical Inference. It also delves into discussions on rate limits for models, user experiences with DeepSeek R1, and accessing reasoning content through the API. This section provides a comprehensive overview of the advancements and discussions in the field of AI within the Latent Space Discord community.

Microsoft's Investment in OpenAI Raises Antitrust Concerns

The FTC expressed concerns about Microsoft's $13 billion investment in OpenAI, fearing it may enhance the company's dominance in the AI market. FTC Chair Lina Khan highlighted the potential for lock-in and disadvantaging start-ups.

Microsoft Researchers on AI System Security

Microsoft researchers concluded in a pre-print paper that AI systems can never be fully secure, highlighting existing security risks and introducing new vulnerabilities. They warned about prevalent threats like gradient-based attacks.

AI Tools for Wall Street Compliance

Compliance firms are utilizing AI to decode trader communications to detect financial crimes, especially with stringent regulatory oversight.

Supreme Court's Decision on TikTok Ban

The Supreme Court upheld a law requiring TikTok to be sold by its Chinese parent or face a ban due to national security concerns, impacting app availability.

Controversy over FrontierMath's Funding

The funding connection between OpenAI and FrontierMath is scrutinized, raising concerns about NDA restrictions and lack of financial transparency for contributors.

This section discusses various topics related to Cohere, including plans for a Konkani language AI model, concerns about Cohere's accessibility, engagement with Cohere API access, feedback on Cohere's interface, model switching and updates. Additionally, the section covers billing questions, challenges with AI behavior, limitations in AI behavior management, versioning for Command-R model, embed job limitations, integration issues with Dify.ai, and a holiday notice for support responses. Furthermore, there is a discussion on Cohere's math performance, limitations of LLMs, and tool usage tips. Lastly, the section includes messages related to a MOOC course confirmation, a document to podcast blueprint by Mozilla AI, and community engagement through open source projects.