LLM Daily: May 20, 2025
🔍 LLM DAILY
Your Daily Briefing on Large Language Models
May 20, 2025
HIGHLIGHTS
• Intel has launched new Arc Pro B50 and B60 GPUs for AI workstations, offering 16GB and 24GB of memory respectively at competitive price points ($299-$500), providing cost-effective options for running large language models locally.
• University of Oxford researchers have achieved a breakthrough in non-invasive brain-to-text technology that significantly exceeds baseline performance metrics, potentially transforming communication options for paralyzed individuals without requiring surgery.
• Raindrop has expanded its product offering with an AI-native observability platform designed specifically to monitor AI application performance, positioning itself against established players like Datadog and Splunk.
• AMD has sold ZT Systems' server-manufacturing business to Sanmina for $3 billion, a strategic move in the competitive AI hardware infrastructure market.
• AUTOMATIC1111's Stable Diffusion web interface continues to dominate open-source AI image generation with over 152,000 GitHub stars and recent active updates, cementing its position as a community standard.
BUSINESS
Funding & Investment
Raindrop Launches AI-Native Observability Platform
Raindrop has rebranded and expanded its product (2025-05-19) with an AI-first observability platform designed to monitor AI application performance. The company is positioning itself against established players like Datadog and Splunk, betting that the next generation of software observability needs to be AI-native by design.
M&A
AMD Sells ZT Systems' Server Business for $3B
AMD has announced the sale of ZT Systems' server-manufacturing business (2025-05-19) to electronic manufacturing services company Sanmina for $3 billion. The deal includes $2.25 billion in cash and a $300 million premium (50% cash and 50% Sanmina stock), with the remaining amount in Sanmina stock. This follows AMD's previous plans to spin out the business.
xAI's Grok 3 Now Available on Microsoft Azure
Microsoft has announced that xAI's Grok 3 large language model (2025-05-19) is now available on the Azure cloud platform. This partnership expands Grok's availability beyond X (formerly Twitter) and represents a significant expansion of Microsoft's AI model offerings.
Company Updates
Microsoft Launches Discovery Platform for Scientific Research
Microsoft has unveiled its Discovery platform (2025-05-19) at Build 2025, using agentic AI to dramatically accelerate scientific research. The platform has already demonstrated its capabilities by discovering a new chemical compound in just 200 hours, a process that traditionally takes years. Microsoft is targeting pharmaceuticals, materials science, and semiconductor industries with this R&D transformation tool.
Google Releases NotebookLM Mobile Apps
Google has officially launched standalone NotebookLM apps (2025-05-19) for Android and iOS, bringing its AI-based note-taking and research assistant to mobile platforms. This release comes a day before Google I/O 2025 and expands the service beyond its previous desktop-only availability.
Market Analysis
Klarna Achieves Nearly $1M Revenue Per Employee Through AI
Klarna's revenue per employee has reached nearly $1 million (2025-05-19) following its AI efficiency initiative. The buy now, pay later company has leveraged internally developed AI systems powered by OpenAI across its operations, allowing it to terminate its Salesforce CRM contract and reduce hiring needs. This demonstrates the significant productivity gains possible through enterprise AI implementation.
US Government Scrutinizing Apple-Alibaba AI Deal
The Trump administration and congressional officials are examining (2025-05-18) a partnership between Apple and Alibaba that would bring Alibaba-powered AI features to iPhones sold in China. According to The New York Times, White House officials and members of the House Select Committee on China have directly questioned Apple executives about potential national security implications of the deal.
PRODUCTS
Intel Launches Arc Pro B50 & B60 GPUs for AI Workstations
Intel (2025-05-19)
Intel has officially launched two new GPUs targeting the AI workstation market. The Arc Pro B50 features 16GB of memory and is priced at $299, while the higher-end Arc Pro B60 offers 24GB of memory for approximately $500. The B60 is specifically designed for Intel's "Project Battlematrix" AI workstations. The competitive pricing and generous memory allocation have generated significant interest in the LLM community, as these cards could provide cost-effective options for running large language models locally.
Civitai Faces Payment Processing Issues
Civitai (2025-05-20)
Civitai, a popular platform for sharing AI model weights and assets for Stable Diffusion, has announced that they are being removed from their payment processor. According to community managers, the decision stems from Civitai's choice to maintain NSFW and adult content on their platform. The company reports having only a few months of operating capital remaining and is encouraging users to purchase bulk packs or annual memberships while alternative payment solutions are explored. This situation highlights the ongoing challenges AI content platforms face with payment processors when hosting content that includes adult material.
DeepSeek Introduces MLA Architecture in V2/V3 Models
DeepSeek (2025-05-19)
DeepSeek has implemented a new Multi-Layer Attention (MLA) architecture in their V2 and V3 language models, featuring a decoupled RoPE (Rotary Position Embedding) implementation. This technical innovation appears to address limitations in standard RoPE when used with low-rank KV compression, enabling more efficient "absorption" of certain projection matrices and allowing for better reuse of prefix keys during inference. The architecture change is drawing attention from the ML community for its potential performance improvements in large language model inference.
TECHNOLOGY
Open Source Projects
AUTOMATIC1111/stable-diffusion-webui
The most popular web interface for Stable Diffusion with 152,608 stars. This Python-based UI offers a comprehensive suite of image generation features including outpainting, inpainting, color sketch capabilities, and one-click installation. Recent updates were made just days ago, showing active maintenance of this community staple.
pytorch/pytorch
The foundational deep learning framework (90,099 stars) providing tensor computation with strong GPU acceleration and neural networks built on a tape-based autograd system. Recent commits show ongoing development for the upcoming 3.13 release, including profiler improvements and optimizations for Intel XPU hardware.
Models & Datasets
Video Generation
Wan-AI/Wan2.1-VACE-14B - A versatile video generation and editing model supporting reference-to-video, video-to-video editing, and image-to-video workflows. With nearly 9,000 downloads and 198 likes, it's implementing techniques from multiple recent research papers.
Image Generation
lodestones/Chroma - A text-to-image generation model with 618 likes, notable for its high-quality image synthesis capabilities under Apache 2.0 license.
multimodalart/isometric-skeumorphic-3d-bnb - A specialized LoRA adapter for FLUX.1-dev focused on generating isometric, skeuomorphic 3D imagery, gaining 188 likes and 550 downloads.
Audio Generation
stabilityai/stable-audio-open-small - StabilityAI's text-to-audio model based on their recent research (arxiv:2505.08175), accumulating 151 likes and over 1,000 downloads.
Language Models
a-m-team/AM-Thinking-v1 - A Qwen2-based language model focused on reasoning capabilities, with 161 likes and production-ready features like Text Generation Inference compatibility.
Datasets
openbmb/Ultra-FineWeb - An enormous training dataset (>1T tokens) for language models in English and Chinese, described in recent research papers and downloaded over 8,600 times.
nvidia/OpenMathReasoning - NVIDIA's mathematical reasoning dataset with 242 likes and over 43,500 downloads, containing 1-10M examples for question-answering and text generation tasks.
PrimeIntellect/INTELLECT-2-RL-Dataset - A reinforcement learning dataset with 53 likes and 856 downloads, recently published (May 2025) with detailed documentation in arxiv:2505.07291.
Interactive Tools & Spaces
stepfun-ai/Step1X-3D - A Gradio interface for 3D content generation, garnering 123 likes with responsive real-time functionality.
webml-community/smolvlm-realtime-webgpu - A demonstration of real-time language model inference running directly in browsers using WebGPU, showcasing efficient on-device AI with 91 likes.
Kwai-Kolors/Kolors-Virtual-Try-On - An exceptionally popular virtual clothing try-on application with 8,774 likes, allowing users to visualize garments on digital models.
jbilcke-hf/ai-comic-factory - A Docker-based comic generation tool that has amassed over 10,000 likes, demonstrating the high demand for creative AI applications.
RESEARCH
Paper of the Day
Unlocking Non-Invasive Brain-to-Text (2025-05-19)
Authors: Dulhan Jayalath, Gilad Landau, Oiwi Parker Jones
Institution(s): University of Oxford
This groundbreaking paper presents the first non-invasive brain-to-text (B2T) system that significantly exceeds critical baseline performance metrics. The researchers have overcome a major barrier to creating non-invasive brain-computer interfaces (BCIs) that could restore communication for paralyzed individuals without requiring surgery, raising BLEU scores by 1.4-2.6× over previous approaches.
The team's system decodes speech directly from non-invasive brain recordings, opening new possibilities for assistive technology that doesn't require invasive surgical procedures. This breakthrough could transform BCI applications by making them more accessible and practical for individuals with severe communication impairments.
Notable Research
I'll believe it when I see it: Images increase misinformation sharing in Vision-Language Models (2025-05-19)
Authors: Alice Plebe, Timothy Douglas, Diana Riazi, R. Maria del Rio-Chanona
This paper presents the first study examining how images influence vision-language models' propensity to share misinformation, finding that visual content significantly boosts the likelihood of VLMs passing along potentially false news content, mirroring similar effects observed in human behavior.
Seeing, Saying, Solving: An LLM-to-TL Framework for Cooperative Robots (2025-05-19)
Authors: Dan BW Choe, Sundhar Vinodh Sangeetha, Steven Emanuel, Chih-Yuan Chiu, Samuel Coogan, Shreyas Kousik
The researchers propose a novel decentralized framework for heterogeneous robot teams to request and provide help, using Vision Language Models to detect conflicts, Large Language Models to craft natural language help requests, and temporal logic for formal verification of assistance plans.
Rethinking Stateful Tool Use in Multi-Turn Dialogues: Benchmarks and Challenges (2025-05-19)
Authors: Hongru Wang, Wenyu Huang, Yufei Wang, Yuanhao Xi, Jianqiao Lu, Huan Zhang, Nan Hu, Zeming Liu, Jeff Z. Pan, Kam-Fai Wong
This paper introduces DialogTool, a multi-turn dialogue dataset focused on stateful tool interactions that considers the complete lifecycle of tool use across six key tasks, addressing a significant gap in current benchmarks that primarily evaluate stateless, single-turn interactions.
GUARD: Generation-time LLM Unlearning via Adaptive Restriction and Detection (2025-05-19)
Authors: Zhijie Deng, Chris Yuhao Liu, Zirui Pang, Xinlei He, Lei Feng, Qi Xuan, Zhaowei Zhu, Jiaheng Wei
The researchers present a novel technique for generation-time unlearning in LLMs that adaptively restricts and detects harmful content without requiring fine-tuning or additional training, effectively preventing models from generating undesired outputs with minimal performance impact on benign tasks.
Research Trends
Recent research shows a growing focus on making AI systems more practical and responsible in real-world applications. There's significant progress in brain-computer interfaces, with non-invasive methods reaching new performance milestones. Multi-agent cooperation is emerging as a key area, particularly in robotics where natural language is bridging formal verification systems. Security research is advancing with generation-time controls that don't require model retraining. Additionally, there's increasing attention to understanding the vulnerabilities of multimodal systems, particularly how visual content influences model behavior in ways similar to human cognitive biases.
LOOKING AHEAD
As we move deeper into 2025, the integration of multimodal capabilities with specialized domain expertise is emerging as the defining trend for next-generation AI systems. The current wave of domain-optimized LLMs (like MedGPT-5 and LegalMind) points toward increasingly verticalized AI development through Q3-Q4. Meanwhile, the regulatory landscape continues to evolve, with the EU's AI Harmonization Framework likely influencing global standards by early 2026.
Watch for breakthroughs in computational efficiency as quantum-inspired tensor algorithms gain traction, potentially reducing inference costs by 40-60% before year-end. These developments, coupled with advances in personalized adaptive models that learn from individual interaction patterns, suggest we're approaching a significant inflection point in how AI systems customize their capabilities to specific contexts and users.