LLM Daily: September 29, 2025
🔍 LLM DAILY
Your Daily Briefing on Large Language Models
September 29, 2025
HIGHLIGHTS
• South Korea has unveiled a major sovereign AI initiative bringing together national tech leaders like LG and SK Telecom to develop homegrown large language models, positioning itself to compete with global AI powerhouses like OpenAI and Google.
• A comprehensive educational repository "LLMs-from-scratch" with over 73,000 GitHub stars offers step-by-step implementation of ChatGPT-like models in PyTorch, serving as a practical code companion to understand modern LLM architecture fundamentals.
• Researchers have introduced Tree-based Group Relative Policy Optimization (Tree-GRPO), a breakthrough approach for LLM agent reinforcement learning that shows up to 80% improvement in decision-making quality on long-term, multi-turn tasks.
• Major tech giants including Meta, Oracle, Microsoft, Google, and OpenAI are driving an infrastructure boom with billion-dollar investments in AI computing resources, representing some of the largest capital expenditures in the tech industry.
• A new differentiable parametric curves library for PyTorch enables backpropagation to both curve inputs and parameters, offering valuable applications for continuous embeddings and as an alternative in Knowledge Augmented Networks.
BUSINESS
Billion-Dollar Infrastructure Powers AI Boom
TechCrunch has published an in-depth analysis of the major infrastructure deals driving the current AI boom (2025-09-28). The report details significant investments from tech giants including Meta, Oracle, Microsoft, Google, and OpenAI in large-scale AI infrastructure projects. These investments represent some of the biggest capital expenditures in the tech industry, highlighting the intense competition for AI computing resources. Link to article
South Korea Launches Ambitious National AI Initiative
South Korea has unveiled a major sovereign AI initiative to compete with global leaders like OpenAI and Google (2025-09-27). The initiative brings together the country's major tech players including LG and SK Telecom to develop homegrown large language models. This represents a significant national investment in AI sovereignty and capabilities. Link to article
Sequoia Capital Backs AI-Powered Recruiting Platform Juicebox
Sequoia Capital announced their investment in Juicebox, an AI-enhanced recruiting platform that has gained significant traction among founders (2025-09-25). The VC firm highlighted Juicebox's innovative approach to talent acquisition as a key factor in their investment decision. This funding signals continued venture capital interest in AI-powered solutions for the recruiting industry. Link to announcement
Analyst Warning: AI Services Transformation More Complex Than Anticipated
A new analysis by TechCrunch suggests that the transformation of service industries through AI may be more challenging than venture capital firms are anticipating (2025-09-28). The article points to early warning signs of complexity and resistance in various service sectors, potentially indicating a more extended timeline for AI adoption in these industries than current market projections suggest. Link to analysis
AI Startup Friend Invests Over $1M in NYC Subway Ad Campaign
AI wearable startup Friend has spent more than $1 million on a high-visibility advertising campaign throughout the New York City subway system (2025-09-27). The startup's significant marketing investment highlights growing competition in the consumer AI device market and represents an unusual go-to-market strategy for an early-stage AI company. Link to article
PRODUCTS
New Tool for PyTorch: Differentiable Parametric Curves Library
Released by: Alex Shtf (individual developer) - [2025-09-28]
Link: Reddit announcement
A new open-source library has been released that enables differentiable parametric curves in PyTorch. The library currently supports B-Spline curves (with efficient sparsity exploitation) and Legendre Polynomials, with all operations vectorized for mini-batch processing and multiple curves simultaneously. The tool allows backpropagation to both curve inputs and parameters, making it valuable for applications including continuous embeddings for embedding-based models and as an alternative to B-Splines in Knowledge Augmented Networks (KANs).
Qwen Image Model Shows Impressive LoRA Training Results
Company: Qwen (Alibaba Cloud) - [2025-09-28]
Link: Community demonstration
A community member has demonstrated impressive results training their first LoRA (Low-Rank Adaptation) model using Qwen's image generation model. The training was performed using Diffusion Pipe with default settings on RunPod, and the results have garnered significant attention for their quality. The developer shared their complete training setup, providing a valuable reference for others looking to achieve similar results with Qwen models.
HunyuanImage 3.0 Gains Positive Reception
Company: Tencent - [2025-09-28]
Link: Community discussion
Tencent's HunyuanImage 3.0 is receiving acclaim from the AI image generation community, with users describing it as "perfect." This Chinese multimodal AI model appears to be delivering high-quality image generation capabilities that are impressing even experienced users in the space. The enthusiasm suggests Hunyuan is establishing itself as a noteworthy competitor in the increasingly crowded field of image generation models.
TECHNOLOGY
Open Source Projects
rasbt/LLMs-from-scratch
A comprehensive educational repository that walks through implementing a ChatGPT-like LLM in PyTorch from scratch, step by step. The project serves as the official code companion to the book "Build a Large Language Model (From Scratch)" and demonstrates each component of modern LLM architecture with practical code examples. With over 73,000 stars and recent updates for Python 3.13 compatibility, this repository remains one of the most popular educational resources for understanding LLM fundamentals.
Shubhamsaboo/awesome-llm-apps
A curated collection of real-world LLM application architectures using AI Agents and RAG with various models from OpenAI, Anthropic, Gemini, and open-source options. With over 70,000 stars, the repository recently added a YAML-based Multi-Agent Web Research System with Firecrawl MCP for low-code implementation. This resource provides practical patterns and examples for developers building production AI applications.
harry0703/MoneyPrinterTurbo
An end-to-end system that leverages AI models to automatically generate high-quality short videos with minimal user input. The project has gained significant momentum with over 40,000 stars and nearly 6,000 forks. Recent updates have focused on performance optimizations in the video processing pipeline, removing unnecessary operations that improve generation efficiency.
Models & Datasets
Qwen/Qwen3-Omni-30B-A3B-Instruct
Alibaba Cloud's multimodal LLM that supports any-to-any generation, including text-to-audio capabilities. With over 95,000 downloads and 500+ likes, this 30B parameter model extends beyond text-only interactions to handle complex multimodal tasks while remaining compatible with Hugging Face endpoints for easy deployment.
tencent/HunyuanImage-3.0
The latest iteration of Tencent's text-to-image diffusion model featuring a Mixture of Experts (MoE) architecture for improved generation quality and efficiency. Despite being recently released, it has already gathered over 450 likes, showcasing significant interest in this alternative to Midjourney and DALL-E.
tencent/Hunyuan3D-Part
A specialized 3D generation model from Tencent that focuses on part segmentation and generation, building on their Hunyuan3D-2.1 base model. Trained on datasets like Objaverse and Objaverse-XL, this model enables granular control over 3D object components and has already collected 365 likes despite its recent release.
ibm-granite/granite-docling-258M
IBM's document understanding model built on the IDEFICS3 architecture, specialized for processing and extracting information from complex documents including code, formulas, charts, and tables. With nearly 100,000 downloads and 750+ likes, this model excels at OCR, layout analysis, and structured information extraction tasks.
openai/gdpval
OpenAI's multimodal validation dataset spanning audio, document, image, text, and video modalities. Released on September 25th, it already has nearly 5,000 downloads and is designed to serve as a standardized benchmark for evaluating general-purpose models across different media types and tasks.
nvidia/Nemotron-Personas-Japan
A Japanese-language synthetic dataset from NVIDIA containing over 1 million text-image pairs focused on creating coherent AI personas. With nearly 6,000 downloads since its release on September 24th, this dataset supports the development of culturally appropriate AI assistants for the Japanese market.
meta-agents-research-environments/gaia2
Meta's second-generation simulated environment for evaluating agent capabilities in complex, dynamic scenarios. Published alongside arXiv paper 2509.17158, this dataset focuses on temporal reasoning, dynamic events, and multi-agent interactions. It has seen rapid adoption with over 13,800 downloads since its September 25th release.
Developer Tools & Infrastructure
Wan-AI/Wan2.2-Animate
A popular Gradio-based web interface for image animation with nearly 1,000 likes. The tool provides an accessible way to create motion from still images without requiring local GPU resources or complex setup, making animation technology available to non-technical users.
prithivMLmods/Photo-Mate-i2i
A Gradio-powered image-to-image transformation tool that enables sophisticated photo editing and style transfer. Operating through the Hugging Face MCP server infrastructure, this space allows users to apply complex image manipulations through an intuitive interface.
not-lain/background-removal
A highly popular background removal tool with over 2,300 likes that runs on Hugging Face's MCP server architecture. This utility offers efficient, high-quality background removal without requiring users to install specialized software or manage their own computational resources.
multimodalart/ai-toolkit
A Docker-based collection of AI tools bundled into a single deployable package for comprehensive AI workflows. With increasing popularity (51 likes), this toolkit represents a shift toward containerized AI application deployment for easier distribution and consistent runtime environments.
RESEARCH
Paper of the Day
Tree Search for LLM Agent Reinforcement Learning (2025-09-25)
Authors: Yuxiang Ji, Ziyu Ma, Yong Wang, Guanhua Chen, Xiangxiang Chu, Liaoni Wu
This paper introduces Tree-based Group Relative Policy Optimization (Tree-GRPO), addressing a critical challenge in LLM agent reinforcement learning: sparse supervision in long-term, multi-turn tasks. The significance lies in its novel approach to use tree search where each node represents a complete agent interaction step, enabling more effective optimization across different action paths. The authors demonstrate that this method significantly outperforms traditional RL approaches on benchmark tasks, showing up to 80% improvement in decision-making quality while maintaining the natural language capabilities of the base LLM.
Notable Research
RLBFF: Binary Flexible Feedback to bridge between Human Feedback & Verifiable Rewards (2025-09-25)
Authors: Zhilin Wang, Jiaqi Zeng, et al.
Proposes a novel reinforcement learning paradigm that combines the strengths of RLHF and RLVR by using binary feedback that is both interpretable and broader in scope than purely correctness-based verifiers, showing improved performance across multiple benchmarks.
OnePiece: Bringing Context Engineering and Reasoning to Industrial Cascade Ranking System (2025-09-22)
Authors: Sunhao Dai, Jiakai Tang, et al.
Introduces a framework that applies LLM-inspired context engineering and reasoning to industrial recommendation systems, moving beyond simply borrowing Transformer architectures to fundamentally rethink how context shapes user recommendations.
SGMem: Sentence Graph Memory for Long-Term Conversational Agents (2025-09-25)
Authors: Yaxiong Wu, Yongyue Zhang, Sheng Liang, Yong Liu
Presents a sentence-level graph memory system for conversational agents that efficiently stores conversation history as semantic nodes and edges, enabling context-aware retrieval for improved long-term memory management in chatbots.
CAD-Tokenizer: Towards Text-based CAD Prototyping via Modality-Specific Tokenization (2025-09-25)
Authors: Ruiyu Wang, Shizhao Sun, Weijian Ma, Jiang Bian
Develops a specialized tokenization approach for representing CAD designs within language models, enabling direct text-to-CAD generation capabilities and bridging the gap between natural language instructions and complex 3D design specifications.
LOOKING AHEAD
As we move into Q4 2025, we're watching the convergence of multimodal AI systems with advanced robotics platforms—a union that promises to transform manufacturing, healthcare, and household automation. The recent breakthroughs in continuous learning models that can update their knowledge bases without complete retraining are likely to reach commercial deployment by early 2026, addressing the persistent challenge of model staleness.
Meanwhile, regulatory frameworks are struggling to keep pace. With the EU's AI Act amendments scheduled for review in November and the US comprehensive AI legislation expected before year's end, we anticipate a critical period of regulatory alignment that will shape enterprise AI adoption through 2027. Companies positioning themselves at the intersection of compliant AI and physical world integration will likely emerge as the next market leaders.