LLM Daily: October 01, 2025
🔍 LLM DAILY
Your Daily Briefing on Large Language Models
October 01, 2025
HIGHLIGHTS
• Periodic Labs has secured a historic $300 million seed funding from investors including Andreessen Horowitz, Nvidia, and Jeff Bezos, marking one of the largest seed rounds in AI history for scientific automation research.
• Zhipu-AI announced their new GLM-4.6 model without plans for a lighter "Air" version, with their previous GLM-4.5 model being recognized as one of the closest open-weight alternatives to commercial models for coding and agentic workloads.
• The "LLMs-from-scratch" repository by Sebastian Raschka has amassed over 73,900 GitHub stars, offering comprehensive educational resources for building ChatGPT-like models in PyTorch from the ground up.
• Researchers have introduced InfLLM-V2, featuring a breakthrough dense-sparse switchable attention mechanism that seamlessly transitions between processing short and long sequences while adding only 0.01% extra parameters to baseline models.
BUSINESS
Former OpenAI and DeepMind Researchers Secure $300M Seed Funding
Periodic Labs, founded by ex-OpenAI and DeepMind researchers, has raised an extraordinary $300 million seed round to pursue scientific automation. The funding comes from an impressive roster of investors including Andreessen Horowitz, Nvidia, Elad Gil, Google's Jeff Dean, former Google CEO Eric Schmidt, and Amazon founder Jeff Bezos. This represents one of the largest seed rounds in AI history, highlighting growing investor confidence in applied AI for scientific discovery. (TechCrunch, 2025-09-30)
OpenAI Enters Social Media Space with Sora App
OpenAI announced it will launch Sora app, a TikTok competitor alongside its upgraded Sora 2 video generation model. The social platform will allow users to create and share AI-generated videos of themselves and friends in a format similar to TikTok's feed. This marks OpenAI's first direct consumer application beyond its API and ChatGPT offerings, signaling the company's ambitions to expand into social media. (TechCrunch, 2025-09-30)
Sequoia Capital Invests in AI Recruiting Platform Juicebox
Sequoia Capital announced its partnership with Juicebox, an AI-powered recruiting platform that has gained significant traction among founders. While the exact investment amount wasn't disclosed, Sequoia highlighted the company's innovative approach to streamlining the hiring process through artificial intelligence. (Sequoia Capital, 2025-09-25)
AI-Driven Startup Operations Featured at TechCrunch Disrupt 2025
TechCrunch Disrupt 2025 will showcase a new trend of startups replacing or augmenting early employees with AI agents. The conference will explore the implications of having AI systems as a company's first "hires" instead of human employees, reflecting the growing integration of AI into core business operations beyond supplementary roles. (TechCrunch, 2025-09-30)
PRODUCTS
Zhipu-AI Announces GLM-4.6 Without Plans for "Air" Version
Official Announcement on X | Zhipu-AI (Chinese AI lab) | (2025-09-30)
Zhipu-AI has announced their new GLM-4.6 model but confirmed they have no plans to release a lighter "Air" version, which would typically be more suitable for local deployment. According to community benchmarks, their previous GLM-4.5 model has been particularly impressive for coding and agentic workloads, with users noting it's among the closest open-weight models to matching closed commercial alternatives. The company continues to be viewed as one of the most promising labs releasing open-weight models.
Microsoft TTS Model Briefly Released Then Removed
Community Discussion | Microsoft | (2025-09-30)
Microsoft reportedly released a high-quality text-to-speech model that significantly outperformed existing offerings, only to quickly remove it from availability. According to user reports, the company "freaked out and removed it immediately once they realized it wasn't meh," suggesting the model may have been inadvertently released before intended or performed beyond expected quality thresholds. This has sparked discussion about Western companies' hesitancy to release cutting-edge AI capabilities compared to Chinese counterparts.
Note: The products section is lighter than usual today, as there were no new AI products featured on Product Hunt and limited product announcements in the provided data sources.
TECHNOLOGY
Open Source Projects
rasbt/LLMs-from-scratch
A comprehensive educational repository for building a ChatGPT-like LLM in PyTorch from scratch. This project offers step-by-step guidance on developing, pretraining, and finetuning GPT-like models and serves as the official code repository for Sebastian Raschka's book. With over 73,900 stars and continued maintenance (most recent update testing Python 3.13 compatibility), it's a popular resource for those wanting to understand LLM architecture fundamentals.
Shubhamsaboo/awesome-llm-apps
A curated collection of LLM applications featuring AI Agents and Retrieval-Augmented Generation (RAG) implementations using OpenAI, Anthropic, Gemini, and open-source models. Recently updated with a YAML-based Multi-Agent Web Research System using Firecrawl MCP, this repository (70,900+ stars) serves as a reference for developers building production-ready AI applications with modern architectures.
openai/openai-cookbook
The official collection of examples and guides for using the OpenAI API, accessible at cookbook.openai.com. With over 68,100 stars, this repository provides code snippets and tutorials for common tasks and was recently updated to work with the latest OpenAI APIs, including Reinforcement from Teacher Feedback (RFT) implementations.
Models & Datasets
Text and Multimodal Models
tencent/HunyuanImage-3.0
Tencent's latest text-to-image model using a mixture-of-experts architecture. Referenced in arXiv paper 2509.23951, this model has gained significant traction with 677 likes despite being recently released.
deepseek-ai/DeepSeek-V3.2-Exp
An experimental conversational model built on DeepSeek-V3.2-Exp-Base with MIT license. Notable for its FP8 compatibility and availability on multiple deployment platforms, the model has already amassed over 5,600 downloads and 408 likes.
Qwen/Qwen3-Omni-30B-A3B-Instruct
A multimodal model from Qwen capable of text-to-audio and any-to-any generation with 30B parameters (using a mixture-of-experts architecture that activates 3B parameters per forward pass). The model has impressive adoption with over 151,600 downloads and 561 likes.
3D Generation Models
tencent/Hunyuan3D-Part
Specialized 3D generation model with part segmentation capabilities, built upon Tencent's Hunyuan3D-2.1 base model. Trained on Objaverse and Objaverse-XL datasets, this model has garnered 444 likes and over 2,300 downloads, showing the growing interest in advanced 3D generation technology.
Animation and Image Models
Wan-AI/Wan2.2-Animate-14B
A diffusion-based model focused on animation generation with Apache 2.0 license. With 578 likes and over 38,100 downloads, it demonstrates substantial community interest in animation generation tools. The model is available in ONNX format for easier deployment.
Datasets
openai/gdpval
OpenAI's multimodal validation dataset supporting audio, documents, images, text, and video modalities. Updated on September 25th, this dataset has already received 141 likes and over 11,100 downloads, highlighting its importance for evaluating multimodal models.
nvidia/Nemotron-Personas-Japan
A Japanese language dataset focused on personas for text generation, with over 7,900 downloads. Released under CC-BY-4.0 license, this dataset includes both text and image modalities and falls in the 1-10M size category.
lmms-lab/LLaVA-OneVision-1.5-Insturct-Data
Instruction data for the LLaVA-OneVision multimodal model combining image and text. Recently updated on September 30th, this Apache 2.0-licensed dataset has seen extensive use with over 47,000 downloads despite having only 35 likes, demonstrating its utility for multimodal model training.
Developer Tools and Interactive Demos
Wan-AI/Wan2.2-Animate
A Gradio-powered demo for Wan2.2's animation capabilities, allowing users to test the animation generation model directly. With 1,198 likes, it's one of the most popular spaces for experiencing generative animation technology.
multimodalart/ai-toolkit
A Docker-based collection of AI tools for multimodal art creation. With 93 likes, this toolkit provides creators with an accessible environment for experimenting with various AI-powered artistic techniques.
not-lain/background-removal
A popular background removal tool with 2,384 likes, implemented as a Gradio interface. The space is deployed with MCP-server compatibility for efficient processing and demonstrates the practical application of image segmentation technology.
ResembleAI/Chatterbox
An interactive conversational AI demo from ResembleAI with 1,507 likes. This Gradio interface showcases voice generation and conversation capabilities, providing users with an accessible way to experiment with voice AI technology.
RESEARCH
Paper of the Day
InfLLM-V2: Dense-Sparse Switchable Attention for Seamless Short-to-Long Adaptation (2025-09-29)
Weilin Zhao, Zihan Zhou, Zhou Su, Chaojun Xiao, Yuxuan Li, Yanghao Li, Yudi Zhang, Weilun Zhao, Zhen Li, Yuxiang Huang, Ao Sun, Xu Han, Zhiyuan Liu
This paper introduces a breakthrough in long-sequence processing for LLMs with a novel dense-sparse switchable attention mechanism. The significance lies in its ability to seamlessly transition between dense attention for short sequences and sparse attention for long sequences without disrupting the conventional pretrain-short/finetune-long workflow, solving a critical computational bottleneck in transformer architectures.
InfLLM-V2 achieves this through a parameter-efficient design that adds only 0.01% extra parameters compared to baseline models, while demonstrating superior performance across various benchmarks. Their switchable attention approach enables smooth adaptation to sequences of varying lengths without the convergence issues that plague existing sparse attention methods.
Notable Research
MARCOS: Deep Thinking by Markov Chain of Continuous Thoughts (2025-09-29)
Jiayu Liu, Zhenya Huang, Anya Sims, Enhong Chen, Yee Whye Teh, Ning Miao
MARCOS introduces a novel reasoning framework for LLMs that generates multiple reasoning paths as a Markov chain of thoughts, allowing models to continuously refine their reasoning and achieve better performance on complex reasoning tasks.
UniAPL: A Unified Adversarial Preference Learning Framework for Instruct-Following (2025-09-29)
FaQiang Qian, WeiKun Zhang, Ziliang Wang, Kang An, Xuhui Zheng, Liangjian Wen, Mengya Gao, Yong Dai, Yichao Wu
This paper reimagines post-training alignment as a unified preference learning problem, proposing an adversarial framework that combines both demonstrated preferences (SFT) and comparative preferences (RL) to address distributional mismatches in traditional sequential alignment pipelines.
Scaling Synthetic Task Generation for Agents via Exploration (2025-09-29)
Ram Ramrakhya, Andrew Szot, Omar Attia, Yuhao Yang, Anh Nguyen, Bogdan Mazoure, Zhe Gan, Harsh Agrawal, Alexander Toshev
The researchers present a novel approach to scaling task generation for multimodal LLM agents by leveraging environment exploration to automatically generate diverse, feasible, and verifiable tasks without costly human annotation.
GLoW: Dual-Scale World Models for LLM Agents Towards Hard-Exploration Problems (2025-09-28)
Minsoo Kim, Seung-won Hwang
GLoW tackles the challenge of "hard-exploration" tasks for LLM agents by implementing dual-scale world models that maintain both global trajectory frontiers of valuable discoveries and local trial-and-error learning, guided by a novel Multi-path Advantage Reflection mechanism.
LOOKING AHEAD
As we enter Q4 2025, multimodal AI systems are rapidly evolving beyond today's capabilities. The integration of large-scale physics simulations with generative AI promises transformative applications in materials science and drug discovery by early 2026. Meanwhile, the emerging "cognitive architecture" approach—combining specialized models into cohesive systems with shared memory and reasoning—is gaining traction among leading labs and startups.
Watch for the first wave of truly autonomous AI research assistants in Q1 2026, capable of designing and running experiments with minimal human oversight. Regulatory frameworks will need to adapt quickly as these systems blur the boundaries between tool and collaborator. The field's focus on alignment techniques continues to intensify as models approach new thresholds of capability.