LLM Daily: June 24, 2025
🔍 LLM DAILY
Your Daily Briefing on Large Language Models
June 24, 2025
HIGHLIGHTS
• Harvey AI's meteoric rise continues with a valuation jump to $5 billion just four months after their previous funding round, underscoring the extraordinary investor appetite for specialized legal AI applications in today's market.
• A developer is launching an educational 50-day series on building tiny language models (15-30M parameters) from scratch using only modest computing resources, democratizing LLM development knowledge for those without access to extensive infrastructure.
• Microsoft Research's groundbreaking paper challenges the fundamental assumption that self-generated feedback always improves LLM reasoning, providing evidence that internal feedback can sometimes harm performance across multiple benchmarks.
• Databricks co-founder Andy Konwinski has committed $100 million of his personal fortune to establish a new AI research fund, representing a significant private investment in advancing fundamental AI capabilities.
• Open-source projects like ComfyUI and Kotaemon continue gaining significant traction, with ComfyUI reaching over 80,000 GitHub stars for its powerful node-based diffusion model interface and Kotaemon offering a developer-friendly RAG-based document interaction system.
BUSINESS
Funding & Investment
Harvey AI's Valuation Jumps to $5B in Just Four Months
Legal AI startup Harvey AI has secured a new funding round that values the company at $5 billion, up from $3 billion just four months after their $300 million Series D. The new round was co-led by Kleiner Perkins and Coatue Management. This rapid valuation growth highlights the accelerating investor interest in legal tech powered by AI. TechCrunch (2025-06-23)
Databricks Co-Founder Pledges $100M for AI Research Fund
Andy Konwinski, co-founder of Databricks and Perplexity, has committed $100 million of his personal funds to create a new institute supporting AI researchers. The fund has already backed a new lab led by Ion Stoica, representing a significant private investment in advancing AI research capabilities. TechCrunch (2025-06-23)
M&A and Partnerships
OpenAI and io Collaboration on AI Hardware Device Revealed in Court Filings
Legal documents from a trademark dispute lawsuit have unveiled previously undisclosed details about OpenAI and Jony Ive's io company working together on a consumer AI hardware device. The lawsuit was filed by iyO, a Google-backed hardware startup developing custom-molded earpieces. TechCrunch (2025-06-23)
OpenAI Removes Promotional Materials for io Deal Following Court Order
OpenAI has pulled promotional content, including a video featuring CEO Sam Altman and Apple designer Jony Ive, related to its $6.5 billion acquisition of the device startup io. The removal appears to be in response to a court order, potentially complicating the high-profile partnership. TechCrunch (2025-06-22)
Company Updates
Salesforce Launches Agentforce 3 with Enhanced AI Capabilities
Salesforce has introduced Agentforce 3, featuring AI agent observability and native Model Context Protocol (MCP) support. The platform provides real-time visibility into AI agent operations and secure interoperability across systems, positioning Salesforce to strengthen its enterprise AI offerings. VentureBeat (2025-06-23)
xAI's Grok Expanding into Spreadsheet Editing
A code leak suggests that Elon Musk's xAI is developing advanced file editing capabilities for its Grok AI, including spreadsheet support. This development signals xAI's intention to compete directly with OpenAI, Google, and Microsoft in the productivity tools space by integrating AI copilots into document editing. TechCrunch (2025-06-23)
LinkedIn's AI Writing Assistant Sees Lower Than Expected Adoption
LinkedIn CEO Ryan Roslansky revealed that the platform's AI-powered writing assistant hasn't gained as much traction as anticipated, despite overall strong AI adoption among LinkedIn users. This indicates potential challenges in user acceptance of AI-generated content suggestions on professional platforms. TechCrunch (2025-06-22)
Market Analysis
AI Regulation Moratorium Bill Advances in Senate
A Republican-led bill to prevent states from enforcing their own AI regulations has cleared a key procedural hurdle in the Senate. This development could significantly impact the regulatory landscape for AI companies, potentially creating a more unified but potentially less stringent national approach to AI governance. TechCrunch (2025-06-22)
PRODUCTS
Building a Tiny LLM from Scratch Series
Developer: Prashant Lakhera (Individual contributor)
Announcement Date: (2025-06-22)
Link: Reddit announcement
A developer is launching a 50-day series starting June 23 that will walk through building small language models (15-30M parameters) from scratch using only a laptop or modest GPU. The daily series will methodically cover all aspects of LLM development including data collection, tokenization, model architecture, training, and evaluation. This educational content aims to demystify LLM development for those without access to massive computing resources.
Chroma Image Generation Model Update
Developer: Silveroxides (Community contributor)
Release Date: (Referenced on 2025-06-23)
Link: Hugging Face model
The Chroma image generation model has received attention for its impressive rendering capabilities, with users sharing examples of its latest v38 and v39 "detail calibrated" versions. The model is available on Hugging Face with accompanying workflow templates for implementation. Community reception has been positive, with users praising the quality of outputs and discussing advanced techniques like latent noise injection workflows to further enhance results.
TECHNOLOGY
Open Source Projects
ComfyUI - 80,520 ⭐
A powerful and modular diffusion model GUI with a node-based interface for advanced image generation workflows. Recent updates include a new resize and pad image node, and improvements to the ImageStitch node allowing white padding. ComfyUI stands out for its flexibility and customization options compared to other stable diffusion interfaces.
Kotaemon - 22,654 ⭐
An open-source RAG-based UI for document interaction that balances clean design with developer-friendly customization. Recent updates have added a CohereReranking parameter, fixed LanceDB retrieval ranking preservation, and addressed vector store query issues. Kotaemon is designed as a comprehensive solution for both end-users and developers working with document-based conversational AI.
BitNet - 20,273 ⭐
Microsoft's official inference framework for 1-bit LLMs, enabling efficient deployment of binary neural networks. Recent commits focus on GPU kernel improvements and documentation updates. BitNet reduces memory and compute requirements while maintaining performance, making it suitable for resource-constrained environments.
Models & Datasets
Nanonets-OCR-s
A fine-tuned Qwen2.5-VL model specialized for OCR and PDF-to-markdown conversion tasks. With over 177,000 downloads, this model has gained popularity for its ability to extract and format text from images and documents.
MiniMax-M1-80k
A new conversational AI model with an 80k context window, referenced in arXiv paper 2506.13585. The model has quickly gained traction with 525 likes and over 10,000 downloads, suggesting strong capabilities for long-context applications.
Mistral-Small-3.2-24B-Instruct-2506
The latest instruct-tuned version of Mistral's 24B model with multilingual support for 25+ languages. Built on the Mistral-Small-3.1-24B-Base architecture, this model offers advanced multimodal capabilities for image-to-text tasks.
Essential Web v1.0
A massive web dataset (between 10B-100B tokens) published alongside arXiv paper 2506.14111. With 137 likes and over 75,000 downloads since its recent release on June 22, this dataset provides high-quality web content for training large language models.
Institutional Books 1.0
A collection of book data with 165 likes and 38,000+ downloads, available in multiple formats including Parquet. The dataset supports various libraries including datasets, dask, mlcroissant, and polars, making it accessible for different analysis approaches.
Developer Tools & Infrastructure
AI Comic Factory
A popular Hugging Face Space with over 10,400 likes that allows users to generate entire comic strips using AI. Implemented as a Docker-based application, it demonstrates the creative potential of AI for visual storytelling.
Kolors Virtual Try-On
A highly popular Gradio-based application (9,118 likes) that enables virtual clothing try-on experiences. This space showcases practical applications of computer vision and generative AI in the fashion retail domain.
AiSheets
A Docker-based application with 283 likes that likely provides spreadsheet-like functionality enhanced with AI capabilities. This tool represents the growing trend of bringing AI functionality to traditional productivity applications.
Chatterbox
ResembleAI's Gradio-based conversational interface with 1,158 likes. Using MCP-server technology, this space likely showcases ResembleAI's voice synthesis capabilities combined with conversational AI for natural interactions.
Multimodal-OCR
A Gradio application with 277 likes focusing on optical character recognition with multimodal capabilities. This space demonstrates the practical implementation of OCR technology using modern multimodal models for improved text extraction from images.
RESEARCH
Paper of the Day
No Free Lunch: Rethinking Internal Feedback for LLM Reasoning (2025-06-20)
Authors: Yanzhi Zhang, Zhaoxi Zhang, Haoxiang Guan, Yilin Cheng, Yitong Duan, Chen Wang, Yue Wang, Shuxin Zheng, Jiyan He
Institution: Microsoft Research
This paper is significant as it challenges the prevailing assumption that self-generated feedback always improves LLM reasoning. Through extensive experiments across multiple benchmarks, the authors demonstrate that internal feedback can sometimes harm performance, contradicting previous research findings. They introduce a more nuanced framework for understanding when and how internal feedback mechanisms help or hinder reasoning, providing important insights for the design of future LLM systems.
Notable Research
From Concepts to Components: Concept-Agnostic Attention Module Discovery in Transformers (2025-06-20)
Authors: Jingtong Su, Julia Kempe, Karen Ullrich
This research introduces a novel method for identifying functional components within transformer models without requiring predefined concepts, revealing that attention modules often align with cognitive functions like deductive reasoning and spatial awareness.
VLN-R1: Vision-Language Navigation via Reinforcement Fine-Tuning (2025-06-20)
Authors: Zhangyang Qi, Zhixiong Zhang, Yizhou Yu, Jiaqi Wang, Hengshuang Zhao
The authors present an end-to-end framework that enables large vision-language models to navigate physical environments using continuous actions based on egocentric video and natural language instructions, demonstrating significant improvements over previous navigation systems.
MM-AttacKG: A Multimodal Approach to Attack Graph Construction with Large Language Models (2025-06-20)
Authors: Yongheng Zhang, Xinyun Zhao, Yunshan Ma, et al.
This paper introduces a novel multimodal approach that leverages LLMs to automatically construct attack graphs from cybersecurity reports containing both text and images, achieving state-of-the-art performance in extracting and organizing cyber threat intelligence.
Measuring (a Sufficient) World Model in LLMs: A Variance Decomposition Framework (2025-06-19)
Authors: Nadav Kunievsky, James A. Evans
The researchers propose a formal framework for evaluating whether LLMs possess a robust world model by measuring their consistency across semantically equivalent prompts while distinguishing between prompts with different meanings, providing a quantitative method to assess model understanding.
LOOKING AHEAD
As we move into Q3 2025, the AI landscape continues its rapid evolution. The emergence of 100T parameter models with improved multimodal reasoning capabilities signals a shift toward systems that can seamlessly integrate and reason across text, vision, audio, and scientific data. Meanwhile, the industry-wide focus on AI alignment and interpretability frameworks is gaining momentum, with several major labs expected to release open specifications for their alignment methodologies by Q4.
Looking further ahead, we anticipate significant breakthroughs in energy-efficient AI hardware by early 2026, potentially reducing computational requirements by 40-60%. This, coupled with advances in few-shot learning techniques, may finally bridge the gap between specialized and general-purpose AI systems—a development that could dramatically accelerate adoption across previously untapped sectors.