Map Features in OpenStreetMap with Computer Vision

Mozilla.ai has developed the OpenStreetMap AI Helper Blueprint, a tool that uses computer vision to automate parts of the mapping process on OpenStreetMap, while keeping humans in control for final verification. The blueprint consists of three stages: creating an object detection dataset from OpenStreetMap, fine-tuning an object detection model, and contributing to OpenStreetMap by running inference and uploading new data.

Read more

Tencent's 'Hunyuan-T1'–The First Mamba-Powered Ultra-Large Model

Tencent Hunyuan-Large is an open-source MoE model with 389 billion parameters and 52 billion activated parameters, making it the largest and most effective Transformer-based MoE model in the industry. The model has achieved state-of-the-art results in various tasks, including text generation, knowledge question answering, and code creation, and is available for fine-tuning and deployment through Tencent's open-source framework.

Read more

CEO of Kubient sentenced for fraud

The CEO of Kubient, an AI ad-tech firm that claimed to be working towards a "world free of fraud", has been sentenced to one year and one day in prison for committing fraud. Paul Roberts pled guilty to creating $1.3 million in fraudulent revenue statements to bolster the company's initial public offering, and was found to have oversold the company's artificial intelligence tool, claiming it could detect 300% more ad fraud than it actually could.

Read more

Most AI value will come from broad automation, not from R&D

Many influential figures in the AI industry believe that the primary economic impact of AI will come from automating research and development (R&D), but this view is not supported by rigorous economic arguments and is likely wrong. The actual economic value of R&D is smaller than people think, and most AI systems will be deployed to perform tasks unrelated to R&D, which will be more economically valuable and drive broad economic growth through automation of labor.

Read more

Meta pirated books to train its AI

Meta, the company behind Facebook and Instagram, pirated millions of books from the online library Library Genesis to train its AI model, Llama 3, after deciding that licensing the books legally would be too expensive and time-consuming. The company's use of pirated material has raised concerns about copyright infringement, and Meta is now facing a lawsuit from authors whose books were included in the library, with the company arguing that its use of the material constitutes "fair use" under copyright law.

Read more

Can AI Compress Like a Genius?

The Kolmogorov-Test (KT) is a new evaluation method for code-generating large language models (LLMs) that assesses their ability to compress data by generating the shortest program that produces a given sequence. Current flagship models, such as GPT4-o and Llama-3.1-405B, perform poorly on the KT, and while training on synthetic data can improve performance, these gains do not generalize well to real-world data, indicating a need for new innovations to achieve better results.

ArXiv

Quantitative Finance: Kronecker-Factored Approximate Curvature Deep Hedging

This paper proposes a novel Deep Hedging framework that integrates Kronecker-Factored Approximate Curvature (K-FAC) optimization with Long Short-Term Memory (LSTM) networks, significantly improving computational efficiency and hedging efficacy. The K-FAC implementation achieves marked reductions in transaction costs and profit and loss variance, and exhibits superior risk-adjusted performance, demonstrating the potential of second-order optimization methods to enhance the practicality of Deep Hedging implementations.

ArXiv

Stop using the elbow criterion for k-means

The "elbow method" for choosing the number of clusters in k-means clustering is flawed and lacks theoretical support, often leading to poor conclusions. Better alternatives are available and should be used instead, with educators and researchers encouraged to abandon the elbow method and adopt more reliable methods for determining the optimal number of clusters.

ArXiv

Revisiting semi-supervised learning in the era of foundation models

Semi-supervised learning (SSL) can be less effective than expected when used with pre-trained vision foundation models (VFMs), with parameter-efficient fine-tuning (PEFT) using only labeled data often matching SSL performance. However, a simple approach that uses PEFT to pseudo-label unlabeled data and then ensembles multiple models to produce more robust labels can be highly effective, providing insights into SSL with VFMs and paving the way for more practical semi-supervised learning.

ArXiv

Difference-in-Differences Designs: A Practitioner's Guide

The Difference-in-Differences (DiD) research design, commonly used in quasi-experimental research, has a well-understood basic form but can be applied in various ways, leading to ad hoc practices. This article provides a framework for organizing and discussing different types of DiD designs and their associated estimators, covering topics such as covariates, weights, and staggered treatments.

ArXiv

Redfly.ai – easily sync your SQL Server Database to Redis on-demand

Redfly.ai is an open-source system that synchronizes databases with Redis, generating a data access layer that integrates data access code with caching, and is designed to improve performance at scale. The system currently supports SQL Server, Redis, Azure Search, and Azure Cloud, with plans to expand to other relational databases and public clouds, and offers a demo and documentation for developers to try out and understand the technology.

Repo

Roblox Releases Cube: Generative AI System for 3D

Roblox has introduced Cube, a generative AI system for 3D intelligence that can generate 3D objects and scenes, with the goal of supporting developers in producing all aspects of a Roblox experience. The Cube 3D model is now open-sourced and accessible to individuals, creators, researchers, and businesses, allowing them to experiment, innovate, and scale their ideas responsibly.

Repo

Now the Kanban for AI Agents UI Is Open Source Too → MIT Licensed

Kaiban Board is a visual interface for managing AI agent teams, providing a Kanban-style interface to create, manage, and share AI agent teams, track tasks and workflows, and collaborate with AI agents. The project is open-source, MIT licensed, and welcomes contributions from the community, with documentation and guidelines available for setup, usage, and contributing to the project.

Repo

OpenAI GPT-4o-transcribe/GPT-4o-mini-transcribe WebSocket example

To set up the project, you need NodeJS v22 and pnpm, then install dependencies, create .env files with required settings such as API keys and URLs, and match the base URL port value. You can start the project by running pnpm dev at the repo root after completing the setup.

Repo

Pocket Flow: 100-line LLM framework. Let Agents build Agents

Pocket Flow is a minimalist LLM framework that is lightweight, expressive, and easy to use, consisting of just 100 lines of code with zero dependencies and vendor lock-in. It allows for the implementation of various design patterns and can be used to build complex LLM applications through a paradigm called Agentic Coding, where humans design and agents code.

Repo