AI News & Launches

Smart Turn is an open-source, community-driven project for native audio turn detection for conversational voice AI, utilizing a Wav2Vec2-BERT model base. It is designed to improve on limitations of VAD-based turn detection by including linguistic and acoustic cues and supports only English currently. The project invites contributions and is licensed under the BSD-2-Clause license.

James Matsuzaki of Recall.ai shares the experience of reducing latency from 12 seconds to 2-3 seconds for their Output Media feature by removing an esoteric Linux process flag, '--die-with-parent'. The post details challenges encountered with Bubblewrap sandboxing and Tokio's threading model in their endeavor.

OpenPipe has used GRPO to achieve performance surpassing R1, o1, o3-mini models on 'Temporal Clue' reasoning game. They discuss task design, hyperparameters, and share a training recipe, dataset, and model weights under the MIT license.

Mistral AI introduces Mistral OCR, a new Optical Character Recognition API offering state-of-the-art document understanding. It processes text, tables, equations, and images with unprecedented accuracy. Ideal for use with RAG systems, it's multilingual and available for self-hosting. Try it on le Chat or la Plateforme.

Mistral OCR introduces the world's best document understanding API, revolutionizing information abstraction and retrieval with Optical Character Recognition technology. The API comprehends media, text, tables, and equations with high accuracy, fostering innovation in document processing.

Israel's military intelligence used intercepted Palestinian communications to create a ChatGPT-like AI tool, raising concerns over bias and inaccuracies in surveillance practices. This advanced model can analyze Arabic dialects intensely, providing insights into vast data collections and potentially impacting Palestinians' daily lives.

Fynix.ai offers a suite of AI-powered tools designed to increase coding efficiency and code quality. Products such as AI Code Assistant and Code Quality Agent integrate into popular IDEs to provide real-time coding aid, automated code reviews, and repository-wide context insight.

Recall.ai provides an integration for Slack huddle data capture, offering bots to record conversations, transcribe them, and manage events. It streamlines meeting management with automated workflows and powers AI-driven applications directly within Slack.

Aider is an AI-powered pair programming tool for developers, enabling them to program alongside large language models (LLMs) directly from their terminal. It integrates with local git repositories to enhance coding productivity by providing features like automatic git commits, compatibility with most programming languages and the ability to work on complex multi-file projects.

Perigon is a data management platform that combines multi-source ingestion, data preparation, and analytics into one seamless workflow. Designed to handle complex data, it offers data integration, real-time data streams, and interactive visualizations to unlock insights.

Arva AI is hiring an AI Product Engineer in London, offering a salary of £65K - £90K GBP with 0.50% - 1.00% equity. The role involves building full-stack features on their AI-powered compliance platform and improving it through feedback loops and AI advancements. The company culture emphasizes fast delivery, customer-first priorities, and transparent communication.

PublicView provides data analytics and visualization tools tailored for the public sector, enabling organizations to gain insights and drive improvements.

Volkswagen AG introduces the ID. EVERY1, the company's most affordable electric vehicle (EV) at €20,000, approximately $21,000. It will be the first model to use the electric architecture developed with Rivian. The ID. EVERY1 is set to offer over 155 miles of range and begin production in two years in Europe.

This quickstart guide provides instructions on how to use llama.cpp portable zip to run llama.cpp on Intel GPU with ipex-llm. It includes prerequisites, step-by-step runtime configuration, and model execution for both Windows and Linux platforms. The guide is intended for running models like DeepSeek-R1-671B-Q4_K_M on Intel hardware, with different prerequisites and steps for Windows and Linux. DeepSeek V3/R1 models can also be run using FlashMoE on Linux platforms, provided the system meets the specified hardware requirements.

An exploration of the ancient Greek yardstick of imitating reality through an art duel between Zeuxis and Parrhasios, and how the Hardham muralist engages with this theme to challenge the viewer's perception.

Datafold is seeking a Senior Software Engineer for their AI-powered data migration tool, the Datafold Migration Agent, offering a salary range of $175K-$245K. Applicants must have 3+ years of experience, proficiency in Python, and a startup mindset. The role entails driving DMA development, designing and scaling backend systems, and collaborating with customers. Datafold, a Series A startup backed by YC and others, prioritizes data quality and provides tools integrated into data engineers' workflows.

Sam Altman, CEO of OpenAI, has reignited his political fundraising activities by hosting an event for Democratic Senator Mark Warner of Virginia. After a period of reduced involvement and a brief dismissal from OpenAI in 2023, Altman is participating in the political scene with a fund-raising lunch on March 20, whose tickets range from $1,000 to $22,000. The event marks his reentry into political support since helping Democrat Dean Phillips in 2022.

Google introduces AI Mode in Search, an experimental feature that enables advanced AI Overviews. It enhances the search experience with more reasoning, thinking, and multimodal capabilities, allowing users to get in-depth assistance on complex queries. Starting with coding and math questions, the feature will expand to more areas. It is currently available for testing in Labs to a limited audience, including Google One AI Premium subscribers.

A group of outliers from the Silicon Valley community, influenced by AI-risk theories and calling themselves 'Zizians', become involved in a series of violent acts, resulting in six deaths. Their descent from rationalist thinkers to extreme behavior illustrates the potential dangers of ideology taken to the extreme.

Lifestack is a web and mobile application designed to optimize your day by combining schedule management with wearable data to plan tasks according to your energy levels, aiming to enhance productivity while maintaining health.

Latitude provides a platform to build self-improving AI agents capable of designing, evaluating, and deploying with your tools and data. Agents operate autonomously, unlike predefined workflows, exhibiting adaptiveness and integrability. Latitude's mission supports AI builders in creating scalable AI agents.

GaliChat is an AI Chatbot that provides 24/7 customer support, lead generation, and support cost reduction by training on specific industry data, and can be set up without coding.

Giles Thomas' latest blog post details his journey through the book 'Build a Large Language Model (from Scratch)' by Sebastian Raschka. He delves into section 3.4 covering 'Implementing self-attention with trainable weights'. The post offers his perspective on understanding the concept of self-attention in LLMs, where inputs are transformed into different spaces to calculate attention scores, leading to the production of context vectors representing the meaning of tokens.

This article by Fred Hebert on Honeycomb's blog, dated March 3rd, 2025, provides insight into the incorporation of AI in technology, focusing on its current limitations and the crucial role humans need to play in the AI loop. Hebert emphasizes that despite the promises of AI, practical deployment necessitates current functionality rather than future potential. The article raises several thoughtful questions regarding the integration of AI tools into systems, the risks of dependency, and enhancing human capabilities instead of replacing them.

Project anon-kode is a terminal-based AI coding assistant that can utilize any model supporting the OpenAI-style API. It is designed to improve and explain code, run tests, and perform other tasks, depending on the chosen model.

Microsoft introduces Dragon Copilot, an AI healthcare assistant offering dictation and ambient listening in clinical settings. The system is developed from Nuance's technology, aiming to reduce administrative burden on clinicians and improve patient care. Microsoft's studies show reduced clinician burnout and enhanced patient experiences.

Trellis provides a powerful AI engine to automate PDF workflows, transforming complex documents into ready-to-use data for operations and engineering teams within seconds, not weeks.

Researchers have developed a new AI method called DNAformer that decodes digital data from DNA strands about 90 times faster than previous techniques. This advancement brings DNA data storage closer to practical use in computing, with potential for long-term preservation, as it's naturally replicable and can remain intact for millennia.

Kive leverages AI to transform the creative process, allowing users to generate on-brand product visuals without needing a studio or photoshoot. The platform is designed to make tasks like social media ideas, ad campaigns, and asset management effortless for brands, agencies, and freelancers.

Opera introduces Browser Operator, an AI agent for browsers, optimizing tasks for efficiency and privacy. It uses local resources for web tasks, and its AI Composer Engine handles instructions in natural language. The user remains in control, providing inputs for sensitive actions, as browsing data is kept private and secure on the user's device.

NeoBase is a cloud-based platform offering secure and private file sharing and storage services. It emphasizes on data privacy and aims to provide an intuitive user experience for managing files in the cloud.

Cardamon.ai is a cutting-edge platform blending artificial intelligence with data analysis to pave the way for innovative tech solutions, specializing in machine learning and complex algorithmic applications.

Dragon Copilot by Microsoft is an AI-powered clinical documentation tool designed to streamline documentation, surface relevant information, automate tasks, and improve clinical workflows across different care settings.

In a blog post on nomadic compute designs for AI workloads, the author discusses building sustainable systems using simple components for flexibility and cost-efficiency. Objects are stored without egress fees, reducing reliance on provider-specific features that impose vendor lock-in. The post further explores production readiness, cloud dependency strategies, and concepts for orchestrating easily replaceable compute resources.

SmartestKid is a Windows desktop AI assistant written in Python, offering voice and text input modes and interfaces with Microsoft Office applications, Images, and the file system. It is open-source with an inviting community for contributions to expand its functionality.

agents-json is a GitHub repository created by Wildcard AI, aiming to describe contracts for API and agent interactions built on top of the OpenAPI standard. It provides a specification to enable AI agents such as language model tools to interact with APIs by using structured contracts. The current version of the agents.json specification is 0.1.0, and the repository contains examples, Python package, and demos for integrating with agents.json.

go-attention is the first pure Go implementation of attention mechanisms and transformer layers, designed for high performance and ease of use. It supports dot-product attention, multi-head attention, and full transformer layers with optimization for the CPU and minimal memory allocations.

Qodo unveils Qodo-Embed-1, a new code embedding model family offering state-of-the-art performance with a smaller size than existing models. Qodo-Embed-1-1.5B achieves a 68.53 score on the CoIR benchmark, and its larger counterpart, Qodo-Embed-1-7B scores 71.5, showing superior efficiency across models.

Article presents a robust Wavefront OBJ parser in C, detailing challenges and techniques in parsing 3D model data, showcasing a technique to handle undefined behavior, and demonstrating the parsing process with sample code.

AI Dev Info provides insights into artificial intelligence and software development, featuring articles, tutorials, and news on latest technologies and trends in the AI programming domain.