Brief

AI tools designed to track meeting participation and contribution are showing bias against non-Western names and accents. These systems, used by companies like Amazon and Meta, are trained on data that underrepresents certain groups, leading to lower accuracy for accented speakers and less recognition of contributions from individuals with non-Western names. While AI has the potential to flatten corporate hierarchies and promote ideas based on merit, its current implementation risks perpetuating existing inequalities. AI

Thinking Machines has unveiled a new class of "interaction models" designed for real-time conversational AI. These models process audio, video, and text in rapid 200-millisecond intervals, eliminating the need for separate turn-detection components. This architecture allows for continuous, interleaved input and output streams, enabling capabilities like speaking while listening and reacting to visual cues without explicit prompts. The system utilizes two co-trained models: a lightweight interaction model for live conversation and a background model for complex tasks like planning and tool use, ensuring low latency for users. AI

A developer built Sakhi, a Hindi voice-to-form application for India's community health workers, in six weeks. The system addresses challenges with unreliable cloud speech-to-text and intermittent connectivity in rural areas. Sakhi offers two modes: a workstation setup using Whisper and Gemma for voice transcription and data extraction, and an offline on-device mode on Android for text-based form filling and danger sign detection. AI

A new Windows desktop application called SEELS has been released, designed for running local Large Language Models (LLMs). Its core feature allows users to correct model responses and use these corrections to train custom LoRA adapters, effectively personalizing the LLM. The app also includes features like voice mode with local STT/TTS, a hardware dashboard, and supports GGUF models, with advanced features planned for future tiers. AI

Researchers at Together AI have found that current state-of-the-art speech recognition models exhibit a significant failure rate, averaging 39% error in transcribing street names, particularly for non-native English speakers who are 18% more likely to be misunderstood. This inaccuracy can lead to substantial real-world consequences, such as increased travel time and costs for services like ride-sharing. The study suggests that a synthetic data generation technique called "cross-lingual style transfer" can improve transcription accuracy by up to 60% with minimal training data. AI

Replit has launched a suite of AI-powered templates designed to streamline developer onboarding and accelerate the creation of AI-driven applications. These templates, available for various programming languages and frameworks, simplify complex setups for tools like vector databases and large language models. Notable examples include templates for Qdrant vector search, comparing Gemini and GPT-4, building AI support agents with OpenAI, and transcribing meetings using OpenAI Whisper. AI

Replit has partnered with OpenAI to integrate advanced AI models into its coding platform. The company is launching a new course on LLMs and GPT, and has introduced beta features powered by OpenAI's Codex model for code explanation. Additionally, Replit is exploring the use of GPT-3 for generating blog content, highlighting the growing synergy between AI and software development environments. AI