Your daily AI dose

Mindstream is your one-stop shop for all things AI.

How good are we? Well, we become only the second ever newsletter (after the Hustle) to be acquired by HubSpot. Our small team of writers works hard to put out the most enjoyable and informative newsletter on AI around.

It’s completely free, and you’ll get a bunch of free AI resources when you subscribe.

Click here to subscribe

Today:

• Ogunlesi Joins OpenAI Board

• MiniMax-01 Goes Open-Source

• OpenAI Unveils Task Scheduler

• Amazon Reinvents Alexa with AI

• Kokoro TTS Debuts on HuggingFace

Ogunlesi Joins OpenAI Board

Adebayo Ogunlesi, a renowned global leader in infrastructure and finance, has joined OpenAI’s Board of Directors. Ogunlesi brings decades of expertise from roles at Global Infrastructure Partners, Credit Suisse, and BlackRock. His appointment strengthens OpenAI’s leadership in advancing AI governance, safety, and infrastructure development. Ogunlesi expressed excitement about contributing to OpenAI’s mission of unlocking AI’s potential responsibly to drive innovation and foster global economic growth.

MiniMax-01 Goes Open-Source

MiniMax has open-sourced its groundbreaking MiniMax-01 series, featuring the MiniMax-Text-01 language model and MiniMax-VL-01 visual multi-modal model. Using innovative Lightning Attention architecture, the models handle up to 4 million tokens, far surpassing competitors. Offering superior efficiency, cost-effectiveness, and multi-modal capabilities, the series aims to drive advancements in AI Agents. Accessible via GitHub and MiniMax's platform, these models encourage research and innovation in long-context AI, marking a new era in AI technology.

OpenAI Unveils Task Scheduler

OpenAI’s new feature, ChatGPT Tasks, brings ChatGPT closer to becoming a personal assistant. Currently in beta, it enables scheduling reminders, notifications, and recurring tasks, accessible across web, desktop, and mobile platforms. This move hints at OpenAI’s broader ambitions, with Tasks potentially laying the groundwork for an AI agent called Operator. By integrating task management into its ecosystem, OpenAI positions ChatGPT as a competitive productivity tool in a crowded market.

Amazon Reinvents Alexa with AI

Amazon is reportedly transforming Alexa into a powerful AI agent, aiming to enhance its capabilities from simple tasks like playing music to acting as a personalized concierge. Powered by generative AI, Alexa will tackle technical challenges like minimizing hallucinations and improving response speed. This shift reflects Amazon's ambition to compete in the evolving AI agent space, as consumers increasingly demand advanced virtual assistants capable of handling everyday tasks and personalized recommendations.

Kokoro TTS Debuts on HuggingFace

Kokoro, a new open-source text-to-speech (TTS) model with 82 million parameters, is now available on HuggingFace. Despite being trained on less than 100 hours of audio and limited to American and British English, it delivers voice quality comparable to commercial services like Eleven Labs. Users can select from 10 voices, but it lacks voice cloning and multilingual support. Licensed under Apache 2.0, Kokoro offers developers accessible, high-performance TTS capabilities.

🧠RESEARCH

The Lessons of Developing Process Reward Models in Mathematical Reasoning

Researchers improved how AI solves math problems by developing Process Reward Models that detect reasoning errors. They found traditional data methods and evaluations were flawed, so they combined multiple techniques to create a more accurate and efficient model. Their new approach outperforms existing models and sets guidelines for future AI research.

Tensor Product Attention Is All You Need

Tensor Product Attention (TPA), a memory-efficient mechanism for language models, addressing the challenge of handling long input sequences. By compactly representing data using tensor decompositions, TPA reduces memory demands and enhances performance. Their new T6 model outperforms standard Transformers, demonstrating superior scalability and quality in sequence modeling tasks.

BIOMEDICA: An Open Biomedical Image-Caption Archive, Dataset, and Vision-Language Models Derived from Scientific Literature

BIOMEDICA, a comprehensive, open biomedical dataset with over 24 million image-text pairs from 6 million articles. Designed to enhance vision-language models (VLMs) in medicine, their BMCA-CLIP models outperform benchmarks in tasks like pathology and radiology, achieving a 6.56% average improvement. The dataset and tools are publicly available for collaboration.

MinMo: A Multimodal Large Language Model for Seamless Voice Interaction

MinMo, a new multimodal large language model with 8 billion parameters, revolutionizes voice interaction by enabling real-time, human-like conversations. Using 1.4 million hours of training data, MinMo achieves state-of-the-art performance in voice comprehension and generation. It supports duplex interaction, nuanced speech control, and delivers low-latency responses for seamless communication.

SPAM: Spike-Aware Adam with Momentum Reset for Stable LLM Training

SPAM (Spike-Aware Adam with Momentum Reset), an optimizer addressing gradient spikes during large language model (LLM) training. These spikes, up to 1000x larger than typical gradients, harm performance and increase inefficiencies. SPAM uses momentum reset and spike-aware gradient clipping to stabilize training, outperforming existing optimizers while improving memory efficiency and scalability.

🛠️TOP TOOLS

BlipCut - AI-powered video translation and localization platform that enables content creators to break language barriers and reach global audiences. 

Prezo - AI-powered platform that combines presentation creation, document editing, and website building into a single, user-friendly interface. 

Supermeme AI - AI-powered meme generator that streamlines the process of creating engaging, shareable content.

Layla AI - Travel assistant designed to revolutionize trip planning and enhance the overall travel experience.

AI Library - Comprehensive platform that houses over 800 neural networks and AI tools designed for content creation and workflow optimization.

📲SOCIAL MEDIA

🗞️MORE NEWS

• Microsoft's AI security team tested over 100 tools, revealing simple attacks often surpass complex methods. Human expertise remains critical for tackling ethical concerns, biases, and evolving AI vulnerabilities in increasingly integrated applications.

• Nvidia backed Taiwanese startup MetAI with $4M in seed funding to advance AI-powered digital twins. MetAI accelerates industrial AI by generating “SimReady” environments in minutes, bridging simulation and real-world operations for robotics and automation.

• Apple has joined the UALink Consortium to develop advanced interconnects for next-gen AI clusters. This initiative aims to overcome connectivity challenges and support the growing demands of AI, enhancing performance and scalability.

• Microsoft and Pearson announced a multiyear partnership to enhance global AI skills. They aim to deliver AI-powered learning tools, personalized education, and certifications, leveraging Microsoft's Azure and Pearson's expertise to prepare the workforce for AI-driven industries.

• Healthcare AI startup Qventus raised $105M in Series D funding led by KKR. Its AI assistant automates surgical preparation, driving rapid growth. Qventus plans to expand applications, hire engineers, and achieve breakeven by 2025.

• AI is aiding the assembly of a record-breaking quantum computer with 1180 ultracold atom-based qubits. Precise laser arrangements ensure qubit alignment, advancing quantum computing towards larger, more accurate systems.