This isn’t traditional business news

Welcome to Morning Brew—the free newsletter designed to keep you in the know on the business news impacting your career, company, and life—in a way you didn’t know you needed.

Note: this isn’t traditional business news. Morning Brew’s approach cuts through the noise and bore of classic business media, opting for short writeups, witty jokes, and above all—presenting the facts.

Save time, actually enjoy business news, and join over 4 million professionals reading daily.

Check it out

Today:

• How DeepSeek Is Changing AI

• OpenAI Launches ChatGPT Gov for Agencies

• Qwen2.5-Max Outperform DeepSeek V3

• YuE: Open-Source Alternative to Suno

• OpenAI Researcher Resigns Over AGI Concerns

• Hugging Face Builds Open Alternative to DeepSeek’s R1

DeepSeek's STUNNING "Sputnik Moment" and Ex-Google CEO's WARNING for the US.

In 2023, Sam Altman doubted that small startups could rival OpenAI's GPT-4. However, DeepSeek, a Chinese company, recently released an open-source AI model trained for $6 million, rivaling OpenAI's best model. 

This breakthrough challenges the dominance of closed-source U.S. models, showing how open-source and cheaper AI can accelerate progress. Industry leaders are calling for a balanced ecosystem to ensure global competitiveness, with open-source models fostering innovation and broader access to AI technologies.

WATCH THE VIDEO ON YOUTUBE

OpenAI Launches ChatGPT Gov for Agencies

OpenAI introduces ChatGPT Gov, a tailored version of ChatGPT for U.S. government agencies. It offers enhanced security, privacy, and compliance, allowing agencies to deploy it in their own cloud infrastructure. With features like GPT-4o, custom GPTs, and administrative tools, ChatGPT Gov supports various government functions, from research to administrative tasks. Over 90,000 government users have already utilized ChatGPT for improving productivity, reducing costs, and enhancing service delivery to the public.

Qwen2.5-Max Outperform DeepSeek V3

Qwen2.5-Max is a large-scale Mixture-of-Expert (MoE) model pretrained on over 20 trillion tokens, designed to push the boundaries of AI performance. It has outperformed other leading models like DeepSeek V3 in key benchmarks, including coding and general capabilities. The model is available via Alibaba Cloud API and can be used in Qwen Chat. With advancements in post-training methods, Qwen2.5-Max aims to further improve intelligence and reasoning for AI models, opening new possibilities for knowledge exploration.

YuE: Open-Source Alternative to Suno

YuE, an open-source AI music tool, offers a free alternative to commercial services like Suno and Udio. Developed by Ruibin Yuan, it allows users to create songs up to five minutes long, with customizable elements such as style, mood, and voice type. YuE uses two AI models for speech and music generation, and can replicate various singing techniques. The tool is available on HuggingFace with demo examples on GitHub.

OpenAI Researcher Resigns Over AGI Concerns

Steven Adler, an OpenAI safety researcher, announced his resignation, criticizing the global race toward AGI (Artificial General Intelligence) as a "very risky gamble." He voiced concerns about AI labs and superpowers' race to develop AGI, emphasizing the dangers this poses. This departure adds to a series of internal conflicts at OpenAI, particularly regarding AI safety, which have surfaced over the past year. Adler's exit follows ongoing tensions within the company and broader industry debates over the risks of rapid AI advancement.

Hugging Face Builds Open Alternative to DeepSeek’s R1

Hugging Face researchers are working on an open-source version of DeepSeek’s AI reasoning model, R1. Unlike DeepSeek’s closed approach, which limits transparency, Hugging Face aims to replicate and fully open-source R1’s architecture and training data. The project, called Open-R1, has gained significant interest, with thousands of developers contributing. If successful, Open-R1 could enable broader AI research and development, benefiting both labs and the tech community.

🧠RESEARCH

Baichuan-Omni-1.5 Technical Report

Baichuan-Omni-1.5 is an advanced omni-modal model designed for seamless interaction across text, audio, and vision. It integrates a high-quality data pipeline, an innovative audio-tokenizer, and a multi-stage training strategy. The model outperforms current competitors, showing strong performance in multimodal medical benchmarks.

Qwen2.5-1M Technical Report

Qwen2.5-1M extends context length to 1 million tokens, improving long-context processing through techniques like long data synthesis and multi-stage fine-tuning. It includes an open-source inference framework with sparse attention and speed optimizations, offering 3x to 7x prefill speedups. Evaluations show strong performance in both long and short-context tasks.

Towards General-Purpose Model-Free Reinforcement Learning

MR.Q, a model-free deep reinforcement learning (RL) algorithm designed for diverse problem settings. It combines model-based representations to linearize value functions, achieving competitive performance across RL benchmarks. MR.Q aims to simplify RL by avoiding the complexities of model-based methods while maintaining strong general-purpose capabilities.

GeoPixel: Pixel Grounding Large Multimodal Model in Remote Sensing

GeoPixel is a high-resolution remote sensing model that supports pixel-level grounding for fine-grained visual understanding. It overcomes challenges in region-level comprehension by using a specialized dataset (GeoPixelD) and a tailored data generation process. GeoPixel outperforms existing models in segmentation tasks, offering improved precision for remote sensing analysis.

Emilia: A Large-Scale, Extensive, Multilingual, and Diverse Dataset for Speech Generation

Emilia is a large-scale multilingual dataset for speech generation, built from over 101k hours of real-world, spontaneous speech in six languages. Using the Emilia-Pipe preprocessing pipeline, it outperforms audiobook-based datasets by capturing diverse speech styles. Emilia's large scale supports advancements in human-like, multilingual speech generation.

🛠️TOP TOOLS

MagicSlides - AI-powered Google Slides add-on that transforms content into professional presentations with remarkable speed and efficiency. 

Flot AI - AI copilot designed to seamlessly integrate ChatGPT and other advanced language models into users’ daily workflows across various applications and websites.

Banter AI - AI-powered phone system that automates customer interactions for businesses

Fix My Code - AI-powered coding assistant developed by UserWay, designed to help developers create more accessible and ADA-compliant websites.

GitFluence - AI-powered Git command generator that simplifies version control workflows for developers.

📲SOCIAL MEDIA

🗞️MORE NEWS

• AI startup Turing tripled its revenue to $300 million in 2024, reaching profitability. With clients like OpenAI and Google, it provides human data trainers to AI labs, aiding in model improvement and data annotation.

• Figure AI is addressing humanoid robot safety by establishing a dedicated safety center. Focused on minimizing workplace injuries, the center will test robots’ stability, AI behavior, and detection capabilities while working alongside humans.

• Block has launched Goose, an open-source AI agent designed to be customizable by developers for various applications, including working with different large language models. Jack Dorsey, Block's head, praised DeepSeek's development approach, signaling strong support for open-source AI tools in the tech community.