DeepSeek V3 Outperforms AI Rivals

PLUS: ARC-AGI-2 Test Stumps AI, Midjourney Expands Into AI Writing and more.

In partnership with

Your job called—it wants better business news

Welcome to Morning Brew—the world’s most engaging business newsletter. Seriously, we mean it.

Morning Brew’s daily email keeps professionals informed on the business news that matters, but with a twist—think jokes, pop culture, quick writeups, and anything that makes traditionally dull news actually enjoyable.

It’s 100% free—so why not give it a shot? And if you decide you’d rather stick with dry, long-winded business news, you can always unsubscribe.

Today:

  • DeepSeek V3 Outperforms AI Rivals

  • Ant Group Advances AI With China Chips

  • OpenAI Enhances ChatGPT Voice Mode

  • ARC-AGI-2 Test Stumps AI

  • Midjourney Expands Into AI Writing

Chinese AI startup DeepSeek has launched DeepSeek-V3-0324, a 641GB model that runs efficiently on high-end consumer hardware like Apple's Mac Studio. Its open-source MIT license challenges Western AI firms' closed models. Using a mixture-of-experts architecture, it activates only relevant parameters, making it faster and more efficient. This release marks a shift in AI deployment, emphasizing open access and efficiency over the traditional high-cost, cloud-based approach.

Why It Matters to AI

  1. Open-Source Disruption – DeepSeek’s MIT-licensed model challenges the dominance of closed AI ecosystems like OpenAI’s, making cutting-edge AI more accessible.

  2. Efficient AI on Consumer Hardware – Running a powerful model locally on a Mac Studio reduces reliance on expensive cloud-based GPUs, redefining AI infrastructure.

  3. China’s AI Surge – The rapid advancement of China’s AI models, now rivaling Western counterparts, accelerates global AI competition and innovation.

Ant Group, backed by Jack Ma, has developed an AI training technique using Chinese-made chips from Alibaba and Huawei, reducing costs by 20%. By leveraging the Mixture of Experts (MoE) approach, their models achieved results comparable to Nvidia’s H800 chips. This breakthrough marks a significant step in China’s AI independence, challenging Nvidia’s dominance and highlighting the growing capabilities of domestic semiconductor technology.

Why It Matters to AI

  1. AI Independence – China’s success in training AI models with domestic chips reduces reliance on Nvidia and Western technology.

  2. Cost Efficiency – A 20% reduction in training costs makes AI development more accessible and scalable.

  3. Competitive AI Innovation – The use of MoE models demonstrates China's progress in optimizing AI performance with limited resources.

OpenAI has upgraded its Advanced Voice Mode in ChatGPT, making the AI assistant more engaging and less interruptive. The update reduces unnecessary interruptions when users pause while speaking and improves responses to be more direct, concise, and creative. Free and paid users benefit, with premium subscribers getting enhanced personality features. These updates come as competition intensifies, with startups like Sesame and major players like Amazon advancing AI voice assistants.

Why It Matters to AI

  1. Improved User Experience – A more natural and responsive AI assistant enhances real-time conversations.

  2. Rising Competition – OpenAI faces increasing pressure from startups and tech giants developing their own voice assistants.

  3. AI Voice Evolution – Advances in conversational AI signal a shift toward more human-like digital assistants.

🧠RESEARCH

This paper explores reducing the number of visual tokens in image-processing models to cut computing costs while maintaining accuracy. The authors introduce a method that selects only the most useful tokens. Tests show that over 50% of tokens can be removed with little impact, suggesting a more efficient approach to image representation.

This paper proposes using AI-generated video as the core of future game engines, enabling endless, interactive content creation. The authors introduce a framework for Generative Game Engines (GGE), highlighting their potential for realism, physics modeling, and player control. This approach could transform game development by reducing costs and expanding creativity.

MAPS is a multi-agent AI system designed to solve complex scientific problems using multiple data types, like text and diagrams. It employs seven specialized agents and Socratic questioning to enhance reasoning and reflection. MAPS outperforms existing models by 15.84%, demonstrating improved problem-solving and adaptability across various datasets.

Bottleneck Sampling is a method to speed up AI-generated images and videos without retraining. By processing at lower resolutions during intermediate steps, it cuts computing costs while keeping quality intact. Tests show up to 3× faster image generation and 2.5× faster video generation with results matching full-resolution methods.

This paper explores using multimodal AI models (MLLMs) as judges for evaluating AI-generated content across different formats, like images, audio, and video. The authors introduce benchmarks to assess these models’ accuracy and fairness. Results show they perform well in understanding tasks but struggle with generation, revealing biases and hallucinations.

🛠️TOP TOOLS

FaceSwapper - AI-powered online platform that offers a suite of advanced image and video editing tools, primarily focused on face swapping technology.

Anakin AI - No-code AI app builder that empowers users to create customized AI applications for automating tasks, generating content, and answering questions.

AI Human Generator - Create hyperrealistic full-body images of people who don’t exist. 

OpenRead - AI-powered interactive platform designed to revolutionize academic research and literature analysis.

Qlip AI - AI-powered platform designed to help content creators efficiently repurpose long-form videos into short, shareable clips for social media. 

📲SOCIAL MEDIA

🗞️MORE NEWS

  • A new AI test, ARC-AGI-2, challenges AI models with unseen pattern-recognition tasks. Most leading models score around 1%, far below humans. The test emphasizes efficiency, aiming to measure true intelligence beyond brute computing power.

  • Midjourney, known for AI image generation, is expanding into text AI. Partnering with NYU, it developed techniques to boost creativity in AI writing. These methods improve storytelling diversity, benefiting content creators, marketers, and AI developers.

  • OpenAI restructured leadership, expanding COO Brad Lightcap’s role while CEO Sam Altman shifts focus to research and products. New promotions include Mark Chen as chief research officer and Julia Villagra as chief people officer.

  • Thrive Capital is leading a $40M investment in AI startup Rogo, valuing it at up to $350M. Rogo, which develops AI software for Wall Street analysts and bankers, uses OpenAI and Anthropic models to automate financial research. Investors see high demand for such tools in finance.

  • Google is rolling out real-time AI video features for Gemini, allowing it to "see" screens and camera feeds. Available to some Google One AI Premium subscribers, this feature enables Gemini to analyze visuals and answer questions in real time.

  • Microsoft is launching 11 AI-powered security agents to automate repetitive cybersecurity tasks, reducing analyst burnout and improving efficiency. These agents handle phishing detection, regulatory notifications, and more, offering configurable autonomy and human oversight.

What'd you think of today's edition?

Login or Subscribe to participate in polls.

Reply

or to participate.