AI Updates on 2025-08-28

Microsoft releases MAI-1-preview, their first foundation model trained end-to-end in-house, now in public testing on LMArena @mustafasuleyman
Microsoft introduces MAI-Voice-1, described as their most expressive, natural voice generation model that can generate a minute of audio in under 1 second on a single GPU @mustafasuleyman
xAI launches Grok Code Fast 1, a speedy and economical reasoning model that excels at agentic coding, available for free on GitHub Copilot, Cursor, and other platforms for 7 days @xai
OpenAI introduces gpt-realtime, their best speech-to-speech model for developers, with the Realtime API officially out of beta and ready for production voice agents @OpenAI
Cohere releases Command A Translate, a state-of-the-art model designed for high-quality translation tasks that outperforms all other models in translation quality @cohere
Tencent AI Lab releases HunyuanVideo-Foley, an end-to-end Text-Video-to-Audio framework for generating high-fidelity audio trained on a massive 100k-hour multimodal dataset @TencentHunyuan

Andrew Ng identifies parallel agents as an emerging important direction for scaling up AI, offering a path to improve results without making users wait longer, as LLM prices per token continue to fall @AndrewYNg
Paul Graham argues that people who are good at programming will use AI to take jobs from those who are mediocre at it, emphasizing that being driven to build things matters more than just technical proficiency @paulg
Karri Saarinen reports that 45% of enterprise customers have enabled Cursor AI agents in Linear within one week of launch, showing rapid adoption of AI agents in product teams @karrisaarinen
Liquid AI achieves over 1 million foundation model downloads through Hugging Face, demonstrating community interest in tiny models designed from first principles @ramin_m_h
Framer raises $100 million Series D at $2 billion valuation, claiming over half a million monthly active users for their no-code website builder @TechCrunch

Heidy Khlaaf from AI Now Institute is recognized on TIME's 100 AI list for addressing safety revisionism of AI in defense and safety-critical contexts where safety has been obscured with little accountability @HeidyKhlaaf
Hamel Husain reports experiencing significant quality issues with GPT-5 in the API, including crazy typos, punctuation errors, and nonsense responses escalating throughout the week @clairevo
Arvind Narayanan observes that more capable models don't necessarily mean more useful products, noting Anthropic's decision to train on user data after holding out for a long time @random_walker

Eugene Yan demonstrates a bilingual recommendation system that combines LLM capabilities with item IDs, creating an LLM-recsys hybrid that can recommend based on user behavior while allowing natural language explanations @eugeneyan
Reid Hoffman highlights Alpha School's use of AI to personalize education, where AI handles pacing and lesson delivery for two hours daily while teachers focus on life skills like teamwork and entrepreneurship @reidhoffman
Google Gemini introduces nano banana image editing capabilities, allowing users to upload multiple pictures and combine them into one image with specific prompts for edits and adjustments @GeminiApp
Qwen releases a Web Dev prompt feature, a design-focused AI assistant for frontend development that helps build websites using React or HTML with TailwindCSS in single-file outputs @Alibaba_Qwen
OpenAI announces new Codex features including IDE extension, cloud-local environment integration, GitHub code reviews, and revamped CLI, all powered by GPT-5 @OpenAIDevs

Ethan Mollick highlights the deep mystery of LLMs: how matrix multiplication for next word prediction manages to simulate human thought well enough to perform human-like tasks, referencing Wolfram's theory about computational irreducibility @emollick
François Chollet explains that reasoning is not a category of tasks but an ability underpinned by compositional generalization, noting that reasoning tasks can be solved without reasoning through memorization @fchollet
NVIDIA's Cosmos Reason tops the physical reasoning leaderboard on Hugging Face, excelling in physical AI applications like data curation, robot planning, and video analytics @NVIDIAAIDev
Berkeley AI Research publishes editorial questioning whether data will solve robotics, with Ken Goldberg arguing that good old-fashioned engineering can close the 100,000-year data gap in robotics @berkeley_ai
A zero-day vulnerability in Chrome was discovered by an AI system, demonstrating AI's capability in cybersecurity research @Sauers_