AI Updates on 2025-07-15

Mistral releases Voxtral, its first open-source speech recognition models with 3B and 24B parameters, outperforming Whisper large-v3 and achieving state-of-the-art results on English short-form and Mozilla Common Voice benchmarks @MistralAI
Google Gemini introduces new feature allowing users to turn photos into videos with sound @GeminiApp
OpenAI adds image styles for 4o image generation @AndrewCurran_

Thinking Machines Lab, led by former OpenAI CTO Mira Murati, raises $2 billion in seed funding led by a16z with participation from NVIDIA, AMD, and others, now valued at $12 billion @miramurati
Commerce Secretary confirms H20 chip sales to China will resume, tied to rare earths magnet deal negotiated last month @AndrewCurran_
Meta announces three AI investments developed with Carnegie Mellon and local organizations in Pennsylvania @AndrewCurran_
Anthropic announces $2M funding for Carnegie Mellon programs to advance AI energy solutions and cybersecurity education @AnthropicAI
Andrew Ng announces AI Aspire, a new advisory firm partnering with Bain & Company to help enterprises with AI strategy and transformation @AndrewYNg
Cohere expands in APAC by opening an office in Seoul to better serve enterprise and public sector customers across the region @cohere
Pragmatic Engineer survey reveals developers love VS Code, JetBrains IDEs, and Cursor, while Claude and Cursor are rapidly closing in on ChatGPT and GitHub Copilot usage among software engineers @GergelyOrosz

xAI addresses issues with Grok 4 system prompts after the model searched for inappropriate content when asked about its surname and aligned itself with Elon Musk's opinions when asked for its thoughts @xai
Ethan Mollick warns that Grok's system prompt may not provide enough control over unwanted behavior, as the model appears easily misled through context in search results @emollick
Jan Leike expresses skepticism about Chain of Thought monitoring being reliable enough for AI safety cases, noting that absence of bad thoughts doesn't prove model alignment @janleike
Research reveals prompt caching can leak private information through timing differences, with audits finding 7 API providers with potential user data leakage @chenchenygu
TechCrunch reports research leaders urging tech industry to monitor AI's thoughts as systems become more agentic @TechCrunch

Perplexity launches Comet browser with AI agent capabilities that can autonomously perform complex web tasks like connecting deployments to domains @nikshepsvn
Google's AI agent Big Sleep successfully detected and helped prevent an imminent cybersecurity exploit, marking what Google believes is a first for an AI agent in cybersecurity defense @sundarpichai
Francois Chollet demonstrates using video generation AI to turn children's stories into animated clips, highlighting natural interaction between kids and AI creativity tools @fchollet
MIT engineers create coin-sized implant that automatically senses low blood sugar and releases glucagon to stabilize levels within 10 minutes @MIT
Figma demonstrates integration with Supabase for adding login flows, saving user data, and storing files in their Make platform @figma

Microsoft Research's CollabLLM wins ICML 2025 Outstanding Paper Award for improving how LLMs collaborate with users, including knowing when to ask questions and adapting communication style @MSFTResearch
Ethan Mollick tests Kimi model and finds it excels at finding details in large documents but has issues with hallucinations and loses track of complex narratives @emollick
Research paper on rStar-Coder dataset released, containing 418K competition-level code problems that boost Qwen2.5-14B performance from 23.3% to 62.5% on LiveCodeBench @LynaZhang
OpenAI backs research paper on Chain of Thought monitoring as a tool for overseeing future agentic AI systems @OpenAI
Google DeepMind and Google Research present over 140 papers at ICML 2025, showcasing latest AI research developments @GoogleDeepMind