AI Updates on 2025-10-25

OpenAI is reportedly training a new music model, which would be their first since Jukebox in 2020, marking a significant shift as they previously avoided legal conflicts with music labels @AndrewCurran_
OpenAI demonstrated a new bidirectional speech model at their London Frontiers event that can translate speech in real-time while still speaking by waiting for complete verbs, with potential launch in coming weeks @btibor91
xAI introduces Mika, the newest Grok Companion, with video content created using Grok Imagine @xai
Meituan releases LongCat-Video, a foundational video generation model with 13.6B parameters supporting Text-to-Video, Image-to-Video, and Video-Continuation generation tasks under MIT license @reach_vb
Odyssey ML announces Odyssey-2, described as representing a totally new capability for AI, launching Monday at 10AM PT @olivercameron

Some OpenAI employees reportedly feel the company is becoming too much of a media lab, though leadership maintains they remain a superintelligence lab at heart with media projects funding core research @AndrewCurran_
Film director Paul Schrader predicts we're two years away from the first AI feature film, aligning with Elon Musk's timeline of fully generated movies being watchable by 2026 and high quality by 2027 @AndrewCurran_
Microsoft's business model differs fundamentally from Google's with only ~5% revenue from ads versus Google's ~80%, explaining their different approaches to developer tools and search @GergelyOrosz
HVAC has surpassed semiconductors, computers/servers, and data centers as the biggest winner of net-new hardware spending since 2022 @a16z
Analysis shows Grok 4 training used less water than a square mile of US farmland uses annually, highlighting efficiency in AI training @a16z

New AI browsers from OpenAI and Perplexity promise increased productivity but come with heightened security risks that users should be aware of @TechCrunch
A high school student in Baltimore County was reportedly handcuffed and searched after an AI security system incorrectly flagged his bag of chips as a possible firearm @TechCrunch
Most people working at the cutting edge of AI appear to have no long-term plan for their unsustainable work habits, raising concerns about burnout in the field @natolambert

Microsoft announces 12 new Copilot features designed to make a difference in real-world applications rather than pulling users away from their tasks @mustafasuleyman
Copilot Mode in Edge serves as an intelligent browsing companion that can read tabs, take actions, and turn browsing history into helpful storylines @Copilot
Claude demonstrates the ability to fetch and navigate its own online documentation when asked questions about itself, showing improved self-referential capabilities @simonw
Giving reasoning models access to data connections for real-time searches and refinement represents a significant leap over traditional RAG systems @emollick
User reports building 5 personal mini-apps in 2 hours with no coding, debugging, or setup required, highlighting the accessibility of modern AI development tools @iVinay

Research shows that creating SVGs activates the same semantic concepts in LLMs as asking them to describe the same objects, revealing interesting insights into AI representation @emollick
Carnegie Mellon researchers win IROS Best Student Paper Award for Neural MP, a generalist neural motion planner that improves success rates by 23%, 17%, and 79% over state-of-the-art sampling, optimization, and learning-based planners @rsalakhu
DeepMind announces progress on AI for materials science, with exciting developments in the AI for Science team @demishassabis
One-year retrospective on KernelBench reveals lessons learned in the journey toward automated GPU/CUDA kernel generation, showing remarkable community progress @simonguozirui
Stanford NLP celebrates its 25th anniversary, highlighting its role in inspiring NLP groups worldwide that led to today's LLMs @stanfordnlp