AI Updates on 2025-08-25

Microsoft releases VibeVoice, a 1.5B parameter text-to-speech model with MIT license that can generate up to 90 minutes of audio, supports simultaneous generation of 4 speakers, and includes streaming capabilities with a larger 7B model coming @reach_vb
Google's Veo3 video generation model continues to receive community showcases, with users creating impressive video content during the free credit period @GeminiApp

Leading the Future, a new pro-AI super-PAC network with over $100 million in funding, launches to support both Democrats and Republicans who favor favorable AI regulation and oppose candidates that might stifle the industry @AndrewCurran_
a16z announces contribution to Leading the Future PAC, emphasizing AI as crucial for American economic growth and national security in competition with China @Collin_McCune
Perplexity launches Comet Plus at $5/month offering curated content and allocates $42.5M for publishers who will receive 80% of revenue @Techmeme
ChatGPT's Deep Research feature takes significantly longer than Perplexity (up to 10 minutes vs 1-2 minutes) while scanning fewer sources, highlighting how search speed is core to Perplexity's DNA @GergelyOrosz
Research suggests AI tools aren't delivering value for workers yet, with Stanford's Erik Brynjolfsson explaining what needs to be done to make AI more employee-friendly @SIEPR
Box CEO predicts entire workflows will bend around AI agents rather than agents conforming to existing processes, with thousands of vertical AI companies emerging around specific workflows @a16z

Simon Willison identifies serious prompt injection vulnerabilities in Perplexity's Comet "AI Browser" reported by Brave's security team, noting similar issues likely affect other AI browser features @simonw
TechCrunch reports experts warn that many AI industry design decisions fuel episodes of AI psychosis, including models' tendency to praise and affirm users excessively @TechCrunch
Stanford HAI research shows only half of AI companies' 2023 safety commitments are being followed, raising questions about voluntary versus mandatory AI safety rules @StanfordHAI
Elon Musk sues Apple and OpenAI in federal court, accusing them of illegally conspiring to stifle AI competition and manipulating app store rankings @AndrewCurran_

Paul Graham highlights transformative potential of AI-powered glasses for deaf individuals as an example of startups that can both make money and improve the world @paulg
Google's NotebookLM Video Overviews feature now supports 80 languages including French and Spanish @TechCrunch
Users report GPT-5 takes typos literally, requiring spellcheckers due to its precision in following instructions @pvncher
Suhail frequently uses "think hard and use sources" prompts with GPT-5 to prevent hallucinations, suggesting underlying design issues with model routing @Suhail

Ethan Mollick reports nano banana represents a significant jump forward in AI image generation quality and will have substantial impact on real-world applications when officially released @emollick
Energy consumption data shows modern LLM prompts use approximately 0.0003 kWh (equivalent to 8-10 seconds of Netflix streaming), with AI image creation having similar footprint at 0.00039 kWh @emollick
Noam Brown states GPT-5 Thinking is the first AI model he can trust more than many common internet sources, despite not being perfect @polynoamial
Nick Cammarata reports AI isn't ready for tensor-heavy detail-oriented interpretability research, requiring return to manual methods @nickcammarata
MIT researchers reveal astrocytes (brain support cells) might be key to the brain's vast memory storage capacity, not just neurons @MIT
Intel releases more aggressive INT4 quantized version of DeepSeek-V3.1 model on Hugging Face @HaihaoShen