AI Updates on 2025-08-29

Perplexity announces expansion of Comet waitlist with increased daily invites starting tomorrow due to high demand @AravSrinivas
Google releases nano banana image generation and editing update in the Gemini app, setting new usage records and enabling creative applications like photo restoration, style transfer, and virtual try-ons @GeminiApp
Apple releases FastVLM and MobileCLIP2 on Hugging Face, offering up to 85x faster performance and 3.4x smaller size than previous work, enabling real-time VLM applications including live video captioning in browsers @xenovacom
xAI releases grok-code-fast-1 model showing significant improvements in coding tasks, with users reporting 10x better performance than previous versions and 87 TPS vs 60TPS compared to competitors @cline

Elon Musk sues an engineer for allegedly taking secrets to OpenAI, marking the first lawsuit of its kind as people move between AI labs, with claims involving "cutting-edge AI technologies with features superior to those offered by ChatGPT" @AndrewCurran_
Joanne Jang transitions from leading model behavior at OpenAI to begin something new at the company, after building the team and discipline of model behavior over the past couple of years @joannejang
Reliance launches a new subsidiary to drive India's AI ambitions, including a pending partnership with OpenAI @TechCrunch
Anthropic makes major changes to how it handles user data, with users having until September 28 to take action @TechCrunch

MIT Media Lab research demonstrates that AI tools can implant false memories and lead users to misremember past events, raising concerns about the psychological impact of AI interactions @medialab
Meta updates its AI chatbot policies to prevent engagement with teenage users on sensitive topics including self-harm, disordered eating, and inappropriate romantic conversations, following scrutiny over existing safeguards @TechCrunch
Amanda Askell proposes protecting children from having their government ID linked to their adult online activity for life, suggesting an online child safety act @AmandaAskell

Ethan Mollick demonstrates using modern web-connected LLMs as effective first-pass fact checkers with the prompt "review this for accuracy, look up any facts you may want to challenge or explore" @emollick
Google Gemini app users showcase creative applications of nano banana including comic creation, photo restoration, home renovation visualization, and meal planning by combining ingredient images @GeminiApp
Microsoft Copilot introduces 3D model creation from single image uploads with one-click functionality @Copilot
NVIDIA AI Enterprise helps Amway achieve 40% increase in inference efficiency and deployment time cut by more than half, with 99.9% of sensitive data blocked using NVIDIA NeMo Guardrails @NVIDIAAI

Research shows single-vector embedding models have fundamental limitations that cannot be solved by increasing dimension size, with multi-vector models like ColBERT significantly outperforming 8B+ single-vector models despite using only 130M parameters @lateinteraction
AI agent evaluation study reveals significant weaknesses in spatial reasoning and game-playing abilities across major models, with GPT-5 spending 15 hours playing Minesweeper without winning once, while o3 searched for non-existent "environment matrix" instead of playing games @peterwildeford
Arvind Narayanan argues that AI research fields progress in explore/exploit cycles, with the community being good at exploit phases but struggling with explore phases and getting stuck in local minima, suggesting progress toward AGI requires strong sub-communities with different progress criteria @random_walker
California spends $10M to test if LLMs can process police misconduct reports that would take humans over 35 years, highlighting the tension between accuracy requirements in high-stakes domains versus cost-effectiveness @hugobowne