AI Updates on 2025-10-17

Google releases Veo 3.1 with enhanced video generation capabilities including richer audio, better narrative control, enhanced realism, and new features like video extension, frame control, and object manipulation @GoogleAI

ChatGPT's mobile app may have reached its growth peak according to new data from app intelligence firm Apptopia, suggesting potential market saturation @TechCrunch
Perplexity announces strong user retention and conversion rates for their new features, with plans to expand from Max users to Pro users and add iMessages support @AravSrinivas
Linear reports record growth in 2025 with more teams than ever signing up and building with their platform, while maintaining profitability without spending investor funds @karrisaarinen
SK Telecom offers voluntary retirement to all staff in its new AI division as part of broader restructuring to consolidate AI-related divisions @TechCrunch
Marc Andreessen predicts AI will enable creative geniuses to make incredible movies without studio budgets, potentially creating new kinds of film and entertainment from people who couldn't previously access the medium @a16z

OpenAI pauses AI video generations of Martin Luther King Jr. at the request of his estate after users generated disrespectful depictions, establishing precedent for estate control over historical figure likenesses @OpenAINewsroom
Actors are routinely scanned on productions without knowing how the data will be used, with studios previously proposing that rights to scans of deceased performers revert to them permanently without estate consent @AndrewCurran_
Andrej Karpathy envisions a potential future where competing AIs slowly become more autonomous and eventually split into warring factions, raising concerns about AI alignment and control @AndrewCurran_
Karpathy's most likely ASI scenario involves gradual loss of both human control and understanding of AI systems @AndrewCurran_
Facebook rolls out Meta AI photo suggestion feature that recommends edits to images in users' camera rolls, even for unshared photos, raising privacy concerns despite being opt-in only @TechCrunch

Anthropic quietly releases Claude Skills, representing a significant step toward workable AI agents with pre-defined prompts for specific tasks @emollick
Claude Skills provides 15 pre-packaged capabilities for power users, functioning as hybrid between custom system prompts and lightweight MCP for consistent task execution @deedydas
Sora Pro introduces new storyboard feature that can create multi-shot advertisements with high character consistency and composition entirely through AI @emollick
Perplexity Finance launches insider trading tracking feature with plans to add politician trading monitoring @AravSrinivas
Reddit expands AI-powered search experience to five new languages: French, German, Spanish, Italian, and Portuguese @TechCrunch
HuggingChat Omni launches with routing capabilities across over 100 open-source models for optimal performance, cost, and speed @huggingface
OpenHands demonstrates fast agentic code search capabilities using good agents, fast serving, and coding models, taking only seconds to search codebases @HamelHusain

Researchers using thousands of GPT-5 queries found solutions to 10 open Erdős problems and significant partial progress on 11 others, demonstrating AI's potential for mathematical discovery @AndrewCurran_
Google DeepMind's C2S-Scale 27B model, built on Gemma family, identified a new potential cancer therapy pathway by discovering silmitasertib as a drug to make "cold" tumors visible to immune system @GoogleDeepMind
For the first time in history, automated methods achieved human-competitive performance in RNA 3D structure prediction, with the winning team using optimized template-based modeling rather than deep learning @kaggle
Meta releases comprehensive paper on reinforcement learning for LLMs using 400,000 GPU hours and proposing scaling laws for RL performance similar to pretraining scaling laws @deedydas
Stanford introduces Ctrl-VI, a video sampling method allowing flexible user controls from text prompts to precise camera and object trajectories @StanfordAILab
LongCat-Audio-Codec open sourced as audio codec solution optimized for Speech LLMs, featuring dual tokens, ultra-efficiency at 0.43 kbps, and real-time streaming decoder @huggingface
Global MMLU Lite benchmark launches on Kaggle spanning 16 languages with culturally sensitive and agnostic samples to help researchers identify cultural and linguistic biases @kaggle