AI Updates on 2025-10-29

OpenAI releases gpt-oss-safeguard models for safety classification, fine-tuned versions of their open models available under Apache 2.0 license on Hugging Face @OpenAI
Cursor announces Cursor 2.0 featuring their first coding model Composer, a frontier coding model that completes tasks in under 30 seconds @cursor_ai
Google announces Gemini Deep Think enhanced reasoning model as part of their AI research partnership funding @GoogleDeepMind
OpenAI launches Pulse feature now available to Pro users on web @OpenAI

OpenAI commits to approximately 30 gigawatts of compute with total cost of ownership of about $1.4 trillion over the years, with goals for automated AI research intern by September 2026 and true automated AI researcher by March 2028 @sama
Anthropic reports 10x growth in run rate revenue in Asia-Pacific region over the past year, with companies like Rakuten, Nomura Research Institute, and Panasonic now using Claude @AnthropicAI
Character AI implements major policy changes requiring users under 18 to no longer engage in open-ended chats with AI, including romantic dialog, while adding stronger age verification and funding an AI safety lab @AndrewCurran_
Early-stage startups increasingly choosing "hip" alternatives like Vercel, Render, Railway, and Supabase over traditional cloud services like AWS for initial hosting and databases @GergelyOrosz
AI coding agents making traditional developer productivity metrics like PR frequency largely meaningless, as they can trivially generate pull requests @GergelyOrosz
NVIDIA's market cap of $5 trillion now exceeds the aggregated stock markets of all countries except the United States, China, and Japan @TechCrunch
Voice-based coding interfaces gaining traction with developers, with Cursor adding native voice mode support and companies like Wispr seeing increased adoption for AI-powered development workflows @GergelyOrosz

Simon Willison warns about security and privacy risks in AI browser agents, stating the risks "feel insurmountably high" until security researchers thoroughly evaluate these products @random_walker
Anthropic research reveals evidence of introspective capabilities in Claude, showing models can sometimes detect injected concepts in their neural patterns, though this works inconsistently and most of the time models fail to exhibit awareness @AnthropicAI
OpenAI's commitment to permanently remain in California was instrumental in gaining Attorney General approval for their for-profit conversion @AndrewCurran_
Concerns raised about AI's impact on social reality and collective sense-making, with warnings about "exponential loneliness" and "exponential interpersonal misalignment" as personal AI capabilities scale @tuhin

Microsoft announces App Builder and Workflow agents in M365 Copilot, allowing users to build apps and automate workflows in minutes directly in chat @satyanadella
Perplexity launches Email Assistant for Pro subscribers with 14-day trial, featuring private drafting and labeling that never logs email content @perplexity_ai
Rocket Mortgage partners with Sierra to transform homeownership experience with AI, focusing on better customer experience rather than just automation @btaylor
NVIDIA Earth-2 enables ultra-fast, high-resolution weather simulations, turning hours of compute into seconds for better disaster preparedness and risk analysis @NVIDIAAI
Google partners with NextEra to reopen the Duane Arnold Energy Center in Iowa specifically to power data centers @TechCrunch
Figma introduces Make kits to integrate design systems with Make, allowing AI to design and build software that matches existing design investments @manosaie

Stanford releases SLP-Helm benchmark testing how AI models diagnose pediatric speech disorders, revealing promises, pitfalls, and bias in AI-assisted speech therapy @StanfordAILab
Research demonstrates AI helping solve a 42-year-old open math problem with expert human guidance, showcasing AI's potential in intellectually challenging academic work @emollick
Google DeepMind develops RL-based system to discover creative chess puzzles, doubling the number of novel puzzles compared to original training data while maintaining aesthetic diversity @TZahavy
New research on training LLMs to discover reasoning abstractions shows that allocating test-time compute to generating abstractions yields greater gains than producing additional solutions @rsalakhu
Study reveals distinct prompts map to unique hidden states inside models, enabling reverse engineering from hidden states back to original prompts @emollick
DeepSeek research suggests new methods for improving AI's ability to remember information @techreview
Quantum computing breakthrough achieves 120 qubit entanglement, the largest entangled state ever achieved on a quantum computer @jaygambetta