AI Updates on 2025-10-27

Anthropic expands Claude for Financial Services with Excel add-in, real-time data connectors to LSE, Moody's, and other financial platforms, plus pre-built Agent Skills for cash flow models and coverage reports @AnthropicAI
Microsoft Copilot introduces long-term memory feature, allowing users to store and recall important information across conversations while maintaining user control over memory management @Copilot
OpenAI updates GPT-5 with input from 170+ mental health experts, reducing inadequate responses in sensitive situations by 65-80% @OpenAI
MiniMax releases MiniMax-M2, a 230B MoE model with 10B active parameters under MIT license, ranking #1 among open-source models on Artificial Analysis benchmarks @reach_vb
Keras 3.12 released with GPTQ quantization API, model distillation API, and PyGrain dataset support across the data API @fchollet

OpenAI proposes building 100 gigawatts of new energy capacity annually and estimates their 5-year infrastructure plans will require 20% of existing skilled trades workforce including electricians and mechanics @AndrewCurran_
Mercor, connecting AI labs with domain experts for model training, reportedly close to raising $350 million at $10 billion valuation @TechCrunch
Amazon's Annapurna Labs, acquired for $350M in 2015, now powers training of Anthropic's Claude models as a cheaper alternative to Nvidia @deedydas
Raghu Raghuram predicts manual labor bottlenecks in data center construction will drive robotics innovation, with infrastructure needs downstream of AI innovation @a16z
Fitbit's Gemini-powered health coach rolls out to Premium subscribers in the U.S. on Android @TechCrunch

Gergely Orosz reports Perplexity started generating fake sources that don't exist, highlighting persistent hallucination issues in LLM products despite previous improvements @GergelyOrosz
New research identifies Pangram as top AI detector with <0.5% false positive/negative rates, effective even on text processed through "stealth" humanizers and new models like GPT-5 @deedydas
Mustafa Suleyman emphasizes AI's value should be measured by daily life improvements: creating, connecting, feeling joy, and chasing ambition @mustafasuleyman

Pinterest tests AI-driven collage feature to help users create outfits from saved Pins using personalized AI-curated boards @TechCrunch
Rocket Mortgage reports clients using their AI Digital Assistant close at rates three times higher than those who don't, powering over 400,000 chats monthly @btaylor
OpenAI introduces ChatGPT text editing feature that can suggest quick edits and update text across documents, emails, and forms @OpenAI
Earth Species Project uses AI to decipher animal languages, potentially sparking a new understanding of interspecies communication @reidhoffman
Odyssey-2 introduces instant, interactive AI video generation at 20FPS that users can interact with in open-ended ways @olivercameron

Cameron Wolfe explains Proximal Policy Optimization (PPO) algorithm used for LLM training, detailing its clipped objective mechanism and actor-critic setup for stable reinforcement learning @cwolferesearch
Ethan Mollick notes that larger AI models are better at understanding intent, making traditional prompt formulas less important while context and goal communication become key @emollick
MIT physicists develop DIGIT imaging method for pinpointing exact locations of tiny light sources down to individual atoms using grid-based mapping @MIT
Glyph framework released by Zai.org scales context length by compressing text into images and processing with vision-language models, reducing computational costs @AdinaYakup
LongCat-Video foundational model from Meituan generates 720p, 30fps videos with unified text-to-video, image-to-video, and video-continuation framework @AdinaYakup