AI Updates on 2025-08-22

Perplexity Max Subscribers can now use GPT-5-Thinking model for reasoning mode queries @AravSrinivas
OpenAI announces medical research breakthrough using GPT-5 with Professor @DeryaTR_ demonstrating its impact @OpenAI
Scale AI announces partnership with Midjourney to license their aesthetic technology for future models and products, bringing beauty to billions @alexandr_wang

Meta signs a six-year $10 billion cloud deal with Google, indicating massive infrastructure investment for AI capabilities @AndrewCurran_
Apple is testing a custom version of Gemini to potentially power the new Siri, with OpenAI and Anthropic still in the race for the contract @AndrewCurran_
OpenAI starts hiring in India and plans to open their first office there later this year, with ChatGPT users growing 4x in the past year in India @sama
Beijing reportedly urging Chinese companies to turn to domestic chips just weeks after Nvidia got approval to sell in China again @TechCrunch
Perplexity CEO claims finance is a vertical where Perplexity is so far ahead of competitors in terms of accuracy, speed, quality, depth and breadth @AravSrinivas

Anthropic releases new research on filtering dangerous CBRN information at pretraining, experimenting with removing chemical, biological, radiological and nuclear weapons data from training without affecting performance on harmless tasks @AnthropicAI
Google DeepMind shares comprehensive methodology for measuring environmental impact of AI, reporting 33x drop in energy use per prompt and 44x reduction in carbon footprint over 12 months @GoogleDeepMind
Ethan Mollick provides data showing AI prompts use minimal resources: Gemini uses 0.00024 kWh and 0.26 mL water per prompt, equivalent to 9 seconds of TV watching and 5 drops of water @emollick

OpenAI's protein engineering project with Retro Biosciences using GPT-4b-micro designed novel variants of Yamanaka factors achieving 50x increase in reprogramming efficiency for extending human lifespan by 10 years @gdb
Google DeepMind's Genie 3 can create interactive 3D worlds from text, photos, or videos, with advanced spatial memory and realistic physics modeling including lighting, gravity, and liquids @demishassabis
Google DeepMind demonstrates AI training inside another AI: SIMA agent learns to navigate environments generated by Genie 3, creating an entire AI-to-AI training loop @alexgraveley
Microsoft Copilot Labs launches new features including 3D modeling, visual chat with real-time animations, web task automation, and AI-powered gameplay @Copilot
Microsoft Copilot Deep Research reports now available worldwide across web, Edge, iOS, and Android, offering 5 free comprehensive research reports monthly @Copilot
Sierra helps CDW build smarter support experience tailored to complex IT procurement teams, demonstrating AI agents' effectiveness for complex B2B use cases @btaylor
Gemini Live will soon support camera sharing with ability to highlight what to focus on, becoming more helpful for visual assistance @GeminiApp

Shanghai AI Lab unveils Intern-S1, a scientific multimodal foundation model that reportedly beats o3 and Gemini-Pro in transforming molecular discovery and natural world reasoning @cgeorgiaw
Qwen-Image-Edit debuts at #2 in Image Editing Arena with ELO 1098, achieving performance on par with GPT-4o while being open weights under Apache 2.0 license @Alibaba_Qwen
Berkeley AI Research introduces CAST method to augment robot datasets with counterfactuals to improve language following in vision-language-action policies @CatGlossop
PyTorch reports 1.7x-2.3x inference efficiency improvement for LLaMA-based encoders using Nested Jagged Tensors, making high-performing LLM encoders more practical for production @PyTorch
Google DeepMind releases Major TOM AlphaEarth Embeddings, a 6 TB prototype dataset for Earth observation and environmental modeling on Hugging Face @mikonvergence