AI Updates on 2025-07-26

Qwen released their updated thinking model with extensive reasoning capabilities, taking 166 seconds to think through complex tasks like drawing instructions @simonw
Google announced Gemini 2.5 Flash-Lite is now stable and generally available for developers and enterprise customers @GoogleAI
Google released Aeneas, a new model designed to help historians interpret, attribute and restore ancient texts @GoogleAI
InternLM released Intern-S1, a 235B MoE multimodal model with 6B vision encoder, trained on 5T multimodal tokens and 2.5T scientific-domain tokens with tool calling capabilities @Xianbao_QIAN

Meta appointed Shengjia Zhao as chief scientist of their AI superintelligence unit @TechCrunch
Perplexity sent out another batch of Comet invites, indicating continued expansion of their AI search platform @AravSrinivas
Windsurf AI reported working with 30% of the Fortune 100 companies including JPMC, Dell, Cisco, Phillips, ServiceNow, and MercadoLibre @sandeepDshah
China's Unitree released a 25kg humanoid robot for $5,900, marking the first time a humanoid robot costs less than a maxed-out MacBook Pro, though with limited 1-hour battery life and basic capabilities @deedydas
Analysis suggests many frontier AI researchers surprisingly don't use AI tools, even the models they train, representing a failure of incentive systems @_xjdr
Software engineers who don't find LLMs useful for coding typically fall into three categories: used them over 2 months ago before improvements like Claude Code, work in esoteric languages/frameworks, or work on large preexisting codebases @deedydas

Future of Life Institute released a safety report card grading leading AI model-makers, with Anthropic scoring highest at C+ while DeepSeek received the lowest grade of F @MIT_CSAIL
Geoffrey Hinton proposed establishing an international community of AI safety institutes to work on techniques for training AI to be benevolent @AndrewCurran_
China's Premier Li Qiang proposed establishing an organization for global AI cooperation and coordination, emphasizing open-source development and sharing advances with developing countries @AndrewCurran_

First controlled study of GenAI in industrial quality control showed engineers using a GPT-3.5 powered troubleshooting system had significant increases in work quality when commissioning new trains @emollick
Google Photos and YouTube now support turning photos into videos using AI, with new Veo effects for transforming selfies into fun videos @GoogleAI
Google launched AI Playground as a new hub for YouTube AI creation features and Opal experiment for building and sharing AI mini apps @GoogleAI
Google Search and Shopping now supports AI-powered virtual try-on for clothes in the US @GoogleAI
NVIDIA demonstrated full ocean emulators being coupled with atmosphere models for the first time, enabling new capabilities in El Niño prediction and seasonal forecasting @NVIDIAAI

Anthropic's interpretability team published multiple research updates including work on automating model auditing, alternative transcoder variants for MLP layers as conditional linear transforms, and announced a new team applying interpretability methods to pressing model behavior questions @ch402
Gemini achieved gold-medal standard performance in the International Mathematical Olympiad, demonstrating significant advancement in mathematical reasoning capabilities @GoogleAI
Huawei showcased their CloudMatrix 384 system incorporating 384 of their 910C chips in its first public appearance at WAIC @AndrewCurran_
Discussion of pretraining as elegant science done by mathematicians versus posttraining as hair raising cowboy research with rapid hyperparameter experimentation, highlighting the different methodological approaches in AI development @tszzl