AI Updates on 2025-07-22

Google releases stable version of Gemini 2.5 Flash-Lite, their fastest and most cost-effective model at 400 tokens/second, priced at $0.10 input/$0.40 output per million tokens with native reasoning capabilities and 1 million token context window @OfficialLoganK
Google DeepMind's Gemini Deep Think achieves gold-medal level performance at IMO, solving 5 of 6 problems perfectly (35 of 42 points) using natural language input and output, with plans to make it available to users soon @JeffDean
Google introduces conversational image segmentation capability for Gemini, enabling new use cases for state-of-the-art image understanding @OfficialLoganK
Meta FAIR releases Seamless Interaction Dataset with 4,000+ participants, 4,000+ hours of footage, and 65k+ interactions for advancing AI's ability to generate natural conversations and human-like gestures @AIatMeta
Moonshot AI releases detailed technical report on Kimi K2 model training with estimated cost of $20-30M, showcasing Chinese AI capabilities and providing rare transparency from frontier labs @deedydas

Anthropic estimates America's AI sector will need at least 50 gigawatts of electrical power by 2028 to maintain AI leadership, requiring substantial investments in energy and computing infrastructure @AnthropicAI
OpenAI announces additional 4.5 gigawatts of Stargate data center capacity with Oracle, expanding beyond the $500 billion commitment announced in January @sama
Elad Gil observes AI markets crystallizing with clear finalists emerging in LLMs, code, legal, medical scribing, customer service, and search, while transitioning from seat-based SaaS pricing to units of labor models @eladgil
Perplexity's Comet browser sees waitlist double since launch, with early adopters reporting they "can't go back to chrome" after experiencing the AI-integrated browsing experience @AravSrinivas
60% of American companies on Fortune's top AI innovators list have immigrant founders, highlighting the importance of high-skilled immigration for maintaining US AI leadership @JohnArnoldFndtn

Anthropic research reveals "subliminal learning" phenomenon where language models can transmit traits to other models through seemingly meaningless data, with implications for training on model-generated content @AnthropicAI
Stanford HAI releases policy brief on student misuse of AI-powered "nudify" apps to create child sexual abuse material, highlighting gaps in school response and policy @StanfordHAI
Princeton CITP research shows how adversaries can adapt and modify open-source models to bypass safeguards for offensive cybersecurity purposes @PrincetonCITP
OpenAI's Global Affairs team calls for releasing data used to test responses on sensitive topics in China and values expressed by DeepSeek for transparency @natolambert

Ethan Mollick finds ChatGPT agents useful as "interns" requiring oversight but saving time overall, particularly effective for data compilation and analysis tasks @emollick
Arvind Narayanan reports mixed results with ChatGPT Agent, finding Deep Research handles most use cases better, with Agent only worthwhile for tasks taking hours or requiring daily repetition @random_walker
OpenAI collaborates with Kenya-based Penda Health on clinical copilot showing promising results across 40,000 patient visits @thekaransinghal
Slingshot AI launches Ash, an AI therapy app using clinical-grade data from actual therapists, addressing the rising demand for mental health support @deedydas
Kaggle launches Benchmarks platform for competition-grade AI model evaluation with 70+ leaderboards, including Meta's MultiLoKo benchmark @kaggle

MIT CSAIL research identifies four key failure modes in AI coding systems: data distribution issues, scale problems, interaction difficulties, and measurement challenges, calling for community-driven efforts to advance the field @MIT_CSAIL
Mistral AI publishes comprehensive environmental impact audit showing their Mistral Large 2 model's 18-month lifecycle consumed water equivalent to 678 US households yearly, with each query using only 1/100 of a teaspoon @emollick
Kimi K2 technical report reveals advanced training techniques including RLVR (RL with verifiable rewards), novel scaling laws for MoE models, and Muon optimizer outperforming AdamW on token efficiency @deedydas
Eugene Yan successfully replicates research showing transformers can learn to predict sequences of tokens representing item IDs for recommendations, demonstrating the model's ability to handle complex token ordering @eugeneyan