AI Updates on 2025-08-21

DeepSeek-V3.1 introduces hybrid inference with Think and Non-Think modes, offering faster thinking capabilities and stronger agent skills with 128K context support @deepseek_ai
Cohere releases Command A Reasoning, their most advanced model for enterprise reasoning tasks, designed for private deployment on less than 2 GPUs with user-controlled token budgets @cohere
ByteDance Seed OSS model with 36B parameters now available on Hugging Face, featuring Apache2 license, native 512k long context, and flexible thinking budget @Xianbao_QIAN
Google announces Veo 3 will be available for free trial in Gemini App, with TPUs being warmed up for the launch @joshwoodward

Anthropic doubles its fundraising target to $10 billion due to high investor demand, significantly increasing from the originally planned amount @AndrewCurran_
Meta reportedly implements a hiring freeze at Meta Superintelligence Labs while working through reorganization that split the AI unit into four new groups @TechCrunch
Research shows 95% of AI pilots fail to achieve sustained P&L impact within six months, though methodology questions remain about the generalizability of findings from 52 convenience-sampled interviews @emollick
Despite 50% LLM adoption among US workers, labor productivity growth remains lower than 2020 levels, challenging claims of 10x productivity gains from AI tools @fchollet
AI demonstrates 92% accuracy vs 72% for experienced lawyers on invoice review tasks, while being 50-100x faster and 99.97% cheaper, highlighting AI's impact on traditional professional services @deedydas
Google reports 33x reduction in energy footprint and 44x reduction in carbon footprint for Gemini Apps text prompts from May 2024 to May 2025, while delivering higher quality responses @JeffDean

Anthropic partners with NNSA to develop nuclear weapons safeguards for AI, creating classifiers that detect concerning nuclear queries while preserving legitimate educational and research uses @AnthropicAI
Mustafa Suleyman warns against seemingly conscious AI, arguing that AI's value comes from being different from humans rather than mimicking human emotions like shame, jealousy, or fear @mustafasuleyman
Anthropic launches three new AI fluency courses co-created with educators to help teachers and students build practical, responsible AI skills, available free to any institution @AnthropicAI

Google launches Gemini for Government platform providing AI tools including NotebookLM and Veo to federal agencies at virtually no cost through partnership with GSA @sundarpichai
Google introduces agentic capabilities in AI Mode for Search, enabling autonomous browsing of multiple sites to find restaurant reservations with real-time availability and direct booking links @GoogleAI
Cursor integrates with Linear to enable AI agents that can be launched directly from issues, creating branches and drafting PRs based on plain language task delegation @cursor_ai
Perplexity launches stock screening for Indian stocks using natural language search, available across web and mobile platforms for both free and paid users @AravSrinivas
Perplexity Comet demonstrates ability to autonomously set up Shopify stores, showcasing advanced e-commerce automation capabilities @AravSrinivas
Runway launches Game Worlds Beta, enabling creation of AI-generated interactive game environments @AndrewCurran_

DeepSeek-V3.1 achieves 66% on SWE-Bench while being 2x cheaper for input tokens and 6x cheaper for output tokens compared to GPT-5, which scores 70-71% on the same benchmark @deedydas
Andrew Ng's Buildathon demonstrates rapid AI-assisted development, with teams building 5 functional products in 6.5 hours using tools like Claude Code, GPT-5, Cursor, and Windsurf @AndrewYNg
Kaggle releases results from first Chess Text Input benchmark where AI models played chess using only text inputs without tools or move validation, establishing Elo-like rankings across 40+ matches per pairing @kaggle
ARC-AGI-3 Preview releases 3 additional games from previously private holdout set, expanding the novelty of public games available for testing AI reasoning capabilities @arcprize
Google DeepMind's Genie 3 creates explorable AI-generated worlds for testing and training AI agents safely, with capabilities for diverse and challenging virtual environments @GoogleDeepMind