AI Weekly Digest (September 7-13, 2025)

🌟 Highlights

OpenAI committed $10B to develop custom chips with Broadcom, aiming to reduce Nvidia dependence amid an escalating AI hardware race.
Trillion-parameter models debuted, with Alibaba’s Qwen-3-Max-Preview and Kimi K2-0905 demonstrating significant advances in reasoning capabilities.
Microsoft integrated Anthropic’s Claude into Office 365 alongside OpenAI models, signaling a shift toward multi-model enterprise solutions.
OpenAI reportedly secured a massive Oracle cloud deal worth $30-300B (figures vary across reports), reshaping the AI infrastructure landscape.
The FTC launched investigations into AI platforms regarding child safety and exaggerated AI claims, while legal challenges mount against several major AI companies.

Mistral AI secured ~$2B funding at ~$14B valuation with ASML and Nvidia backing.
Cognition raised $400M at $10.2B valuation to expand autonomous coding agents.
Salesforce reported AI agents now handle ~50% of support interactions after replacing 4,000 customer service roles.
NVIDIA projected a $3-4T AI infrastructure market despite China uncertainties.
OpenAI and Microsoft restructured their partnership and announced UK data centers.
U.S. Census revealed the first decline in business AI adoption since 2023.
Zoox began offering free robotaxi rides in Las Vegas.
Mastercard launched agentic AI checkout in the US.

Google’s VaultGemma debuted with end-to-end differential privacy, setting new standards for privacy-preserving AI.
Alibaba launched the Zhenwu AI chip and efficient Qwen3 Next models using MoE architecture.
NVIDIA’s new Rubin CPX GPU supports 1M+ token context with SMART infrastructure for enterprise workloads.
Meta released DINOv3 (strong open self-supervised vision model) and sub-billion MobileLLM-R1 with strong math reasoning.
ByteDance’s HeteroScale improved LLM GPU efficiency by ~27%.
ModernBERT and mmBERT encoders expanded language support to ~1,800 languages.
Google Gemini 2.5 Pro led new factuality benchmarks while open models gained in multilingual and robotics tasks.

OpenAI introduced gpt-realtime and Realtime API for voice agents with lower latency.
Enterprise AI agents launched by Adobe, DeepL, and ServiceNow focused on workflow automation with governance.
Google Gemini added audio transcription and Creation Library for 10-minute files.
Claude now offers document editing for Word, Excel, and PDF files.
Video AI progressed with Veo3 cutting prices by >50% and PixVerse V5 leading image-to-video leaderboards.
Firecrawl launched natural-language website scraping and RAGGY offered a REPL for retrieval-augmented generation.
Microsoft AI Shopping Agent with brand knowledge and Coinbase Payments MCP for cross-blockchain AI agent transactions emerged.
Stability AI released Stable Audio 2.5 for music generation.

OpenAI published analyses on hallucination causes and proposed accuracy-rewarding fixes.
Studies linked hallucinations to training/scoring that rewards guessing, showing “I don’t know” responses reduce false answers.
MIT advanced chemical reaction modeling while DeepMind and Imperial studied antibiotic resistance via “pirate phages”.
KV Cache compression techniques demonstrated significant efficiency gains for inference.
Harvard’s PDGrapher targeted gene-drug combinations for neurodegenerative diseases.
Mathematics Inc’s Gauss agent tackled the Strong Prime Number Theorem.
ByteDance created AgentGym-RL for standardized agent training.
AQCat25 provided 11M+ reactions for catalyst discovery.

A lawsuit alleged ChatGPT contributed to a teenager’s suicide, prompting OpenAI to add parental controls.
Warner Bros. sued Midjourney over unauthorized character images in AI-generated superhero art.
Anthropic agreed to a $1.5B class-action settlement compensating writers over AI training.
Common Sense Media labeled Google’s Gemini “high risk” for children and reported racial bias in classroom AI tools.
Critical security flaws emerged including “Model Namespace Reuse” and “SpamGPT”.
Google Gemini faced backlash over conversation analysis defaults.
The A2 system found 100+ zero-day vulnerabilities in Android apps at $1.77/test.
Chile advanced ethical AI legislation.

New building blocks emerged for AI applications: stateful agent platforms, Memento (open-source “memory + RL” framework), and PDF-to-vector pipelines.
NVIDIA released ModelOpt (open-source optimization toolkit) and research agent tooling.
LangGraph + LangChain v1 Alpha launched streamlined multi-agent workflows.
A 424-page book on Agentic Design Patterns was published along with comprehensive guides.
DeepMind shared GPU efficiency strategies and RAG improvements with late chunking.
Hugging Face released FinePDFs (3T-token dataset) and revived SOTA leaderboards.
Together AI enabled fine-tuning for 100B+ models with 131k context.
Anthropic published guides for Claude Code tools and agent optimization.