AI Updates on 2025-09-12

Baidu releases ERNIE-4.5-21B-A3B-Thinking model, now the top trending text-generation model on Hugging Face with 21B total parameters, 3B active parameters per token, and enhanced 128K long-context understanding capabilities @Baidu_Inc
Cursor releases new Tab model trained with online reinforcement learning, making 21% fewer suggestions while achieving 28% higher accept rate for suggestions @cursor_ai
Google Research releases VaultGemma, an open model trained from scratch with differential privacy, presenting scaling laws for differentially private language models @GoogleResearch
Qwen releases Qwen3-Next-80B-A3B model with day-0 support from SGLang for speculative decoding and vLLM for efficient inference with accelerated kernels @Alibaba_Qwen

OpenAI and Microsoft sign non-binding MOU for OpenAI's transition to public benefit corporation, with the nonprofit's equity stake exceeding $100 billion @AndrewCurran_
25% of Linear workspaces now use AI agents, with 50%+ adoption in enterprise, mainly using Cursor, Devin & Codegen coding agents directly tasked from Linear to fix bugs and improvements @karrisaarinen
Hugging Face partners with multiple providers to bring hundreds of state-of-the-art open models directly into VS Code and GitHub Copilot, offering open weights models with competitive pricing and seamless switching @ClementDelangue
Parahelp raises Series A funding, with top AI companies including Perplexity, Replit, Bolt, and HeyGen using their AI customer support agent platform @snowmaker
Cresta creates breakthrough advertisement built 100% with AI in 5 weeks, from scripting to video generation and voices, demonstrating AI's potential for content creation @cresta

California Senate passes SB 243 requiring AI companion operators to implement safety protocols and holding companies legally accountable, potentially making California the first state with such regulations @TechCrunch
Google's AI crawler cannot be blocked separately from its web crawler, allowing the search giant to use content for AI training without publishers' consent @TechCrunch
Anthropic collaborates with US Center for AI Standards and Innovation and UK AI Security Institute to test models like Claude Opus 4 and 4.1 for vulnerabilities before deployment @AnthropicAI

Ethan Mollick discusses how AI systems are shifting from collaborative tools where users shape the process to systems where users become supplicants receiving opaque outputs @emollick
Replit builds their own computer use model for browser testing after finding Claude and GPT-5's Computer Use models too slow and expensive, achieving up to 15x faster performance @amasad
Qwen Code releases v0.0.10 & v0.0.11 with new features including subagents for task decomposition, Todo Write tool for task tracking, and "Welcome Back" project summaries @Alibaba_Qwen
Paul Graham reports a founder can write 10,000 lines of code in a day with AI assistance, noting this equals 500 lines per hour which is achievable in verbose languages @paulg

Research reveals LLM Hacking where using LLMs as data annotators can produce any desired scientific result, raising concerns about research validity @joabaum
OpenAI's reasoning models have evolved from thinking for seconds with o1-preview a year ago to current models that can think for hours, browse the web, and write code @polynoamial
Analysis of GPT-5 on AssistantBench shows higher precision and lower guess rates than o3, challenging OpenAI's claims about hallucinations and model calibration @PKirgis
Physical Intelligence robotics models work with only 1-second context length, relying on current world state rather than memory to execute complex multi-minute plans @dwarkesh_sp
Sergey Levine predicts fully autonomous household robots within 5 years, citing LLMs' common sense and prior knowledge as game-changing scaffolding for robot models @dwarkesh_sp
Meta's vLLM disaggregated implementation improves inference efficiency in latency and throughput compared to their internal stack, with optimizations being upstreamed to the vLLM community @PyTorch