AI Updates on 2025-09-08

Alibaba releases Qwen3-ASR, an all-in-one speech recognition model supporting 11 languages with auto language detection, custom context support, and under 8% word error rate even with background music @Alibaba_Qwen

OpenAI is backing an AI-generated feature-length animated movie called Critterz with a $30 million budget and 9-month production timeline, set to debut at Cannes in May 2026 @AndrewCurran_
Databricks confirms another $1 billion funding round at a $100 billion valuation, just months after raising $10 billion @TechCrunch
Cognition Labs raises funding led by Founders Fund with participation from Lux Capital, 8VC, and others for their AI coding agent Devin @TechCrunch
Chinese robot maker Unitree files for a $7 billion IPO with over $140 million in revenue, holding 70% global market share in robot dogs and becoming the biggest public humanoid robot company @deedydas
AI startup founders face extreme time pressure with approximately 6 months or less to find product-market fit before potentially having to fold or sell due to the revolutionary nature of AI technology @GergelyOrosz

Anthropic endorses California's SB 53 bill, advocating for transparency-based governance of powerful AI systems rather than technical micromanagement, while emphasizing the need for thoughtful AI governance today rather than reactive measures tomorrow @AnthropicAI
François Chollet warns that as AI-generated content floods the internet and humans increasingly rely on generative AI, future models will inevitably be trained mostly on AI-generated content, leading to culture becoming "slop remixed from slop" @fchollet
Sam Altman observes that AI Twitter and Reddit now feel "very fake" compared to a year or two ago, attributing this to real people adopting LLM-speak, extreme hype cycles, engagement optimization, and potential astroturfing @sama

Perplexity launches Perplexity for Government, offering zero data usage, fully secure access to premium AI models for U.S. Government use without contracts or licenses @perplexity_ai
Google's AI Mode in Search expands to five new languages: Hindi, Indonesian, Japanese, Korean, and Brazilian Portuguese, using a custom version of Gemini 2.5 for culturally relevant search experiences @sundarpichai
Google DeepMind introduces RoboBallet, an AI system that can choreograph up to 8 robot arms working together without collisions, outperforming traditional methods by approximately 25% in task and motion planning @GoogleDeepMind
Gemini App now supports audio file uploads, addressing the number one user request for file type support @joshwoodward
Cognition Labs CEO demonstrates how Devin AI is used internally for project planning, bug fixes, deepwiki research, and serving as first line of defense for engineering questions @clairevo

Research reveals a clear performance gap between online and offline reinforcement learning algorithms for LLM training, with online methods like PPO handling out-of-distribution data more robustly than offline methods like DPO, though the gap can be minimized through semi-online approaches @cwolferesearch
Ethan Mollick tests GPT-5 Pro on creating compelling D&D puzzles, finding significant improvements in puzzle coherence compared to GPT-4 and Claude 3 Opus, though single-prompt approaches still struggle with extraneous details and weird justifications @emollick
Paul Graham discovers that GPT-5 is reliably bad at monograms, unable to solve any correctly even after being told it's wrong and asked to think longer for better answers @paulg
Hugging Face releases FinePDF, the largest publicly available PDF dataset with 3 trillion tokens across 475 million documents in 1,733 languages, achieving performance nearly on par with state-of-the-art HTML collections @rohanpaul_ai
François Chollet proposes that AGI will be "an algorithmic encoding of the process of Science itself" rather than an individual mind, describing science as a program synthesis process that produces symbolic models @fchollet