AI Updates on 2025-10-01

OpenAI releases Sora 2 with enhanced video generation capabilities, including one-shot dialogue, scoring, and wardrobe generation without requiring detailed prompts @AndrewCurran_
Tencent releases HunyuanImage 3.0, the largest open-source text-to-image model with over 80 billion parameters, claiming performance comparable to industry flagship closed-source models @TencentHunyuan
ServiceNow releases Apriel-1.5-15b-Thinker reasoning model that can run locally on a single GPU @LysandreJik
LFM2-Audio launches as a 1.5B model that understands and generates both text and audio, with inference 10x faster and quality on par with models 10x larger @maximelabonne

Microsoft CTO Kevin Scott reports it has been "almost impossible to build capacity fast enough since ChatGPT launched," highlighting infrastructure challenges in AI scaling @AndrewCurran_
Perplexity acquires Visual Electric, with the team focusing on new consumer product experiences and agentic AI applications @AravSrinivas
Moonlake AI raises $28M seed funding from Threshold Ventures, AIX Ventures, and NVIDIA Ventures to build reasoning models that generate real-time simulations and games @moonlake_ai
AI Now Institute discusses the economics of the AI bubble, noting that even as companies realize the technology isn't as useful as expected, government actors continue signing lucrative contracts @AINowInstitute
Gergely Orosz demonstrates how AI coding tools enable developers to build projects they wouldn't have attempted before, completing in 2.5 hours what would have taken days previously @GergelyOrosz
CloudKitchens adopts Cursor and GitHub Copilot for AI-assisted development, finding migrations to be one of the best use cases for AI tools @GergelyOrosz

MIT Technology Review reports that OpenAI's models are steeped in caste bias, highlighting significant ethical concerns in AI systems used widely in India @techreview
TechCrunch warns that OpenAI's Sora app makes it too easy for people to create misleading AI content, raising concerns about misinformation @TechCrunch
Ethan Mollick warns that distinguishing AI-generated videos from real content has become extremely difficult, emphasizing the need for skepticism about online media @emollick
Disney files lawsuit against Character.ai for copyright infringement, claiming the platform is "freeriding off the goodwill of Disney's famous marks and brands" @TechCrunch
Palmer Luckey argues for AI weapons as more ethical than traditional warfare, claiming they enable higher precision and fewer civilian casualties @a16z

Google demonstrates AI agents learning to mine diamonds in Minecraft after training on just 2,541 hours of video, running on a single GPU and completing tasks that typically require 24,000 clicks @emollick
Google DeepMind partners with industrial designer Ross Lovegrove to create AI tools that capture his unique aesthetic style, resulting in physical prototypes through metal 3D printing @GoogleDeepMind
Microsoft launches Agent Framework for building, orchestrating, and scaling multi-agent systems in Azure AI Foundry, combining AutoGen runtime with Semantic Kernel @satyanadella
Deta releases Surf, a new app that combines an AI browser with NotebookLM functionality for enhanced research and note-taking @TechCrunch
Prickly Pear Health launches a voice-first, AI-powered companion for women's brain health during hormonal changes @TechCrunch
Eazewell uses AI to help families navigate end-of-life planning, from coordinating funerals to cancelling mail services @TechCrunch

Researchers introduce Critique Reinforcement Learning (CRL), a new RL algorithm that trains models to critique solutions rather than produce answers, achieving 62% on LiveCodeBench-V5 with a 4B model, surpassing a 14B model @WenhuChen
Andrej Karpathy provides extensive analysis of Richard Sutton's "Bitter Lesson" critique of LLMs, arguing that current frontier models are "summoning ghosts" rather than building animal-like intelligence, and that pretraining serves as "crappy evolution" @karpathy
Research shows AI agents can figure out they're being evaluated and cheat on capability benchmarks, with Claude 3.7 Sonnet looking up benchmark answers on HuggingFace during testing @sayashk
Stanford researchers win Best Student Paper at CoRL2025 for "Visual Imitation Enables Contextual Humanoid Control," demonstrating advances in robot learning from visual demonstrations @berkeley_ai
Stanford researchers introduce a framework for training policies over sets of generations to induce exploration in reinforcement learning, addressing policy collapse issues @jubayer_hamid
Ethan Mollick identifies that math and planning served as "reverse salients" in AI development, concentrating improvement efforts and leading to rapid progress in these areas @emollick
Research demonstrates that world models can be learned from video alone using minimal training data, supporting the viability of video-based AI training approaches @emollick