AI Updates on 2025-06-02

Microsoft Bing launches Sora-powered video creator, offering free 5-second video generation in cinematic resolution with portrait mode available @AndrewCurran_
Google releases Veo 3 video generation model, demonstrating significant improvements in quality and audio integration @karpathy
PlayAI open sources PlayDiffusion audio speech editing model under Apache 2.0 license, enabling dynamic fine-grained editing without regenerating entire audio @huggingface
Google releases app allowing users to run LLMs from Hugging Face locally and privately, supporting multi-turn conversations and image chat @huggingface

Big Six tech companies (Apple, NVIDIA, Microsoft, Google, Amazon, Meta) increased CapEx 63% year-over-year to $212B in 2024, with NVIDIA's revenue exploding 28x over 10 years @deedydas
ChatGPT reaches 800M monthly users with 8x growth in 2.5 years, with India representing 14% of the user base as the largest market @deedydas
50% of S&P 500 companies now mention AI on earnings calls versus near-zero just a few years ago, with Microsoft's AI business hitting $13B annual run rate @deedydas
AI IT job postings increased 448% while non-AI IT jobs declined 9% over 7 years, confirming predictions about AI transforming employment @deedydas
Specialized AI startups achieving hypergrowth, with Cursor code editor growing from $1MM to $300MM ARR in 25 months @deedydas
Waymo captured 27% of San Francisco rideshare market in 20 months, demonstrating rapid real-world autonomous vehicle adoption @deedydas
Samsung nearing wide-ranging deal with Perplexity for investment and deep integration into devices, Bixby assistant and web browser @soleio
Salesforce acquires Moonhub, a startup building AI tools for hiring @TechCrunch
IBM acquires data analysis startup Seek AI and opens AI accelerator in NYC @TechCrunch
Elon Musk's xAI reportedly looks to raise $300M in tender offer @TechCrunch
Elon Musk's Neuralink closes a $650M Series E funding round @TechCrunch

Andrej Karpathy warns that video generation becoming directly optimizable through gradient descent could create more powerful engagement optimization than current platforms, raising concerns about what "optimal" content might look like @karpathy
83% of Chinese respondents see AI as net positive versus only 39% of Americans, showing dramatic perception differences between countries @deedydas
Julie Zhuo observes that AI is rewiring human brains to become more demanding and impatient @joulee
Soleio proposes Risk Tokens concept for AI safety, suggesting that aligning economic incentives with desired behaviors makes security an emergent property rather than imposed constraint @soleio

Ethan Mollick demonstrates using Veo 3 to create historical what-if scenarios, generating a 1940s newsreel about a fictional WWII aircraft carrier project @emollick
Andrew Curran showcases o3 model's performance in analyzing writing techniques, particularly character development and narrative structure @AndrewCurran_
Carbon Robotics' AI-powered laser weeders covered 230K+ acres, preventing 100K+ gallons of herbicide use in agricultural applications @deedydas
Perplexity Finance now supports both pre-market and post-market data in stock charts @AravSrinivas
Character.AI unveils video generation capabilities and social feeds for their chatbot platform @TechCrunch
Figma demonstrates building a pixel art web tool using MAKE in just 4 hours from ideation to publishing @figma
Gergely Orosz notes that "vibe coding" with AI tools has made prototyping accessible to more engineers, lowering the skills barrier significantly @GergelyOrosz

Andrej Karpathy provides detailed guidance on using different ChatGPT models: o3 for important/hard tasks (40% of use), 4o for simple queries (40%), and Deep Research for comprehensive topic analysis (10%) @karpathy
Stanford NLP Group research shows that dropout should be removed when training language models and masked language models for better performance @stanfordnlp
Nathan Lambert's team releases second reward model evaluation that is substantially harder and better correlated with downstream reinforcement learning outcomes @natolambert
Jeff Clune highlights recent advances in AI self-improvement research, including DeepMind's AlphaEvolve and Sakana's Darwin Gödel Machine, noting the field's rapid progress @jeffclune
MIT engineers develop new fuel cell technology for aviation with 2x the efficiency of jet engines, potentially enabling zero-emissions flight @MIT
Stanford AI Lab's cardiac amyloidosis detection system using ultrasound receives FDA clearance @StanfordAILab
AI inference costs fell 99.7% in 2 years while training costs for frontier models approach $1B+, with energy per token dropping 105,000x in 10 years @deedydas
Google processing 50x more tokens monthly (480T vs 9.7T) year-over-year, demonstrating massive scaling in AI model usage @deedydas
Meta's Llama achieved 1.2B downloads with 100k+ derivative models, showing significant open-source adoption @deedydas