AI Updates on 2025-06-09

Google introduces Veo 3 Fast in Gemini App and Flow, offering >2x faster video generation with 720p resolution and serving optimizations @joshwoodward
Apple announces new generation of LLMs for Apple Intelligence features and introduces Foundation Models framework giving developers direct access to on-device foundation language models @ruomingpang

OpenAI reaches $10 billion ARR including consumer, enterprise and API revenue, nearly double from last year, with internal targets of $125 billion by 2029 @AndrewCurran_
Stripe reports payment volume from customers who signed up in 2025 is tracking 116% ahead of the same week last year, potentially influenced by AI adoption @patrickc
Stripe engineer reports their payments foundation model for fraud detection cut missed credit card fraud by up to 5x in some cases, demonstrating rare "instant wins" from AI deployment @random_walker
AI Engineer roles offer massive ROI for software engineers wanting to work at startups, with the transition being surprisingly easy compared to traditional ML engineering @GergelyOrosz

AI Now Institute argues we should focus less on debating how "good" technologies like ChatGPT are and more on whether the AI industry's unaccountable power is good for society @AINowInstitute
Eric Jang expresses concern about protesters vandalizing Waymo vehicles during LA riots, arguing it undermines public support for causes when autonomous vehicles make people feel safer @ericjang11
Hamel Husain suggests the cultural norm should shift from being ashamed of using AI to being ashamed of NOT using AI, advocating for celebrating AI-assisted achievements @HamelHusain

UK government deploys Extract system using Gemini to help council planners make faster decisions, turning complex planning documents into digital data in 40 seconds @GoogleDeepMind
Barclays scales Microsoft 365 Copilot to 100,000 employees, making Copilot the UI for Barclays AI across the organization @satyanadella
Microsoft partners with AgeUK to use speech-to-text technology for monitoring and scaling their Telephone Friendship Service supporting 4,500+ older adults @Microsoft
Deedy demonstrates comprehensive AI video creation stack using 10 tools including ChatGPT, Midjourney v7, Veo 3, and others to create Hollywood-grade 2-minute trailer @deedydas
Cameron Wolfe explains AI agent capabilities from basic LLM tool use to autonomous systems that can run asynchronously and take concrete actions on users' behalf @cwolferesearch

New research shows modern AI models perform better when allowed to "think" rather than being instructed to "answer directly," with structured answers not being problematic @emollick
Medical study demonstrates doctors using custom GPT-4 produce significantly more accurate diagnoses than doctors with Google/PubMed, though AI alone matches doctors + AI performance @emollick
Research suggests Tower of Hanoi reasoning limitations in LLMs may be due to training constraints on thinking time rather than fundamental reasoning inability @emollick
Stanford researchers present General User Model (GUM) that learns user habits and preferences from everyday computer use to anticipate needs across any context @oshaikh13
MIT CSAIL data reveals ByteDance Seed's Seedance 1.0 leads video generation models including Google's Veo 2 in both text-to-video and image-to-video generation @MIT_CSAIL
New research on generative reward modeling shows inference-time scaling approaches are becoming the dominant direction for reward modeling systems @natolambert
NVIDIA releases Nemotron-Personas dataset providing high-quality synthetic training data reflecting real-world demographics while complying with privacy standards @NVIDIAAIDev