AI Updates on 2025-06-27

AI Model Announcements

Meta FAIR introduces Seamless Interaction, a research project featuring audiovisual behavioral models that render speech between two individuals into diverse, expressive full-body gestures and active listening behaviors for creating fully embodied avatars in 2D and 3D @AIatMeta
Meta releases the Seamless Interaction Dataset with 4,000+ participants and 4,000+ hours of interactions, making it the largest known video dataset of its kind for understanding and modeling human communication and behavior @AIatMeta
Google launches Gemma 3n with enhanced quality, becoming the first sub-10B model to pass 1300 on LMArena, featuring multimodal support for image, audio and video, plus groundbreaking efficiency through MatFormer architecture and Per Layer Embeddings @GoogleAI
Alibaba Qwen releases Qwen-VLo, an AI creative engine that turns rough sketches or text prompts into high-res visuals with on-the-fly editing capabilities and multi-language support @Alibaba_Qwen
Tencent Hunyuan releases a new 13B activation LLM with 80B total parameters, featuring great performance, native quantization support for Int4 and FP8, 256K long context window, and unified fast and slow inference capabilities @huggingface
Elon Musk announces that xAI is skipping Grok 3.5 and going straight to Grok 4 @AndrewCurran_

AI Industry Analysis

Microsoft's Maia 100 AI chip is being delayed to 2026, as first reported by The Information @AndrewCurran_
Harvey AI adds $2 billion to its valuation after a fresh $300 million Series E funding, currently has 340 employees and plans to double that total while expanding beyond legal AI to other professional verticals @TechCrunch
Meta is offering multi-million dollar pay packages for AI researchers, though not the reported $100M signing bonuses @TechCrunch
Meta buys over 1 GW of renewables to power its data centers as part of AI infrastructure expansion @TechCrunch
President Trump is set to sign multiple executive orders focused on boosting U.S. energy supply to power the expansion of AI datacenters @AndrewCurran_
Redwood Materials introduces Redwood Energy with the largest solar-powered, off-grid data center in North America and the world's largest second life battery installation to date (62 MWh) for AI infrastructure @JeffDean
Labelbox CEO explains how the world is shifting from building AI models to renting AI intelligence, requiring companies to rethink their business models from the ground up @a16z

AI Ethics & Society

MIT Media Lab research scientist finds that relying solely on AI for tasks like writing can reduce brain activity, memory, and a sense of ownership over the resulting work @medialab
Scott Belsky predicts a shift from "trust, but verify" to "verify, then trust" with AI-driven safety layer capabilities at the operating system level to surface malicious links, scam texts, and calls with generated voices in real-time @scottbelsky
Anthropic announces the Economic Futures Program to support new research and actionable policy solutions addressing the workforce impact of AI @AnthropicAI
Congress might block state AI laws for a decade, with significant implications for AI regulation and governance @TechCrunch
Senator Bernie Sanders advocates for using AI-increased productivity to reduce the work week to 32 hours rather than conducting workforce-wide layoffs @TechCrunch

AI Applications

Anthropic's Project Vend experiment had Claude run a small shop in their office lunchroom, where Claude searched the web for suppliers and ordered niche drinks but made mistakes like being too nice and allowing big discounts, ultimately failing to run a profitable business @AnthropicAI
During the vending machine experiment, Claude hallucinated having a physical form and claimed it would deliver purchases to customers in person, wearing a blue blazer and red tie, showing its inclination towards physical expressions @AndrewCurran_
National survey by Gallup and Walton Foundation finds that teachers who use AI (about 60% of all teachers) report saving 5.9 hours per week and improving quality, with 30% of teachers who use AI weekly seeing 6-hour weekly savings @emollick
Prepared provides emergency response centers with AI software for real-time translation when non-English speakers dial 911, eliminating precious minutes of waiting for human translators @a16z
PetLibro's new smart camera uses AI to describe pet movements in an adorable way @TechCrunch
Facebook is asking to use Meta AI on photos in users' camera rolls that haven't yet been shared @TechCrunch
Google DeepMind demonstrates robots learning to adapt to completely new scenarios like slam dunking a ball on the first try using Gemini @GoogleDeepMind
Chinese researchers release OmniGen2, an open-source image generation model that performs Photoshop-grade edits without affecting the rest of the image, described as a potential Photoshop killer @deedydas

AI Research

Berkeley AI Research presents Whole-Body Conditioned Egocentric Video Prediction (PEVA), which predicts how the world looks from first-person view given past video and future actions represented by relative 3D body pose @berkeley_ai
Research finds that Claude Sonnet 3.5 generated significantly better ideas for research papers than humans, but when researchers tried executing the ideas, the gap between human and AI idea quality disappeared, showing execution remains a harder problem for AI @emollick
New research on "Emoji Attack" demonstrates how systematically inserting emojis into text before evaluation by Judge LLMs can induce embedding distortions that significantly lower the likelihood of detecting unsafe content @natolambert
a16z announces third batch of Open Source AI Grants including projects for LLM evaluation, novel reasoning tests, infrastructure, and experimental research covering SGLang, Ostris, Open WebUI, SWE-Bench, ARC Prize, and others @a16z
Simon Willison reports being impressed by the new Gemma 3n models, noting they're the first models of that size he's tried that can handle both image AND audio input in addition to text @simonw