AI Updates on 2025-06-13

Meta releases Sonata, a significant advancement in 3D self-supervised learning that addresses geometric shortcuts and provides robust 3D point representations for perception applications @AIatMeta
Stability AI optimizes Stable Diffusion 3.5 using TensorRT and FP8 quantization, delivering 2.3x faster generation with Large model and 1.7x faster with Medium model, plus 40% lower VRAM requirements @StabilityAI
Hugging Face releases Ming-Omni, an open-source GPT-4o rival with unified perception and generation capabilities, supporting text, image, audio, and video inputs with only 2.8B active parameters @Tu7uruu
Tencent releases Hunyuan 3D 2.1, the first fully open-source, production-ready PBR 3D generative model with cinema-grade visuals and PBR material synthesis @TencentHunyuan
NVIDIA releases Nemotron-Personas dataset with 600k personas grounded in real-world data, built with compound AI systems for synthetic data generation @NVIDIAAIDev

US Army launches Detachment 201, recruiting senior tech executives including Palantir CTO Shyam Sankar, Meta CTO Andrew Bosworth, OpenAI CPO Kevin Weil, and Thinking Machines' Bob McGrew to serve part-time as Lieutenant Colonels in Army Reserve @AndrewCurran_
Meta makes a $14.3B bet on Scale AI with major leadership changes, as Scale co-founder Alexandr Wang slots into Meta's team, signaling broader AI race dynamics @TechCrunch
Hugging Face announces going all-in on PyTorch, consolidating their user base and focusing efforts on PyTorch to simplify the transformers library and remove bloating @PyTorch
Amazon joins the nuclear energy trend by purchasing 1.92 GW for AWS, following other tech giants investing in nuclear power for AI infrastructure @TechCrunch
Perplexity Finance shows strong user engagement and query growth, with CEO Aravind Srinivas positioning it as an alternative to expensive finance products like Bloomberg Terminal with better user experience @AravSrinivas

A 10-year federal ban on state AI regulation is moving through Congress, which would roll back existing protections and halt future AI safeguards according to AI Now Institute @AINowInstitute
Ethan Mollick warns about an under-rated privacy risk of LLMs: their ability to find valuable information in large piles of recorded content that previously couldn't be sorted through, making everyone's recorded social media content searchable @emollick
Simon Willison publishes extensive analysis of design patterns for securing LLM agents against prompt injection attacks, providing six design patterns to protect tool-using AI systems @simonw
Research reveals misalignment between what workers want AI to automate versus what AI experts believe it can automate, highlighting the need for active human involvement in shaping AI's role in work @emollick

Google DeepMind partners with Darren Aronofsky's Primordial Soup to create ANCESTRA, the first film using Veo generative video model alongside traditional filmmaking, debuting at Tribeca Film Festival @GoogleDeepMind
Reddit user reports ChatGPT saved his wife's life by correcting a doctor's fatal misdiagnosis, with comments filled with similar life-saving AI stories @deedydas
OpenAI adds Canvas download functionality, allowing users to export documents as PDF, docx, or markdown, and code files in appropriate formats like .py, .js, .sql @OpenAI
MIT researchers develop photonic AI hardware accelerator for 6G wireless signal processing, performing machine-learning computations at the speed of light for real-time edge device data analysis @MIT
Google tests Audio Overviews for Search queries, expanding AI-generated content formats beyond text @TechCrunch

Follow-up study debunks Apple's "Illusion of Thinking" paper, showing that AI models succeed when format allows compressed answers, proving earlier collapse was a measurement artifact due to token limits rather than reasoning failures @deedydas
New research on machine unlearning shows that distilling a conventionally "unlearned" model creates a model resistant to relearning attacks, making real machine unlearning possible @Turn_Trout
Stanford releases BountyBench, the first framework to capture offensive and defensive cyber-capabilities of AI agents in evolving real-world systems @StanfordAILab
Meta releases Reading Recognition in the Wild dataset featuring video, eye gaze, and head pose sensor outputs, the first egocentric dataset with high-frequency eye-tracking data at 60 Hz for wearable device applications @AIatMeta
Research paper "ReasonMed" introduces a 370K multi-agent generated dataset for advancing medical reasoning, with a recipe that may generalize beyond medical tasks @communicating
NVIDIA explains reasoning models as a rising class of AI designed to go beyond traditional LLMs by thinking out loud and following structured, intentional logic, making them ideal for agentic AI systems @NVIDIAAI
PyTorch releases ParetoQ quantization technique delivering state-of-the-art results across bit-widths, showing 1.58-, 2-, and 3-bit quantization offer better size-accuracy trade-offs than 4-bit for large language models @PyTorch