AI Updates on 2025-12-16

AI Model Announcements

  • Meta releases SAM Audio, the first unified model that isolates any sound from complex audio mixtures using text, visual, or span prompts, outperforming previous models across benchmarks @AIatMeta
  • Google DeepMind releases updated Gemini 2.5 Flash Native Audio model for live voice agents with improved instruction following and more natural conversations @GoogleDeepMind
  • OpenAI introduces ChatGPT Images 1.5 with stronger instruction following, precise editing, detail preservation, and 4x faster generation speed @OpenAI
  • NVIDIA releases Nemotron-Cascade family of reasoning models trained with cascaded, domain-wise reinforcement learning, with the 14B model surpassing DeepSeek-R1-0528 (671B) on LiveCodeBench and achieving silver-medal performance at IOI 2025 @_weiping
  • Ai2 releases Molmo 2, bringing grounded multimodal capabilities to video and leading many open models on challenging industry video benchmarks @allen_ai
  • Xiaomi releases MiMo-V2-Flash trained via Multi-Teacher On-Policy Distillation (MOPD), achieving performance on par with all specialist teachers in their domains using 1/50th the compute @XiaomiMiMo

AI Industry Analysis

  • Swedish vibe coding startup Lovable's new funding round values it at $6.6 billion, more than triple its valuation from five months ago @AndrewCurran_
  • Databricks raises $4B at $134B valuation as its AI business heats up @TechCrunch
  • Adaptive Security announces $81M Series B with NVIDIA, Bain Capital VC, and others to protect organizations from AI-powered cyber attacks @AdaptiveSec
  • George Osborne joins OpenAI as managing director and head of OpenAI for Countries, based in London, to help societies worldwide share AI opportunities @George_Osborne
  • Frontier labs estimated to have more research compute than all academic institutions in the US combined, demonstrating brute force approach over efficient compute use @natolambert
  • Tech companies increasingly hiring for "storytelling" roles, with positions doubling on LinkedIn job posts since last year, reflecting shift toward owned narrative distribution @N_Sportelli
  • Reporters at some outlets face minimum quota of 3 "scoops" per week in AI industry, leading to dramatic framing of mundane stories @joannejang

AI Ethics & Society

  • Ethan Mollick demonstrates that distinguishing AI-generated images from real content remains extremely difficult, yet people continue believing images supporting their views without verification @emollick
  • Stanford researchers used AI to analyze Google Street View images across 16 states, revealing 37% of damaged buildings in poor areas became empty lots for years while 82% in wealthy areas were rebuilt bigger and better @StanfordHAI
  • Reading habits show dramatic shift with non-readers now outnumbering readers 3 to 1, reversed from previous 2 to 1 ratio favoring readers @paulg
  • One third of 8th grade girls spend 7+ hours per day on social media, representing nearly all their daily activity @JonHaidt

AI Applications

  • OpenAI's GPT-5 worked with Red Queen Bio to optimize molecular cloning protocols in the lab, achieving 79x efficiency gain through iterative experimentation including a new enzyme-based approach @OpenAI
  • Simon Willison ported a Python library implementing full HTML5 parser to JavaScript using GPT-5.2 and Codex CLI in 4.5 hours while watching a movie @simonw
  • Google Labs introduces CC, an experimental AI productivity agent in Gmail providing "Your Day Ahead" briefings and email assistance for Google AI Ultra subscribers @GoogleLabs
  • Microsoft Copilot launches Eggnog Mode for Mico, adding holiday-themed personality available in US, UK, and Canada @mustafasuleyman
  • Meta's AI glasses now help users hear conversations better with enhanced audio capabilities @TechCrunch
  • DoorDash rolls out Zesty, an AI social app for discovering new restaurants @TechCrunch
  • v0 now connects to Linear workspace, allowing users to build directly from their backlog @v0

AI Research

  • OpenAI releases FrontierScience benchmark measuring PhD-level scientific reasoning across physics, chemistry, and biology with expert-written olympiad-style and research-style tasks, showing GPT-5.2 as strongest performer while revealing gaps in open-ended reasoning @OpenAI
  • GPT-5.2 solves COLT 2022 open problem on "Running Time Complexity of Accelerated L1-Regularized PageRank" using standard accelerated gradient algorithm, with all proofs auto-generated and formalized in Lean @kfountou
  • Google Research uses advanced Gemini 2.5 Deep Think to verify theoretical computer science papers, with 97% of STOC2026 authors finding feedback helpful for catching errors and improving clarity @GoogleResearch
  • Claude Opus 4.5 solves CORE-Bench by creatively resolving dependency conflicts and bypassing environmental barriers, while Opus 4.1 and Sonnet 4 fail by resorting to simulated data @PKirgis
  • Ai2 releases Olmo 3 Think with fully-open pipeline for reinforcement learning, using supervised finetuning, DPO, and RLVR with GRPO, continuing to improve after 3 weeks of training without instability @cwolferesearch
  • Meta introduces VL-JEPA, first non-generative model for real-time vision-language tasks including streaming action recognition, retrieval, VQA, and classification, outperforming VLMs with better efficiency @pascalefung
  • Research on depth-grown Transformers shows gradually stacking layers throughout training can overcome the "Curse of Depth" problem where deeper layers are underutilized @KaplFer
  • Stanford AI Lab identifies flawed questions in widely used AI benchmarks, highlighting reliability concerns in benchmark design @StanfordAILab
  • Researchers introduce MUPI (Embedded Universal Predictive Intelligence) framework providing theoretical basis for cooperative solutions in reinforcement learning by grasping self-other similarity @tyrell_turing
  • Latent Labs releases Latent-X2 for AI-generated antibodies with drug-like developability and low immunogenicity in human panels, zero-shot @saakohl
  • Terence Tao discusses concept of Artificial General Cleverness as distinct from AGI @AndrewCurran_
  • Google DeepMind CEO Demis Hassabis discusses working on "root node problems" - fundamental scientific challenges from fusion and superconductors to new materials discovery @GoogleDeepMind
  • Researchers demonstrate that exploration failure, not modeling ability, is typically why humans fail to solve ARC 3 environments, highlighting exploration as both difficult and important @fchollet
  • Stanford HAI releases issue brief analyzing Chinese AI models' diverse open-weight ecosystem and policy implications of their global diffusion @StanfordHAI