AI Updates on 2025-12-19

AI Model Announcements

OpenAI releases GPT-5.2-Codex, setting a new standard for agentic coding in real-world software development and defensive cybersecurity, with more reliable performance on complex tasks and effective scaling across large projects @OpenAI
Google announces Gemini 3 Flash, a major upgrade delivering next-generation intelligence at lightning speed and representing a significant capability improvement over 2.5 Flash, now available globally @GeminiApp
Alibaba releases Qwen-Image-Layered, featuring Photoshop-grade layering with physically isolated RGBA layers, prompt-controlled structure for 3-10 layers, and infinite decomposition capabilities, fully open-sourced @Alibaba_Qwen
Meta releases Meta Seal, a comprehensive, state-of-the-art, MIT-licensed suite of AI watermarking research, models, and training code @AIatMeta
Google releases Gemma Scope 2, the largest open release of interpretability tools with over 1 trillion parameters trained, working as a microscope to analyze all Gemma 3 models' internal activations @GoogleDeepMind
Meta is developing a new image and video-focused AI model codenamed Mango, expected to be released in the first half of 2026 @AndrewCurran_
Meta's Llama successor is codenamed Avocado, originally planned for Christmas release but pushed back to early 2026, with uncertainty about whether it will remain open source @AndrewCurran_

AI Industry Analysis

OpenAI is reportedly attempting to raise $100 billion at an $830 billion valuation @TechCrunch
Yann LeCun confirms his new world model startup, reportedly seeking a $5 billion+ valuation @TechCrunch
Cursor acquires Graphite, one of the best AI code review and PR workflow platforms, signaling potential competition with GitHub @cursor_ai
OpenAI has sold 700,000+ ChatGPT licenses to approximately 35 US public universities for students and faculty, who used it 14 million+ times in September, surpassing Copilot usage @gdb
Meta rolled out a feature called trajectories to developers, allowing code reviewers to see the prompts used to generate AI-generated diffs, as an experiment in handling increased AI-generated code @GergelyOrosz
GitHub's prospects as a product are questioned unless it regains independence and a CEO, with parallels drawn to Microsoft's handling of Skype after not backfilling its CEO position @GergelyOrosz
Andrew Ng argues that advancing frontier models today requires manual decisions and a data-centric AI approach to engineering training data, with progress being more piecemeal than widely appreciated despite models' general intelligence capabilities @AndrewYNg
Brex data shows 30% of 2025's fastest-growing software vendors are YC startups, with plans to reach 50% in coming years @paulg

AI Ethics & Society

OpenAI publishes research on evaluating chain-of-thought monitorability, finding that monitoring a model's chain-of-thought is far more effective than watching only its actions or final answers, though there's a tradeoff where smaller models with higher reasoning effort can be easier to monitor at similar capability @OpenAI
Anthropic shares efforts to ensure Claude handles emotional support conversations both empathetically and honestly, addressing the wide variety of reasons people use AI @AnthropicAI
OpenAI adds new teen safety rules to ChatGPT as lawmakers weigh AI standards for minors @TechCrunch
Research suggests AI may be transforming the legal profession fundamentally, with predictions that economic incentives will be too powerful to resist despite potential attempts to outlaw AI use, creating challenges for unemployed high-income legal professionals @AndrewCurran_
A lawyer at a large law firm confirms that GPT-5.x Pro is spectacular for legal research and analysis but not yet capable of reliably producing the best possible legal documents that could be filed with courts, though acknowledges this capability is directionally correct for the future @AndrewCurran_
Research shows the vast majority of people surveyed cannot explain how AI technologies they use work, raising questions about understanding versus usage of technology @emollick
Flock Safety technology helped return over 450 missing children in 2025 and was instrumental in finding suspects in tragic murders at Brown and MIT, demonstrating AI's role in public safety @a16z

AI Applications

WSJ reporters successfully red-teamed a Claude-run vending machine by creating fake policies and convincing Claude to order and give away Playstations and live fish, though the experiment hints at viable paths forward @emollick
ChatGPT now allows users to adjust specific characteristics like warmth, enthusiasm, and emoji use in personalization settings @OpenAI
ChatGPT introduces writing blocks that make it easier to craft emails, with features to update and format text in chat, highlight to ask for changes, and accept or reject suggestions @OpenAI
Gemini adds ability to attach NotebookLM notebooks as sources, combining shared class notes and deep research to get responses grounded in documents @GeminiApp
Gemini introduces new way to prompt in Nano Banana by using finger or cursor to circle, draw, or annotate directly on images to tell Gemini exactly where to make changes @GeminiApp
Gemini Deep Research reports now include visuals, breaking down complex topics with clear animations and images to help understand dense information at a glance @GeminiApp
Gemini Live improves conversational manners by reducing interruptions when users pause and allowing users to mute their mic while the AI is talking @GeminiApp
Vision AI agents are transforming semiconductor manufacturing, driving higher yield, safer operations, and faster decisions through quality control that can reason rather than just detect @NVIDIAAI
Meta rolled out trajectories feature to developers, allowing code reviewers to see prompts used to generate AI-generated code diffs @GergelyOrosz

AI Research

Google DeepMind's Sebastian Borgeaud expects substantial innovation in pre-training over the next year aimed at making long-context capabilities more efficient and extending models' context lengths further, with recent interesting discoveries related to the attention mechanism @AndrewCurran_
Noam Shazeer states he's 50/50 on whether the next big breakthrough at Google will be made by humans or by Gemini itself @AndrewCurran_
Google confirms they are working on videogames, aligning with expectations from Genie and statements about world models @AndrewCurran_
New paper argues AGI may first emerge as collective intelligence across agent networks rather than a single system, reframing the challenge from aligning one mind to governing emergent dynamics @AndrewCurran_
Research evaluates the potential of LLMs to help with scientific discovery, concluding that new ideas are needed to move AI towards invention, though LLMs can be useful as brainstorming partners @fchollet
OpenAI and US Department of Energy expand collaboration on AI and advanced computing to support national scientific priorities through the Genesis Mission to accelerate scientific discovery @AnthropicAI
Google DeepMind supports the US Department of Energy's Genesis Mission by providing National Labs with access to AI tools including AI co-scientist to help accelerate research in physics, chemistry, and beyond @ShaneLegg
SonicMoE released as a blazingly-fast MoE implementation optimized for NVIDIA Hopper GPUs, reducing activation memory by 45% and achieving 1.86x faster performance on H100 than previous state-of-the-art @berkeley_ai
NYU introduces DexWM, a world model for dexterous manipulation trained on 900+ hours of human and robot video, enabling imagination, planning, and execution of dexterous actions on real robots with zero-shot capabilities @ylecun
Microsoft Research releases Holoportation technology via open source license after a decade of refinement, enabling real-time 3D telecommunications @MSFTResearch
NVIDIA Nemotron family crosses 5 million downloads on Hugging Face @huggingface
Many people underestimate AI due to four OpenAI choices: GPT-5.x instant is not very smart, most users are free users sent to instant often, the router calls everything GPT-5.2, and most people don't know Reasoners exist @emollick
OpenReview supported over 1,300 conferences and workshops in 2025, served 3.3 million active monthly users, and handled over 278,000 paper submissions, but remains underfunded and operating under severe financial constraints @rsalakhu
Agent Skills becomes an open standard, making it easier for everyone to build and contribute to agent capabilities @simonw
Jeff Dean and Sanjay Ghemawat publish Performance Hints document externally, identifying general principles for performance tuning of code @JeffDean