📰 AI News Daily — 20 Dec 2025

TL;DR (Top 5 Highlights)

• Google rolls out Gemini 3 across Search, NotebookLM, and partners; Flash boosts speed and cost, powering third‑party research like Antigravity’s in‑browser “computer.”
• OpenAI reportedly targets up to $100B in new funding at ~$830B valuation; Disney signs a three‑year deal to use Sora for fan‑made character videos.
• NVIDIA Blackwell plus optimized vLLM delivers ~33% more tokens per dollar; vLLM joins the PyTorch Foundation to accelerate open inference efficiency.
• Google expands SynthID to detect AI‑generated video in Gemini; California enacts a landmark deepfake law, raising industry bar for authenticity and compliance.
• U.S. Department of Energy partners with OpenAI and Periodic Labs to pair AI with national‑lab experiments; OpenReview wins a $1M lifeline amid urgent sustainability calls.

🛠️ New Tools

• Bold Google FunctionGemma — compact, open model for function calling and on‑device agents; cuts latency and cloud dependency, enabling faster, privacy‑preserving assistants and embedded workflows on phones and edge devices.
• Bold Gemma Scope 2 — interpretability suite with sparse autoencoders and layer transcoders; helps researchers examine circuits and failure modes, improving safety and debuggability for small and mid‑size models.
• Bold Microsoft Agent Lightning — drop‑in reinforcement learning for existing agents; upgrades decision quality without rewrites, reducing experimentation friction for production agent teams.
• Bold Meta Seal — open watermarking suite for generative media; simplifies provenance tagging and verification pipelines, helping platforms meet disclosure requirements and combat manipulation at scale.
• Bold Jetson Field Kit — ruggedized pack with Orin Nano and sensors; streamlines edge AI deployments for conservation, robotics, and inspection where connectivity and power are limited.
• Bold Google Opal (in Gemini) — natural‑language app builder with visual step editors; lets teams spin up custom mini‑apps and workflows rapidly, rivaling low‑code tools with tighter AI integration.

🤖 LLM Updates

• Bold Gemini 3 Flash — tops coding evaluations (SWE‑Bench Verified, Vals), delivering strong code reasoning at lower cost; speeds CI fixes and strengthens agentic coding workflows for production teams.
• Bold Gemini 3 vs. GPT‑5.2 — trade wins across Toolathlon, ECI, GSO, and ALE‑Bench; no single leader, encouraging multi‑model routing for robustness, latency, and cost control.
• Bold AI2 OLMo‑3.1‑32B‑Think — public reasoning trials open; transparent “think” traces support research into long‑chain reasoning, evaluation, and debuggable deliberation.
• Bold Parallel decoding + MoE gains — a rewritten MoE stack halves memory and nearly doubles training speed; “Jacobi Forcing” enables causal parallel decoding, lowering long‑response latency.
• Bold Tokenizers + structure — first tokenizer scaling law and DSR‑Bench show structural reasoning gaps; tool use and scaffolding remain vital for order, hierarchy, and connectivity tasks.
• Bold Inference efficiency — NVIDIA Blackwell with optimized vLLM yields ~33% more tokens per dollar; vLLM joins the PyTorch Foundation, accelerating community improvements to serving stacks.

📑 Research & Papers

• Bold Allen AI video reasoning stack — releases models, datasets, and a benchmark for agentic video understanding; standardizes evaluation of temporal reasoning and planning over complex scenes.
• Bold Meta PE‑AV — open audiovisual perception engine; unifies sound and vision for richer multimodal agents in robotics, accessibility, and safety applications with reproducible training recipes.
• Bold DexWM — learns dexterous manipulation from human videos; advances robot skills with less teleoperation, narrowing sim‑to‑real gaps for practical manipulation tasks.
• Bold Stanford ARTEMIS — AI agent beats human pentesters on vuln discovery at ~$18/hour; still needs human oversight, pointing to hybrid, safer security workflows.
• Bold Clinical LLM safety — study confirms prompt injection vulnerabilities in healthcare; urges sandboxing, tool isolation, and regulatory oversight before bedside deployments.
• Bold Scholarly integrity — analysis warns generative tools can inflate low‑quality papers; recommends updated peer review processes and automated integrity checks to preserve trust.

🏢 Industry & Policy

• Bold Google Gemini 3 — rolling across Search (new generative UI), NotebookLM upgrades, and third‑party integrations; signals broader, more reliable deployment of Gemini for mainstream consumer and enterprise workflows.
• Bold OpenAI — reportedly seeking up to $100B at ~$830B valuation; Disney inks a three‑year Sora deal for fan‑made character videos, expanding AI video into entertainment pipelines.
• Bold DOE partnerships — Department of Energy teams with OpenAI and Periodic Labs to fuse AI with supercomputing and real‑world experiments, accelerating discovery and setting safety standards in national labs.
• Bold OWASP — publishes first AI agent security risks list; gives developers a shared threat model for data poisoning, adversarial prompts, and tool misuse, improving baseline defenses.
• Bold Content authenticity — Google expands SynthID to video verification inside Gemini; California enacts a deepfake law, raising compliance pressure on platforms and advertisers handling synthetic media.
• Bold Infrastructure scrutiny — OpenAI’s $7B, 1.4 GW Michigan data center approval sparks local protests; highlights growing regulatory, environmental, and community considerations for hyperscale AI builds.

📚 Tutorials & Guides

• Bold François Chollet — makes his deep learning text freely available online; an accessible, high‑quality resource for practitioners leveling up fundamentals and modern techniques.
• Bold NVIDIA NeMo Agent Toolkit course — hands‑on training for production‑ready agents; covers tool use, safety, and orchestration, accelerating enterprise deployments.
• Bold Jeff Dean & Sanjay Ghemawat — share performance‑engineering principles from inside Google; practical guidance for squeezing latency and cost out of large‑scale systems.
• Bold Agentic timelapse tutorial — step‑by‑step guide for generating photorealistic home renovation timelapses; demonstrates multi‑tool orchestration and iterative refinement in creative workflows.
• Bold Kenton Varda podcast — explores code modes and model‑centric programming; clarifies where LLMs fit in developer tooling and how to design resilient AI‑first software.
• Bold 2025 must‑knows — curated primer on RL, RLHF variants, test‑time scaling, neuro‑symbolic methods, and new hardware; frames key concepts shaping next‑wave AI systems.

🎬 Showcases & Demos

• Bold Agentic renovation — an AI agent produces smooth, photorealistic home‑renovation timelapses; showcases coordinated planning, vision models, and iterative editing for consumer‑friendly design previews.
• Bold NitroGen — open‑source generalist agent plays 1,000+ games across genres; highlights transfer learning and tool routing for broad, zero‑shot competence.
• Bold Kling Motion & Kling.ai 2.6 — cinematic video generation and high‑speed anime action; creators report festival‑worthy results, raising bar for accessible, stylized storytelling.
• Bold Runway GWM‑1 — next‑gen video models show stronger motion coherence and editability; tighter hooks into creative tools hint at faster, iterative production cycles.
• Bold Antigravity + Gemini 3 Flash — in‑browser research “computer” gets faster and more capable; demonstrates practical gains from low‑latency, high‑throughput inference in everyday workflows.
• Bold DexWM in action — dexterous robotic motions learned from human video; compelling progress toward affordable, adaptable manipulation skills in the real world.

💡 Discussions & Ideas

• Bold Open science — leaders push for sustainable infrastructure (OpenReview fees, $1M pledge) and argue open models plus scaffolding provide transparency, control, and community verification.
• Bold Human‑centered AI — roadmaps emphasize test‑time scaling, RL, and neuro‑symbolic approaches (Yejin Choi et al.); aim to align systems with human goals instead of raw benchmark chasing.
• Bold Infrastructure lessons — Temporal shines for long‑running agents; caution against brittle serverless backends. Best practice: build context before generation and move to learned context management.
• Bold Rethinking scaling — symmetric inductive biases can rival brute‑force data growth; efficiency breakthroughs like FlashAttention have saved massive global compute, reshaping research priorities.
• Bold Geopolitics — China’s accelerating research and Hainan’s digital zone shift perceptions of innovation leadership, intensifying global competition for talent and standards.
• Bold Path to HLAI — Yann LeCun expects gradual progress over 5–20 years; argues LLMs alone won’t yield real‑world intelligence, urging exploration of new representations and learning paradigms.
• Bold UX paradox — the “vending machine” effect: boundless generative options can frustrate undecided users; better guidance, defaults, and constraints improve satisfaction and outcomes.

Source Credits

Curated from 250+ RSS feeds, Twitter expert lists, Reddit, and Hacker News.