AI just dropped three serious upgrades pushing research, collaboration, and agentic systems to the next level. From autonomous analysts to exam-beating AI and open agent models, here’s what’s new:

🔍 Gemini Deep Research Agent — An Analyst That Works for You
Google’s Gemini Deep Research Agent autonomously plans, searches, reads, and synthesizes multi-step research into cited reports, turning complex analysis into an “analyst-in-a-box” for developers and enterprises.

🏆 Zoom AI Breaks Humanity’s Last Exam — Meetings Get a Brain
Zoom’s AI hits new reasoning highs on Humanity’s Last Exam, powering smarter summaries, real-time translation, and agentic workflows that turn calls into actionable intelligence hubs.

🤖 NVIDIA Nemotron 3 — Open Models Built for Agentic AI
NVIDIA launches Nemotron 3 (Nano, Super, Ultra), open-weight agentic models with massive context, higher throughput, and efficient MoE design, built for scalable, transparent multi-agent systems.

AI isn’t just responding anymore, it’s researching, reasoning, and running workflows end-to-end.

Google Launches Gemini Deep Research Agent

Google's Gemini Deep Research Agent, powered by Gemini 3 Pro, autonomously handles complex multi-step research by planning queries, searching the web (and optionally private files), reading content, filling knowledge gaps, and synthesizing detailed, cited reports—perfect for market analysis or competitive intelligence. Developers activate it via the Interactions API with background execution (background=true) and agent 'deep-research-pro-preview-12-2025', polling for completion since tasks take minutes; streaming provides real-time thought summaries and text deltas for progress tracking. Recent upgrades deliver state-of-the-art benchmarks like 46.4% on Humanity's Last Exam (beating prior models), improved web navigation, and upcoming consumer rollout in Gemini apps, Search, and NotebookLM, making it a cost-effective "analyst-in-a-box" alternative to low-latency chat.​

Zoom AI Tops Humanity’s Last Exam

Zoom unveils groundbreaking AI achievements on "Humanity's Last Exam"—the world's toughest benchmark testing expert-level reasoning across 100+ disciplines, where Gemini Deep Research Agent scores 46.4% (beating GPT-5 Pro's 38.9%) while Zoom's custom AI agents push multimodal understanding boundaries for real-time collaboration. This breakthrough powers Zoom's AI Companion with advanced meeting summaries, real-time translation rivaling Google's, and agentic workflows that autonomously handle action items, transcribe discussions, and generate insights—transforming video calls into productive intelligence hubs. Developers access via Zoom API integrates Gemini 3 Pro's deep research capabilities for custom bots, enabling secure enterprise deployments that analyze calls, detect sentiment, and auto-schedule follow-ups, positioning Zoom as the AI-first communications platform amid 2025's agentic AI race.

Nemotron 3: NVIDIA’s Open Agentic AI Models

NVIDIA Nemotron 3 is a new family of open AI models—Nano, Super and Ultra—built to power transparent, efficient agentic AI across industries, with open weights, data and recipes for maximum customizability. Nemotron 3 introduces a hybrid latent Mixture‑of‑Experts architecture that delivers leading accuracy while dramatically boosting throughput, with Nano offering roughly 4x the throughput of Nemotron 2 Nano and up to 1 million‑token context for multi-agent workflows at low inference cost. Super (~100B parameters) and Ultra (~500B, with up to 50B active per token) target high-end reasoning and large-scale multi-agent systems, using NVFP4 training, long-context support, and reinforcement learning across diverse environments to cut memory needs and improve long-horizon reasoning for complex, specialized agents.​

Hand Picked Video

In this video we'll look at GPT-OSS, OpenAI's unexpected open source model that rivals O3 performance, features built-in web search, and how to test it yourself locally.

Top AI Products from this week

• AI Motion Designer by Agent Opus - For creators who want to create polished motion graphics but don't know how to use After Effects. Agent Opus' AI Motion Designer turns a prompt, text, image, and any idea into professional animations.

• NexaSDK for Mobile - NexaSDK for Mobile lets developers use the latest multimodal AI models fully on-device on iOS & Android apps with Apple Neural Engine and Snapdragon NPU acceleration.

• QualGent - QualGent is the enterprise-grade AI QA agent that helps you test apps at the speed of thought. Describe tests in plain English or connect your app context. QualGent creates tests and runs them on emulators or real iOS/Android devices with self-healing reliability.

• xPrivo - xPrivo is a free, open-source, private AI chat assistant that focuses entirely on keeping you anonymous. You never need to create an account, even with the PRO membership. You can either self-host it or use it directly via the website.

• Varchive - Varchive is a showcase of apps, websites, and experimental projects built with AI assistance, from passion projects to enterprise-grade apps. Varchive itself is built and maintained with assistance from Cursor and Codex. We've populated the archive with a few real-world apps—now it's time to submit yours.

• Stakpak 3.0 CLI - Stakpak is a fully open source DevOps agent written in Rust that helps developers secure, deploy, and operate production infrastructure from the terminal or in GitHub Actions You can run it locally, bring your own keys, or use it with self-hosted models, while keeping safety built in from day one. Stakpak is designed to work reliably with real production infrastructure. Try it now: curl -sSL| sh

This week in AI

• Sample level debugging for MLOps - Bridge the gap between aggregate metrics and real-world edge cases by deeply inspecting per-sample behavior, fixing mislabeled data, and validating models before deployment.

• Google Disco AI Browser - Google's experimental Disco browser uses Gemini 3 to turn open tabs into custom interactive apps via GenTabs. Plan trips, meal prep, or garden layouts with AI-generated tools that remix your browsing context.

• Worlds First Mortis AI Game - Codex Mortis claims "world's first fully playable game created 100% through AI." Vampire Survivors-style bullet hell built with Claude Code, ChatGPT art, PIXI.js in 3 months. Steam demo out now.

• World App Update - New World App adds World Chat (E2E encrypted messaging for verified humans), borderless payments via virtual accounts in 18 countries, Earn (15-18% APY), 100+ assets, and Mini Apps in chats. Most-used self-custody wallet.

• US Tech Force Launch - Trump admin launches Tech Force to hire 1,000 early-career AI/software specialists for federal agencies. 2-year roles ($130K-$200K) focus on AI modernization, partnering with Microsoft, Google, OpenAI. Apps open now.

Paper of The Day

Researchers introduce an error-driven prompt optimization framework for a Code Generation Agent using on-premises SLMs like Qwen3 4B on financial tabular data (TAT-QA). By clustering prediction errors via HDBSCAN and iteratively adding domain-specific rules (e.g., "'percentage change' results 'percent' scale"), they boost exact-match accuracy from 59.96% to 70.82%, surpassing GPT-3.5 Turbo while ensuring privacy.

To read the whole paper 👉️ here