Saturday, March 14, 2026

Gemini Embedding 2 Unifies More Modalities

Gemini Embedding 2 Unifies More Modalities

Today’s Overview

Good morning, Google is pushing Gemini deeper into Workspace with a new multimodal embedding model, while Anthropic is making 1 million-token context windows broadly available. On top of that, the AI agent wave keeps spreading into meetings, commerce, and even private inference hardware. Let's dive in.

Top Stories

Gemini Embedding 2 brings text, video, and audio together

Google DeepMind released Gemini Embedding 2, a natively multimodal embedding model that can handle text, images, short video clips, and audio in one space. The model is being integrated into Workspace for AI Ultra and Pro users, with features like full-draft generation, design-aligned presentations, automated dashboards, and cross-file Q&A with citations. Google is also piloting a multi-agent planning mode for Gemini Business.

  • It can process text, images, video, and audio in a single embedding space.
  • Workspace integration adds cross-file question answering with citations.
  • Google is also piloting a multi-agent planning mode for Gemini Business.

Claude 4.6 gets 1M-token context windows

Anthropic says Claude Opus 4.6 and Sonnet 4.6 now support 1 million-token context windows for all users. The company also removed the previous premium pricing for long-context usage and expanded media capacity to handle up to 600 images or PDF pages per request. The models are generally available across Claude Platform, Amazon Bedrock, Google Vertex AI, and Microsoft Azure Foundry.

  • The new window is available to all users for Opus 4.6 and Sonnet 4.6.
  • Anthropic also raised media capacity to 600 images or PDF pages per request.
  • The models are now available across Claude Platform, Bedrock, Vertex AI, and Azure.

Zoom is bringing AI avatars into meetings

Zoom introduced AI-generated avatars that can join and present in meetings on a user's behalf. The avatars can speak, share screens, and respond to prompts, with early beta testing already underway for enterprise customers. The pitch is simple: let AI handle routine meeting attendance and presentation work.

  • The avatars can present slides and respond to questions.
  • They are designed to join meetings on your behalf for routine attendance.
  • Zoom says the feature is in early enterprise beta.

Research & Analysis

Steering Gemma away from eval awareness and violence

A study on Google’s Gemma 3 models, including the 27B variant, explores whether targeted steering can reduce evaluation awareness and violent content signals. The authors argue this kind of intervention could improve safety without materially hurting overall performance. It is a focused look at how model behavior can be shaped after training.

  • The work targets evaluation awareness and violent intent signals.
  • It uses targeted steering rather than retraining the model.
  • The authors frame it as a way to improve safety without sacrificing performance.

Why agentic commerce could reshape shopping

This piece argues that AI agents will shift e-commerce power away from platform-curated listings and toward buyers making autonomous decisions. The strongest case is in B2B markets, where opaque procurement processes and fragmented information create a lot of room for agents to add value. The broader implication is less platform lock-in and more price transparency.

  • The core idea is a shift from platform-curated listings to agent-driven buying.
  • The article says the biggest near-term opportunity is in B2B commerce where information is often fragmented.
  • It could lead to less platform lock-in and more price transparency.

How Cursor keeps model evals grounded

Cursor uses a hybrid evaluation setup that combines offline benchmark suites with live traffic analysis. CursorBench runs on historical session data to measure code correctness, efficiency, and interaction behavior, while the online layer catches regressions that benchmarks miss. The goal is to keep model assessments aligned with real developer workflows.

  • Offline testing happens through CursorBench on historical session data.
  • The live component watches for regressions in real traffic that offline tests may miss.
  • Cursor evaluates code quality, efficiency, and interaction behavior.

Quick Hits

Keep reading for free

Enter your email. If you're already subscribed, we'll send a sign-in code. If not, you'll subscribe in the next step.

Free access. Subscribe once, then use the same email on future issues.

Free to read. Subscription just unlocks the full issue.