Explore what's new in AI

Google Search information agents monitor topics and send synthesized updates

These agents are basically background monitors with triggers and urgency logic, which turns Search into an alerting layer instead of a one-shot query tool. Useful for ops, markets, and news tracking.

Hermes Agent patches cold starts and tool-call latency with four speed fixes

The release shows concrete wins across cold start, agent loop overhead, and subprocess polling, which is exactly where long-running agents feel slow in real use.

Gemini Daily Brief builds a morning digest from Gmail, Calendar, and Tasks

This is a practical agent pattern: synthesize inbox, calendar, and tasks into one prioritized view. It reduces the startup cost of getting oriented every morning.

Index launches Shapley-based revenue sharing for content used by agents

By paying sources based on estimated marginal contribution, Index is trying to make agent consumption measurable and monetizable. That could matter if AI answer engines keep scaling.

Google redesigns Gemini with Neural Expressive and inline Live

Google is making the app feel more like a fluid assistant than a chat box, with stronger motion, richer UI, and faster live switching. Good design work can change how often people stay in the loop.

Project Genie grounds Google's world model in Street View locations

Grounding simulated spaces in real Street View data makes the world model much more useful for exploration and planning. It's a notable step toward real-world interactive environments.

Google Stitch lets teams prompt UI mockups and refine them by voice

Stitch is trying to make UI generation feel like working with a design partner, then handing off to build systems. Good signal for prompt-to-UI workflows becoming more operational.

xAI brings Grok imagine into Paper for faster image exploration

Paper is adding a faster image model for rapid generations and edit chains. That is the kind of thing creators will actually iterate on instead of waiting for one perfect render.

Interfaces adds animation review and watchback to its launch issue

The product is aimed at a real pain point, figuring out why motion feels wrong. Recording and replaying animation states can save a lot of guesswork in UI debugging.

Google Docs Live turns voice into real-time document editing

Voice-first editing turns Google Docs into a conversational input surface, which is a meaningful productivity shift for drafting and structured note capture. It also hints at more multimodal Workspace flows.

Google Pics uses object segmentation for cleaner image edits

Pics can segment objects and understand how they relate, so edits become targeted instead of brute-force. That's a nicer workflow for business visuals and quick marketing assets.

Google Search box gets AI-powered multimodal suggestions

Search is moving from autocomplete to intent shaping, with text, images, files, videos, and tabs all in play. That changes how users compose queries and how agents can help.

TanStack Start adds Deferred Hydration for visible and idle UI work

You can now delay hydration until a component is visible, interacted with, or idle. That's a practical way to cut wasted work in modern React apps.

Gemini app passes 900 million monthly users and doubles in a year

Google's consumer AI surface is now massive, which explains why every new feature ships in app, search, and workspace simultaneously. It's a real distribution moat.

Android Halo puts agent progress at the top of the screen

Google is building persistent agent status for Android, which could make long-running tasks less opaque. The idea is to keep progress visible without leaving the current app.

Ask YouTube turns video search into a conversational answer surface

YouTube search is becoming an interactive answer surface, with structured text plus relevant videos. That could change how people discover tutorials and how creators get surfaced.

Perplexity says Rho cut weekly meetings 90 percent with Computer

This is one of the few case studies with hard numbers, 120 work hours saved in 12 weeks. It's a good signal that computer-use agents can reduce coordination overhead.

Google formalizes agentic commerce with UCP, AP2, and Universal Cart

Google is defining both the shopping language and the payment controls for agents, which is a big step toward commercial autonomy. The guardrails are as important as the convenience.

Dwarkesh and Eric Jang revisit AlphaGo as a model for tractable search

The AlphaGo intuition still matters: if the model can reject hopeless branches early, search becomes tractable. That is a useful mental model for modern agent evaluation.

ComfyUI adds a Claude partner node for prompt rewriting and image critique

This makes Claude a reusable node inside visual workflows, which is a nice pattern for multimodal pipelines. Prompt cleanup and structured text handoff are obvious wins.

Google Flow Tools lets creators build and remix custom video effects

The new Flow tools aim at custom creative workflows, not just one-off generation. That's a useful signal that AI video tools are becoming programmable media apps.

Google's Gemini for Science adds tools for papers, hypotheses, and code

Google is turning Gemini into a lab workflow, with tools for literature review, hypothesis generation, and turning research goals into code. That's a useful step beyond generic chat.

Google cuts AI Ultra to $100 and drops the top tier from $250 to $200

Google is bundling agentic features, storage, and media tools into a lower entry tier, which should increase adoption pressure for the whole ecosystem. Pricing is now part of the platform strategy.

Karpathy joins Anthropic's pretraining team to speed up frontier R&D

Anthropic is signaling a stronger pretraining push, with Karpathy saying he will help use Claude to accelerate research on Claude. Good talent and compounding R&D loop signal.

Anthropic's computer use blog maps production reliability fixes

Anthropic is now documenting the failure modes that matter in prod, click accuracy, thinking effort, context length, and replayable demos. That's the kind of operational guidance teams need.

Thinking Machines offers $100k grants for human-AI interactivity research

The grants put real money behind better human-AI interfaces, not just model scale. Expect more work on feedback, controllability, and interaction design.