Explore what's new in AI

The Efficiency Frontier trims LLM context token spend by 25%

The paper says the right context strategy can cut token spend by about 25% at similar quality. Reuse changes which method wins.

Firecrawl launches Monitoring for URL change alerts

Firecrawl's Monitoring feature turns a URL into an alertable target. Useful if you want agents to watch pages instead of polling them by hand.

Lngram introduces latent n-gram memory for multimodal LLMs

Lngram moves n-gram memory into latent space so Transformers can offload static lookup. The pitch is better long-context efficiency without rewriting the backbone.

xAI opens Grok Build 0.1 in public beta with API pricing

Grok Build is now in public beta on the API console with published pricing. That makes the coding-agent stack easier to evaluate and budget.

minWM brings real-time interactive video world models into one framework

minWM is an open framework for real-time interactive video world models. It points at a broader stack for simulating and controlling visual futures.

Browser Use adds stealth browsing through Browser Use Cloud

Browser Use Cloud adds persistent auth and stealth browsing so agents can get through sites that normally block automation. That's a useful new hammer.

OpenAI retires o3 and GPT-4.5 from ChatGPT

OpenAI is retiring o3 and GPT-4.5 from ChatGPT, not the API. That forces users to migrate favorites inside the product UI.

Moondream grounds prompts to pixel-accurate objects

Moondream's grounding demo maps natural-language prompts to exact pixels. That's the kind of visual precision agents need for real UI and image tasks.

GitHub Sane Defaults applies branch protection in one command

GitHub Sane Defaults automates branch protection, auto-merge, and rulesets in one command. It's a simple way to enforce repo hygiene at scale.

LangChain Deep Agents v0.6 makes harness profiles first-class

Deep Agents v0.6 turns harness profiles into a first-class knob. The idea is to tune behavior around the task, not keep reprompting by hand.

ComfyUI adds OpenRouter model access inside workflows

ComfyUI's OpenRouter integration lets workflows call multiple models directly. That's a practical way to swap providers without rewiring the app.

OpenAI Codex can search prior chat content for old threads

Codex can now search prior chat content, which matters once one task turns into a maze of side threads. It makes long-lived sessions easier to mine.

OpenRouter adds server-side apply_patch for model file edits

OpenRouter's apply_patch lets models propose edits as validated diffs. That moves file changes toward a safer, more auditable protocol.

OpenAI and Terence Tao say AI lowers research friction

The gain is not answers, it's making weird ideas cheap enough to test. Tao's point is that AI lowers the cost of exploring promising false starts.

John Schulman says renderers should be a standalone LLM library

Schulman argues renderers sit beneath the model and should be shared infrastructure. Getting rendering wrong hurts caching, safety, and train-test match.

Google turns Gemini Omni into sketch-to-motion demos

Gemini Omni is being pushed as a creative editor, not just a video model. The demos show sketch-to-motion and prompt-driven editing workflows.

TanStack AI adds React Native support for mobile streaming

TanStack AI on React Native lets mobile apps stream agent output without a separate stack. Useful if you want the same UI logic on web and phone.

LiteParse runs PDF parsing in the browser with WASM

LiteParse pushes PDF parsing into WASM so it can run in browsers, workers, and edge runtimes. That's a cleaner ingestion path for RAG pipelines.

Google Flow Agent batches scene edits and asset organization

Flow Agent can brainstorm, batch scene variants, and organize final assets. That makes it more than a demo, it's a real creative workflow layer.

Google makes Gemini Spark available to AI Ultra subscribers

Spark is now live for US AI Ultra users, giving Google a 24/7 agent that can work while you steer. It pushes Gemini closer to an always-on assistant.

ChatGPT adds table of contents for long-running chats

Long chats now get a clickable TOC once they pass five replies. Small UX work like this matters when agents turn conversations into projects.

Salesforce says a 231-day Claude Code migration shipped in 13 days

Salesforce says one migration landed in 13 days, with 21 endpoints and full test coverage. That's the kind of agentic ROI enterprise teams actually care about.

VS Code Remote Sessions let GitHub mobile steer agent work

VS Code Remote Sessions let you reach agent work from GitHub mobile. That makes background coding loops much easier to steer away from the laptop.