Explore what's new in AI

Claude's computer-use guide says pre-downscale images and trim context

The guide gives practical browser and computer-use tips, including image sizing and effort tuning, to cut wasted tokens and failures.

ClickUp Brain makes model choice feel native across 14 frontier models

One dropdown now spans Anthropic, OpenAI, and Google, and the UI adapts to the chosen model. That's a cleaner pattern for multi-model apps.

Dr Vivian Balakrishnan shows a Raspberry Pi AI assistant stack

The talk shows a Pi 5 pipeline with ingest, extract, and synthesize steps, plus WhatsApp, Whisper, and Obsidian components.

Linear Agent lets teams ask how features work without interrupting engineers

This is a concrete agent product that answers implementation questions from code context, which is the right shape for product and eng planning.

A non-technical EA shipped a six-skill Claude workflow for podcast outreach

The post is a concrete blueprint for turning one manual process into a reusable agent pipeline with context files, tests, and review.

Pika MCP lets agents call creative models through skills and presets

The workflow turns trending video and image formats into reusable agent actions, which is how these stacks get adopted.

AI SDK moves telemetry off mandatory OTel and into Node diagnostics

The new setup is easier to customize and avoids extra dependency baggage, which should make instrumentation simpler to adopt.

OpenCode 1.50.1 adds pinned sessions for worktree multitasking

You can keep each PR or task as a pinned slot and pick up exactly where you left off, which is handy for agent-heavy workflows.

Andrew N Carr's markerless capture tracks motion from webcam video

The system estimates landmarks from regular video, making multi-person capture and camera pose tracking much cheaper to prototype.

GitHub Copilot explains the harness behind its VS Code agent

The post is a useful look at the layer that shapes agent behavior, from tool calls and memory to offline benchmarking and session control.

Codex debugged a printer queue after three turns of feedback

A real office failure becomes a tiny computer-use benchmark, and the agent found the broken path faster than manual trial and error.

Codex teaches piano by turning prompts into a working practice loop

The demo shows a concrete computer-use workflow, plus prompts and code, so others can reproduce the setup instead of copying a screenshot.

Pi removes tools from its coding agent and ships a no-tools mode

The experiment forces the model to reason with fewer crutches, a useful stress test for how much the harness itself matters.

v0 Browser Use lets generated apps open themselves and patch flows

The agent can browse the app it built, inspect screenshots, critique UX, and keep iterating without a human in the loop.

Vercel gates source maps behind auth for production debugging

You can inspect shipped code without exposing source files, which removes a common tradeoff for teams shipping browser bundles.

Vercel CLI adds native curl syntax and SSO-protected agent deployments

The CLI can now test deployments with auth, while Vercel also keeps agent traffic behind org SSO so teams can safely hit internal apps.

Codex on ChatGPT mobile turns your laptop into a satellite device

The mobile workflow lets Codex keep working while the laptop stays closed, which is a real shift for long-running agent tasks.

Yacine says every Codex prompt should start with /goal

A simple goal header can keep sessions focused and easier to recover when the agent starts looping. Small prompt conventions can pay off fast.

PwC's Is Grep All You Need? compares grep and embeddings in coding agents

The paper argues the right harness can make simple grep-style search match or beat embeddings on coding-agent tasks. That is a big retrieval design signal.

X promises monthly algorithm release notes on GitHub

X is opening a changelog for feed ranking changes, and the DeepWiki readout shows dwell time is explicitly rewarded.

Dwarkesh Patel's AlphaGo lecture frames RL credit assignment in modern agents

The lecture compares MCTS with policy gradients and shows where research agents help, then still hit the hard parts of search and credit assignment.

Figure runs Helix-02 for 50 hours and sorts 63,000 packages

The livestream claims human-parity throughput on real package sorting and automatic resets when the policy drifts. It is a rare long-duration autonomy demo.

Zero debuts as an agent-friendly systems language

The language pitches explicit capabilities, JSON diagnostics, and typed safe fixes. It is clearly designed to be repaired by agents, not just humans.

Gemini Spark screenshots show Google adding skills and tool use

The UI hints at a more agentic Gemini, with a skills creation flow and tool-aware reasoning surfaced on mobile and desktop.

TanStack Start inlines route-aware CSS during SSR to improve FCP

Inlining only the current route's CSS cuts blocking requests, a small SSR optimization that should show up in first paint and perceived speed.