Explore what's new in AI

Higgsfield CLI packs ad-video modes, face training, and 30-plus models

The CLI wraps a broad media stack into one binary, from UGC ads to product imagery and scheduled generations. It is a useful sign that creative pipelines are becoming composable.

LangChain's DeepAgents CLI v0.0.30 adds tracing, MCP, and a ready-to-code loop

The new CLI surfaces LangSmith tracing, loads MCP tools, and stays in a long-running code loop, which makes the subagent workflow much easier to debug.

OpenAI's Codex plugin for Claude Code adds review and rescue commands

The plugin gives Claude Code read-only review, adversarial review, and rescue commands powered by Codex. It is a practical cross-tool bridge for multi-model workflows.

Crabbox 0.5.0 adds browser leases and authenticated WebVNC

Crabbox 0.5.0 adds desktop and browser leases, authenticated WebVNC, and Windows and WSL2 support. It makes remote CI boxes much more usable for hands-on work.

Anthropic's Orbit leak points to a proactive assistant across Gmail and Slack

The screenshots suggest Orbit will summarize activity from Gmail, Slack, GitHub, Drive, and Figma before you ask. That would push Claude toward an always-on work assistant.

Pauliusztin's GraphRAG guide treats agent memory as graph traversal

The guide argues GraphRAG is a data modeling problem, not a retrieval trick, and shows a memory pipeline built around graphs and MCP tools.

Nanowhale shows an agent fully pretraining a tiny DeepSeek model

nanowhale is a tiny DeepSeek model that was fully pretrained by an agent. It is a compact proof that agentic training loops can produce real model artifacts, not just code.

TickerTrends shows Codex passing Claude Code in downloads

TickerTrends shows Codex overtaking Claude Code in download volume on April 30. The gap widens fast enough that distribution momentum may be shifting, not just model quality.

badlogic's pi comment extension opens the last assistant reply in an editor

The pi comment extension pulls the last assistant message into your editor, lets you annotate it, then writes it back. That is a clean workflow for reviewing agent output.

TheTuringPost explains vector embeddings, latent space, and RoPE

The guide walks through embeddings as vectors, similarity, and latent space, then ties them back to search and generation. It is a solid refresher for RAG-heavy systems.

Codex goal mode can ping you later about bugs and flaky tests

Codex goal mode can set future reminders for unresolved threads, bugs, and flaky tests. That pushes long-running agent work closer to a real follow-up system.

OpenAI's Codex limit rings show remaining usage for any pet

The limit ring overlay makes Codex usage visible around any pet without patching the app. That is a neat way to surface quota pressure before an agent gets cut off.

Redis adds array indexing and grep-style text search

Redis's new array and grep-style search features widen what you can query without leaving the database. That matters for agent memory and retrieval systems.

OpenAI ships ChatGPT for Intune to managed enterprise devices

OpenAI now has an Intune-managed ChatGPT app for enterprise iPhones and iPads. It brings voice, file handling, and writing help into environments that need device control.

Nous Research makes Trinity-Large-Thinking free for a week

Nous is giving Trinity-Large-Thinking free for a week inside Nous Portal. That makes it easy to test a newer reasoning model in Hermes Agent without paying for inference.

aidenybai's localterm runs a terminal in the browser

localterm turns a browser tab into a terminal, with tabs, split panes, and quick app switching. That makes it easier to keep shell work in the same place as your dev stack.

OpenAI gives Codex users 10x rate limits through June 5

A temporary bump gives heavy Codex users a much larger runway before the June 5 reset. It is a useful window for long agent sessions and batchy workflows.

RAVEN validates 118 exoplanets from NASA's TESS star catalog

The pipeline turns a false-alarm-heavy search into a discovery engine, confirming 31 new worlds and cutting uncertainty 10x.

Microsoft Build puts multi-agent patterns and MCP on the VS Code agenda

The Build lineup shows Microsoft treating agent UX, MCP, and verification as core VS Code topics. That is a strong signal for the next wave of dev tooling.

Hermes Agent adds shareable achievement badges from the dashboard

The dashboard can mint shareable PNG badges for agent accomplishments, turning tool usage into a lightweight social proof loop.

Corey Ganim's monthly second-brain audit catches stale facts before they spread

A five-minute monthly checklist hunts contradictions, stale pages, and uncited claims before the knowledge base poisons itself.

Perplexity builds a Digest feature for connected-source summaries

The interface hints at an aggregated summary layer pulling from connected sources inside Perplexity Computer. That would make proactive research and context briefing much smoother.

OpenCode adds GUI chat, inline diffs, and provider switching mid-conversation

The new GUI layer trims friction for terminal-first agent work, including diffs, queues, and provider switching. It makes OpenCode easier to use as a daily driver.

Ollama adds Claude Desktop support via third-party inference

You can now point Claude Desktop at Ollama's cloud models without swapping apps. It widens the local-first path for people who want Claude workflows with non-OpenAI models.

OpenAI launches a $10B deployment JV with private-equity firms

OpenAI is building a distribution and deployment arm with PE firms, not just a model vendor. That shifts enterprise AI selling toward workflow rollout and change management.

xAI rolls out Grok connectors with custom MCP support

Grok connectors now reach GitHub, Notion, Linear, Google, Microsoft, and custom MCP endpoints. It is a concrete step toward making Grok operate inside real work apps.