Explore what's new in AI

OpenAI Podcast says GPT-5.4 Pro helped solve a 60-year Erdos problem

A real math breakthrough is stronger than a benchmark win, and it hints research agents are moving from summarizing to discovering.

ChatGPT Search reaches 159.1M monthly active recipients in the EU

OpenAI's DSA disclosure shows Search has become a major European surface, not a side feature, with user growth accelerating hard.

Pika launches person-shaped creative agents for visual production

Pika is moving from one-shot prompts to agents with voice, face, and taste, which is a new UX for video creation workflows.

gogcli 0.14.0 adds encrypted backups and resumable Gmail exports

The release turns Google data into local encrypted backups with Markdown mirrors, which is genuinely useful for ops and migration work.

Prism-Reranker ships a four-size open reranker family for agentic search

Reranking is still one of the easiest wins in search and RAG, and this release gives builders a small, fast, MIT-licensed option.

SSL turns unstructured SKILL.md files into typed agent skill graphs

The paper makes skills easier to search and safer to review, which matters once skill registries get too large for plain text.

Skill Retrieval Augmentation shows agents still need help choosing skills

The SRA-Bench numbers suggest retrieval alone is not enough, because models often load skills whether or not the task needs them.

Claude Code adds push notifications when long tasks finish or need input

This small UX win has outsized impact, letting people stay in flow while long-running agent jobs progress in the background.

Codex inside PostHog lets agents write queries and verify them in one loop

Putting the browser inside the coding agent keeps analysis, execution, and follow-up work in one context, which is exactly where these systems shine.

VS Code gets Agent Plugins for skills, commands, MCP servers, and hooks

The extension bundles a lot of the agent plumbing in one install, lowering the friction for teams standardizing workflows in the editor.

steipete runs Codex on every commit to catch regressions and security bugs

Using the agent as a continuous reviewer found a real bug in 10 minutes, a strong pattern for pre-merge safety gates.

Firecrawl ships /parse to convert PDFs and spreadsheets into clean AI-ready data

A dedicated parse endpoint saves teams from building their own document-cleanup pipeline, and the Rust parser is claimed to be 5x faster.

browser_use opens Browser Harness for agentic web automation

This is a useful building block for teams that want browser-level task execution without wiring a full cloud product around it.

OpenKB turns raw docs into an interlinked wiki without a vector DB

Instead of re-chunking on every query, OpenKB compiles knowledge into persistent pages and cross-links, making large corpora easier to navigate.

Matt Pocock updates Skills for GitHub, Jira, Linear, or local files

The workflow now abstracts the issue tracker itself, so one skill package can follow the same loop whether you use cloud tools or plain files.

Anthropic says Claude Code is hunting papercuts and moving to a no-flicker renderer

The team is explicitly cleaning up UX bugs before making the renderer default, a sign that stability is now part of the moat.

Dwarkesh Patel clips Adam Marblestone on reward functions and sample efficiency

The thread frames sample efficiency as a loss-function problem, a useful lens for post-training, synthetic data, and agent learning.

OpenCode hits 150,000 GitHub stars as the agentic CLI keeps growing

The star milestone shows the open-source agent CLI has moved from curiosity to default infrastructure for a large slice of devs.