Monday, March 24, 2026

Sora is dead. OpenAI's video generation product announced its shutdown today, which didn't stop OpenAI from also unveiling shopping inside ChatGPT via something called the Agentic Commerce Protocol. Meanwhile, Anthropic shipped auto mode for Claude Code, a safer alternative to --dangerously-skip-permissions that uses a safety check that reviews each action. On the local inference front, people are now running massive AI models on a regular Mac by loading only what they need from disk. And three solid Show HNs today, including sub-second video search.

Big Moves

Goodbye to Sora

OpenAI is shutting down Sora, their video generation tool. 987 points on HN, which tells you people had feelings about this one. Sora launched to enormous hype, and now it's gone. The video generation space has gotten crowded and Sora never really found its footing after the initial demo reel.
Hacker News

Auto Mode for Claude Code

Anthropic introduced auto mode, a new permissions mode in Claude Code. Instead of approving every action or using --dangerously-skip-permissions, Claude now makes permission decisions on your behalf with a safety net: a separate AI reviews each action for safety before it runs. Simon Willison has the breakdown. This is a good middle ground — the constant permission prompts were the main reason people reached for the dangerous flag in the first place.
Simon Willison

OpenAI Puts Shopping in ChatGPT

ChatGPT now does product discovery and side-by-side comparisons, powered by something OpenAI is calling the Agentic Commerce Protocol. Merchants can integrate directly. So ChatGPT is a shopping assistant now. The name is a lot.
OpenAI Blog

Things People Built

Show HN: SentrySearch — Sub-second video search with Gemini

Gemini can now understand video, and someone immediately built a sub-second video search tool on top of it. 407 points on HN. This is the kind of thing that was a multi-month research project two years ago and is now a weekend Show HN.
Hacker News

Show HN: ProofShot — AI coding agents that check their own UI

ProofShot gives AI coding agents the ability to screenshot and verify the UI they generate. Simple idea, obvious in retrospect: if the agent can see what it built, it can catch its own mistakes. 140 points.
Hacker News

Show HN: AI Roundtable — 200 models debate your question

Throw a question at 200 models and watch them argue about it. Fun to poke at, and occasionally useful for seeing where models disagree.
Hacker News

Tools & Platforms

FastMCP

FastMCP is a framework for building MCP (a standard for connecting AI to your apps) servers. Clean getting-started docs. If you've been meaning to build one and wanted something more structured than rolling your own, this is the on-ramp.
Hacker News

Sandboxing AI Agents, 100x Faster

Cloudflare announced Dynamic Workers for safely running AI agent code in isolation. The pitch: spin up a contained environment for each agent action, 100x faster than traditional approaches. As agents get more autonomy (see: Claude Code auto mode above), keeping each action safely walled off matters more.
Cloudflare Blog

Worth Reading

So Where Are All the AI Apps?

From Answer.AI (Jeremy Howard's lab). 428 points on HN. The question is genuine, not rhetorical — where are the apps that actually use these capabilities well? Good discussion in the comments about what's missing between "impressive demo" and "thing people use every day."
Answer.AI

Also

Today’s Sources