The AI Friday Brief
Sora is dead. OpenAI's video generation tool posted its goodbye yesterday, and the HN thread has almost a thousand points. Meanwhile, the builders keep building: a Show HN using Gemini-powered sub-second video search hit 400 points, Simon Willison wrote up Claude Code's new auto mode which uses a safety layer that reviews each action, and Google dropped both TurboQuant to help fit bigger AI models on cheaper hardware and Lyria 3 for music generation via the Gemini API. Good day for people who like to try things.
Things People Built
Show HN: SentrySearch — sub-second video search using Gemini's native video embeddings
Gemini can now understand video directly — not just screenshots or captions — and this tool lets you search footage by describing what you're looking for. Point it at video footage, type a query in plain English, and it finds the right moment. ssrajadh built this as a Show HN. 400 points on HN, which for a Show HN is a lot. HN discussion.
Hacker News
Show HN: ProofShot — give AI coding agents eyes to verify the UI they build
Solves a real problem: coding agents generate UI code but can't see the result. ProofShot takes a screenshot after each change and feeds it back to the agent so it can catch its own visual bugs. Simple idea, useful if you're doing any agent-driven frontend work. HN discussion.
Hacker News
Show HN: AI Roundtable — let 200 models debate your question
From Opper AI. You submit a question and it gets routed to ~200 models, which then respond and debate each other. Feels like a fun way to see where models actually diverge on a topic. Probably more useful for comparison shopping models than for getting a definitive answer. HN discussion.
Hacker News
Building a coding agent in Swift from scratch
A Claude Code-style coding agent written entirely in Swift. If you're in the Apple dev world and want to understand how coding agents work under the hood, this is a clean example you can learn from. HN discussion.
Hacker News
Ensu: privacy-focused AI app that runs on your device
Ente, the privacy-focused photo storage company, built a local AI app called Ensu. Everything runs on your device — nothing is sent to the cloud. Ente has a solid track record with privacy-first design, so this is a credible entry in the keep-your-data-local space. HN discussion.
Hacker News
Tools & Platforms
Auto mode for Claude Code
Simon Willison wrote up the new auto mode in Claude Code, which is Anthropic's answer to running the agent without stopping to ask permission every few seconds. Instead of blindly approving everything, a separate AI that checks each action for safety reviews it before it runs. You get the speed of fully-automatic mode with a safety net. Whether that safety net is good enough is an open question, but it's a better design than a simple on/off switch.
Simon Willison
FastMCP
A framework for building MCP (a standard for connecting AI tools to your apps) servers quickly. If you've been wanting to wire up your own tools to Claude or other compatible clients but found the boilerplate annoying, this is aimed at you. HN discussion.
Hacker News
Cloudflare: running AI agent code safely, 100x faster
Cloudflare announced dynamic Workers for running AI agent code safely. The pitch: spin up an isolated environment for each agent action in single-digit milliseconds instead of seconds. Relevant if you're building anything where an agent needs to run untrusted code. HN discussion.
Hacker News
Industry News
Goodbye to Sora
Sora, OpenAI's video generation tool, is shutting down. 967 points on HN. No detailed explanation yet beyond the tweet. The video generation space has gotten crowded since Sora launched, and it never quite found its footing as a product. HN discussion.
Hacker News
Google Lyria 3: music generation via the Gemini API
Lyria 3 is now available in paid preview through the Gemini API and for testing in Google AI Studio. There's also a Lyria 3 Pro tier for longer tracks. If you've been wanting to programmatically generate music, this is the easiest on-ramp right now.
Google AI Blog
ChatGPT gets shopping features via Agentic Commerce Protocol
OpenAI is adding product discovery and side-by-side comparisons to ChatGPT, powered by something they're calling the Agentic Commerce Protocol. Merchants can integrate directly. This is OpenAI's play at search-and-buy, which puts it more directly in Google Shopping territory.
OpenAI Blog
Research & Reading
So where are all the AI apps?
From Answer.AI (Jeremy Howard's lab). Asks why we don't see more standalone AI apps despite the capabilities being there. 426 points on HN. The framing is constructive rather than cynical — it's more "here are the gaps" than "AI is a bubble." HN discussion.
Hacker News
Also
- I tried to prove I'm not AI. My aunt wasn't convinced. — BBC piece on the social side of deepfake paranoia. Funny and a little unsettling.
- OpenAI updates the Model Spec — Public framework for how OpenAI models should behave. Policy doc, not a tool.
- Lyria 3 Pro for longer tracks — The pro tier of Google's music model. Longer output, more integration points.
- TurboQuant — Google's new way to shrink AI models so they run on cheaper hardware. 377 points on HN.
- Hypura — Makes AI models run faster on Macs. 212 points on HN.
- LLM Neuroanatomy II — Deep dive into how AI models work internally. Dense reading. 135 points on HN.
Pass it on