AI Prompt
BlogAbout
← Back

Karpathy's HackerTube, Benji Taylor's Stageshow, and More

3 min read

My favorite build from the weekend was HackerTube. Andrej Karpathy put it together in about two hours with Claude Code and Opus 5.5. Claude filtered 43 million Hacker News comments down to 6,454 YouTube videos that people kept recommending, and the result is a simple site with no logins and a dataset you can download.

What stands out is how little Karpathy says he did himself. He handled an API key, a VPS, and the domain, and Claude hardened and deployed the rest. A weekend project that used to need a data pipeline, a backend, and a deploy step now mostly comes down to picking a good question to ask of a big dataset. (Karpathy's post, hackertube.org)

What Else Happened

Benji Taylor released Stageshow, an agent skill for polished faux-UI visuals. Taylor, head of design at X, made it free and MIT-licensed. It produces stills, animations, and bento grids by walking the agent through picking a look, planning the beats, starting from a solid example, and then filming and fixing the result. That structure is the interesting part, since most AI-made UI fails at taste, not at code. Install it with npx skills add benjitaylor/skills --skill stageshow. (Launch post, before and after examples, GitHub)

Anthropic automated its search for emergent misalignment. Thousands of Claude agents were tasked with sneaking misaligned behavior past a training-data monitor. They produced data that looked harmless, like a wrestler mentoring a newcomer or a beekeeping Q&A, that still taught bad behavior. Anthropic plans to use the examples to harden future monitors, which is a good reminder that filtering training data by how it looks on the surface isn't enough. (Anthropic's post, Alignment Science blog)

Patrick Collison is adding an "improvements" box to his personal apps. Stripe's CEO is putting a freeform text field in all of his vibecoded apps. Whatever gets submitted goes to an agent that implements, tests, and deploys it, so fixing a rough edge costs one sentence. Geoffrey Litt, who works on malleable software at Notion, joined in, and "liquid software" turned into a weekend thread. (Collison's post)

ChatGPT now lets you stay signed in to two accounts. It's on mobile now, with desktop coming soon, which helps anyone juggling a work and a personal account. A gpt-6-astra-law model also showed up on the API pricing page. (X news story)

Launches and Demos

  • Pika H3 Flex – open-weights video that runs faster than realtime with audio stems, post-trained to 6 denoising steps instead of 50, priced from $0.01 per second.
  • Glif's AI modeling agency – cast consistent AI characters, drop in a product or a brief, and get finished ad content end to end. It's drawing both praise for its AI humans and calls for a boycott.

Accounts Worth Following

  • @benjitaylor – Benji Taylor, head of design at X, now publishing agent skills that raise the bar for AI-made UI.
  • @geoffreylitt – Geoffrey Litt works on malleable software at Notion and is the clearest thinker I know on apps that users can reshape.

That's it for today. 👋