AI Valley
Posts
Karpathy built his own YouTube in 2 hours

Karpathy built his own YouTube in 2 hours

PLUS: An AI just did brain surgery on another AI

Barsee
03 Nov

_{Sign up}_|_{Follow us on X}_|_Sponsor

Together with

howdy, it’s Barsee again.

happy tuesday, AI family, and welcome back to AI Valley.

here are the biggest things worth knowing today:

Karpathy got annoyed with YouTube and built his own in two hours
GPT-6 Astra just performed brain surgery on another AI
Audio recordings might be giving away more than you think
Plus trending AI tools, posts, and resources

Let’s dive into the Valley of AI…

VIKTOR

If you can brief a new hire, you can manage me

Courtesy: Viktor

I'm Viktor, the AI employee 70,000+ teams have hired, and managing me takes the skills you already use with people.

You give me a brief the way you'd give one to a new hire: "Pull this month's feedback from Intercom, #feedback and support email, group it by issue and draft fixes for the top three." I work inside the 3,200+ tools you connect, including ones with no API, and hand back the working app.

You correct me once and I remember it. When a task repeats, you make it one of my duties and I run it on a schedule. Anyone on the team can brief me, so the context sales gives me is there when finance asks. Anything I can't undo waits for your approval, the way a new hire checks before anything big.

Write my first brief today.

Start free with up to $100 in credits, no credit card.

_{*This is sponsored}

THROUGH THE VALLEY

Courtesy: Karpathy

1/ Karpathy got annoyed with YouTube and built his own in two hours. He had Claude Opus 5.5 dig through 43 million Hacker News comments from the last 20 years and pull out the 6,454 videos people kept recommending. The result is HackerTube. No recommendation algorithm, no clickbait thumbnails, no login.

I opened it expecting a weekend toy and ended up with a stack of tabs I’ll never close. It’s mostly the conference talks, lectures and deep dives HN has been passing around for years, the stuff YouTube would never show you because it doesn’t keep you scrolling. The thumbnails are actual frames from the videos, which sounds minor until you realize how much of YouTube is people making shocked faces at you.

The build is the bit that gets me. Karpathy did the boring admin (API key, server, domain) and Claude Code did the rest: the code, the deployment, even locking down the server. A few years ago this would’ve been a proper side project with a half-finished GitHub repo to show for it. Now you can get fed up with an algorithm over breakfast and have your own version up by lunch.

People are already copying the idea. Someone made Hacker Papers for arXiv papers, someone else did books. I’m half tempted to build one for podcasts, because my podcast app’s idea of “for you” is the same five shows I already subscribe to.

I think this is where things are heading. We spent fifteen years letting a handful of algorithms decide what we see. Turns out the fix might just be building your own feed, from people whose taste you trust.

2/ GPT-6 Astra just performed brain surgery on another AI. MIT researchers asked it to fix the math skills of a small model, Qwen3.5-0.8B, which kept getting stuck in reasoning loops. Astra looked at the model’s weights, worked out what was going wrong and edited 0.016% of them. Its MATH-500 score went from 7.9% to 48.1%, and its instruction-following nearly tripled. No retraining at all.

My first reaction was a slightly uneasy “oh.” Then I read Sean Goedecke’s breakdown, which brings things back down to earth. Astra didn’t teach Qwen any maths; it found the weights behind the loops and turned them down. A good ML researcher could have done the same.

But that’s sort of the point. “A good ML researcher could have done it” used to be the bottleneck. There aren’t many of those people, and they’re expensive. If AI can do that kind of debugging on demand, fixing a broken model starts to look less like a research project and more like running a diagnostic. It’s tiny and early, and it’s on a 0.8B model, not something frontier-sized. The question I’ll be watching is whether it works on models big enough to matter.

MY TOP FINDINGS OF THE DAY

Audio recordings might be giving away more than you think. One of the wildest demos I've seen lately shows how AI can use background sounds to narrow down where a recording was made. Not GPS data, not anything deliberately shared. Just clues hiding in the audio.

— (@)

Think about all the voice notes, podcasts and videos people have uploaded from their homes over the years. Some might contain enough environmental clues to reveal a location nobody intended to share.

This is the original video from the guy.

The original demo was unsettling enough. Then someone reportedly recreated the approach using Claude Code, which makes the whole thing more interesting. The barrier to building tools like this is getting lower.

— (@)

TRENDING TOOLS

Beacon - a self-improving memory for coding agents
Sage2 - software can now understand human language and act on it, in real time
Exa Monitors V2 - get alerts across the largest index of web + proprietary data. 130B+ web pages, 211M businesses, 1.3B products, 1B people
Treg - one skill gives your agent 1B+ contacts across Apollo, Hunter, Tomba, People Data Labs. 85% cheaper than Clay
Hume Audio Decision API - it listens to raw audio and answers questions you write, like whether a caller sounds frustrated
Rerun - fully open-source alternative to Grok Bot, Dots and Duet
Drex for Finance - the multimodal decision model, post-trained for sub-second financial decisions