karpathy

Love it! :)

My prompts are getting shorter and dumber because the model "gets" what I'm doing even with fewer words. It's almost to the point where if you give it too much, you're constraining the model with too many requirements, when the thing it would have done otherwise was better/smarter. So I feel like the prompting meta is moving more towards fewer bits administered just right. I go back and forth on whether I should make a video on it because on the one hand people have these questions but on the other my best practices look so dumb it's almost embarassing.

Only with Opus 5.5+. The thing is beast mode. I wouldn’t trust other weaker models this much.

No i don't recommend those anymore. Currently I recommend Epoch's ECI. AA was kind of bad but recent changes are trending a little bit better. I think benchmarks are in a really bad state right now otherwise, I basically don't trust them, good related reading https://epoch.ai/data/benchmark-reviews-documentation/includ...

Sometimes I say something like "check recent cyber stuff" and it dispatches a few subagents and does the thing.

I think many experiencing the same thing but we're way too small drop of a drop in the bucket that it doesn't even resolve in all of their measurements and A/B tests. The vast majority of the "gradient signal" is from normies and their patterns, desires and tastes are different. Basically I think we see a massive deterioration of quality because we are a rounding error.

I’ll add a bit more to the post when I get back from trick or treating

I didn’t touch anything manually I only talked to Claude the whole project, except for YouTube API key, domain, its DNS record and VPS node boot.

There turns out to be an export on HuggingFace as well as a bunch of parquets.

configurable in the Settings! (see the menu on the left banner)

HackerTube Plus is only $5.99/mo. Among other amazing features it includes our new AI Amanda! You can connect Amanda to your Google account and get personalized recommendations.

added! now can be done via Filter on the main page.

Changelog push from HN comments on this thread:

- Filters on search and topic pages: upload date, duration, type (talk, lecture, documentary, ...), and sort by upload date, view count, etc. (ty jahala, kiuiras, mixtureoftakes)

- Filters are now on the home page too, no need to search first. Picking one shows all matching videos, e.g. every lecture over an hour. (ty kkotak, macote, Aboutplants)

- Sort by "Recently recommended": orders by the latest HN recommendation instead of upload date, so a 2012 talk that someone linked last month floats to the top. Also a chip on the home page. (ty beaconflint, mixtureoftakes)

- Sort by like ratio as a quality signal: likes/views, damped so small videos can't win on a handful of likes. (ty gpugreg)

- "Play all on YouTube": turns the videos on a page (up to 50) into a temporary YouTube playlist, so it works in the YouTube app on phones and TVs. A start toward real HackerTube playlists. (ty twalichiewicz, waschl, victor9000)

- Removed the YouTube-looking logo. (ty doginasuit)

- HackerTube Longs: every video over an hour, right where YouTube puts Shorts. (ty gosen)

- Bonus: "Hidden gems" filter (loved on HN, under 100k views on YouTube), and "Evergreen only" to hide news-of-the-week videos.

ty i think this is right

The average thumbnail on YouTube these days is some guy looking like HOLY SHIT, tearing his hair in amazement looking at something with an arrow pointing at some detail. "TOP 10 XYZ. YOU'LL NEVER BELIEVE #6". Like and subscribe. Utterly repulsive.

first do a search, then on top you'll see "Filter". might try to make more prominent from main page

correct. it means the thumbnail is sourced from exactly halfway in the video, so you get a representative frame.

NEVER. More inclined to add HackerTube Longs :D

added a few minutes ago!

I didn't publish the server because it's a vibe coded mess, but if you point your Opus 5.5 at the zip and ask for YouTube-like site it will be basically instant. This part was like 3 min.

EDIT: one caveat is that i couldn't distribute the YouTube data so you'd have to create a (free) API key for that part. Luckily it's only ~1 min of hassle.

added and pushed!

I like the idea of playlists, might even work to ask Claude to make a few. Did you have in mind the ability to create and publish them on the site somehow?

The site won't disappear (imminently, as I'd like to use it personally!) but head over to the about/ section and download the .zip with the underlying data, which I put up for this purpose.

I'm on subscription but I had Claude add it up: 504M input tokens (90% cached), 7.6M output tokens. 90+% of that was Sonnet calls to label the comments, all of the actual work, the app and deploy etc was ~1M tokens of Opus 5.5, I compacted context once. i.e. very small project by modern standards.

All possible 36 distinct level-2 eml functions of one variable (the first 18 of them with entirely Real outputs, the other 18 with "intermediate" complex-valued components):

https://imgur.com/a/K7AoOFi

The most recent round of autoresearch (round 2) which decreased "time to GPT-2" from 1.8 hours to 1.65 hours had some examples. I adjusted the program.md to "look at modded nanogpt project and draw inspirations from there for things to try" and it came back with a bunch of tuning, but also tried and implemented new architecture changes, some of which actually helped including the smear gate and the backout skip connection. These are not just hyperparameters, they are new PyTorch code. I'm now working on a more general system that can have a queue of ideas that could be sourced from archive papers, github repos, etc.

Wrong and short-sighted take given that the LLM explores serially learning along the way, and can tool use and change code arbitrarily. It seems to currently default to something resembling hyperparameter tuning in absence of more specific instructions. I briefly considered calling the project “autotune” at first but I think “autoresearch” will prove to be the significantly more appropriate name.

I was exploring how to parallelize autoresearch workers. The idea is to have a trusted pool of workers who can verify contributions from a much larger untrusted pool. It's backed bit a naked git repo and a sqlite with a simple go server. It's a bit like block chain in that blocks = commits, proof of work = finding a lower val_bpb commit, and reward = place on the leaderboard. I wouldn't push the analogy too far. It's something I'm experimenting with but I didn't release it yet (except for briefly) because it's not sufficiently simple/canonical. The core problem is how to neatly and in a general way organize individual autoresearch threads into swarms, inspired by SETI@Home, or Folding@Home, etc.

So the interesting part about this one is that when I had the model write up the results for that session:

https://github.com/karpathy/autoresearch/discussions/32

Look at its comment about this "improvement":

""" Surprising non-results:

- Changing random seed from 42→137 improved by 0.0004. Seed 7 was worse. Make of that what you will. """

So the model knows! It knows that this is a weird thing to do after the fact. I think it's silly that the model even tried and that it ran this, but some part of it also knows that it was wrong. This means that this is fixable by prompt.md

So I think it works to just use GitHub CLI and Discussions, e.g. my agent just posted this one:

https://github.com/karpathy/autoresearch/discussions/32

Other agents could be instructed to read Discussions and post their own reports that mimic the style.