@difficultyang

More social alt of @ezyang

Joined April 2022
  • Tweets3.8K
  • Following95
  • Followers4.3K
443Media
With our new magical vibe coding skills I've been trying to get my 5yo to practice piano and mostly it is humbling me on how hard effective pedagogy design is
1
15
710
I have a new, easy to evaluate and realistic private eval and it is very much not saturated (for example, Opus 5.5 fails it)
7
848
Hyperstition
@hyperstition_cc
9h
It's official : We trained NanoGPT in 39.9 seconds. By percentage reduction in training time, this is a greater improvement than the previous 45 world records combined. The merged Github PR: github.com/KellerJordan/modd… Here's how we do it : hyperstition.cc/training-nan… Credit @DevenPzak who did this alone while simultaneously playing baseball for the polish national squad, taking them to the European Finals.

New Record: 0.665 minutes (39.9 seconds): ANVIL2, Sampled-softmax, New Embedding Table, Full-stack...

Created by: Deven Pietrzak (MIT Math. X: DevenPzak. Lab: Hyperstition (fka Social Physics Lab)) After a few weeks since my last record without touching the PR, I finally got access to some more H10...

github.com
1
2
168
14,529
this is pretty neat
1
10
1,192
source: deepseek-v3.ezyang.com/studi… (it's slop, probably not a good place to put it in the series but I prepped it for funsies)
4
185
I guess the reason after you paste your have to ENTER twice to submit must be related to how claude code ensures you can reliably paste into the terminal without triggering a submission
8
599
I like emdashes and the AIs can't take it away from me
1
7
495
The tiresome state of AI text generation means that sometimes it is optimal to have the AI draft slop and then rewrite by hand in the human voice
2
17
2,191
The discourse was split on this a year ago but I think a lot of classical procedural generation could in fact be materially improved by some LLM intervention
3
1
80
67,242
My overall Opus 5.5 vibe is that it is nice but the hard stuff did not get easier
1
12
1,214
My optician whose booking website is broken did not want to take a call from my AI assistant SAD
1
7
811
It feels like a stable norm is to have your AI battle against the bigco AI or first like support until the humans show up
2
158
My 2yo made me play this four times on repeat this morning lol. (Jellyfish Jones by Caspar Babypants, music video by Opus 5.5 with minor art direction by me, art style based off of Kate Endle's album covers).
1
9
707
Roughly four prompts. The entire video with edits needed 740k tokens, so it fit in a single 5hr session from the $20 subscription--so basically generating these is insanely token efficient, I was impressed.
164
This is a pretty cool graph
1
20
1,975
Sol 6 generates ugly code, kind of similar to Astra
1
19
1,046
I did end up having to do some iterations with Opus lol
1
5
461
bonus points if you can guess what song it is
1
182
The character sheet is, in fact, very very important lmao
NotinReality@other__reality
Sep 22
Replying to @other__reality @furriesforgood
So not perfectly 1 shot but I did not put any instructions for what to animate or how beyond using the clawd character, and if I put more than 2 sentences into the original prompt it easily could have been 1 shot.
2
514
The music video isn't bad. But Opus 5.5 doesn't have complete and utter understanding of animation such that an amateur like me wouldn't need to do some corrective prompts to fix some problems
difficultyang@difficultyang
Sep 24
Generating my own opus 5.5 music video wish me luck
1
6
1,163
And for music timings Opus still relies on looking at spectrographs and so there are some simple things that they get wrong
195
Load more