rohan anil @_arohan_

At this point, every government should make envs and evals for what they want in their models so model devs can aim to hill climb that. Similar to various civics exam.

3 1 52 3K 6

View Details

@bitcoinmalayali @Felix_Josemon @KeralaPH At some point @anmol01gulati put a bunch of low resource lang tokens to gemini/gemma tokenizer. This 10-100x number of vocab items, particularly affected was malayalam.

1 0 3 100 0

View Details

Peter Steinberger 🦞 @steipete

14 hours ago

Google fired the guy that made the google workspace cli, because he made the google workspace cli. Lucky me, Google can't fire me. gogcli.sh

Justin Poehnelt @JPoehnelt

21 hours ago

Two months ago I was fired by Google for creating the Google Workspace CLI. It went viral, hit #1 on Hacker News, gained thousands of GitHub stars and many thousands of actual users in just a couple days. It was an incredible, confusing journey, from directors and leaders asking

489 735 11K 3.0M 4K

133 314 6K 945K 3K

View Details

rohan anil @_arohan_

14 hours ago

@JPoehnelt ❤️‍🩹

0 0 9 4K 0

View Details

rohan anil @_arohan_

19 hours ago

Holy moly guacamole 🥑 Nvfp4 brah

Bryce, the CUDA Colonel @blelbach

20 hours ago

nvfp4 in the #1 submission is clearly a reference for a tiny quantized NVFP4 model that does QR factorization!

2 1 71 18K 27

3 2 61 13K 23

View Details

rohan anil @_arohan_

22 hours ago

@DimitrisPapail This is why you look like looped models huh arxiv.org/abs/2606.00605

0 0 11 1K 6

View Details

levi @levidiamode

a day ago

171/365 of GPU Programming Finally got below 2ms on the GPU Mode QR challenge. Still quite far away from CUDA colonel @blelbach, so I'm excited to read everyones' writeups once the competition is over. I especially wonder how #1 is utilizing NVFP4 (or is the submission name a false flag?). Every lower precision attempt has not panned out for me so far unfortunately. Very curious where the top 3 will converge in a week from now... If you're still debating whether to participate, would highly recommend! You learn so much from just trying stuff and being in the discord. And @modal gives you $30 in monthly compute for free. Maybe we can get a bit more compute for the next challenge if we all ask @charles_irl nicely 😁😁

levi @levidiamode

2 days ago

170/365 of GPU Programming Just going through the motions of QR factorization and Gram Schmidt today while concepts around orthogonality, norms and projections. It's been a while since I studied Linear Algebra, so it's been fun reviewing some older concepts and learning new

5 1 75 10K 34

8 3 96 17K 51

View Details

rohan anil @_arohan_

a day ago

@Dorialexander Its not that much and particularly given many of them >128k tbh

1 0 4 734 0

View Details

rohan anil @_arohan_

a day ago

It will be exciting times when we start collaborating on standardization around every part of the training tech stack starting from the tokenizer. It’s not that clear why open models use slightly different tokenizers. That would be a good accelerant.

11 0 102 10K 4

View Details

rohan anil @_arohan_

a day ago

My personal expectations is there are a few algorithmic moves to give hardware + model codesign moats.

1 1 12 2K 0

View Details

Aravind Srinivas @AravSrinivas

2 days ago

GLM is the kind of model that revives serious interest in open source AI. It passes the blind test relative to the frontier models on the median production grade knowledge worker task. It’s affordable to serve. And is a sub trillion parameter model, meaning it has a lot of potential to go beyond matching the frontier at the median level of difficulty to also doing it for the long tail. Plenty to look forward to!

85 93 2K 104K 230

View Details

rohan anil @_arohan_

a day ago

Whats your favorite linear algebra operation? My colleague said SVD must be made by god himself because it’s a miracle in 2d.

24 2 190 21K 63

View Details

Karl Mehta @karlmehta

2 days ago

Ilya Sutskever says the AI scaling era is running out of road because pre-training is about to hit a finite-data wall. "At some point though, pre-training will run out of data. The data is very clearly finite." "Then, okay, what do you do next? Either you do some kind of a souped-up pre-training... or you're doing RL, or maybe something else." "In some sense, we are back to the age of research." His timeline is the part most AI investors and founders should hear: "From 2012 to 2020, it was the age of research. Now from 2020 to 2025, it was the age of scaling... But now the scale is so big." "Is the belief really that... if you had 100x more everything would be so different?" "I don't think that's true. So it's back to the age of research again, just with big computers."

107 96 928 197K 463

View Details

Shashank Joshi @shashj

2 days ago

An update. A US official tells me that Sen. Warner misunderstood the NSA director Gen. Rudd in this case. Rudd did use the 'hours, not weeks' wording, but the use of Mythos in this context was—as widely assumed—part of a red-teaming effort, i.e. testing the security of internal networks. The official also told me that the agency's red-teams no longer have access to Mythos, because their authority for accessing it was under Project Glasswing.

Shashank Joshi @shashj

3 days ago

This now widely circulated claim is based on a line I wrote last week (economist.com/briefing/2026/…). I accurately quoted Mark Warner, vice chair of the Senate intelligence committee, saying that the NSA chief had told him that Mythos “broke into almost all of our classified systems,