sankalp

45.9K posts

sankalp banner
sankalp

sankalp

@dejavucoder

ai and side quests. i post well, just follow. i consult on ai engineering stuff | seeking post-training, auto-research adjacent, evals related work atm

bangalore, india Katılım Ekim 2021
682 Takip Edilen25.9K Takipçiler
Sabitlenmiş Tweet
sankalp
sankalp@dejavucoder·
my latest blog post "auto-research with codex: how I achieved a 212x faster kernel over baseline with codex in GPU Mode's qr_v2 problem" is up now. in this post, i talk about my approach towards auto-kerneling on the QR decomposition problem. sankalp.bearblog.dev/autoresearch/
sankalp tweet media
English
14
29
406
46.1K
sankalp retweetledi
Michael Lutz
Michael Lutz@Michael_J_Lutz·
I recently placed 5th in GPU MODE's qr_v2 competition with a batched QR decomposition kernel 280x faster than torch.geqrf on a B200. My autoresearch swarm ran ~2k experiments over two weeks. Read the full writeup to learn about the Householder QR algorithm, kernel optimization tricks, and a glimpse of agentic kernel research in 2026. ml-mike.com/writing/qr_v2/
Michael Lutz tweet media
English
3
13
50
2.6K
Jeffrey Emanuel
Jeffrey Emanuel@doodlestein·
I swear, these models love whimsy. In the middle of doing this totally pragmatic system maintenance work for me across my fleet of machines, Opus randomly referred to two of my machines, named ThinkStation01 and ThinkStation02 (ts1 and ts2), as "the twins":
Jeffrey Emanuel tweet media
English
3
2
44
3.6K
sankalp
sankalp@dejavucoder·
nyc divergent thinking sf convergent thinking
English
0
0
3
674
Eiso Kant
Eiso Kant@eisokant·
[request for papers] What's the most interesting research in multi-agent/multiplayer RL? Where it's not just hierarchical orchestrator <> subagents but training for actual collaboration?
English
13
3
45
10.3K
sankalp
sankalp@dejavucoder·
@novasarc01 i yearn for another first elden ring playthrough. played dark souls 3 last month though.
English
1
0
3
87
sankalp retweetledi
prasanna 👾
prasanna 👾@prasannakotyal·
@dejavucoder If you look at datacenter GPUs, data movement happens at 3 main levels: 1. On-GPU memory bandwidth HBM bandwidth tells how quickly the GPU's memory can supply data from HBM via mem controllers/cache to the GPU's compute units. This is per GPU and is generally fast.
English
1
1
1
1.2K
Trash Panda 🦝
Trash Panda 🦝@trashpandaemoji·
@dejavucoder With codex desktop I use a heartbeat schedule/automation to keep the thread running even if it stops.
English
3
0
55
5.6K
sankalp
sankalp@dejavucoder·
let's say you are running a /goal and it pauses because your 5 hour limit runs out, it would be great if it could resume automatically when the session resets.
English
70
20
1.3K
89.1K
sankalp
sankalp@dejavucoder·
when you will meet people who do distributed training, you will notice they are very humble. such is the nature of distributed training.
English
6
2
75
3.6K
sankalp
sankalp@dejavucoder·
you can just use claude to filter and skim through to see if it looks interesting. extension of that would be to build a log of file where it tracks your preferences. can extend furether to build karpathy style knowledge base and then use the agent to match against recently read articles etc.
English
0
0
0
47
Shobhit Shrivastava
Shobhit Shrivastava@shri_shobhit·
@dejavucoder I want to train an AI to look at a content and tell me whether I should it read it or not :)
English
1
0
2
166