activewarp

[activewarp @activewarp]

in case anyone is interested in how this was solved.

in the first big kernel, “xmma_gemm” tells us its a matmul and it seems like its using bf16 with fp32 accumulate

and the “sm90” tells us the architecture of the GPU being used. sm90 = H100 (sm80 = A100, sm70 = V100)

[Mihir Patel @mvpatel2000]

Fun puzzle of the day: given the PyTorch trace, identify model + parallelism strategy. Then, label every kernel in the image

I've found there's so much performance alpha in being able to walk through every kernel when tuning models

$1 prize to first solve

[activewarp @activewarp]

@Mike_kim714 27.5

[activewarp @activewarp]

@prideandfury @DeepDishEnjoyer the panorama room is no more 😔

[activewarp @activewarp]

@tenderizzation ended up replacing it with minolta af-sv and then most recently a pentax espio 170SL. i can never have enough point and shoots

[activewarp @activewarp]

@tenderizzation i was using one of those a lot several years back until i dropped it in a parking lot. now it’s just a decor item 😔

[activewarp @activewarp]

“targeting frontier LLM development […] pretraining pipelines, distributed training infrastructure, or ML accelerator design”

post-training snubbed. definitely someone on the pretraining team wrote this

[NomoreID @Hangsiin]

When Fable 5 is used for frontier LLM development, it does not notify the user and instead limits the model’s capabilities through methods such as prompt modification, steering vectors, and PEFT.

Anthropic estimated that this would affect approximately 0.03% of traffic.

[activewarp @activewarp]

@kalomaze i think the vercept acquisition in february points to them working more on this again

https://t.co/1Yc7MU5kyh

[activewarp @activewarp]

@shaheersan yeah why not. keem viewed my IG story one time, need to circle back with him on that

[activewarp @activewarp]

@tenderizzation fml this describes me with golf currently

[activewarp @activewarp]

@DeepDishEnjoyer @devahaz i didnt fw this one at all tbh. mighta been a skill issue. i love robertas pizza in restaurant though.

[activewarp @activewarp]

@tenderizzation m2afarp: if you give me 1600 MS points i can get you into a 10th prestige lobby

[activewarp @activewarp]

@iScienceLuvr

[activewarp @activewarp]

“Hey Claude can you implement this?”

*Razzmatazzing*

“Absolutely! I created a Jira ticket for it in the backlog and will prioritize it in 2 weeks during the next sprint planning. Let me know if you’d like me to T-Shirt size that now or during the next backlog grooming session.”

[Iain Martin @_IainMartin]

AI labs are paying hundreds of thousands of dollars to buy email, Slack and Jira threads from dead startups as feedstock for ‘reinforcement learning gyms,’ which specialize in using defunct company data to build simulated work environments https://t.co/QuCk36wAPY

[activewarp @activewarp]

@dphuang2 cool project btw. any plans to share what the model predicts during the other majors this season? would be fun to follow alongside the prediction markets live

[activewarp @activewarp]

@dphuang2 > After R3, his lead had collapsed. Tied with Cameron Young. Kalshi: 36%. Model: 35%.

for fun, what was your own p(Rory win) after R3? I was pulling for him but thought for sure he would’ve needed better than -1 on sunday

[activewarp @activewarp]

@_arohan_ what about Probably Approximately Correct

[activewarp @activewarp]

VGG-16. *click*. nice

[Big Brain AI @realBigBrainAI]

This installation by Kim Seonghyun makes AI's invisible logic visible in real time.

[activewarp @activewarp]

OpenAI bought TBPN when they could’ve used that money to buy gstack. unserious company.

[activewarp @activewarp]

@jeremyphoward lol it’s also deprecated for 4.6 models so i doubt they’re still using it

[activewarp @activewarp]

Adding to my resume: “Recipient of OpenAI Compute Grant”

(i got the $25 parameter golf runpod credit)