jason @jvmncs

❤️s/RTs are randomized and differentially private. github.com/jvmncs Brooklyn Joined February 2010

Tweets

3K
Followers

993
Following

1K
Likes

9K

Charles 🎉 Frye @charles_irl

a day ago

We reverse-engineered Flash Attention 4.

18 64 651 90K 409

Download Image

over the past 4 days, I read the really great LLM scaling book from google DeepMind/JAX and took some (very abridged) notes it's a long book, but I think this provides an okay tl;dr of the plot at some level of technical background ekzhang.substack.com/p/abridged-not…

4 22 268 18K 314

Download Image

jason @jvmncs

2 months ago

anecdotally, gpt-5 has been a huge improvement over Sonnet for me in @AmpCode for a mostly autonomous Rust rewrite of a large-ish Java project, gpt-5 is churning through a long backlog of half-assed tasks that Sonnet tried to claim were done. at half-price, bc no oracle needed

0 0 3 167 1

Luis Capelo @luiscape

2 months ago

We just launched GPU memory snapshotting on @modal_labs in alpha. Speed up cold boots by up to 12x 😇 If you're deploying AI models, a huge amount of cold boot time comes from loading model weights into GPU memory. This makes it difficult to scale GPU resources up and down…

14 43 341 76K 136

Download Image

jason @jvmncs

2 months ago

Society if Claude would stop estimating 18 month timelines when we both know he's about to write it in an afternoon

0 0 1 101 0

Download Image

jason @jvmncs

2 months ago

one of the better parts of AI image generation is seeing stuff like this in technical writing. huge upgrade from matplotlib!

0 0 0 80 0

Download Image

jason @jvmncs

2 months ago

why hasn't anyone pretrained a deepseek-style MoE at FP8 optimized for 8xH100 inference?

0 0 2 137 0

jason @jvmncs

3 months ago

wondering why Gödel was underhyped during his life, and then you read something like this in every article about his non-incompleteness work

1 0 1 183 0

Download Image

Micah Goldblum @micahgoldblum

3 months ago

🚨 Did you know that small-batch vanilla SGD without momentum (i.e. the first optimizer you learn about in intro ML) is virtually as fast as AdamW for LLM pretraining on a per-FLOP basis? 📜 1/n

28 114 834 391K 647

Download Image

jason @jvmncs

3 months ago

MCP is supposed to be more than just tool calling, yet most MCP servers AND clients don’t implement the rest of the protocol. this seems like a problem

1 0 2 188 0

thebes @voooooogel

3 months ago

6 12 184 12K 10

Download Image

jason @jvmncs

3 months ago

achievement unlocked

0 0 0 125 1

Download Image

Charles 🎉 Frye @charles_irl

4 months ago

Dozens of teams have asked my advice on running LLMs. How fast is @deepseek_ai V3 with vLLM on 8 GPUs? What's the max throughput of @Alibaba_Qwen 2.5 Coder with SGLang on one H100? Running & sharing benchmarks ad hoc was too slow So we built a tiny app, the LLM Engine Advisor

27 81 594 85K 587

Download Video

Modal @modal

4 months ago

B200s and H200s, now available on Modal 🥩 (brought to you by Colin Weld, Member of Technical Staff @ Modal!)

4 4 121 13K 13

Download Image

jason @jvmncs

4 months ago

good evening to everyone except the person who taught Gemini 2.5 Pro to litter its code with slop comments

0 0 1 73 0

Pablo Samuel Castro @pcastr

4 months ago

Mind the GAP! we've had a few works proposing techniques for enabling scaling in deep rl, such as MoEs, tokenization, & sparse training. @g_sokar and i looked further & found a bit more clarity into *what* enables scaling, leading us to simpler solutions (see GAP in figure)! 1/

6 18 122 11K 104

Download Image

Sara Hooker @sarahookr

5 months ago

It is critical for scientific integrity that we trust our measure of progress. The @lmarena_ai has become the go-to evaluation for AI progress. Our release today demonstrates the difficulty in maintaining fair evaluations on @lmarena_ai, despite best intentions.