anshuman @athleticKoder

maximizing shareholder value; mostly ml here; heyyanshuman.com Joined February 2020

Tweets

10K
Followers

20K
Following

979
Likes

9K

anshuman @athleticKoder

14 hours ago

@sundarpichai a technical report plis plis plis, on pretraining and post training techniques.

0 0 0 1K 0

View Details

anshuman @athleticKoder

14 hours ago

@oneill_c good read x.com/athleticKoder/…

pretty much what MLA does. but that's an architectural choice to be taken before pre-training. borrowed from MLA, but applied to any model out there to compress the effective KV Cache. I wonder if there is an open recipe about it. and vLLM should figure out a way to support it.

1 0 5 2K 3

1 0 1 1K 2

View Details

anshuman @athleticKoder

14 hours ago

Charlie O'Neill @oneill_c

17 hours ago

1/ You can shrink a language model's KV cache by 200×, in a single forward pass, and it still answers correctly. At 256k context that's 36 GiB of cache down to ~360 MiB, with no change to the base model. Here's how we did it 👇

18 71 727 63K 679

1 0 5 2K 3

View Details

anshuman @athleticKoder

6 days ago

@osanseviero Gemma4 is genuinely very very strong base model. been working on something huge at work and can't stop appreciating how good pretraining of gemma4 is.

0 0 3 1K 2

View Details

anshuman @athleticKoder

2 weeks ago

@MilksandMatcha @tyler_fong_ Congratsss Sarah!

0 0 1 1K 0

View Details

Anindyadeep @anindyadeeps

2 weeks ago

We have released the biggest protein data collection on Hugging Face, guys! We have been working on this for more than 3 weeks now, starting from curating the raw data, doing a lot of filtering, splitting the datasets, sharding them, and doing a lot of analysis. Everything is summarized in our recent blog post.

LiteFold @try_litefold

2 weeks ago

LLMs got FineWeb, The Pile, RedPajama, Dolma. Protein ML got per-paper supplementary tables and FTP mirrors scattered across a dozen institutions. Today we're releasing AminoWeb on @huggingface : 29 cleaned, ML-ready protein datasets, ~7.5 TB total. Sequence, structure,

11 32 143 46K 84

38 72 360 87K 201

View Details

anshuman @athleticKoder

2 weeks ago

@1littlecoder @nitprashant @emergentlabs hehe thanks for the tag. but not looking for a carrier pivot from core ML to growth hacking.

0 0 0 40 0

View Details

anshuman @athleticKoder

3 weeks ago

@lindensli this is surprisingly simple

0 0 0 411 0

View Details

anshuman @athleticKoder

3 weeks ago

@dexhorthy @darin_gordon you might be smart enough if i just post raw ideas. but not everyone on this app/planet is smart enough :D

2 0 1 55 0

View Details

anshuman @athleticKoder

3 weeks ago

@dexhorthy @thdxr yup, these model feel so good at times and then slop at times. if the taste in writing is so bad, imagine code quality in a vibe coder's codebase. one model to rule them (tasks) all is essentially a flawed approach if we talk about taste

0 0 0 27 0

View Details

anshuman @athleticKoder

3 weeks ago

@thdxr @dexhorthy :/

anshuman @athleticKoder

3 weeks ago

@paulabartabajo_ @dexhorthy i do agree with you there. i just have a lot of running thoughts and so much information to share but time has always been the bottleneck. gpt 5.5 has bad taste for writing. especially those recurring templates and style that pisses reader off at times. funnily enough when i

0 0 0 130 0

0 0 0 65 0

View Details

anshuman @athleticKoder

3 weeks ago

i do agree with you there. i just have a lot of running thoughts and so much information to share but time has always been the bottleneck. gpt 5.5 has bad taste for writing. especially those recurring templates and style that pisses reader off at times. funnily enough when i first asked gpt 5.5 to put together a draft it created five thousand single lines as article. it was so so bad.