DeepMind closed AlphaFold, so @gcorso.bsky.social + Jeremy Wohlwend built Boltz. Now its the preferred solution for many scientists. They have succeeded with open weights because they focused on community. Now they're launching a new PBC. What do you use Boltz for? Links 👇

Tired of creating a new notebook and writing AnnData/pandas/numpy/??? just to peek at a H5AD file? Me too! So I created a DuckDB extension that can read AnnData files.

Working with a large set of AnnData files on S3 is tough if you don't have the metadata indexed, so I created a tool to help. github.com/honicky/annd... It uses partial downloads to dramatically speed up extracting the metadata without downloading the whole file. `pip install anndata-metadata`

GitHub - honicky/anndata-metadata: A Python library and CLI tool for extracting metadata from AnnData .h5ad files, both locally and on S3

A Python library and CLI tool for extracting metadata from AnnData .h5ad files, both locally and on S3 - honicky/anndata-metadata

github.com

I promised to explain how you can build your very own, custom GenePT embeddings. Here's another Lab Note: learning-exhaust.hashnode.dev/lab-notes-cu... Enjoy!

Gene Embeddings, LLMs, and a $6.83 Experiment That Might Matter

Better Gene Embeddings Through Prompt Engineering (Or, at Least, We Tried)

learning-exhaust.hashnode.dev

Bio data is amazing!!! Don't be intimidated! learning-exhaust.hashnode.dev/lab-notes-co...

GenePT vs scGPT - what I learned in a few days

In the spirit of “Learning in Public,” and “learning exhaust,” I’m going to start adding “Lab Notes” blog posts that chronicle little discoveries or failures that I have along the way to a larger goal...

learning-exhaust.hashnode.dev

Data products often scale cost like hardware, iterate like software, and scale performance like... data products. learning-exhaust.hashnode.dev/data-product...

Data Products are Different

Why you have to manage data products differently than software and hardware.

learning-exhaust.hashnode.dev

Oh man, writing up the Cerebras paper on Weight Streaming was a rabbit hole filled with parallel algorithms. learning-exhaust.hashnode.dev/one-thing-i-... @picocreator.bsky.social @eugeneyan.com @swyx.io @latentspacepod.bsky.social

Weight streaming might work well on GPUs!

I wonder why it isn't a thing...

learning-exhaust.hashnode.dev

First post to bsky: I've been trying to post a bit more frequently and in smaller bites, so I'm going to try to pick one interesting thing I've learned from papers I read, and then write a quick post about them. Here's my first one: learning-exhaust.hashnode.dev/one-thing-i-...

Embeddings are task specific

Modern training techniques for embedding models mean that we should probably include a prompt or fine tune to a specific type of task.

learning-exhaust.hashnode.dev