Sameer Soi

@pitdesi @Jason People forget that this was normalized (I heard this kind of casually racist banter in the open in western PA in the 90s and aughts). We might look back and see the 2010s were a blip.

@Jason My replies x.com/pitdesi/status…

my DMs

Non-Indian homie couldn’t remember the name of a jalebi and called it the “Indian honey pretzel” Astounding, no limits to creativity

@FanBoyShi @vsbuffalo Definitely heard about its use at 10X

I'm curious — is Rust actively used by any biotechs or genomics companies? If so, in what areas?

@thotmanifestor @InjeelJiday @SashaGusevPosts And for a more thorough edification on the matter: ncbi.nlm.nih.gov/pmc/articles/P…

@thotmanifestor @InjeelJiday @SashaGusevPosts Now tell me what % of variation is explained by those PCs. Not to mention sampling can produce patterns in PCA that can easily be over interpreted.

The Li + Durbin 2011 paper showing that you can infer an entire population history from just a single genome is still one of the most mind-blowing results in genetics I have ever seen. nature.com/articles/natur…

@Katie_Krause @CA_EDD @GovPressOffice @latimes No, never. I was able to figure out my issue by consulting my company’s HR provider as well as Internet forums. I never got through to a human even once.

I phone @CA_EDD all days of the week x hours of operation to ask about PFL. Each time the call queue is full. Instead of arranging a callback as one would expect in 2024, the call is unceremoniously dropped. Each time. @GovPressOffice is this the best CA can do for new parents?

@mnp_47 @vsbuffalo But you like the languages you like and those tastes change too! I used to love R the switched to Scala and then python. Now I look at R code and get a mild concussion.

@mnp_47 @vsbuffalo This functional style is employed by the popular dplyr syntax in R, which many people find readable. Ofc it wasn’t invented there (see OCaML, Scala, F# etc). The resulting code reads more like a narrative than typical procedural code.

Rust is such a beautiful language for working with genomic data. Clean syntax, functional programming, and extremely fast.

@BauerKahan would appreciate knowing your thoughts on how new parents should be supported by @CA_EDD

@sanjaypartly You stink. Am I doing it right?

@Codie_Sanchez (Assuming they died due to aging and not randomly)

@Codie_Sanchez The exact same plot could be generated if people in the trades literally died (or retired) and no one replaced them. You would need to see if the number of people in trades were increasing at the same time.

The trades aren't dying. They're getting younger.

@joe_pickrell A tardy April’s Fools’ Day?

"The genetic architecture of dog ownership: large-scale genome-wide association study in 97,552 European-ancestry individuals" (no common variants associated with the phenotype of 'Do you own a dog?') academic.oup.com/g3journal/adva…

@dhh @NZXT @AMD @PyTorch @lightning @Neovim @codeiumdev @humanscale Eco Sit, @audioengineusa P4 speakers, @hifimanofficial Sundara headphones and @HermanMiller chair are crucial to the full experience as well!

@dhh Custom built desktop @NZXT case @AMD Ryzen 12-cores, 96GB RAM running Ubuntu; python + @PyTorch + @lightning all edited through @Neovim with some help from @codeiumdev. Pure joy.

The best computer, the best programming language, the best editor is the one you can't wait to work with in the morning.

@dan_biderman @cherrvak @brianltrippe Fair. A lot of benchmarking in bio-ML is about out-of-domain generalization. Often we are interested in how predictions about proteins we don't know much about will fare. This could align with your result that LoRA is a good regularizer that helps with generalization.

@sameersoi @cherrvak @brianltrippe Will read. “Outperforming” is a rather strong statement but it made me curious

People think LoRA is a magic bullet for LLMs. Is it? Does it deliver the same quality as full finetuning but on consumer GPUs? Though LoRA has the advantage of a lower memory footprint, we find that it often substantially underperforms full finetuning. However, it forgets less of the base model’s capabilities. In this work, we exhaustively explore this trade-off and provide practitioners a clear view of the difference between the methods. arxiv.org/abs/2405.09673

@dan_biderman @cherrvak @brianltrippe Right from the abstract "On the PPI prediction task, we surprisingly find that PEFT models actually outperform traditional fine-tuning while using two orders of magnitude fewer parameters."

@dan_biderman @cherrvak @brianltrippe Here's a good ref: biorxiv.org/content/10.110… Similar in spirit to how we are doing it in our application area.

@dan_biderman @cherrvak @brianltrippe (I can't answer that but I use LoRA for finetuning antibody PLM's on activity signals)

@sameersoi @cherrvak Do people use Lora in protein modeling with diffusion models? @brianltrippe