gravitas_deficiency comments on 'It's Literally the Gulag': Furious Meta Employees Speak Out on 'Soul-Crushing' AI Jobs

[–] gravitas_deficiency@sh.itjust.works 5 points 5 months ago* (6 children)

Oncological genomic sequencing cross-correlated with clinical patient outcome data to improve cancer treatment targeting based on real world patient outcome data. Personally, I work on data pipeline stuff. The mission is fantastic; the management and C-suites are a massive pain in my ass. But when are they not?

permalink

fedilink

source

parent

hideshow 6 child comments

[–] flying_sheep@lemmy.ml 3 points 5 months ago (5 children)

I hope pipeline stuff is better than when I started. I (not so) accidentally played a big part in revolutionizing my field by building anndata and scanpy. Then I tried doing it again by writing sequencing stuff in Rust, but was too early / too bad, but now that's happening too!

permalink

fedilink

source

parent

hideshow 5 child comments

[–] gravitas_deficiency@sh.itjust.works 3 points 5 months ago* (4 children)

I’m personally playing with the idea of putting together a POC to transition our “foundation” data warehouse from PSQL to a graphDB, because the extensibility and maintainability of our current system is fucking awful. Like, some upstream entity gets a version bump and there’s like 5 systems we have to go through and add columns to various tables (usually 2-4 such modifications in each workflow - and it doesn’t help how my manager fucking loves to slice work up so much that it becomes a massive pain to integrate, instead of telling one Eng “integrate this new data element” and have it done, soup to nuts, in a week or two across the whole ecosystem) and occasionally fuck around with joins and so on every single time there’s a new piece of data we want to integrate. And we have no capability to scan back historically and evaluate our holistic state at some particular time index, which can be really helpful for some applications.

Anyways, I’m fucking swamped at work so haven’t touched that at all, but I’ve wanted to explore that idea for well over a year and a half at this point.

permalink

fedilink

source

parent

hideshow 4 child comments

[–] flying_sheep@lemmy.ml 3 points 5 months ago (3 children)

I have little experience with graphdb, but a lot of experience with the pain you're describing. Maintaining schemas is a pain, maybe if you don't need the performance, you can go that route!

permalink

fedilink

source

parent

hideshow 3 child comments

[–] gravitas_deficiency@sh.itjust.works 3 points 5 months ago (2 children)

The thing that interests me about it is that it will be a lot more trivially interrogable by ML stuff (bespoke ML specifically, not LLM), which could glean an absolute shitload of interesting insights for us.

I am an enormous fucking Luddite for a whole swath of reasons when it comes to LLMs, but ML outside of that context can be immensity powerful when employed correctly.

permalink

fedilink

source

parent

hideshow 2 child comments

[–] flying_sheep@lemmy.ml 2 points 5 months ago (1 child)

For sure, my lab has been doing that for a long time.

How is graphdb more ML-friendly?

permalink

fedilink

source

parent

hideshow 1 child comment

[–] gravitas_deficiency@sh.itjust.works 1 point 5 months ago

If you’re doing PSQL (or any typical relational DB flavor), there’s a lot more complexity in terms of understanding the shape of the data, what joins to what, how to optimize queries, etc. Graph DBs are gonna be easier for a model to explore, since they can just do stuff like “I want to see tests with samples that have reactivity to mutation ABC on chromosome 14 over a threshold of X”, which is a lot easier for an ML agent (or less experienced developer, or even a molecular biologist with limited CS/DB experience) to just intuitively evaluate correctly using the syntax of GraphQL than it would be trying to do a shitload of joins between 6 or 7 tables in PSQL.

permalink

fedilink

source

parent