Just like everyone else, I'm absolutely fucking thrilled to see Artificial Intelligence showing up all over the place. A lot of commercial websites now have slow new AI functions that are essentially "ctrl+f" in your web browser - the kind of improvement is assume Elon Musk and his fans are most excited about.
now leasing office space for HHS bioinformaticians to log into AWS to do their work...cause that's a cost savings over them logging into AWS from spaces that cost the HHS $0! .... Though my favorite thing this week is the stated attempt to put AI into all government databases.. ..which...run...on.. .COBOL.... if you aren't familiar, when I took Fortran 77 in highschool (the 77 was that version of the language was finalized) my teacher would make fun of COBOL being old and useless. Totally going to seamlessly integrate with AI. No worries there. 
If you're thinking something like "oh no, did peer reviewers see the words 'large language models and multi-omics' in the abstract and just accept it without ever looking at it?" you're probably a cynical jerk. Geez.... I mean, that's exactly what I thought, and that's why I spent my morning trying to see if this LLM could actually do things (other than generate "Internal Server" errors, which, to be fair, it generates an awful lot of)
While I'm ranting you can absolutely buy this at BestBuy right now. Just in case your "smart washer" from 2018 that is the single worst appliance you've ever owned in your entire life needs downgraded for 5x what a non-smart washer/drier that works costs!
With that lead in!
If you're thinking something like "oh no, did peer reviewers see the words 'large language models and multi-omics' in the abstract and just accept it without ever looking at it?" you're probably a cynical jerk. Geez.... I mean, that's exactly what I thought, and that's why I spent my morning trying to see if this LLM could actually do things (other than generate "Internal Server" errors, which, to be fair, it generates an awful lot of)
Short summary? There is some value here, I think. Particularly if you're not a big fan of reading, but you're patient enough to ask an algorithm to dig through some results for you, and you aren't patient enough to dig through said results yourself.
For a bad example - I went to a pancreatic cancer dataset and I asked Dr. BioRight how often KRAS mutations show up in the cohort. Since it's the most mutated gene in that dataset, I figured it would skim the abstract and give me the answer. It didn't. ChatGPT400 or whatever it's called now will do that for you, that's not why you want to use this tool.
I got frustrated, saved this blog post in my "do not push the publish button, Ben" folder and moved on. I went back because it popped up in my newsfeed and then I grabbed a random dataset and started asking it stuff about the source data in the study.
And this is where things get really interesting. In this study with a 65 patient cohort I just asked what the most commonly mutated genes were. It appears to generate R commands and run them?
That's actually super legit! Slow....but it really sounds impressive. Now - I'm not checking to see if this is actually right, but you can see where this might be an asset.
You don't go to an AI to make art for you because you know how to make art. You don't ask ChatGPT to write something for you if you want it written well and accurately. And you don't want to ask Dr. BioRight 2.0 about a cancer study that you reanalyzed yourself and contributed to a paper on. But if you are interested in skimming publicly available data for stuff you want to dig into depth on, it is faster than pulling the manuscript source data, figuring out which table is the one you want, and looking for it yourself. And that can probably be of use at times.
No comments:
Post a Comment