in a fast moving field there are always caveats. not everyone is hooked up to AI twitter
English
samar plok
5.4K posts
samar plok
@greebletheory
the bad news is we’re the cavalry
St Louis, MO Katılım Aralık 2023
1.6K Takip Edilen183 Takipçiler
dont really want to get into the weeds but critiquing applied LLM papers with “this doesnt use the latest model” is a thought terminating cliche and im tired of it
English
samar plok retweetledi
Men aged 16-29 should be close reading as many lyric poems as possible
English
lol is it bad that i legitimately thought to try the human expert approach as an undergrad
cant really be experts if they are like me lol
Jake P. Taylor-King@wildtypehuman
@parmita Oh no, the model outperforms a bachelors student who tries the first thing that comes to mind
English
@alexanderfuxi if anything, having a too involved workflow here might bring up too many possibly ambiguous issues. bedtools overlaps on known genes is probably enough for a small task
English
NO, THIS ARE NOT QUESTIONS THAT KEEP PHD UP AT NIGHT.
It's a good demonstration/eval that demonstrate LLMs can synthesize expertise from different domains and write code to efficiently solve problems, though.
Give a cell biologist an IGV and they will be perfectly fine at this
Chubby♨️@kimmonismus
Anthropic just dropped a benchmark that should make every scientist pay attention. BioMysteryBench puts AI models through 99 real bioinformatics challenges, using raw, messy datasets from actual research, think unprocessed DNA sequences and clinical samples. However: these aren't textbook problems with neat answers. They're the kind of open-ended puzzles that keep PhD students up at night. The results are exciting. Claude's latest models (4.7) solve the majority of tasks that trained human experts can handle, and on 23 problems that a panel of five domain experts couldn't crack, Claude Mythos Preview nailed 30% of them. How? By combining knowledge from hundreds of thousands of papers and layering multiple analytical strategies when uncertain, essentially doing what a room full of specialists would do, but faster and in a single run. Genentech and Roche independently confirmed this trajectory with their own CompBioBench, where Claude Opus 4.6 reached 81% overall accuracy and 69% on the hardest questions. Two separate benchmarks, same conclusion: AI is no longer just keeping pace with biologists, it's pulling ahead on some of the hardest problems.
English
samar plok retweetledi
Can my parents stop getting older while I figure things out for the next 5 years it’s actually so unfair
English
samar plok retweetledi
guy who beats the "you're a leftist but you send your kid to a private school" rap by moving to an affluent suburb and sending their kids to their local public school
English
samar plok retweetledi
Young men from ages 16-29 should be reading as many tweets about Hemingway, Steinbeck, Fitzgerald, Kerouac, and Whitman as possible
English
samar plok retweetledi
samar plok retweetledi
samar plok retweetledi
@AlejoFraticelli would be much easier to have the llm write a synthetic read simulator and then generate fake results that way
English
Would an LLM be able to generate fake FASTQs from a designed effect (say let’s start with bulk RNAseq to make it easier - and let’s say I want to show that macrophages are less inflammatory upon imaginary treatment X)? Could it make the FASTQ indistinguishable from real sequencer data (except for the sequencer ID)? Could this be happening already right now?
Any thoughts? @lpachter
English
@InVitroFuture the problem as always is that homeschooling depends on the home, and some homes are just quite evil
English
it's pretty annoying when people who hate homeschoolers like me (raised by progressives with postsecondary degrees, in the 10-20% who actually took the SAT, 2400) try to use us as an argument for the kind of homeschooling that thinks Pride flags and BLM are bad things
Smirkley@Smirkley
Only teachers are smart enough to educate children. Education majors, training to become certified teachers, score an average of 1029 on the SAT. While homeschooled students (taught by parents with no formal teaching credentials) score 1190.
English
twitter’s algorithm has a way of concentrating and distilling your biggest fears into Phone Sludge bc the only things i’ve got on the tl all day are how everything is abt to be automated and how top percentile grants arent being funded
Jason Shepherd@JasonSynaptic
Indeed. I have a 3%ile scored R01 still in limbo at the NIA...with no word if it will get funded. PIs are at breaking points...they need to be able to plan for personnel for one!
English
samar plok retweetledi
@curiouswavefn I like this idea in theory, but in practice the many-routes-to-failure problem in drug development is likely to mean that advanced AI models will not only fail, but fail for reasons we (and the models) won't be able to learn from.
English
samar plok retweetledi
kaiser just announced a new medically assisted dying program in daly city for people who can’t get frontier lab jobs
English