Biomni: A General-Purpose Biomedical AI Agent • Hacker News Board

Re: Biomni: A General-Purpose Biomedical AI Agent

#41

Post by spwa4 » Fri, Jul 11, 2025, 12:17 PM UTC

greazy wrote: post #27

andy99 wrote: post #13

I'm sure they've thought of this but curious how it fared on evaluations for supporting biological threats, ie elevating threat actor capabilities with respect to making biological weapons. I'm personally sceptical that LLMs can currently do this (and it's based on Claude that does test this) but still interesting to see.

Creating a biological weapon requires a whole bunch of unique and specialised skills, equipment, safety measures (so you don't infect/kill yourself/your people) and even multidisciplinary skill sets. Take for example the Kameido (Japan) incident by the Aum Shinrikyo cult/religious group [1]. Same group which committed the Sarin attack [2]. > The use of an attenuated B. anthracis strain, low spore concentrations, inef…

> Creating a biological weapon requires a whole bunch of unique and specialised skills, equipment, safety measures

I just tell some investors our god tells me to do that.

> too quick, it kills the host without spreading and authorities will notice, too slow, same problem: authorities will notice

The current authorities are Trump's authorities and don't believe in vaccines, have said in an interview Covid is either a Jewish or Chinese conspiracy (and "they made themselves immune"), and that the "disease epidemic" needs to end (this last one is easy to misunderstand. Kennedy Jr. doesn't believe any particular disease is an epidemic. Epidemics of that sort don't exist according to him. That people believe they get sick, THAT is the epidemic that must be stopped)

Re: Biomni: A General-Purpose Biomedical AI Agent

#42

Post by govideo » Fri, Jul 11, 2025, 9:45 PM UTC

panabee wrote: post #40

govideo wrote: post #38

Earlier quoted context omitted.

I'd love to hear more of our thoughts re open questions in biomedical ML. You sound like you have a crisp, nuanced grasp the landscape, which is rare. That would be very helpful to me, as an undergrad in CS (with bio) trying to crystalize research to pursue in bio/ML/GenAI. Thank you.

Thanks, but no one truly understands biomedicine, let alone biomedical ML. Feynman's quote -- "A scientist is never certain" -- is apt for biomedical ML. Context: imagine the human body as the most devilish operating system ever: 10b+ lines of code (more than merely genomics), tight coupling everywhere, zero comments. Oh, and one faulty line may cause death. Are you more interested in data, ML, or biology (e.g., pred…

ML first, then Bio and Data. Of course, interconnectedness runs high (eg just read about ML for non-random missingness in med records) and that data is the foundational bottleneck/need across the board.

Interesting anecdote abt Stanford doctors annotating QA question!

Each of your comments get my mind going... I'm going to think about them more and may ping you on other channels, per your profile. Thanks!

Re: Biomni: A General-Purpose Biomedical AI Agent

#43

Post by panabee » Sat, Jul 12, 2025, 3:26 AM UTC

govideo wrote: post #42

panabee wrote: post #40

Earlier quoted context omitted.

Thanks, but no one truly understands biomedicine, let alone biomedical ML. Feynman's quote -- "A scientist is never certain" -- is apt for biomedical ML. Context: imagine the human body as the most devilish operating system ever: 10b+ lines of code (more than merely genomics), tight coupling everywhere, zero comments. Oh, and one faulty line may cause death. Are you more interested in data, ML, or biology (e.g., pred…

ML first, then Bio and Data. Of course, interconnectedness runs high (eg just read about ML for non-random missingness in med records) and that data is the foundational bottleneck/need across the board. Interesting anecdote abt Stanford doctors annotating QA question! Each of your comments get my mind going... I'm going to think about them more and may ping you on other channels, per your profile. Thanks!

More like alarming anecdote. :) Google did a wonderful job relabeling MedQA, a core benchmark, but even they missed some (e.g., question 448 in the test set remains wrong according to Stanford doctors).

For ML, start with MedGemma. It's a great family. 4B is tiny and easy to experiment with. Pick an area and try finetuning.

Note the new image encoder, MedSigLIP, which leverages another cool Google model, SigLIP. It's unclear if MedSigLIP is the right approach (open question!), but it's innovative and worth studying for newcomers. Follow Lucas Beyer, SigLIP's senior author and now at Meta. He'll drop tons of computer vision knowledge (and entertaining takes).

For bio, read 10 papers in a domain of passion (e.g., lung cancer). If you (or AI) can't find one biased/outdated assumption or method, I'll gift a $20 Starbucks gift card. (Ping on Twitter.) This matters because data is downstream of study design, and of course models are downstream of data.

Starbucks offer open to up to three people.