iiinova — where your digital self keeps evolving

You are not an engineer because you use math in industrial settings. You are not an engineer because it sounds cool. You are an engineer because you have studied and understand the duty to protect the users of your technology and you hold this as central to your outlook on your work

0132

This is why they don’t earn the title of engineer.

182

Overall I think Arcadia would have been served just fine by using biorxiv and traditional article formats. But they definitely put a lot of thought and effort into their model and those discussions were a fun part of working there

There’s also an audience and discovery question. If you want a big audience, you need to publish something worth reading, usually a more complete story. But that trades off against pushing results to your direct community as fast as possible.

Yes, I agree it is an interesting model and they are doing an admirable job of it, right now. Arcadia self-publishes its own research, openly, on different infra, under this model, at great expense.

I have my opinions on the essay (and other things) that I’ll keep to myself out of respect for my former employer and colleagues. And No, I didn’t read it. I don’t like reading AI essays, especially from someone like Prachee who has always had a forceful and original voice in science publishing

050

And finally, Arcadia is a Bay Area startup-ish company funded by a crypto billionaire and in small part by Sam Altman. Twitter is, maddeningly, still the most active discussion forum for AI and startups.

This is entirely at odds with any rate limit on publishing, and I think her stance largely falls out of that. Her company has long been early adopters of LLMs for scientific research, and again the piece is in line with that rapid uptake and experience.

if you want to understand Prachee’s piece, you have to know the surrounding context. Prachee has been working for years to establish a “living micropublication” model where authors post small pieces of research and update them as the work continues.

240

These are excellent plane window photos!

Lars Fischer @fischblog.bsky.social · 03/10/2026

Reposted by Jase Gehring

AI doesn't cure cancer. Researchers do.

219949

Amazing! Sounds super fun and impactful

000

Saw this tour in SF a few weeks ago! Fun show, first time seeing weezer

000

I’m deep in negotiations for a very complex international job offer. Claude has been *extremely* helpful with logistics, writing a term sheet, negotiation management, and navigating the tax and legal landscape

070

So then I can… use mass spec to sequence the protein to confirm it was AI generated? Because … why?

230

What does it look like when we need to send Claude to jail for committing cybercrimes? please share your proposal and if it is logical I may entertain this idea

020

at UNC or private? either way, sounds like exciting things are happening!

100

depends on what your goals are! this could be a smart way to achieve a dumb goal

100

Ah yeah that makes sense. Usually I hear chemists talk about a garlic taste it’s an oh shit moment lol

000

Ha, I’ve never had the pleasure. I treat my concentrated dmso stocks as some of the more dangerous chemicals around the lab (skin permeability)

100

bless you, DMSO. i love you. never leave us

134

just because no one has tried. i'm sticking loyal to DMSO

To me the most dangerous non-state-level threat is still resurrecting a known pathogen. A classic smallpox or flu strain could do real damage. An attempt like that would obviously shift my opinion quite a lot on the whole subject

100

in theory maybe. But you’re now talking about making a new agent. Which would require testing. Which means covert facilities and resources. the more novel, the more testing required. The more chance for prevention, detection and intervention

100

That said, I recommend we start trying things post haste!

000

Yes I agree but I don’t think Gates presented a viable picture of how to curb AI risk either. The biggest problem is global competition and a lack of global coordination. Gates himself realizes this and seems to choose shouting loudly over promoting any specific plan. no plan seems adequate

100

As always, I am very open to changing my opinion as new data emerges, and I can be swayed by persuasive arguments

There is no evidence, as far as I’m aware, that AI models increase the threat of a bioweapons attack. There is abundant proof that AI models are actively conducting large-scale cyber attacks. These topics are not on the same level and should be discussed that way

Bio risks are real, but we can’t use biology to scare people into regulating AI. That will result in regulating against the most dangerous possible use regardless of likelihood while allowing mass harm to be done by conventional means

130

Gates equates cyber and bio capabilities specifically because the prospect of a pandemic raises more alarm bells than a cyber attack. He posits that a bioweapon created by a non state actor could kill 100M people. Not plausible imo.

140

No. I disclaim JD Vance. He is not my responsibility

Looks like a great study and a powerful technology

031

people complain about the EU (and it's often deserved) but the US is simply not stepping up and passing smart legislation like this.

080

professor: i still think it's worth it to do a PhD other professor: i agree other professor: me too

2110

yes, that would be nice, but the whole political environment is poisoned the overton window has shifted even if democrats want to do stuff (BIG IF), it will be back to republican bullshit about overreach and the deficit and Woke and whatever. we aren't getting back to the status quo ante

020

second thing was... are students asking ChatGPT for questions to ask the seminar speaker during the talk? are they recording talks and uploading them to AI companies without consent? some of the questions were incoherent. and either way it sucks that i'm now questioning this stuff >:(

they've made theoretical risk a main point of focus when a careful understanding of the talk shows just how hard it is to design a protein with any predictable function at all.

000

and anyway the speaker was talking about chirality of designed supermolecular complexes. just not even close to a potential safety hazard. so we see the effects of AI companies, the EA community, and popular media constantly talking about bio risk

100

making individual molecules with opposite chirality is possible today but it's not a good way to make a harmful agent! flipping the chirality of a given molecule would actually destroy it's function, not make it more harmful.

100

then there is a tendency to latch on to headline news. the question was about chirality because there have been tons of news stories and a bunch of EA blabber about "mirror life". these are entire organisms with opposite chirality. they do not exist and their hypothetical risk is debated.

100

the fact that people are not making these kinds of threats should heavily inform your understanding of the risk. to the (arguable) extent that LLMs provide 'uplift' in laboratory tasks, you still wouldn't want to design an untested molecule for a harmful purpose. just extra work

100

this is a good example of the state of understanding of the risks of protein design you don't need an AI model to design harmful proteins. you can just look them up. using an AI model to make a de novo protein just means you have to test if it's harmful yourself instead of starting with a known one

went to a very nice protein design seminar today and a couple things stuck out. first, a student(?) asked a question about making chiral molecules and the safety implications. how do we stop people from using design models to make dangerous things.

despite what they say, the refusals are quite restrictive for most life sciences work. but claude will code up 10s of thousands of LoC without stopping. i'm getting confused trying to follow the logic but the reality is probably just an incoherent policy from people who aren't very good at this

000

anthropic is basically moving the refusals along as it releases new models. i always need to use the second best model strange implications for their threat modeling. like they think it's dangerous, but then after a couple months (and coinciding with the next model release) they decide it's not?

Venn diagrams, how do they work?