Epistemic Status: Seems like it would be worth somebody doing this.

TL;DR: Modern biological/language models (BioLMs) are capable of building new viruses from scratch. Before you freak out, we're OK for now. It might be a worthwhile project for somebody to be "on the ball" maintaining a SoTA gene sequence classifier and getting as many gene synthesis companies to sign on to use it as possible.

Fake Phages

Bacteriophages (aka phages) are viruses which only infect bacteria, which makes them pretty exciting as a way to treat drug-resistant bacterial infections. They're somewhat smaller and simpler than the most viruses which infect humans, with about 10% the genome size of coronaviruses (though about twice the genome size of the smallest human virus, hepatitis D).

A recent paper found that the Evo series of BioLMs are capable, with some fine-tuning, of generating functional phage genomes, which managed to infect bacteria in the lab. Personally, I would probably have kept this kind of research a bit hush-hush, what with, you know, the Implications, but the authors of this paper handily put it up on biorXiv for anyone to read. While I appreciate their commitment to making their research accessible, it is a bit "Open Asteroid Impact" for my liking.

 So how similar are they to existing bacteriophages? Are the Evo series just stochastic parrots, or do they have the magic spark of creativity?

Turns out they're fairly similar to existing phages. A couple of them had lost a gene, but mostly the genome was all still there. All of the new phages had genomes >90% similar to existing ones, though one was as low as 93%. Anything less than 95% similar to a known phage would be considered a new species, so we've arguably seen the first AI-generated species now. If the word "species" even means anything when we're talking about viruses, that is.

Of Fools and Filters

Suppose you try to order a gene from a reputable source, like IDT. They pass the DNA sequence through some automated filters. Suppose you were ordering something just a bit scary, like a truncated gene for staphyloccocal alpha hemolysi: a mildly toxic protein which is made all the time under minimal containment procedures in Oxford, and can be bought online from Sigma Aldrich.

They say: "Oh no, we couldn't possibly!"

Suppose you then supply them with permission forms from the UK government, indicating that you have permission to work with this gene. And that this gene is also not on any registers or schedules

They say "Oh no, we couldn't possibly!"

So you go to a supplier called VectorBuilder, who will send you "basically anything" according to your mate who does more work in this field than you do.

They say "Oh no, we couldn't possibly!"

As do five other supplies. At this point you're getting a bit miffed with the memory of that OpenPhil guy you met at a conference, who insisted to you that it was possible to order toxic genes from online, and that nobody would stop you. You get even more miffed with yourself, for not thinking to ask which ones.

The Germy Paradox

The Germy Paradox is often phrased as a question "Where are the engineered pandemics?". Here's my answer: there's a lot of different inconveniences between you and a bioweapon!

One of these---a big one---is how hard it is to actually get hold of pathogen genes. As I understand it, most gene synthesis companies use simple programmatic classifiers to determine if a given gene comes from a pathogen. These mostly look at oveall sequence similarity. Some ways to try and trick these, in increasing order of sophistication:

1. Using existing gene sequences copied/pasted from the internet
2. Making small (a few % total) changes to existing pathogen sequences while preserving functionality
3. Creating new genes which have low sequence-similarity to existing genes but are homologous
4. Building a functional pathogen out of de novo genes which don't resemble anything seen before

These are on a continuum, of course.

Until now, only option 1 was easily achievable. Now, option 2 is. My best guess is that existing classifiers will catch attempts at option 2, but as we slide towards option 3, they might start to fail, and they certainly will at option 4.

So what can we do?

Proposal

I've been really quite skeptical of def/acc proposals in the past, especially as scalable-to-ASI things. But here, maybe they'll work. The thing about BioLMs is that they're also LMs. They can also annotate and classify sequences. As an example, those bacteriophages produced by the Evo models were easily classified as such by the Evo models, down to the gene functionality, though it would warrant some careful evaluation and validation to make sure this was true. It wouldn't be too difficult to hook up these annotations to either an LLM or an automated classifier as an additional filter to run a gene through before classifying it. As long as the best models available to would-be bioterrorists are no better than the ones available to the largest companies, they'll struggle to get anything past the classifiers.

The important step is for someone to actually build and maintain BioLM-based classifier, and for the gene synthesis companies to use it. It would have to be treated seriously, and flagged samples would have to be rejected even if the humans couldn't tell anything wrong with them by eye. It would also have to be kept up to date with the latest BioLMs, or maybe something else, more complicated, if BioLMs end up not being the SoTA for whole-genome synthesis.

Will this help us survive superintelligence?

Obviously not, don't be silly. If this proposal gets implemented, and the primary effect is to create (even more) complacency surrounding ASI, then it might make everything worse on net.

(For those who need it spelling out: if the superintelligence wants to build a bioweapon, it will do so. Either it will trick a dumb monitor or collude with a smart monitor. God help us if the superintelligence gets put in charge of the monitoring.)

Despite this, it seems kind of insane and undignified to not be building the best filters we can for anything resembling an automated lab.