246
AI creates first synthetic viruses
(thelemmy.club)
Why go for the cures right?
Paywalled source
https://ft.trib.al/Uk6VCRG
I don't think so. I don't have access to the research, but almost assuredly it is a custom neural network for specifically things like this. An LLM would never be able to describe in detail the virus to a point that researchers would be able to fabricate it. That's just not how that works.
You don't understand. This isn't a debate. They used Evo1 and Evo2 which are both LLMs. That is just a fact. You don't understand what LLMs are. I'm not trying to be mean or rude, it's simply true if you think an LLM couldn't do this, because it literally did.
Edit: sorry, that probably came of extremely rude and short tempered, and I shouldn't treat you that way.
LLMs are not only chat-bots, and in fact so not even need to "speak" any language you or I would recognize. Evo 1 & 2 are LLMs built off of raw genomic data. Literally the raw A, C, G, T pattern of the genetic code itself. Trillions of nucleotides. So the output of the LLM wasn't "Here's what a bacteriophage would look like!" It was "AT,CG,AG...." Raw genetic code which can then be synthesized.
That thing looks like it's a 'billion" residues long. Granted I have only rudimentary understanding of proteins, as I did my master years ago on thermodynamics, but using enzymes as the system.
How can you be sure anything in there will fold correctly. That you can design and describe a synthetic virus that in theory will work, but actually making it?
Edit: Ah, normal GMO, just used the models to tell them what to modify. Not a complete synthetic virus. Probably just did some normal injection of new lines in existing viruses.
Okay so, am I crazy here or does that mean it's just a machine learning neural network. Which is what LLMs are, but LLMs are specifically designed for language. That's their whole purpose. If this was not designed for language isn't it by definition not a LLM.
Yes, language -- a LLM trained only on English writing will have zero ability to respond to a question in Spanish. These LLM are trained in the language of genetics (e.g. ACTG).
It's like putting an internal combustion engine in a car or a boat. One navigates land and the other water, but the engine can actually just be the same one.
You're mixing up colloquial usage with scientific usage. Scientists are notoriously bad at giving things names (thanks dark matter...) All LLMs are neural networks. All neural networks are AI. Not all AI are neural networks. Not all neural networks are LLMs. This is a case of nested hierarchies, like phylogenetic taxonomies. You are a bony fish.
So yes, LLMs are literally just a form of neural networks, but the language they learn on does not have to specifically be English, or German. And this "Language" was genetic code. So Evo1 and Evo2 Literally speak genetic code. That is their language, since it is what they were trained on. They are predictive text models that can predict what the next genetic sequence will/should be.