> I cannot wait to ask it how to make nuclear weapons, psychedelic drugs
Your town's university library likely has available info for that already. The biggest barrier to entry is, and has been for decades:
- the hardware you need to buy
- the skill to assemble it correctly so that it actually works as you want,
- and of course the source material, which has a high controlled supply chain (that's also true for drug precursors, even though much less than for enriched uranium of course).
Not killing yourself in the process is also a challenge by the way.
AI isn't going to help you much there.
> to write erotica.
If someone makes an LLM that's able to write good erotica, despite the bazillion crap fanfics it's been trained upon, that's actually an incredible achievement from an ML perspective…
It can bridge the gap in knowledge and experience though. Sure, I could find some organic chemistry textbooks in the library and start working from high school chemistry knowledge to make drugs, but it would be difficult and time consuming with no guide or tutor showing me the way.
Methheads making drugs in their basement didn't take that route. They're following guides written by more educated people. That's where the AI can help by distilling that knowledge into specific tasks. Now for this example it doesn't really matter since you can find the instructions "for dummies" for most anything fun already and like you said, precursors are heavily regulated and monitored.
I wonder how controlled equipment for RNA synthesis is? What if the barrier for engineering or modifying a virus went from a PhD down to just the ability to request AI for step by step instructions?
You're vastly underestimating the know-how that's required for doing stuff.
Reproducing research done by other teams can be very difficult even if you have experimented people in your lab, and there are tons of stuff that are never written anywhere in research papers and at still being taught in person by senior members of the lab to younger folks: it's never going to happen in the training set of your LLM, and you'd then need tons of trial and errors to actually get things working. And if you don't understand what you're even trying to do, you have zero chance to learn from your mistake (nor does the LLM, with your uninformed eyes as sole input for gaining feedback).
Your town's university library likely has available info for that already. The biggest barrier to entry is, and has been for decades:
- the hardware you need to buy
- the skill to assemble it correctly so that it actually works as you want,
- and of course the source material, which has a high controlled supply chain (that's also true for drug precursors, even though much less than for enriched uranium of course).
Not killing yourself in the process is also a challenge by the way.
AI isn't going to help you much there.
> to write erotica.
If someone makes an LLM that's able to write good erotica, despite the bazillion crap fanfics it's been trained upon, that's actually an incredible achievement from an ML perspective…