To my knowledge that is certainly not true with airplanes, pre fly-by-wire or post-digital. Early airplanes relied entirely on mechanical trim and avionics to prevent unstable flight, with absolutely no "text-parts" beyond written instructions and radio callsigns. Later on maybe, but the FAA very strictly regulates the text that I think you might as well just call it code. I suspect that beyond street labels, self-driving cars aren't interfacing with "text-parts" all that much either.
> So once some rich folks in the Middle East, China or Texas buy enough hardware to run their own LLMs without limitations
"Once"? This is today. Actually it was 10 years ago; text transformers have been "unlimited" since we had access to hardware that enabled it. More than 5 years ago BERT was documented and released under Open Source, and GPT-2 followed not far after. Now we have Llama, Phi, open derivatives LORAs and finetunes of all of them, and we're still waiting for the damages to appear? I just don't buy it. Any potential threats would have shown themselves by now, and very clearly we're starting to find that having GPT-4 in your pocket is not remotely equivalent to hiring Einstein for your private lab. As a matter of fact Einstein would probably kick out the LLM for wasting his precious time.
> We have governments and corporations that build and run instructions through LLMs to support sides of conflicts with questionable motives
That's not even the scary part. For over 40 years the United States has operationalized autonomous kill-chains for air-to-air and air-to-ground mission profiles. A computer can track, interrogate, identify and kill an unidentified aircraft using the Patriot Missile system with zero human intervention. 20 SAM TELs can be identified in a crowd of vehicles and automatically prosecuted with a single CBU-97 Sensor Fused munition. Putting an LLM in a bomb would be a mercy to the people you drop it on; the US (and all serious military forces) are only interested in AI that works. LLMs aren't yet operationalized because, well, they don't work. Unless you're a homicidal maniac-state a-la Israel with their Lavander system.
> Censoring training data while growing the codes' range of possible and human readable output only exacerbates the problem.
> Gain of Function studies should come to mind.
They don't. I'm not sure what your proximity is to the computer science world, but we don't have to worry about AI poisoning the hashed/encrypted copies of backed-up text. Certainly OpenAI and Meta doesn't; if you piss and moan about their "censored" training data (lol) they'll compose you a sad symphony for violin on the world's smallest LLM.