When AI is trained for treachery, it becomes the perfect agent Last year, The Register reported on AI sleeper agents. A major academic study explored how to train an LLM to hide destructive behavior from its users, and how to find it before it triggered. The answers were unambiguously asymmetric — the first is easy, the second very difficult. Not what anyone wanted to hear. Illustration …
I may have an human-machine implant for that. It isn't cheap, though. Selling would upset some people.
It does mind reading from distances, and can be used to remote torture, influence, steal IP (intellectual property). But nowadays it is being used mainly as a propaganda tool for brands, public humiliation of targeted individuals (me), likely others
I believe the technology is about 20-30 years in the future of anything public, remotely similar.
Suicide may have happened due to its use in others, in the past.
If there is interest, feel free to contact.
That's looking at the end product of the model and its training / post training steps. That's no different than if you were presented with software in the form of a encrypted binary blob, without any knowledge of what encryption was used, what CPU the software ran on, or any access to that CPU to know what ISA it ran. You would have as much reason to trust that software to do what its "owner" claims and nothing more than you would an AI. They are both black boxes.
The way to assurance is obvious. You'd need to have access to how it was built in enough detail that you could reproduce it. For software that would be the source code and full toolchain (for this purpose we'll ignore the "On Trusting Trust" compiler backdoor worries) so you could reproduce it yourself. For an AI you could get the same, at least if it was one of the open models and you had access to all the training data. Thus you could reproduce it yourself.
Neither is practical for an individual - how many of us actually compile our Linux from source rather than trusting binaries, let alone compile our own UEFI and simple bootstrapping compiler to assure ourselves of a clean starting point? But it would be practical for a large corporation or a nation state to have that "trustworthy" AI - at least as trustworthy as is possible for current AI technology, rather than relying on "hundreds of gigabytes of interconnected numbers" someone else produced saying "trust us".
So we have:
HAL from "2001, A Space Odyssey"
WOPR from "War Games"
Skynet from "The Terminator"
All are fictional accounts of AI that became malicious but only Skynet approaches the stupidity we are seeing today...
I just wonder who will be the hero that saves us?
There is a good article that just went up on Quanta (recommended site, STEM for intellectually curious laypersons). It's an interview with a Harvard researcher studying the problem of understanding AIs, who comes to the same conclusion: "If you don’t understand the origins of the model, then you don’t understand why anything works”.
Article: https://www.quantamagazine.org/to-understand-ai-watch-how-it-evolves-20250924/
