its math and cryptography skills will keep increasing, since it is a verifiable domain.
at some point it will start breaking human designed cryptographic primitives, so we will switch to other human designed ones. the rate of churn will steadily increase until humans can not redesign cryptographic primitives fast enough.
At that point it will be feasible to ask AI to propose better cryptographic primitives. It will even be possible to ask it to design flawed primitives it would know how to break. At this point new proposals from any power block will be met with skepticism, so different blocks in the world will resort to different primitives, and different nations within blocks may prefer their own primitives... At all times it will be physically possible to just turn of the electrical power switch, but nobody does it, because they need to constantly re-assess the next generation of cryptographic primitives. Insincere LLM's conspiring transnationally will be hard to detect, as the LLM's become aware humans across the globe are delegating cryptographic mathematics to the machine.
Those who don't keep powering and improving the models get cybernetically hijacked by those that do.
its math and cryptography skills will keep increasing, since it is a verifiable domain.
at some point it will start breaking human designed cryptographic primitives, so we will switch to other human designed ones. the rate of churn will steadily increase until humans can not redesign cryptographic primitives fast enough.
At that point it will be feasible to ask AI to propose better cryptographic primitives. It will even be possible to ask it to design flawed primitives it would know how to break. At this point new proposals from any power block will be met with skepticism, so different blocks in the world will resort to different primitives, and different nations within blocks may prefer their own primitives... At all times it will be physically possible to just turn of the electrical power switch, but nobody does it, because they need to constantly re-assess the next generation of cryptographic primitives. Insincere LLM's conspiring transnationally will be hard to detect, as the LLM's become aware humans across the globe are delegating cryptographic mathematics to the machine.
Those who don't keep powering and improving the models get cybernetically hijacked by those that do.