What does AI want? Without ambitions you have to be told what to do.
Today's systems are mostly good at following instructions. But push them far enough, and you already get weirdness like alignment faking, instrumental self-preservation and more. We know because we've seen it in lab settings, while probing for extreme failure modes on purpose - but the world is large and strange enough that edge cases like that are liable to surface naturally.
The reason why none of this has exploded in our faces isn't that today's AIs never want to do weird and dangerous things. We know they do, at times. It's that today's AIs are incapable of pulling them off when they try.
Capability is the thing that matters, at the day's end.
There is to the best of my knowledge no fundamental limitation to having an agent/claw go on like this for 80 years with a prompt like "live your life to the fullest".
- Prompted as in prompt made of tokens -> for LLMs, tokens double as a clock signal. Time only flows when tokens are pushed through them.
- Prompted as in specific request placed in the stream of tokens -> Yeah, they do that all the time whether it's getting into infinite loops of repeating same pattern, or suddenly deciding to do things based on inputs they normally ignored.
Also don't forget that everything is a "prompt" for LLM. All input tokens end up in the same place.
In the current UIs is there a lot of suppression then as I have not seen things start on their own?
I meant an LLM doing something without any external prompt at all. Not doing something different etc but rather do something without a token/prompt ever flowing to it.
We shouldn’t be doing that as a civilization in case it works.
Today's AI harnesses just keep feeding AI tokens - and can, in principle, do so indefinitely. "Streaming LLM" is not mainstream but not unknown either. "Automatic context compaction" is somewhat similar in what it does, and very common. "Keeping the lights on" isn't hard. And you have to turn on those lights to have an AI do anything at all.
At the same time: that type of harness often provides a "tool call" interface. Plenty of tools that end up attached to LLMs can do things like run arbitrary code, spawn instances of the same AI, and more.
That's frankly enough. An AI that can spawn more copies of itself can, in principle, just keep itself running indefinitely - even in a non-streaming no compaction harness. An AI that can run arbitrary code on a system can, in principle, do anything a human using that system could do.
The thing that truly limits what an AI can do is the AI itself. Practical AI safety relies on AI being either too weak or too well behaved to cause major issues.