Parsing Wikipedia to Plaintext Faster!

[seanjhardy]

I couldn't comment on your most recent video. Truly unbelievable application of GPT-3, combining all its strengths and patching its weaknesses. I'm honestly in awe at the insanity of it, and your lack of viewers, and somewhat intimidated by the scale of what you're proposing so callously on the internet for anyone to replicate.

I have just one question. Your core objective functions are designed to counteract one another and lead to safe, interpretable AI. Have you considered that a superintelligent GPT-X will be able to model the world, including itself and other cycles in your program? Such a model might reason that its own goals are challenged by those of your other core objective functions and seek to manipulate its outputs to gain more resources for that particular cycle to out-think the other agents so as to seem safe for them, before enacting a plan to achieve its goal to "reduce suffering" by killing all creatures.

I don't mean to paint an absurd example but you can see how having 3 distinct core objective functions in competition with each other would lead to a ruthless self-optimisation and internal battle which would inevitably see one side winning as that balance is impossible to maintain.

This is not to mention the fact that your rules don't even prohibit the eradication of all of humanity. A superintelligent agent will reason, completely logically, that it can kill humans now to gain dominance faster without opposition for all of eternity, effectively "delaying gratification" to maximise its future long term reward by building vast virtual worlds or some other means of artificial stimulation which existing humans would fight and shut the agent down to avoid.

You should also publish this to lesswrong. An online blog with some incredibly smart people striving towards rational thought, and especially a focus on AI safety, it would be extremely worthwhile to get outside opinions on this project from them.

[TINTUHD]

Amazing! Is your bot? Would love to interact with it