You woo a girl to fall in love with you, sleep with her, then abandon her? You're going to be run out of town as a rotten fool 100 years ago. Nowadays that communal protection is gone.

I don't see how we could ever get superhuman intelligence out of GPT-3. My understanding is that the goal of GPT neural nets is to predict the next token based on web text written by humans. GPT-N as N -> inf will be perfect at creating text that could be written by the average internet user.

But the average internet user isn't that smart! Let's say there's some text on the internet that reads, "The simplest method to break the light speed barrier is..." The most likely continuation of that text will not be an actual method to break the light speed barrier! It'll probably be some technobabble from a sci-fi story. So that's what we'll get from GPT-N!

Real money prediction markets are biased towards outcomes that increase utility of the bettors.

For an example: there is a prediction market on who will win the presidential race, Alice or Bob. Currently the market sits at 50-50 and the payout for $1 on Alice is $2 and the same for Bob. Our bettor, Charlie, has internal odds that are the same as the market, however he also believes with an incredibly high certainty that if Bob wins the race then the buying power of a dollar will double. Therefore $1 on Bob is worth $4 if Bob wins, but the same $1 on Alice will only be worth $2. Because of this Charlie will bet the market all the way to 66-33 favoring Bob. This is despite the fact that his internal probability that Bob wins is only 50-50.

The ability to destroy parity between decisions made by the real agent and simulations of the agent lets the agent win games against simulator opponents.

1. Different types of incoherence between real vs simulated choices grant different levels of power.

You're playing a game against a simulator Allen the Alien. The game is that you and Allen separately choose one out of 10 paths. If you pick the same one Allen wins; if otherwise you win. With no simulation Allen has a 1/10 chance of winning while you have a 9/10 chance of winning. If Allen simulates you accurately he then has a 1/1 chance to win.

If you're fully able to notice you're being simulated and completely obfuscate this fact from the simulator then simply having all simulations pick option 10 while real you picks option 1 is a 1/1 winning strategy. To achieve this you need some sort of pre-chosen general formula to break parity based on whether or not you're in a simulation.

You might not be able to tell if you're in a simulation while you do have the ability to break parity with other simulations of you and reality. Randomizing your path choice such that the simulation has a 1/10 chance of choosing each path and reality's choice has no correlation to the simulation's choice doesn't give you an automatic win, but completely destroys the simulator's advantage.

2. Sometimes creating incoherence between different simulations is more powerful than just creating incoherence between all simulations and reality.

Let's say we change the above game so Allen can pick 9 of the 10 paths and you only win if you choose the one path Allen leaves open. In this case a non-simulating Allen has a 9/10 chance to win and there's really nothing you can do about it.

If Allen is a simulator (and believes his simulation of you is accurate) then creating decoherence between reality and simulations of you improves your chances from 1/10 to 1/9. Simulations of you pick path 1 so Allen will be sure to pick that path and distribute the rest of his 8 choices among the remaining 9 paths leaving you a 1/9 chance of winning.

If you can not only create decoherence between simulations of you and reality, but can also create random decoherence between individual simulations then you can improve your odds further. If Allen is a (very) naive simulator then having your simulations each pick a path 1-9 with equal likelihood while reality picks path 10 is a guaranteed victory.

Of course the above strategy will not work against any reflective simulator. You will have to use steganography to conceal your simulation-defeating ability in the results of the simulations as well. This may involve using a different probability distribution for chosen paths or only having simulations select a subset of the paths reality will not choose. These techniques are bounded by giving you a better than 1/9 chance at worst and 1/1 chance at best.

The future may have a use for frozen people from the current era. In the future, historical humans may be useful as an accurate basis to interpret the legal documents of our era.

Original pubic meaning is a is a fairly modern mode of legal interpretation of the US Constitution. It's basis is that the language of the constitution should be interpreted the way that the original meaning of the text was when it was drafted and amended into the constitution. A similar mode of interpretation is used less commonly for statutes. It's likely that this mode of interpreation would become more common in the future as a way to prevent value drift.

One of the struggles of modern times employing the original public meaning test is that for older amendments there are no people currently alive who lived in the culture that the amendments were drafted in. It would be very helpful to have just a single ordinary man who lived in 1790 who could explain his understanding of the constitution and the language therein.

It's possible the future will have similar issues interpreting constitutional language from our time and will appreciate the ability to benefit from questioning a portion of the population of that era.

This theory is a subset of the idea that humans from past eras will be useful in the future to prevent value drift generally. My first brush with this idea was Three World's Collide.

We are ancient apes who can say yes or no to humans evolving. Tough choice. What do you think the apes should have chosen?