The link for the AI crafting a super weapon seems to be broken. Here is a later article that is the best I could find: https://www.digitalspy.com/videogames/a796635/elite-dangerous-ai-super-weapons-bug/

22mo

Thanks! Link changed.

Although this isn’t a direct answer, I think there’s something that changed recently with chat gpt such that it is now much better at filtering out illegal advice. It appears to be more complex than simply running a filter over what words were in the prompt or what words are in chat gpt’s output. By recent, I mean in the last 24 hours, and many tricks to “jailbreak” chat gpt no longer work.

It gives the impression that they modified the design of it to train on not providing illegal information.

43mo

It feels to me like the update today made it even better at filtering out
answers that OpenAI doesn't want it to give.
It seems to me like the run basically on:
"Have an AI that flags whether or not a prompt or an answer violates the rules.
Mark the text red if it does. Offer the user a way to say that text was marked
wrongly as violating the rules."
This then gives them training data they can use to improve their filtering.
Given how much ChatGPT is used this method will allow them to filter out more
and more of what they want to filter out.

14mo

Hmm, that's interesting. Thanks Peter!

I was thinking something similar, but I missed the point about the prior. To get intuition, I considered placing like 99% probability on one day in 2030. Then generic uncertainty spreads out this distribution both ways, leaving the median exactly what it was before. Each bit of probability mass is equally likely to move left or right when you apply generic uncertainty. Although this seems like it should be slightly wrong since the tiny bit of probability that it is achieved right now can't go back in time, so will always shift right.

In other words, I

It’s worth noting that gradient descent towards maximum entropy (with respect to the Wasserstein metric and Lebesgue measure, respectively) is equivalent to the heat equation, which justifies your picture of probability mass diffusing outward. It’s also exactly right that if you put a barrier at the left end of the possibility space (i.e. ruling out the date of AGI’s arrival being earlier than the present moment), then this natural direction of increasing entropy will eventually settle into all the probability masses spreading to the right forever, so the ...

Does this hide the text? (Sorry just testing things out rn)

Wow

Ok so you can hide stuff by typing >! on a new line

Yep that's right! And it's a good thing to point out, since there's a very strong bias towards whatever can be expressed in a simple manner. So, the particular universal Turing machine you choose can matter a lot.

However, in another sense, the choice is irrelevant. No matter what universal Turing machine is used for the Universal prior, AIXI will still converge to the true probability distribution in the limit. Furthermore, for a certain very general definition of prior, the Universal prior assigns more* probability to all possible hypotheses than any other type of prior.

*More means up to a constant factor. So f(x)=x is more than g(x)=2x because we are allowed to say f(x)>1/3g(x) for all x.

Here's some mantras I have:

That which you are aware of, you are free from.

And some variation of:

Truth comes knocking. You say "go away, I'm looking for the truth." It goes away, puzzling.

The above I rediscovered recently through reading Zen and the Art of Motorcycle Maintenance.

As stated by others, there are counter examples. An important class of counter examples I can think of is when you want to pick up on mental attitudes or traits that likely only the best have–think "You are the average of your 5 closest friends."