It should be obvious that the current most powerful country and the biggest villain of the 20th century are outliers.
Chatgpt tells me the median country founded in the 18th century lasted 200 years. But I’m not sure about definitions. It also points out that you could plausibly consider the US Civil War as the country undergoing a regime change. It also expressed skepticism that the founders actually believed it would only last 50 years.
That said, it really is quite striking how bad people are at predicting the future. For instance, when the Founding Fathers created the US constitution, most of them didn’t expect the country to last more than fifty years. And, when the Nazis came to power, they expected to last for a millennium, calling themselves the “Thousand-Year Reich,” but, ultimately, their rule only lasted for twelve years.
You’re picking two huge outliers as examples of failed predictions. But most countries didn’t last as long as the US, and if their founders had similar predictions they would have been right.
One thing to understand is that this task is very different for current AI harnesses, for similar reasons as counting letters in strawberry was difficult for a previous generation. It needs to repeatedly run a model with different context, which has a long lag, so it prefers to return results more quickly. If you give it some guidance as to how to call agents, it will do better.
“Use a fast agent to generate the words and ranking at each step”
Adding this at the end seems to work. It does replace one bit at the end to avoid hanging dialogue, which seems reasonable.
https://chatgpt.com/share/6a79f2d2-3980-83eb-aa59-873658f756f7?ogimg=plain
It did it in batches so I can’t fully verify each batch was done properly sequentially. I would assume if you specify that each word be done by a different agent it would work fully.
Well sure. If you get something that’s Bayesian evidence of a simulation, then Solomonoff as a formalization of Bayes should then upweight simulations.
I don’t see how our physics could be simpler to simulate in a way that isn’t just even simpler to compute directly, though.
Re footnote, if you have a simple simulator you still need to specify all of our physics, so it’s going to end up more complicated than other models unless the simulator theory actually explains complex parts of our observations simply. Even if it does you can probably get those same explanatory algorithms more simply without a simulation. E.g. if the simulation cares about observers, you can find a simpler algorithm that cares about observers directly without simulation overhead.
The example of a pill that defeats breathalyzers is in tension with the mind reading point. Why are breathalyzers good but mind reading bad? You don’t explain.
If there’s X Boltzmann brains, each of them takes log2(X) bits to specify, which carries a 1/X probability penalty, that is exactly cancelled by the large number of brains. So you just never get the overwhelming confidence in Boltzmann that presents a problem.
Conceptually, the simplest algorithm for predicting your experiences is not going to be “simulate a world of Boltzmann brains, and choose BB number 198573859385 …..” because then all of the ordered experiences you are having need to be encoded into that selection, and you lose the ordering simplicity.
I don’t buy this distinction.
Every thread going on from you will observe something in Roulette too. (The threads that die don’t go on from you, then.)
When predicting your future experiences, you automatically condition on the fact that there will be future experiences, so they won’t always match objective probabilities. And you can also adjust for some versions of you having more experiences than others.
Sure, but then why is it strange to do the same with duplication? You’re shifting your relative probability of seeing A, by creating extra versions of you that see A.
I don’t think the epsilon chance of observing now really helps. It makes the current probabilities meaningful but you still have no particular reason to insist that the current probabilities (“what happens if I were to observe this today”) matches the future probabilities (“what happens if I observe it in the future”), given that the observations are correlated.
Re death, it just doesn’t seem different to me. If I can magically increase my probability in outcome A by committing to suicide in not-A worlds, I can also increase my probability in A by duplicating myself in A worlds. You’ve set up a formalism that allows the former and not the latter, then you’re pointing to a contradiction. But they’re both about as mysterious and settled the same way. The first isn’t controversial and the second shouldn’t be either.
(The exact probabilities still depend on SSA vs SIA vs others. But the fact that they won’t match the objective probabilities seems fine.)
Note that the math isn’t showing up for me, might be a formatting issue.
There’s a fatal flaw in the logic here.
Note that the martingale is defined over observations and expected observations.
But then you talk about the prior probability of heads. This is not meaningful, since she does not observe the coin flip.
Here’s a consistent probability assignment that meets the martingale criteria.
On Sunday, she estimates probability ~1 of waking up tomorrow in a room. She doesn’t have a probability on a particular room or of a coin result as of Monday because she doesn’t observe that on Monday. She further estimates a probability of 2⁄3 for Tuesday room 1 and 1⁄3 for Tuesday room 2.
On Monday, she confirms that she woke up. She continues to have the same 2⁄3 and 1⁄3 probability assignment across the two rooms possibilities for Tuesday, since nothing has changed.
Alternatively, she can consistently maintain that each room is 1⁄2 likely.
I’ve got a different scenario with no duplicates. A fair coin is flipped today, if it’s heads, Roulette Rory gets killed with 75% probability before learning what the coin says. If it’s tails Rory survives. We ask Rory what her probability is. She says today her probability is 50⁄50, but if she learns about results she is 80% confident in tails.
Why is it ok for Rory to expect her future self to be 80% confident in something that she’s currently 50% confident in, but it’s not ok for Beauty to do the same? In both cases, the number of observers is varied in a way that’s correlated with the observation. That’s all that’s happening here.
SALP has blown up. Still unclear what happened, but CNBC reported they sold the entire public equities book to one buyer before the market opened today.
If you specify everything about the experiment, then you have added complexity on both sides. But note that things like “a fair coin will be flipped” is not actually possible to specify under SI, which works with deterministic Turing machines. It’s not clear how to specify something that works like that.
And we mostly care about these problems because of the relevance to more realistic versions like doomsday, where we do not have the specification and so you do have asymmetrical complexity. In any real world scenario, you have uncertainty over exactly what is going on, and some of the possibilities are simpler than others and get more weight. But you never really exclude possibilities like “Omega is lying to me about how this works” or “it’s all a simulation”, they just become improbable if you have significant evidence against.
Well a world with N people, with N large and not simple like a power of 2 or something, is going to be log2 N extra bits to specify over the same world with one person. Then an extra log2 N bits to locate a specific observer.
But since you might be any one of the N people, the second one cancels out and you’re left with the inherent complexity of the world. So log2 N, but if N is a simple number it can be much less. For N power of 2, it can be log log N. And you can have stuff that grows much faster than powers of 2, eg. N=Graham’s number doesn’t take many bits to specify.
In the standard SB question, not significantly, because the scenario is so contrived and the number of people under tails isn’t that high.
But in versions with thousands of people, yes.
There’s also the Solomonoff Induction location issue, where larger worlds get penalized if it takes many bits to locate observers within them.
I agree with Andrew Gelman that in many cases we should report probabilities with fewer significant digits.
Really, I think it would be just fine to specify win probabilities to the nearest 10%, which will register shifts of 0.4% in expected vote share. Probabilities to the nearest 10%: if it’s good enough for the National Weather Service, it’s good enough for me.
He quotes a comment that disagrees:
I disagree with Andrew that 65.7% is too precise. One reason is that intrade prices are quoted to 3 digit precision — viz., ranging from $0.00 to $10.00, with bid-ask spreads as low as 1 cent. So, for example, if I believed Nate’s formula was more accurate than Intrade, and the current quote was $6.56 bid, $6.57 ask, I would like Nate to provide more precision in order to determine whether or not to buy ‘Barack Obama to be re-elected on 2012′. Even with today’s low interest rates, I would still need Nate to forecast at least a 65.74% probability in order to believe purchasing ’Obama 2012′ at $6.57 would outperform a 0.90% money market rate over the next 18 days.
This is absurd to anyone who has traded election markets seriously. You do not trade a 0.1% edge on a market like this.
If Alice said “I need to provide a point estimate, so I think it’s 36.49284727693% likely”, we would all recognize it as absurd. So we’re just haggling over the price. In general, outside of areas where you have enough data to calibrate well, nearest 5% is all that’s meaningful.
This is separate from the probability range Bob offers, which should only be done when you specify what you mean.
You do better with leverage at the expense of increasing risk, including total loss in worlds that are only slightly off from what you expect.
You seem to be talking about what people with hundreds of billions collectively can do. Capacity is a real constraint at that point.
It’s also likely some of the best investments won’t want that much capital. Even if you’re right on the 60%, some of that can come from new private companies you can’t invest in or can only invest after they’ve grown a lot.
You’re filtering based on data that isn’t available at the time of the prediction. Number of traders at resolution isn’t known, and resolution time isn’t known for all markets at market creation. So calibration will have a lookahead bias.
You should ideally filter only based on info available at the time of the predictions you’re grading.